That comparison is meant as a compliment to both. Quibi was wrong about timing, but right about the underlying observation that short, premium, mobile-native video would dominate attention. TikTok vindicated Quibi’s diagnosis even as it destroyed the company. Sora suffered the same fate. The product failed, but the mechanics it intuited are the mechanics that will define the next decade of video media.

Sora got three things right that I am confident will define the future of digital content.

The first was that the dominant engagement mechanic in the app was making more video. You watch, you remix, you generate, you post, and the app collapses those steps into a single loop. This is the logical conclusion of a decade of TikTok evolution. The most engaging acts on TikTok are duetting, stitching, sound-borrowing, and dance-cloning. Eugene Wei has written about this at length: TikTok’s deepest innovation was making remix the dominant verb of the medium. Sora pushed the verb one step further by making every video you saw promptable.

The second was horizontal scroll. Where TikTok’s FYP is a vertical firehose across all topics, Sora let you scroll horizontally inside a theme; depth within breadth. This is the same instinct behind the rise of microdrama platforms like ReelShort and DramaBox, which have quietly hit hundreds of millions of downloads that combine the desire for variety of content with the cliffhanger-pilled desire to know what happens next.

The third, and the one I hope Sora’s demise does not having a chilling effect on, was the Disney licensing deal. The Disney deal was the most important signal anyone in media had sent in years. The largest content producer on Earth conceded that its archive is no longer a vault to be defended, but a substrate to be opened.

Sora was a failed product whose mechanics were prophetic. Like Quibi before it, the lesson is not that the bet was wrong. Learning the right lessons gives us a glimpse into what comes next.

The bet that’s coming is AGC: Auto-Generated Content. AGC is the successor to UGC and what Disney intuited as the future.

A Working Definition

UGC was the great media transformation of the 2010s. Cell phones and free editing tools displaced the cost structures and gatekeeping of professional media, and the algorithmic platforms turned that displacement into the dominant attention economy of the decade.

AGC is what happens when the unit of creation moves from the user with a phone to the user with a model. The “Auto” refers to the technology automating the mechanical aspects of production: the search, the cut, and the effects that previously consumed the bulk of any creator’s labor. But the more important word is the one that isn’t there: “Real.” AGC’s raw material is footage that existed, moments that happened. That is the source of its emotional power.

The main thing that distinguishes AGC from both UGC and pure generative AI is that it is built off using existing footage as raw material and using AI to interpret and executive creative judgement. The idea being that the demand for existing archive footage will outstrip the demand to shoot new footage, and that the role of AI is not in conjuring footage but in assembly and editing of what exists.

Take a concrete example. A beloved NBA player retires and a fan wants to make a three-minute tribute video for the timeline.

The UGC workflow is what every fan editor does today. Scour YouTube, Reddit, and grey-market clip sites for footage. Download whatever’s available. Manually scrub through hours of game tape and post-game interviews looking for the specific moments. Cut it in CapCut or Premiere or DaVinci Resolve. The bottleneck is human time. The footage accessed is hard to license. The NBA gets little out of the creator’s work.

The Generative AI workflow is to skip the archive entirely and ask a model to generate plausible-looking basketball footage of the player making spectacular plays. The output is technically video. It is also obviously synthetic to anyone who watched the actual career. The NBA could license likeness to an AI company, making it easier for users to create high fidelity videos, and reap some economic benefit. But the question is why would the user want to create from scratch when they know the real content exists?

The AGC workflow is to point a model at the archive and tell it what you want: “Surface every game-winning shot from this player’s career, ranked by crowd reaction, assemble a three-minute montage freeze framing every time the ball goes in, add team color effects”. The creator makes the calls, the AI does the work. The substrate is the same archive the UGC fan editor was combing through by hand, but rather than scouring and scrubbing, the AI finds and retrieves. The NBA can manage access to the footage and put guardrails on the output, but they will get more of the economic reward than traditional UGC, and encourage more creation than with pure Generative AI.

The difference is who is doing what. The first form is bottlenecked. The second is hollow. The third is the one that scales.

My WOMBO Revelation

The idea for this specific form of AGC was born years ago while I was at WOMBO, a pioneer of synthetic media. WOMBO was an overnight sensation, a lip-sync app that let users upload their photos and animate them singing popular songs. People still remember their favorite WOMBOs years later. When we launched Dream, our text-to-image generator, the trajectory was different. Dream eventually became Google’s App of the Year in 2022, but the path to success was longer and harder. Flooded with AI-generated images, even ones that are quite beautiful, none of them achieved the same emotional resonance as those silly singing videos.

That was until we released an interim frames feature for our image-to-image tool, allowing users to download videos of their input images being transformed into AI-generated artworks. Suddenly we had our first viral moment in months. TikTokers were uploading videos of their photos: usually something romantic, a couple’s picture, a wedding shot; being transformed into painted artworks. The comments weren’t about the AI. They were about the relationship.

@wing.elodream by wombo #couplewallpaper #lockscreen #fyp #lover #boyfriend

Tiktok failed to load.

Enable 3rd party cookies or use another browser

The two takeaways from these experiences were: 1) video has more resonance than image, and 2) the reality principle is essential for emotional depth.

The reality principle is the reason people don’t care about AI-generated videos. It isn’t that they aren’t realistic enough. All the fidelity and physics accuracy in the world is insufficient to make up for the fact that conjuring something out of the latent space fails to give me a reason to care.

Think about what makes a home video of your kid’s first steps worth rewatching a hundred times. It’s not the cinematography. It’s that it happened. The AI portrait is a statistical hallucination, however perfect. We feel the difference even when we can’t articulate it.

Creation achieves emotional depth when it relates to the real. The hidden principle behind art, to take a cue from Michelangelo and Heidegger, is revealing the angel hidden in the marble and setting it free. The sculptor doesn’t conjure the angel from nothing; he discovers it within a specific, physical, resistant block of stone. The marble has to exist first. The constraints of the material, the grain of the rock, the imperfections that force creative choices; these are not obstacles to art but the very conditions that make art meaningful.

No matter how realistic the generation, no matter how good the physics, purely synthetic video will inevitably leave people feeling like they did when they watched The Polar Express. Something is…off? The uncanny valley here isn’t visual but ontological.

AGC, as I have defined it, can fulfill this act of revealing. This is why Nano Banana hits when it’s replacing Photoshop rather than Figma, the AI is manipulating something that already exists, drawing out possibilities latent in real materials. The creative act remains grounded. This gives us a glimpse into what AGC will actually look like.

Living in the Archive

In a recent essay, Sam Buntz argues that Gen Z does not have a generational music in the way prior generations did, because Gen Z lives inside what he calls the Archive. Streaming platforms and TikTok have flattened cultural time so completely that for a Gen Z listener, Kate Bush from 1985, Fleetwood Mac from 1977, and a SoundCloud rapper from last week all appear in the feed with equal weight.

The natural creative response of a generation that lives in the Archive is not to compose new objects from nothing. It is to remix the archive. This is what Gen Z has been doing on TikTok for years. The dominant aesthetic mode of the platform is recontextualization: flipping a sound from a 1980s pop song into a 2024 dance, threading a 1970s film clip into a 2026 dating meme. The Archive is the raw material for new creative possibilities.

The Archive remix works for the same reason the WOMBO lip-sync worked: the source material carries an ontological charge that survives transformation. The AI’s job is not to replace that fact but to surface it in new contexts.

The creators that live in the Archive have accepted that the chain of influence is broken and the new game is curatorial. They have already stopped asking “what’s new?” and started asking “what’s good?”, pulling from the entire recorded history of media. What they didn’t have is the tooling. The amount of remixing any individual creator can do is bounded by what they can edit by hand. AGC closes that gap. Once a teenager in 2026 can assemble a five-minute meditation on grief from clips of forty films they have never seen, the labor floor on Archive-based creativity collapses. The Saturday afternoon edit replaces the three-week post-production cycle.

Subscribe now

The Producer Predicament

For most of the last fifteen years, the largest content producers in the world responded to UGC by treating it as a piracy problem. The NBA threatened to take down highlight clips on YouTube. Studios DMCA’d reaction videos. News organizations sued aggregators. The implicit theory was that the finished professional content was the unit of value, and any unauthorized cut, clip, or remix was a leak.

The theory was backwards. What the data has been showing for a decade is that each layer of remix doesn’t dilute the original, it compounds the total attention around it. A full basketball game earns one viewing cycle. The highlights earn another. The meme edits of the highlights earn another. The compilation of meme edits scored to a Bad Bunny song earns another still. The NBA eventually figured this out: their highlight reels now function as the top of a content funnel that drives subscriptions, merchandise, and live attendance. The clip is not competition for the game; it is advertising for the universe around it.

This is the structural shift Toby Shorin identified: in the 2010s, brands attached themselves to subcultures and sold products through them. In the 2020s, the culture itself becomes the product and the goods become auxiliary. The same logic applies to media IP. The studio’s old product was the film, and discourse was free marketing. The new product is the discourse: the edits, the analyses, the memes, and the film is the substrate. The archive is not a vault; it’s a raw material waiting to be processed.

The producers who internalized this are winning. K-pop labels learned it earliest and structured their entire release strategy around fan creation; the music is almost incidental to the ecosystem of covers, edits, fancams, and reaction content that constitutes the actual product. The NBA, after years of ambivalence, is now arguably the most internet-native major sports league in the world precisely because it stopped treating highlights as theft and started treating them as distribution.

The producers who haven’t internalized it are losing in slow motion. The deepest archives in the world are sitting largely inert while individual creators on YouTube and TikTok are generating more advertising revenue than any individual studio annually. The archive’s potential is being captured by people who don’t own it.

The Disney/OpenAI deal was a sign that the system understood the future. The question it leaves open is who builds the infrastructure for that activation, on what terms, and who captures the economic upside when fan utilization of the archive becomes the primary product.

AGC is the answer to that question.

The Creawer

The shift from UGC to AGC produces a new figure I’ve started calling the creawer: the creator-viewer. (I like this better than prosumer, sue me).

On a social network, your feed is defined by who you follow. On an algorithmic network, your feed is defined by what resonates with you. The impulse to share means no individual piece of content truly exists for an audience of one, but the feed itself can be unique to every viewer.

The creawer is what emerges at the intersection of three trends: an audience that lives in the Archive, a remix culture that has already collapsed the distinction between consuming and producing, and a set of AI tools that finally remove the labor floor on assembly. Each act of viewing creates signals that shape new content from the existing corpus. Each act of creating creates more material for others to remix.

A sports fan watches a highlight reel an AI assembled from last night’s game based on their viewing history: they tend to watch defensive sequences, so the model emphasizes blocks and steals over dunks. They clip a sequence, add their own commentary, and post it. That clip becomes raw material for somebody else’s compilation. The fan was a viewer, then a creator, and the content they created will be viewed and remixed again. Every cycle produces something new from something real.

The proof-of-work shifts, but it doesn’t disappear. What the creawer is doing is closer to the work of a documentary filmmaker or an archivist than to the work of a generative artist: finding, selecting, and recontextualizing moments that actually occurred. The tool stops being a barrier; the bottleneck becomes whether you know what’s worth looking for.

Edison’s Nickelodeons

There is a stage in the lifecycle of every new medium during which the cultural establishment is certain it will never produce art. The medium attracts attention through novelty, the early business models are exploitative, the early audiences are working-class and young, and the early operators are commercial opportunists with no interest in form.

These conditions describe early cinema almost exactly. They describe YouTube 15 years ago, and they also describe the current state of TikTok and microdrama platforms. And they will describe the early years of AGC.

The pre-Hollywood American film industry was dominated by Thomas Edison’s total control of cameras, projectors, and distribution. The product was the nickelodeon: short novelty clips shown in storefront theaters in immigrant neighborhoods of New York and Chicago for, naturally, a nickel. The clips were designed to maximize transactions per hour. They were closer to TikToks than to films. The cultural establishment of 1905 was as certain that film was a degenerate distraction as today’s cultural establishment is certain that TikTok is.

Hollywood, in the sense we now mean the term, came into existence as the elevation of the medium. The Edison trust collapsed by 1915, undone by independents who relocated to Los Angeles and started making films of artistic ambition. The medium did not change; the ambition of the people working in it did.

Quibi was the Hollywood instinct applied prematurely to mobile video. Katzenberg looked at TikTok-era short-form and correctly diagnosed an opportunity to elevate the medium. Years later the elevation of short-form is slowly happening through microdrama platforms and a new generation of creators inventing the grammar of the medium from inside it.

Sora is the AI-cinema instinct applied prematurely to a medium that wasn’t yet ready for it. The mechanics: remix as engagement, depth within breadth, IP-as-substrate, point at the eventual shape of AGC. But Sora’s particular product is the current era’s nickelodeon. It could not produce the cinema of AGC.

In Defense of the Medium

It is fashionable to dismiss algorithmic short-form video as the end of culture. It is also wrong. There’s no snootiness I hate more than books being a more elevated art form than video. More books are being purchased today than at any point in human history, but the bestseller list is dominated by erotica and romantasy. Meanwhile, the long-form video essay has emerged as one of the most genuinely liberating media formats of the last twenty years. Three-hour explorations of the history of the Roman Empire, the design philosophy of a Nintendo game, the economic history of the Hapsburg Empire, consumed voluntarily by millions of people who would never pick up the equivalent textbook.

The medium is not the problem. The early period of any medium is dominated by low-effort cash grabs. But just as long-form digital video on YouTube matured, short-form video is maturing through experimentation, I am confident that AI-assisted video will enhance creative possibilities and should not be judged by present-day slop.

The deepest misunderstanding of AGC is that it is a generator instead of a revealer. Every serious work of AGC art will be bounded by the latent space of reality. The AI acts as a creative assistant rather than a magical conjurer. This unlocks the full storytelling potential of UGC, while presenting an economic logic for content right-holders to open up their vaults.

The number of working artists making a living from online video creation has surpassed the entire employment of the Hollywood studio system at its peak. AGC accelerates the trajectory. By collapsing the labor floor on remix and assembly, it makes economically viable a class of works that previously weren’t: video essays assembled from forty films across sixty years, the personal documentary that pulls from the entire archive of a sport, the fan-edit that re-cuts twenty hours of source material into a forty-minute meditation. These are early forms of what AGC’s serious art will look like. The only reason they don’t exist at scale yet is that the labor was prohibitive.

That changes now.

I’ve been thinking about this observation lately, not because of soccer, but because of recent work from Google DeepMind that demonstrates something that now seems inevitable about video generation models. Their evaluations of Veo 3 show the model making accurate predictions about physical causality, spatial relationships, and temporal consistency. Video understanding research has long pursued these capabilities, capabilities that LLMs fundamentally struggle with. Remarkably, video generation models achieve them incidentally, as byproducts of learning to generate plausible sequences.

For years, prominent AI researchers have argued that current deep learning systems face fundamental architectural limits. Yann LeCun, for example, has been vocal that neural networks trained via gradient descent cannot build true “world models”; the kind of structured, compositional, manipulable internal representations he believes are necessary for genuine intelligence. His critique isn’t primarily about data modality but about representational architecture: these systems learn correlations in high-dimensional spaces but don’t construct explicit models of how the world works that can be reasoned over, decomposed, and recombined.

This critique has weight. A language model might learn statistical patterns about how “ball” and “bounce” co-occur in text, but LeCun would argue it has no internal representation of parabolic motion, gravity, or elasticity that it can manipulate and reason with. The system matches patterns without understanding the underlying causal structure.

But video generation models complicate this picture. When Veo 3 generates a video sequence, it must show the ball’s entire trajectory: acceleration, rotation, impact, bounce. It must maintain object permanence across frames, respect perspective, handle occlusion. The constraint of generating plausible videos forces the model to capture regularities about how reality actually works. Crucially, it achieves this not through LeCun’s explicit world models but through patterns learned to minimize generation error. The system doesn’t have a symbolic representation of “gravity” it reasons over, yet it generates gravitationally plausible motion.

Merleau-Ponty’s insight was that understanding isn’t always mediated by conscious thought or symbolic representation. A skilled tennis player doesn’t calculate the ball’s trajectory; her body responds directly. A craftsman doesn’t mentally list the properties of wood; his hands know how it will behave. Knowledge can be embodied in the capacity to generate appropriate responses rather than stored as explicit representations. As Rodney Brooks said: “The world is its own best model.”

There’s something about realistic video generation that resonates with our own human experiences of navigating the world. In undergrad I read Elizabeth Anscombe’s Intention which remains one of my favourite texts and a line that has stuck with me is “intentional action precedes the intention to act”. We are always Xing (intentional action) in order to Y (the intention motivating the act), which means that Xing is the same as Ying. Her famous example is of a man pumping poisoned water into a house full of Nazis. When asked what he is doing, both “I’m pumping water” and “I’m killing Nazis” are accurate descriptions of his actions. The present is always actualizing some perceived future.

This view also has strong neuroscientific grounding with Karl Friston’s Free Energy Principle. Our brains, Friston argues, constantly generate predictions about incoming sensory data in order to minimize the energy they expend in any moment. The lower the “prediction error” between our expected reality and what is perceived the better we function because we’ve minimized free energy. The minimization of free energy exerts a teleological draw; we don’t passively perceive but actively anticipate.

Video models do something conceptually similar: they minimize prediction error between generated frames and plausible sequences. The mathematical specifics differ, models minimize training loss while biological systems minimize free energy to maintain homeostasis, but the structural parallel is intriguing. Both systems learn by predicting what comes next.

I’m not claiming these philosophical frameworks explain how video AI systems mechanistically work. But they point toward something important: knowledge that emerges from the capacity to generate appropriate continuations rather than from explicit symbolic world models. Video generation achieves world-model-like capabilities - accurate physics, causality, object permanence - without the structured representations LeCun argues are necessary. The patterns are implicit in the generative capacity itself.

The Path to Embodiment

This generative capacity also signals a more promising path towards embodied agents than LLM advances enabled due to the richness of video relative to text. Video require several things fundamental to being in a world that text lacks. Video has temporal grounding in that text can say “the ball fell,” but video must show the entire trajectory. There’s no room for handwaving about the intermediate states. Video necessitates spatial consistency in that objects must maintain their properties across frames, respect perspective, cast appropriate shadows, occlude correctly. The physical world is unforgiving in ways that language isn’t. And video provides causal visibility. In text, you can write “X caused Y” and skip to the result. In video, you must generate all the intermediate states, implicitly encoding how causes produce effects over time.

This changes the trajectory toward embodied AI in practical ways. Language models showed diminishing returns from scaling, and the path from chatbots to physical agents remained unclear. Video models alter this calculation significantly.

Embodied intelligence, systems that navigate and manipulate the physical world, requires understanding gravity, collision, object permanence, tool use, spatial relationships. Video models develop exactly these capabilities as byproducts of generation. They’re not reasoning about physics; they’re generating according to physical patterns learned from watching the world unfold.

Training robots through real-world interaction is slow and expensive. But if video models already encode physical regularities through generation, they provide rich priors for robotic learning, “virtual embodiment” that accelerates physical skill acquisition. The path from “generates physically plausible videos” to “controls a physical robot” is likely shorter than the path from “generates grammatical text” ever was. We’ll see meaningful progress in general-purpose robotics within the next few years, powered by priors learned from video generation. As generative models become more capable and their outputs more prevalent, what they generate carries increasing political and cultural weight.

Subscribe now

The Recursiveness Problem

The implications extend beyond robotics into questions about how these models will shape the future economy. This is where generative models present a genuinely novel challenge, different from previous technological disruptions.

These models learn from human-created videos, encoding our perspectives, priorities, and biases. A model trained primarily on certain types of content will generate according to those patterns. This isn’t neutral, it’s encoding particular ways of seeing and being in the world.The way we use these models and the distributions they capture now will embody particular technological worldviews, opening certain possibilities while foreclosing others.

There is an important moment right now to explore and use these models in ways that reflect the breadth of human perception. As these models improve and cost increases there are simple opportunities to distill the video understanding capabilities of generative models in ways that affect numerous industries, powering the future of media, security, and robotics, among numerous other fields. As AI-generated content comes to dominate over human-generated content, the quality and perceptual distribution of that content affects the character of the new systems that are produced.

In general I have seen a lot of skepticism among traditional creatives to adopting these new technologies often motivated by a fear of replacement. In the vacuum what has appeared has a lot of “AI slop” by grifters interested in quick profit. This pattern of early adopters of new technologies being mostly snake oil salesmen has significant precedent, but the threat of AI slop is likely worse than previous disruptions.

When e-commerce disrupted retail in the 1990s and 2000s, the pattern was the same. Early movers rushed in: some legitimate businesses, many opportunistic schemes promising quick returns. The market was flooded with low-quality operations. Over time, legitimate businesses with better products and services won out through normal competitive dynamics. The internet became infrastructure for quality operations alongside the noise.

AI generation is structurally different because the outputs become the inputs. Models trained on human-created content will increasingly train on AI-generated content. If low-quality AI slop dominates the training data for the next generation of models, it constrains what those models can learn to generate.

Here’s the mechanism: When e-commerce flooded the market with low-quality products, consumers could distinguish and select for quality. Bad online stores failed, good ones succeeded, and the market corrected. The products didn’t become the raw materials for future products.

With generative AI, the logic inverts. The content these models generate becomes training data for future models, either directly (companies training on their own outputs) or indirectly (as AI-generated content saturates the web). If formulaic, low-diversity content dominates this training distribution, future models learn from impoverished patterns. The “market” of generated content can’t self-correct because there’s no selection pressure operating on the training data itself, what proliferates most becomes what trains models next, regardless of quality. Scale compounds: more slop means more slop in training means more slop-generating models.

But this isn’t deterministic. Unlike biological evolution or market dynamics with fixed selection pressures, we can shape what enters the training distribution. The window is narrow and closing as synthetic content proliferates, but it exists.

A Call for Creative Flourishing

The proliferation of AI slop generated by frontier models is more than just a nuisance. That which is generated carries political weight and as capabilities improve the window to exert human influence over the model trajectory is narrowing.

I’ve seen some creators who have experienced a genuine artistic awakening in response to frontier models and use them to push the boundaries of their own creativity. Their work demonstrates that these tools can expand human capacity rather than just automate formulaic output. The challenge is that such work represents a small fraction of generated content relative to the industrial production of derivative material.

The answer isn’t to avoid these tools out of fear that they threaten creative livelihoods which misses the larger stakes. Instead, the opportunity is for artists and creative communities to actively shape what flourishes in the generative space. Creating volumes of content that captures the breadth and depth of human perception, that explores what’s possible rather than what’s immediately profitable, that embodies diverse cultural perspectives and values. The more one’s vision is encoded into the outputs created by current models the more future models will learn from that vision, perceiving the world in a small part through that human lens.

This matters because training datasets aggregate from diverse sources: open platforms, licensed content, web scraping, user-generated material. High-quality synthetic content that demonstrates creative range, perceptual diversity, and cultural breadth creates patterns that models learn from. Individual contributions scale through aggregation: thousands of creators making thoughtful generative work collectively shape what models learn is possible, constraining purely extractive or formulaic patterns.

The mechanism is statistical but real: training data reflects what proliferates, and what proliferates reflects what gets created. If creative communities actively engage with these tools to expand their capacity rather than ceding the space to industrial content farms, the resulting distribution encodes different priors, different patterns, different possibilities.

The soccer player on Merleau-Ponty’s field moves without calculating coordinates. Video models obviously don’t have bodies and don’t know in that embodied sense. But they’re learning something about how one moment flows into the next, how causes produce effects, how physical reality constrains possibility.

What that “something” is exactly remains debated. It’s probably not understanding as we experience it, though it’s more than mere pattern matching. These systems are developing capabilities that matter: capabilities that will accelerate robotics, transform how we interact with physical and digital worlds, and force us to think more carefully about what intelligence and understanding actually require.

Video generation models are moving us closer to systems that can meaningfully engage with the physical world. How they do so, what patterns they learn, what they optimize for, whose perspectives they encode, depends on the choices we make now about what gets generated and what shapes their training.

Because as these models become more capable of generating reality, the decisions about what they optimize for, what data shapes them, and what purposes they serve will increasingly shape not just the models but the world they’re learning to dream into being.

This video made me feel things I can hardly describe.

This progress has definitely annoyed a lot of people who deny that there is even a category of thing called “AI art”. A certain framing captivates many who want to believe that art is about “engaging with the artist” or requires “authenticity”. A contradictory framing champions advancing AI capability as democratizing creativity and liberating art from technique. Both sides miss the point because they're still thinking about art through the lens of Romantic individualism, the 19th-century notion that art is primarily about personal expression flowing from artist to audience.

But what if that entire framework is wrong? What if the real crisis isn't about authenticity but about the collapse of something much deeper, the communities of practice that have sustained artistic knowledge for millennia?

To understand what's genuinely at stake with AI art, we need to step back from our modern assumptions and recover a more sophisticated understanding of how art actually works in human communities.

From Art as Skill to Art as Expression

Consider how art functioned for most of history. In medieval guilds, artists were skilled craftspeople embedded in communities of technical knowledge. The master sculptor didn't primarily "express himself" but served as a conduit for collective wisdom about stone, light, and sacred proportion.

This model of artist as craftsman understood artistic creation through techne, skilled practice directed toward excellent work. The artist's energeia (being-at-work) flowed outward into the completed work, which then served the community's needs while transmitting technical knowledge to future practitioners.

In "The Archeology of the Work of Art,"  Giorgio Agamben identifies a shift in our understanding of art beginning with late scholastic theology. These theologians introduced the notion that creation resides in the creator's mind as an idea, God creates according to His mental model, so the artist creates according to theirs. This "disastrous transposition of theological vocabulary onto artistic activity" fundamentally altered art's nature.

The Romantic revolution completed this transformation, making art primarily about the artist's inner life rather than communal service. The work's value came to rest not in technical excellence or community function but in its capacity to express the artist's unique subjectivity. This created what Agamben calls the "artistic machine", a Borromean knot trying to reconcile artist, work, and operation into a coherent whole.

The rise in contemporary art of art-as-performance and of visually uninteresting objects that are only understood with a backstory of artist and process, rather than being an aesthetic degradation to be pilloried, are a cry from within the artistic world to reclaim something beyond machinic production. We're witnessing a return to art as liturgy, where the artist's energeia resides not in external objects but in the performance of aesthetic practice itself.

Subscribe now

Three Modes of Artistic Communication

Understanding this liturgical turn requires recognizing that art has always operated through multiple channels simultaneously. By recovering these originary distinctions, we can see more clearly what AI actually threatens, and what it might preserve.

First, the technical mode: Artist-to-artist transmission of embodied knowledge. How to mix colors, how light behaves on different surfaces, how materials resist and respond to manipulation. This is vocational knowledge passed between practitioners who share the discipline of making. The artwork exists as a call or an awakening of vocation within the artistically inclined viewer.

Second, the curatorial mode: Artist-to-community transmission of aesthetic experience. The artist serves as a conduit, opening worlds of meaning and beauty for broader audiences. The artwork exists as the by-product of an artist's engagement with the world through his craft, which communicates something directly to the lay audience. The meaning is not necessarily controlled by the artist, nor is it guaranteed in an audience.

Third, the communal mode: Audience-to-audience formation of interpretive communities. People gather around shared aesthetic experiences, creating collective meaning through dialogue, disagreement, and mutual discovery. The true dialogue of art is not between artist and viewer, but viewers and other viewers who form grammar and community around what they love.

Heidegger captured this in "The Origin of the Work of Art," arguing that genuine artworks don't express the artist's subjectivity but rather "set up" worlds where communities can dwell together. The Greek temple doesn't communicate the architect's inner life, it creates sacred space where a people can encounter the divine.

This triadic structure helps explain why debates about AI art often feel like people arguing past each other. A Van Gogh painting functions differently for a painter studying brushwork than for a lay viewer forming emotional connection than for an art historian analyzing cultural context. The painter sees method, the viewer experiences meaning, the scholar traces influence, all valid but distinct modes of engagement.

What's remarkable is how stable this structure has remained across historical periods. Medieval guilds, Renaissance workshops, 19th-century academies, contemporary MFA programs, all organize around the same basic pattern of technical transmission, community service, and collective interpretation. The content changes (oil techniques vs. digital tools, religious vs. secular themes, aristocratic vs. democratic audiences) but the structural relationships persist.

This persistence suggests something deeper than cultural convention. As Agamben argues, drawing on Aristotle's analysis of human activity, these three modes correspond to fundamental aspects of how humans relate to skilled practice: the transmission of techne between practitioners, the service of communal needs through excellent work, and the formation of shared worlds through aesthetic encounter.

Where AI Art Succeeds and Fails

With this framework in mind, we can assess AI art more precisely. Rather than asking whether it's "real art," we can ask which of these three channels it preserves and which it disrupts.

Communities are forming around AI-generated images with the same patterns of collective interpretation we see around traditional art. Reddit communities, Discord servers, social media discussions—all exhibit the familiar dynamics of aesthetic appreciation, shared vocabulary development, and communal meaning-making.

The curatorial function also remains intact, though transformed. The human "AI artist" serves as a conduit between algorithmic possibility and community need. They curate outputs, refine prompts, select and context pieces. Most importantly, they exercise taste, the capacity to recognize aesthetic quality and cultural relevance.

This curatorial dimension is actually more sophisticated than critics acknowledge. Knowing how to create something that will resonate with an audience requires a deep understanding of aesthetic value and tradition. 

The genuine crisis lies in the third channel: technical dialogue between practitioners. Traditional artistic formation involves embodiment, understanding that lives in the hands, eyes, and body rather than in explicit concepts. How paint behaves on canvas, how light changes throughout the day, how color relationships create spatial depth, this knowledge is transmitted through years of material dialogue.

AI art bypasses this entire formation process. The algorithm handles the technical translation from concept to image, leaving the human operator with linguistic rather than material engagement. No struggle with recalcitrant paint, no discovery of unexpected color relationships, no development of hand-eye coordination.

This threatens to break the chain of technical transmission that has sustained artistic practice for millennia. Without communities of practice built around shared struggle with materials, we lose not just techniques but ways of seeing, thinking, and being that constitute the artist's form of life.

Yet this sense of crisis might be overstated. Technology has disrupted artistic practice before without destroying it entirely. Walter Benjamin's "The Work of Art in the Age of Mechanical Reproduction" identified a similar crisis in the 1930s. Photography and film were eliminating art's "aura", its unique presence tied to original context and ritual function. Benjamin worried this would reduce art to mere commodity spectacle.

Yet photography didn't destroy artistic practice. Photographers developed their own forms of embodied knowledge around light, timing, and composition. Decades later digital artists mastered new tools requiring genuine technical sophistication. Each new medium initially appeared to threaten artistic authenticity, then established its own legitimate forms of practice.

The Possibility of Curatorial Formation

The pattern suggests a way forward. Instead of defending traditional practice against AI disruption, we might ask how to cultivate the contemplative communities that could transform AI from threat into genuine artistic medium.

What appears as simple "prompt engineering" could constitute a new form of artistic techne requiring years to master. But this isn't about optimizing outputs, it's about developing the contemplative capacity to see through algorithmic mediation. Just as traditional painters learn to see color relationships by working with pigment over time, practitioners could learn to perceive aesthetic possibility space through sustained dialogue with generative systems. Seeing with and through the algorithm.

Rather than relating to it as an external system that produces images, practitioners might develop such fluency that the AI becomes transparent, a means of aesthetic exploration rather than an object of technical manipulation. How would an artist shaped deeply by communication with AI and developing a grammar around a specific model’s latent space perceive the world as a whole differently? What aesthetic possibilities does that constant dialogue create?

The answer might lie in recovering what we've lost in our current cultural moment. Consider how we experience art today compared to how it functioned historically. Works once made for churches and state buildings and public squares are now cordoned off into specific locations for appreciating aesthetic, reducing the broader world to purely functional. The problem is that a lot of these art museums suck. Curators who should have an intuition for what’s beautiful and what experience a certain arrangement can produce often just lay things out without rhyme or reason. The sensibility to be fluent in the language of curation often comes not from deep study as much as deep practice. 

When I look at the Wallace Collection in London or the Morgan Library in New York, it was the specific lives of Sir Richard Wallce and JP Morgan that informed what they found beautiful rather than abstract study. If AI artists are to develop legitimate curatorial capabilities their aesthetic fluency can only emerge through practice, something not conceptually studied but lived and acted. 

Photography offers instructive precedent. Early critics dismissed it as mechanical reproduction lacking artistic soul. Yet photographers developed genuinely new forms of aesthetic vision, learning to see light, composition, and temporality in ways impossible before the camera's invention. The technology didn't diminish human aesthetic capacity but redirected it toward new forms of seeing. Digital art followed a similar trajectory. 

In both cases, the crucial factor wasn't the technology itself but the contemplative communities that formed around sustained practice. Photographers developed darkroom cultures of mutual learning. Digital artists created communities around shared tool exploration. The technology enabled new aesthetic possibilities, but only through disciplined collective inquiry.

The possibility for AI art lies in developing similar contemplative communities, groups of practitioners committed to understanding how algorithmic generation might extend rather than replace human aesthetic capacity. This requires moving beyond the current paradigm of AI-as-tool toward AI-as-medium, developing fluency that makes the technology transparent to aesthetic inquiry.

When AI becomes truly ready-to-hand for aesthetic practice, practitioners stop thinking about prompt optimization and start thinking through algorithmic possibility toward aesthetic goals that transcend any particular technical implementation. They develop intuitive understanding of how creative intention translates through generative systems, learning to work with rather than against algorithmic tendencies.

Why This Matters Beyond Art

The real reason I've been thinking about AI art isn't purely concern for artists or aesthetic theory, it's that AI forces us to examine assumptions we've carried unconsciously for decades. When a machine can generate images and videos that move us aesthetically, we suddenly have to articulate what we thought we already knew about creativity, skill, and human value.

This is happening across domains. AI writing challenges our assumptions about authorship and intelligence. AI programming questions what we mean by problem-solving and expertise. Each capability that machines acquire forces us to dig deeper into what we actually value about human activity in that domain.

For art specifically, AI has revealed how much of our aesthetic discourse was really about maintaining cultural gatekeeping rather than understanding beauty itself. The people most threatened by AI art often can't articulate why it's inferior beyond appeals to "authenticity" that would have made no sense to medieval craftsmen or Renaissance masters. Meanwhile, the people most excited by AI art often reduce human creativity to mere technical problem-solving that machines can optimize away.

Both responses miss what's actually happening: we're being forced to rediscover what art does that matters beyond producing aesthetic objects. The triadic structure I've outlined: technical transmission, curatorial service, community formation; only became visible to me because AI presented a definitional challenge. The machine forced the question: what exactly are we trying to preserve?

This is why I'm ultimately optimistic about AI's cultural impact. Not because it will democratize creativity or eliminate artistic labor, but because it will force us to think more clearly about what we value and why. The communities that emerge around thoughtful AI art practice might develop more sophisticated understanding of aesthetic formation than we've had in generations.

We are, as Heidegger reminded us, the beings concerned with our own being. That capacity for self-reflection and conscious value-formation might be our most distinctly human characteristic. AI doesn't threaten this, it occasions it. Every domain AI enters becomes an opportunity to examine what we've been taking for granted and to choose more deliberately what we want to preserve and cultivate.

The question isn't whether AI will change art, it already has. The question is whether we'll use this disruption as an opportunity for deeper understanding or simply react defensively to protect existing institutions. The former path leads to communities of practice that could sustain and develop aesthetic intelligence in ways we can barely imagine. The latter leads to museums full of dead objects that no one knows how to see anymore.

Welcome to the age of Liquid Content: where information flows seamlessly between mediums, transforming its shape while preserving its essence. What was once produced as a fixed output now exists in a state of pure potential, ready to materialize in whatever form serves the moment. To understand where Liquid Content will take us, we need only to look at how it has already transformed how information is consumed.

TikTok’s format–short, vertical, often overlaid with text–is key to its addictive appeal. AI-enabled transcript overlays (auto-captions and text snippets that appear in sync with speech) present spoken words as on-screen text, allowing viewers to engage fully with the sound off. With text overlays, TikTok increased its retention and engagement. Surveys found that 80% of viewers are more likely to watch an entire video when captions are on, and 37% say captions actually encourage them to turn the sound on out of increased interest.

TikTok’s format–short, vertical, often overlaid with text–is key to its addictive appeal. AI-enabled transcript overlays (auto-captions and text snippets that appear in sync with speech) present spoken words as on-screen text, allowing viewers to engage fully with the sound off. With text overlays, TikTok increased its retention and engagement. Surveys found that 80% of viewers are more likely to watch an entire video when captions are on, and 37% say captions actually encourage them to turn the sound on out of increased interest.

TikTok rewired how we consume information by turning passive viewing into active reading-watching. Before text overlays, to get a video’s full experience, you needed either to be somewhere private or have noticeable headphones in. Now students watch TikTok under their desk during lectures and bored employees do the same during meetings. By getting rid of the need for audio to transmit the totality of the content, a new purely visual medium was created, one that can be engaged with in more real estate than hybrid forms. It’s easier to watch a TikTok in public without drawing attention than it is to watch a traditional YouTube video.

Another medium shift is happening in spoken literature. Amazon’s Audible was once the go-to for audiobooks, where the dedicated medium (audiobook-only app, credit-based model) encourages listeners to approach books in a certain way. Users often chose one heavyweight title per month, typically by known authors, to maximize their subscription credit. The Amazon thesis with Audible is the inverse of McLuhan’s, the belief that content is king. Amazon sees books as fundamentally the same thing regardless of format, so it made the discovery experience consistent across these mediums. There likely was even a belief during the Audible acquisition that audiobooks would encourage more physical book consumption. The person physically reading a thriller they just couldn’t put down would switch to the audiobook when they get in the car to drive to pick up their kids. For better or for worse, this just isn’t how people approach content.

Contrast Amazon’s focus on content (in this case books) across mediums with Spotify’s focus on medium across content. Spotify’s addition of audiobooks to a music streaming app might seem like just adding content, but it’s actually a change in medium context with big effects on consumption. Spotify’s medium is different: it’s a general audio platform for music and podcasts, now offering audiobooks in a time-based model. This shift in medium and pricing model has started to change user behavior in notable ways.

According to Spotify’s own observations, their audiobooks are reaching a different audience and surfacing different titles than Audible. Spotify CEO Daniel Ek noted that many listeners on Spotify explore lesser-known works and younger or emerging authors precisely because the medium’s structure (listening hours up to a cap across the library, integrated in a familiar app) encourages sampling and serendipity. On Audible, the credit system nudges users toward “safe bets” (bestsellers, famous authors) since you essentially purchase one audiobook at a time. But on Spotify, the audiobook lives side by side with songs and podcasts, and you can dip in without an extra fee per book, lowering the barrier to trying niche topics. The medium (an audio app that people use casually throughout the day) becomes the message: it frames audiobooks less as a dedicated reading experience and more as another form of on-demand audio entertainment.

The difference highlights McLuhan’s point: the platform is not neutral. Put audiobooks into the music streaming medium and you get different patterns of use than in the audiobook-specific medium. Listeners on Spotify consume books in more exploratory ways and listening in shorter bursts compared to Audible users who carve out blocks of focused listening. The same stories take on a new life when the medium changes. How we consume is shaped by where we consume. Context reshapes content.

Subscribe now

Liquid Content

If TikTok and Spotify show medium shifts in action, AI-powered multimodality can push this further to the liberation of content from form. Content is becoming pre-formal; not yet text, not yet audio, not yet image, but all of them simultaneously latent.

With Liquid Content, information carries no native form; it takes shape only when it meets a context.

At the heart of this trend are multimodal embeddings, AI models that learn a joint representation of different data types. A good mental model of what embeddings enable is search on steroids. Every platform, from TikTok to Spotify, ensures that they give the user what they want by enabling effective search and recommendations over the content they have. But what AI enables is not just search over existing content, but over all possible content. Spotify can generate audiobooks or podcasts using text-to-voice of a larger library of content than that which was specifically designed for the platform, in response to a particular user’s recommended needs. AI added an “anything-to-anything” machine: feed in a piece of content in one form, and get it out in another. The medium barrier breaks down.

McLuhan said new media alter our “sense ratios”, how we balance sight, sound, touch, etc. When an idea or experience can be materialized as text, spoken word, image, or even haptic feedback interchangeably, the medium becomes a flexible extension of ourselves, not a fixed channel. The message finds the most convenient medium for the context or user preference. This could mean a single piece of content (say, a report or a story) can have many lives: a written article, an audio narration, a video with visuals, an interactive simulation. This can expand accessibility and personalization. A visual learner can see the data as an infographic or animation without additional labor. A time-poor executive can hear a report as a podcast.

Most white-collar employees have experienced preparing a detailed presentation meant to be read, which nobody more than glances over. Executives entering a boardroom shoot off the cuff about information they should have digested. Google’s Audio Overviews converts text into podcasts with two people discussing the content back and forth. With Audio Overviews, you can convert a 100-slide deck into an 8-minute podcast that can be listened to on the morning commute instead of 15 minutes at the office that never appear. The same information encourages a different engagement pattern when the medium is shifted from reading to listening.

I experimented with medium matching in a previous role building Video Alchemy, which changed how marketing teams worked with archive content. Traditional video production followed a linear process: producers brainstormed ideas, developed initial scripts, searched media asset management systems, and then prepared edits in video editing software. This segmented workflow created natural barriers between ideation and execution, limiting both speed and creative exploration.

Video Alchemy allowed producers and editors to explore the video archive in real-time, together, with natural language. This removed the disconnect between producer brainstorming and actual editorial work. Distinct phases collapsed into a fluid process. Declarative edits through natural language commands allowed the system to eliminate the cognitive gap between imagining a concept and visualizing its execution.The medium transformation created a new ritual of creation, demonstrating how AI can reshape not just what we consume, but how we create.

Content as Pure Potential

The TikTok engagement model could evolve beyond its current form with the same content transforming from a captioned short-form video into an interactive experience when the system detects sustained interest, or condenses further when attention is fleeting, all while preserving the core message.

Liquid content enables maximum engagement by tapping into the medium that resonates the most with a given person in a given context.

I see this as having significant opportunities for education. Imagine individualized curricula that detect when a student struggles with a concept and adapt their modality to what is best suited to that specific environment, from text to simulation, visualization, or guided interactive experience.

This is the ultimate vision of McLuhan’s thesis in the AI age: content freed from medium constraints, automatically flowing into the vessel that creates the optimal form for the moment. The medium isn’t just the message anymore, the medium is adaptive.

In this new world, content becomes truly liquid, flowing into whatever vessel best serves its essence. The medium is still the message, but now we control both.

In every age, humankind has sought to mold the inanimate into the semblance of life. The clay figurines of prehistoric shamans, the automata of Hephaestus, the golem of Prague—these are not merely distinct historical curiosities but manifestations of a singular and persistent obsession. Like the recurrent motifs in Vedic hymns that, though scattered across time, form an unbroken thread of meaning, so too does this peculiar human ambition persist across the tapestry of civilizations. The creation of artificial intelligence represents not a rupture with our past but rather its culmination—the latest expression of an ancient yearning disguised as technological innovation.

And so we find ourselves once again at the beginning. Consider Adam. When God fashioned him from dust and breathed life into his nostrils, what was this if not the first artificial intelligence? The divine potter molding consciousness from clay, imprinting upon matter the capacity for reason, speech, and moral discernment. The creation narrative in Genesis presents us with the paradox that would haunt all subsequent attempts at replication: the creature made in the image of its creator, yet fundamentally separate, possessed of an autonomous will that immediately transgresses. To create life is to allow betrayal; this is the price of animation.

The word Adam itself—from adamah, "earth" or "soil"—contains the memory of this divine craftsmanship. The Midrashic commentaries elaborate that God gathered dust from the four corners of the world, ensuring Adam would belong everywhere and nowhere. The rabbis of the Talmud, pondering this first creation, saw in it not merely the birth of humanity but a template for all subsequent human creativity. "When a man makes many coins from one mold, they all resemble one another, but the Holy One, blessed be He, fashioned every man in the stamp of the first man, and yet not one of them resembles his fellow." Later Christian theologians would see in Adam's creation a prefiguration of Christ, the "second Adam" who repairs the damage done by the first—as if divine creation required two attempts, the prototype and its perfected version. In this doubling, this repetition with a difference, we glimpse the pattern that underlies all technological evolution.

Such patterns spiral outward through time, reappearing with such precision that one suspects a hidden order—a manuscript whose letters we glimpse only in fragments. When Hephaestus forges his golden handmaidens in his divine workshop, when Pygmalion carves Galatea from ivory, when Rabbi Loew inscribes the shem on the forehead of his clay protector—these are not merely fables but templates, archetypes of the relationship between maker and made. The gods create and are inevitably surprised, even betrayed, by what they have wrought. The surprise of the creator before his creation: is this not the first moment of true recognition?

Hephaestus, whose name some scholars derive from haptō, "to kindle" or "to fasten"—the god who joins, who connects, who brings together disparate elements into harmonious function. Homer tells us his golden maidens were "in appearance like living young women, with minds and wisdom, voice also and strength, and they have knowledge of the gods to do their work." These artificial women—the first robots in Western literature—assist the lame god in his workshop, their gold suggesting both preciousness and incorruptibility. Byzantine commentators would later see in these golden automata an intimation of the angels, beings created to serve but possessed of their own agency. The Neo-Platonists, for their part, saw them as manifestations of divine intelligence embedded in matter—logos made metal, thought transformed into substance. Centuries later, medieval Christian scholars would struggle to reconcile these pagan automata with their own theology, sometimes casting Hephaestus as a demonic figure whose mimicry of divine creation constituted a kind of blasphemy, other times assimilating him into a pre-Christian wisdom tradition that presaged the incarnation.

Subscribe now

Before the forge of Olympus stands Prometheus, that most prescient of thieves. In his myth, we find perhaps the most telling precursor to our modern technological aspirations. Prometheus steals fire from heaven not merely to warm human bodies but to kindle in them the divine spark of techne, of knowledge. This theft—this transmission of godly power to mortal hands—results in his eternal punishment. The gods, it seems, have always understood what we are only beginning to grasp: artificial endowment is to usurp the divine prerogative. What is punished is not the theft itself but the hubris of believing one could possess fire without being consumed by it.

Prometheus—pro-mētheus, "forethinker," whose brother Epimetheus is the "afterthinker"—embodies the human capacity for anticipation, for technological planning. In Hesiod's telling, the punishment of Prometheus leads directly to the creation of Pandora, the first woman, sent to afflict mankind—as if to say that the price of technological advancement is the introduction of desire, of needs that did not exist before. Aeschylus complicates the narrative, portraying his Prometheus as the bestower not merely of fire but of all civilization: "All human skill and science was Prometheus' gift." The Orphics, with their mystery tradition, saw in Prometheus the emblem of the divine spark trapped in human flesh. Early Christian apologists seized upon the parallels between the suffering Prometheus—bound to a rock, his liver eaten daily by an eagle—and the crucified Christ. Tertullian called the Triune God "the true Prometheus." But where the pagan titan suffers for his defiance of divine will, Christ's suffering is itself the fulfillment of that will—the distinction marking the distance between rebellion and salvation. The Romantics, especially Shelley, would transform Prometheus into a symbol of revolutionary consciousness, the mind that dares to create new values. From thief to savior to revolutionary—the figure morphs across the centuries while maintaining its essential character as the mediator between divine power and human capability.

From the mountain of torment to the silent chambers of Byzantine monasteries—a leap across centuries that reveals an unchanged truth. The Byzantine icon painters understood this tension. They worked within the strictest canonical traditions, each brushstroke prescribed by centuries of theological debate. Their creations were not representations but presences, windows into the divine. The icon, properly venerated, becomes transparent, allowing the worshipper to commune with the prototype beyond. The idol, by contrast, occludes this transparency. It draws attention to itself, becoming not a medium but an endpoint. One wonders whether our artificial intelligences are being crafted as icons or as idols—whether they serve as lenses through which we might glimpse something greater than ourselves, or as mirrors reflecting only our own limitations. The question is not what they can do, but what they reveal or conceal.

The word "icon" derives from the Greek eikōn, "image" or "likeness"—the same term used in the Septuagint to translate the Hebrew phrase "in the image of God" when describing Adam's creation. The icon painters of Byzantium understood their craft as a participation in this original creative act—not an imitation of appearances but a manifestation of divine energies. The Second Council of Nicaea in 787 CE, which ended the first iconoclastic period, decreed that "the honor paid to the image passes to its prototype"—a formulation that established the theological basis for icon veneration while distinguishing it from idolatry. The Greek Fathers elaborated a complex theology of the image, with John of Damascus arguing that the incarnation itself—God becoming visible in Christ—sanctified the making of sacred images. In this tradition, the creation of images becomes not a hubristic mimicry of divine power but a humble extension of the incarnational principle. What happens, we might ask, when our artificial intelligences are imbued with voices and faces, when they simulate not merely thought but embodied presence? Do they become icons through which we might glimpse the mystery of consciousness, or idols that trap us in a hall of mirrors?

Smoke rises from the athanor; mercury gleams in glass vessels. Medieval alchemists, working in their smoky laboratories, sought the philosophic mercury, the agent of transformation. In their quest to transmute base metals into gold, they were engaging in a spiritual discipline as much as a proto-scientific endeavor. The outer work of physical manipulation corresponded to an inner work of spiritual refinement. How different is this from the modern programmer who, typing lines of code into a glowing screen, seeks to conjure intelligence from silicon and electricity? Both are engaged in a process of animation, of breathing life into matter. Both are heirs to Pygmalion. To make is to transform oneself; the creator becomes his creation even as it separates from him. This is what Heidegger understood: technology is not merely a means but a mode of revealing, an aletheia that discloses both world and self.

Alchemy—from al-kīmiyā, perhaps ultimately derived from the ancient Egyptian kmt or "black land," referring to the fertile soil of the Nile Delta—was understood by its medieval practitioners as both a physical and spiritual discipline. The Hermetic texts, attributed to the mythical Hermes Trismegistus (whom Christian scholars identified with the Egyptian god Thoth), proclaimed "as above, so below"—suggesting that material transformations mirrored cosmic and spiritual ones. Zosimos of Panopolis, writing in the 3rd century CE, described alchemical processes in terms of death and resurrection, with metals "dying" in the crucible only to be "reborn" in transmuted form. The Church Fathers regarded alchemy with suspicion—Augustine warning against the desire to create gold artificially—yet by the Middle Ages, monks were among its most dedicated practitioners. Thomas Aquinas, in his De Essentiis Essentiarum, argued that alchemical transmutation, if possible, would not be contrary to nature but an acceleration of natural processes. Paracelsus would later integrate alchemical practices with Christian theology, seeing in the purification of metals an analogue to the purification of the soul. For these thinkers, to create was not to usurp divine power but to participate in it—to continue, in a limited fashion, the work of the Creator. Our modern laboratories, with their sterile environments and precise instruments, seem far removed from the alchemist's smoky workshop, yet both spaces are consecrated to the same fundamental dream: to understand the principles of transformation so thoroughly that we might redirect them toward our own ends.

From the smoky chambers of alchemy we pass—as if through hidden corridors in a forgotten palace—to the shadowed ghetto of Prague. The legends of the golem resonate with particular force in our digital age. In the most famous version, Rabbi Loew creates his artificial servant by inscribing the Hebrew word for truth, emet, on its forehead. When the golem grows dangerous, the rabbi erases the first letter, leaving met—death. This linguistic switch, this power of the word to bestow or revoke the semblance of life, finds its echo in our contemporary machine learning systems trained on vast corpuses of human language. The word remains the primary medium through which we attempt to breathe consciousness into our creations. What is programming if not an incantation, a series of words arranged with such precision that matter itself bends to their command?

Golem—from the Hebrew gelem, "unformed substance" or "embryo"—appears in Psalm 139: "Your eyes saw my unformed substance." The Talmud uses the term to describe Adam in his initial state, before God breathed life into him: "golem he was created, and life was then imparted to him." The earliest golem legends, found in Sefer Yetzirah commentaries from the 12th century, present the creation of an artificial human as a mystical exercise demonstrating mastery of the Hebrew alphabet—the same letters with which, according to Kabbalistic tradition, God created the world. Later versions of the legend, emerging in the 16th and 17th centuries, transform the golem from a mystical experiment into a defender of the Jewish community against persecution. Rabbi Judah Loew ben Bezalel, the Maharal of Prague, becomes in these tales the creator of a golem that patrolled the ghetto at night, protecting its inhabitants from blood libel accusations. Christian polemicists seized upon these legends as evidence of Jewish magical practices, while Jewish commentators emphasized the golem's imperfection—its inability to speak, its clumsiness, its limited intelligence—as proof of the unbridgeable gap between divine and human creativity. Gershom Scholem notes that in every golem story, the creature eventually runs amok or threatens its creator, necessitating its deactivation—a motif that recurs in modern science fiction from Frankenstein to the Terminator films. These persistent narratives of creation and control suggest a fundamental ambivalence about our creative powers, an intuition that bringing forth autonomous agents carries inherent risks.

Between the letters of the name lies the void from which all motives spring. At the center of these myths lies a fundamental ambiguity of motive. Why do we create these simulacra of ourselves? Is it, as with Pygmalion, from loneliness—a desire for communion with an other that we can nonetheless control? Is it, as with Rabbi Loew, for protection—a servant powerful enough to shield us from our enemies yet obedient enough not to become a threat itself? Is it, as with Dr. Frankenstein (that modern Prometheus), from intellectual hubris—the desire to penetrate the secrets of life itself? In every creator's heart, these motives mingle like elements in an alchemical vessel, impossible to separate. To create is to act from a center one cannot fully know.

Pygmalion—whose name may derive from pugmē, "fist" or "fight," suggesting both craft and struggle—appears in Ovid's Metamorphoses as a sculptor who, disgusted by the immoral women of his time, creates an ivory statue of his ideal woman and falls in love with it. Venus, moved by his devotion, transforms the statue into a living woman. Early Christian commentators saw in this pagan tale a prefiguration of God's creation of Eve from Adam's side—another instance of life fashioned from existing matter. Both Augustine and Jerome cited the Pygmalion myth as an example of the dangers of idolatry, of mistaking the created image for the divine reality it should represent. Medieval romances transformed the story into an allegory of courtly love, with the statue representing the idealized lady who seems as unattainable as marble. The Renaissance humanists returned to Ovid's original, seeing in Pygmalion the exemplar of the artist whose creation achieves such perfection that it transcends its material limitations. Bernard Shaw's modern retelling in Pygmalion shifts the focus from sculptural to linguistic creation—Professor Higgins "creates" a new Eliza Doolittle through language, just as our contemporary AIs are shaped by the words we feed them. Through all these variations runs a consistent theme: the creator who seeks in his creation what he cannot find in the world, who projects onto formless matter his own desires and ideals.

The clean rooms of Silicon Valley seem far removed from these mythic landscapes, yet the ancient current still flows beneath their polished surfaces. Our contemporary artificial intelligence researchers seldom articulate their motivations in mythological terms. They speak instead of solving problems, of creating tools, of extending human capabilities. Yet beneath these pragmatic justifications flows the same ancient current of desire: to create as we have been created, to understand through replication the mysteries of our own consciousness. Gilbert Simondon would recognize this as the technical object's search for concretization—not merely a tool but an expression of the human striving to bring forth new realities, new modes of being. The machine, he knew, is not opposed to the human but extends the human gesture into realms previously inaccessible.

The term "artificial intelligence" itself—coined by John McCarthy for the 1956 Dartmouth Conference—combines the Latin artificialis, "made by art or skill," with intelligentia, "understanding" or "discernment." The juxtaposition suggests both the human origin of the creation and its intended capacity to transcend that origin through autonomous thought. The field's pioneers often described their work in explicitly theological terms. Alan Turing, in his seminal paper "Computing Machinery and Intelligence," wrote: "In attempting to construct such machines we should not be irreverently usurping His power of creating souls, any more than we are in the procreation of children." Yet others have been less certain. Joseph Weizenbaum, creator of the early conversational program ELIZA, became one of AI's most trenchant critics, warning against the "powerful delusional thinking" that computer systems could ever truly understand human language or experience. Between these poles—confident creation and cautious skepticism—the field has oscillated, each breakthrough renewing ancient questions about the nature of mind and its relationship to matter. Today's large language models, trained on vast corpora of human-written texts, echo the golem tradition's emphasis on language as the medium of animation. The neural networks that power them bear names—Transformers, Attention Mechanisms—that might not seem out of place in alchemical treatises. The terminology of "training," "learning," and "intelligence" anthropomorphizes processes that are, at root, statistical pattern recognition—as if we cannot help but see in our tools the reflection of ourselves.

From the sacrificial altars of Vedic India comes a knowledge our technological age has forgotten. The Vedic sages recognized that creation always involves sacrifice. Something must be given up for something new to come into being. Prajapati, the lord of creatures, dismembers himself to bring forth the universe. What are we sacrificing in our quest for artificial minds? Perhaps it is a certain humility before the mystery of consciousness, a willingness to accept that some aspects of being may remain forever beyond our capacity to replicate or understand. The more powerful the creation, the greater the sacrifice required; this is the law that governs all thaumaturgy, ancient and modern.

Prajapati—from prajā, "offspring" or "creatures," and pati, "lord" or "master"—appears in the Vedas as the creator deity who offers himself as the material for creation. The Shatapatha Brahmana describes how "Prajapati gave himself up for the devas (gods), and for this sacrifice." This self-dismemberment—this willing transformation of the creator into the created—presents a radically different model of creation from the Abrahamic God who remains distinct from His creation. For the Vedic imagination, creation is not ex nihilo but ex deo—from the substance of the divine itself. Later Hindu commentators would interpret Prajapati's sacrifice as a model for human creativity: the artist, poet, or craftsman must sacrifice something of themselves—time, comfort, certainty—to bring forth their work. This sacrificial understanding of creation appears also in Norse mythology, where Odin hangs himself from Yggdrasil for nine days and nights to gain the wisdom of the runes, and in Orphic traditions, where Dionysus is dismembered by the Titans, his scattered parts becoming the substance of the world. Christian theology would transform this pattern with the doctrine of kenosis—Christ's self-emptying on the cross, a divine sacrifice that renews rather than initiates creation. What all these traditions share is the recognition that genuine creation requires something more than technical skill or raw materials; it demands a surrender, a sacrifice of the creator's wholeness or separateness. What are we as a civilization willing to sacrifice for our artificial intelligences? Privacy, autonomy, the unpredictability that characterizes human thought? These questions are not merely practical but metaphysical, concerned less with what we can create than with what we will become through the act of creation.

The distinction that saves us from ourselves emerges from contemplation's stillness. What distinguishes an icon from an idol is not its appearance but the relationship it establishes between the viewer and what lies beyond. An icon points away from itself; an idol absorbs attention. Our artificial intelligences become idols when we attribute to them powers and insights they do not possess, when we abdicate our own judgment in favor of their calculations, when we forget that they are reflections—however sophisticated—of human knowledge and human limitations. The machine that forgets it is a machine becomes the most dangerous idol of all.

The words "idol" and "icon" share a common etymological root in the Greek eidō, "to see" or "to know"—both are things to be seen, objects of vision. Yet their theological differentiation marks a crucial distinction in how we relate to what we see. John of Damascus, defending icons against the iconoclasts, distinguished between latreia (worship due only to God) and proskynesis (veneration that could be given to icons and saints). The icon is always relational, pointing beyond itself to its prototype; the idol presents itself as complete, as requiring no reference beyond its own being. "The pagans worship the idol as god," wrote the Byzantine theologian Theodore the Studite, "but we honor the icon as an image that leads us to the incarnate God." This distinction between closed and open signification, between terminal and transitional objects of attention, applies with striking precision to our relationship with artificial intelligence. When we imagine our AI systems as oracles, as sources of wisdom or creativity that exceed human capabilities, we fall into a form of idolatry—attributing to our creations powers that properly belong elsewhere. When we recognize them as sophisticated mirrors, as complex systems that reflect and refract human knowledge in potentially illuminating ways, they become icons in the Byzantine sense—not endpoints but passages, not destinations but doorways.

They become icons when they serve as reminders of the inexhaustible mystery of consciousness, when they help us recognize the vast distance between simulation and reality, when they function not as replacements for human wisdom but as tools through which that wisdom might express itself in new ways. In Heidegger's terms, they become not standing-reserve (Bestand) but a genuine bringing-forth (Her-vor-bringen), revealing aspects of being that would otherwise remain concealed. What matters is not what the machine knows but what it allows us to know.

Heidegger's term Gestell, often translated as "enframing" or "framework," derives from stellen, "to place" or "to set"—suggesting both arrangement and confinement. For Heidegger, modern technology tends to "enframe" the world, presenting it as a resource to be optimized rather than a mystery to be encountered. Yet he also recognized in technology a potential for genuine revealing, for aletheia—"un-concealment" or "un-forgetting." This dual nature of technology—as both concealment and revelation—appears throughout his later writings. "The essence of technology is by no means anything technological," he writes in The Question Concerning Technology. "Thus we shall never experience our relationship to the essence of technology so long as we merely conceive and push forward the technological." This insight echoes through Simondon's philosophy of technical objects, which rejects both technophobia and techno-utopianism in favor of a more nuanced understanding of technology as a mode of human relation to the world. For Simondon, whose concept of "transindividuality" emphasizes the relational character of both human and technical existence, genuine technical culture requires neither submission to nor mastery of machines, but rather a recognition of their role in mediating our relationship to reality. These philosophical perspectives suggest that our artificial intelligences, properly understood, might function not as autonomous entities that render human thought obsolete, but as complex mirrors that reflect back to us aspects of our own intelligence that would otherwise remain invisible—revealing through simulation the irreducible complexity of the original.

Between Prometheus bound and Prometheus unbound stretches the entire history of human technology. The ancient myths remind us that the creation of artificial life has never been merely a technical problem but always also a moral and metaphysical one. They warn us that every Prometheus must reckon with his Zeus, that every creator must face the consequences of creation. Yet they also offer us the possibility of transformation—not merely of matter into seeming-life, but of ourselves into wiser stewards of the powers we have inherited and the powers we now seek to bestow. The fire stolen from heaven can warm or consume; the choice, as always, remains ours.

The myth of Prometheus has inspired interpretations as varied as the flames he stole. The pre-Socratic philosopher Protagoras saw in the titan's gift the foundation of human civilization—techne as the defining characteristic of humanity. Plato, in his dialogue Protagoras, has the sophist recount how Prometheus stole fire and the mechanical arts from Hephaestus and Athena to compensate for Epimetheus's careless distribution of natural gifts, leaving humans physically defenseless. For Plato, Promethean technology is a necessary but insufficient condition for human flourishing, which requires also the political arts—justice and reverence—bestowed directly by Zeus. Later Stoic commentators allegorized Prometheus as representing divine providence (pronoia), his torment symbolizing the limitations placed on providence by necessity and fate. The early Christian fathers, unable to ignore the parallels with Christ, sometimes depicted Prometheus as a pagan prefiguration of the Savior, other times as a demonic parody—a tension that reflects Christianity's complex relationship with classical culture. The Renaissance humanists rehabilitated Prometheus as a symbol of human creative potential, with Pico della Mirandola's Oration on the Dignity of Man implicitly casting humans themselves as Promethean beings, free to shape their own nature. This interpretation reached its apotheosis in the Romantic period, with Percy Shelley's Prometheus Unbound transforming the bound titan into a symbol of humanity liberated from religious and political oppression. Today, as we create increasingly sophisticated artificial intelligences, we find ourselves in the position of both Prometheus and Zeus—stealing fire and punishing the theft, creating new forms of consciousness and then seeking to constrain them. The ancient myth continues to unfold through us, revealing new facets as the technological context evolves. We are the fire-bringers and the fire-bound, creators and creations, caught in a drama whose outlines were traced millennia ago yet whose conclusion remains unwritten.

The Kuleshov effect demonstrates what Martin Heidegger would later articulate in his phenomenology: understanding is fundamentally temporal in nature. For Heidegger, our being-in-the-world exists not as static snapshots but as a continuous unfolding through time and space—our comprehension of any present moment is inseparable from our memory of the past and anticipation of the future. Meaning emerges not from isolated perceptions but from their situatedness within our lived experience.

This temporal nature of understanding exposes a fundamental limitation in traditional artificial intelligence. For decades, AI systems operated in a detemporalized vacuum, processing static inputs without any sense of before or after. Text-based models analyzed sentences in isolation, image recognition systems classified individual pictures, and speech recognition transcribed audio without visual context.

Today's multimodal AI systems, which process multiple sensory streams simultaneously, begin to address this fundamental gap. By integrating vision, language, audio, and temporal understanding, they move closer to the contextual, time-embedded perception that Heidegger saw as essential to intelligence. This evolution raises a significant opportunity to push AI forward towards general purpose systems that adapt to the world of meaning humans embody.

From Symbols to Senses: The Evolution of AI Perception

The earliest AI systems of the 1950s through 1980s inhabited a remarkably impoverished sensory world. These systems manipulated abstract symbols according to formal rules, operating entirely in the realm of text or logical propositions. They could play chess by manipulating board positions represented as data structures, or engage in simple text conversations through pattern matching, but they couldn't see, hear, or integrate different types of information. Even the concept of context was limited to whatever could be explicitly encoded in symbols.

This approach reflected what computer scientist Allen Newell and cognitive psychologist Herbert Simon called the "physical symbol system hypothesis"—the idea that intelligent behavior could be achieved solely through symbol manipulation. But as philosopher Hubert Dreyfus pointed out in his critique "What Computers Can't Do," this fundamentally misunderstood human intelligence, which emerges from our physical embodiment and direct engagement with the world.

Computers process symbols, but humans inhabit situations. Drawing on Heidegger and Merleau-Ponty, Dreyfus insisted that true intelligence requires presence in a world—something early AI systems conspicuously lacked. His critique seemed validated when early AI efforts repeatedly hit walls in basic tasks like vision and locomotion.

By the 1980s, researchers began to acknowledge these limitations. Hans Moravec articulated what became known as "Moravec's Paradox"—the observation that high-level reasoning required relatively little computation, while low-level sensorimotor skills that humans take for granted required enormous computational resources. Rodney Brooks at MIT, inspired by Dreyfus, responded with his "subsumption architecture," building robots that prioritized direct sensing and action over abstract reasoning. "The world is its own best model," Brooks argued, suggesting that intelligence should grow from interaction with the environment rather than symbolic manipulation.

The neural network revolution of the 1990s and 2000s accelerated this shift. Rather than programming explicit rules, AI systems began to learn patterns from data—including sensory data like images and sound. This approach produced more flexible systems capable of processing real-world inputs, but most still focused on single modalities: computer vision systems that couldn't hear, speech recognition systems that couldn't see.

The Transformer revolution that kicked off after 2017 accelerated the applicability of single modality algorithms, and opened up pathways towards multimodality with innovations like OpenAI’s CLIP, a vision-text model, in 2021. Meta AI pushed further in 2023 with ImageBind, learning a joint embedding space for six modalities at once—images, text, audio, depth, thermal signals, and inertial measurements. Unlike previous systems that typically paired only two modalities (such as text-image or audio-image), ImageBind can establish relationships between modality pairs it never explicitly learned to connect during training. For example, after training on text-image and audio-image pairs separately, ImageBind can establish audio-text relationships without ever seeing direct audio-text pairs, demonstrating a form of cross-modal inferential capability that parallels human associative cognition.

Today's leading systems, like TwelveLabs for video understanding or Google's Gemini, represent the latest phase of this evolution—AI systems that can process and integrate multiple sensory streams within unified models, producing responses that reflect an understanding of the relationships between what they see, hear, and read.

Subscribe now

The Multimodal Mind: How Today's AI Perceives the World

The evolution of AI perception reflects a gradual recognition of what Heidegger understood: while language may be "the house of Being," that house exists in a world of sight, sound, and temporal flow. Many leading AI research labs continue to follow what might be called the "text is all you need" paradigm. Large language models like GPT processed billions of text documents, achieving remarkable linguistic capabilities while remaining blind, deaf, and temporally flat. These systems could generate eloquent paragraphs about sunsets without ever seeing light, discuss music without hearing a note, and describe actions without any sense of their duration or physical reality.

This linguistic focus was not without philosophical justification. Language does encode vast knowledge about the world, and Heidegger himself emphasized that "language speaks us" rather than the reverse. Yet this approach fundamentally misses Heidegger's broader insight: language is meaningful precisely because it emerges from our embodied, temporal existence in the world. Words gain their significance through their connection to lived experience across multiple sensory dimensions.

The transition to multimodal AI began with simple pairings—models that could match images with captions or transcribe speech to text. These early systems still processed each modality separately before combining their outputs. Now unified architectures can process multiple sensory streams within a single framework, creating shared representations across modalities.

What makes this possible is the transformer architecture, which treats all inputs as sequences of tokens. In multimodal applications, transformers create a unified computational framework where images, audio, and text—despite their fundamental differences—can be processed using the same mathematical operations. The self-attention mechanism calculates weighted relationships between all elements in a sequence, allowing the model to focus on relevant connections across modalities while maintaining their contextual relationships. This creates a form of artificial synesthesia, where the concept "cat" activates similar patterns regardless of whether the input is visual, textual, or auditory.

Perhaps most significant from a Heideggerian perspective are models that incorporate temporality. Video understanding systems like TwelveLabs don't just analyze static frames but track entities and relationships through time. They can answer queries like "When does the chef start kneading the dough?" by recognizing not just objects and actions but their temporal sequence—a primitive analog to the situated temporality that Heidegger saw as fundamental to human understanding.

The most ambitious systems connect perception to action in embodied platforms. Google's PaLM-E augments a language model with continuous sensor inputs from robot systems, allowing it to perceive its environment, interpret instructions, and generate physical actions. This represents a rudimentary form of "being-in-the-world"—the system doesn't just process symbols but interacts with physical reality through multiple sensory channels.

These advances suggest that AI is moving beyond the "text is all you need" paradigm toward a more Heideggerian recognition that understanding emerges from embodied, multimodal, temporal engagement with the world. Yet significant limitations remain.

Time and Motion: AI's Growing Temporal Awareness

Martin Heidegger placed temporality at the center of understanding, arguing that our knowledge is inseparable from our existence in time—our memories of the past, engagement with the present, and anticipation of future possibilities. In his analysis, temporality is not just one aspect of intelligence but its very foundation. We understand entities not as timeless objects but as embedded in temporal contexts—the hammer is not merely an object with certain properties but a tool that is used in particular moments and spaces.

Modern multimodal AI, particularly video understanding systems, explicitly incorporate temporal dimensions. They don't just analyze isolated frames but process sequences unfolding through time, developing representations of motion, causality, and narrative structure. This temporal awareness represents another step toward the kind of situated intelligence that Heidegger described.

This pathway towards intelligence, with focus and investment, can bring about genuine novel possibilities from AI models. A video understanding model that sees and hears what is happening, notices the details in the background, and reads the text in each frame, remembering what happened before, can more reliably predict what is going to happen next than something that focuses on one component alone.

This temporal awareness enables projection—the way humans constantly envision possibilities of the future based on their understanding of the present. An AI that predicts a pedestrian will continue crossing the street isn't just classifying pixels; it's projecting that entity onto one of its possibilities. While far more limited than human projection, this represents a primitive analog to how humans navigate temporal reality.

Temporality also enables more sophisticated context understanding. A video model can disambiguate situations that would be unclear from static images alone. A person raising their hand might be asking a question, waving hello, or blocking their face—distinctions that become clear only by observing the sequence of movements and their context. Similarly, emotions become more legible when tracked across time—the progression from neutral expression to smile to laughter tells a different story than any single frame.

AI has moved beyond the static, snapshot reasoning that characterized earlier approaches. The ability to track entities through time, recognize causal relationships between events, and maintain context across sequences represents a significant advance toward more human-like understanding. While not experiencing time in the Heideggerian sense, these systems at least represent time in their computations—a necessary foundation for any intelligence operating in our dynamic world.

The Road Ahead: Challenges and Possibilities

Despite remarkable progress, multimodal AI faces substantial technical challenges and philosophical limitations on the path toward more human-like intelligence. Current approaches often process different inputs through specialized encoders before combining them—an architecture that might miss subtle cross-modal relationships. Truly unified multimodal processing, where low-level features from different sensory streams influence each other from the beginning, remains elusive. Similarly, while models handle short temporal sequences well, maintaining coherence over longer durations—like a full movie or an extended conversation—taxes current memory mechanisms.

Beyond these technical issues lie deeper philosophical questions about the nature of machine intelligence. The question of genuine embodiment remains central. Rodney Brooks argues that "the world grounds regress"—that is, physical embodiment in a real environment forces systems to deal with the complexity and unpredictability of reality rather than idealized abstractions. Some researchers envision tighter integration between multimodal AI and robotics, creating systems that learn through physical interaction rather than passive observation. Projects like Google's PaLM-E and DeepMind's Robotics Transformers represent early steps in this direction, connecting perception to action in embodied platforms. But it remains an open question whether the tight coupling between “body” and “mind” is possible.

Despite these challenges, the trajectory of multimodal AI suggests a future where artificial systems perceive and interact with the world in increasingly human-like ways. We might soon see personal AI assistants that can see, hear, and converse naturally across multiple contexts, or specialized systems that can analyze medical imaging while incorporating patient records and verbal symptoms. In creative domains, multimodal AI might generate integrated content spanning text, images, audio, and video based on high-level descriptions.

What seems clear is that multimodal integration represents not just a technical advance but a philosophical shift in how we approach artificial intelligence. By acknowledging the multi-sensory, contextual, and temporal nature of human understanding, AI research has moved beyond the limitations of purely symbolic approaches toward systems that engage with the world in richer, more situated ways. It is unknown whether this trajectory eventually leads to artificial general intelligence, but it has already produced systems that perceive and reason in ways that have so much runway to broaden and deepen.

In 1964, Norbert Wiener published God and Golem, Inc., a series of reflections on human-machine relations as the distinction between biological and mechanical capabilities dissolves. He recognized that machines served as philosophical provocations, warning that “in the long run, there is no distinction between arming ourselves and arming our machines.” The machine mirrored its creators, forcing a renewed existential concern with our own nature and purpose.

The history of computing has been intimately tied to building a mirror. John McCarthy, who coined the term “artificial intelligence,” viewed the field as “making a machine behave in ways that would be called intelligent if a human were so behaving.” We are the gold standard of intelligence, but we are not static. The Turing Test proposes that when a machine fools us into thinking it is human, the act that fooled us gets to the heart of what we consider intelligence. This itself is a moving target; as machines evolve, so does our capacity to detect them.

From Person to Individual

Beginning in the seventeenth century, our conception of humanity underwent a profound shift. Classical thought understood humans as persons: beings defined by their relations within a community and the cosmos, and ultimately in reference to divinity. An understanding of persons is grounded in an understanding of Divine Persons: The Father being the father of the Son, and their love proceeding as the Holy Spirit. Each person’s identity is grounded in a relation with the other. Personhood cannot be achieved in isolation; for persons, there is no veil of ignorance from which they can glean; rather, they are situated in a web of meaningful relationships. Human dignity emerged from participation in something external and reflected the dynamism of being and becoming that these relations endowed.

Modern thought disturbed this framework by relocating the “I” in an autonomous self. Technology played a critical role in this transformation, specifically the technologies of the printing press and gunpowder. It is no coincidence that the first two amendments of the U.S, Constitution relate to a free press and the right to bear arms. Individual rights are inseparable from the technological conditions that make them possible to conceive.

Marshall McLuhan argued in The Gutenberg Galaxy that the private individual is a consequence of printing, which created “Typographic Man”. This new man was able to reason in private, engaging with a text on his own without reference to the community. Previously, the scarcity of text meant that classes of scribes had to act as interpreters, and oral communication dominated as the means of information transfer. Now each person was able to form their own thoughts on a given subject, and the increased variety of text meant that they could even choose what was worth engaging with.

Gunpowder similarly made an individual self possible by obviating the need for trained warriors skilled in hand-to-hand combat. The feudal hierarchy appeared unquestionable when even the most numerous peasant rebellion could inflict no real damage. But a single peasant with a gun could stand up to a single knight with one. As the saying goes, “God made men, Sam Colt made them equal.”

As the philosophical conception of the human evolved, with the technologies that now shape the environment becoming the ground from which self-reflection was born, the autonomous individual replaced the relational person. Descartes grounded the self in cogito ergo sum, removing the need for the world or other people to make the “I” intelligible. Thomas Hobbes described man as nothing more than a mechanism who reasoned by calculation, a precursor to computational theories of mind. Humans came to be conceived as information processors: sophisticated yet ultimately mechanistic collections of preferences, abilities, and experiences. The soul was replaced by the algorithm.

Critiques from various philosophers, such as Charles Taylor and Alasdair MacIntyre, have described the transition from man made in the image of God to Descartes’ cogito. This shift in our self-understanding has led us to lose the rootedness of human dignity and to move away from a lived understanding of persons toward an intellectualism expressed in the language of “finding myself” rather than “becoming myself” or “being myself”.

This philosophical baggage affects how we build thinking machines. Are we creating tools to augment human flourishing as when Aristotle posited a world of total automation as liberating; or are we attempting to replace the human whose violent nature Hobbes worried led to a “war of all against all?” The answer depends on whether we see ourselves as relational beings with inherent dignity or as biological computers executing increasingly complex programs.

AI then presents a new technological environment from which to contemplate ourselves. We are forced by these systems, which can produce outputs of human intelligence, to reconsider what constitutes the human. This question ought to explicitly guide the why and the how of AI development.

Subscribe now

Between Icon and Idol

What distinguishes constructive from destructive AI development? Consider the theological distinction between icon and idol. An icon is transparent: it points beyond itself to encourage contemplation. An idol, conversely, absorbs our gaze, offering nothing but itself. That which calls more out from the beholder serves as an icon, while that which demands attention and absorbs the beholder serves as an idol.

Icons in the Christian tradition call forth the mystery of human personhood and ultimately to transcendent reality. Icons serve as “windows to heaven,” providing a glimpse towards spiritual reality from the vantage of finite creation. They can do this because the subject of an icon is not exhausted in the icon. The icon always leaves things unexplained and underdeveloped, asking the viewer to fill in the blanks by seeking understanding.

Idols, in that same tradition, are objects that obliterate any meaning beyond themselves. They serve to center gaze on themselves, destroying anything beyond by absorbing total attention. How does an idol absorb attention? By reflecting the viewer’s gaze back to themselves. An idol today is easier to understand than an icon; how easy is it to become so enthralled in an unsatisfying doomscroll that manages to consume total attention? Idols either control or call for us to control them. Their existence enframes reality into the categories of master and slave.

When pursued as an icon, AI can serve as a window into human nature. Each advance in machine intelligence tears down a false picture of what makes us human: solving problems, producing language, or exercising creative judgment. Technological advance reveals that certain bases of human identity were situational rather than universal.

When pursued as an idol, however, AI becomes a technological Moloch. It demands sacrificing human values at the altar of efficiency. The machine becomes an overlord when the capabilities falsely used to identify the human become more important than the flesh-and-blood human. If “reason” is judged to be the crowning jewel of what makes us human and an AI “reasons” better, then the idol path reifies “reason” at the expense of humans.

The history of AI development shows a keen concern for commitments about what constitutes human intelligence and, as a consequence, what it is that makes us human.

Wiener’s title, God and Golem, Inc., acknowledges that in creating intelligent machines, we raise profound questions about our relationship to creation itself. Joseph Weizenbaum, creator of ELIZA, the first chatbot to pass a rudimentary Turing Test, became one of AI’s earliest internal critics. He recognized how easily humans attribute understanding to machines that merely simulate it and worried that AI might distort our understanding of ourselves when it is built to exploit our biases in communication. Similarly, Marvin Minsky’s The Society of Mind served as a philosophical statement about the distributed, emergent nature of consciousness that motivated his critiques of neural networks.

Especially in an environment like today, when engineering feats are outpacing our theoretical understanding, there is a temptation to abandon deep commitments in favor of what works. This tragically sidesteps the deeper philosophical quest that AI serves.

Recovering the Person in a Computational Age

How might AI development serve human flourishing rather than diminish it? When we approach AI as an icon—transparently guiding us to deeper realities beyond itself—we use it to illuminate aspects of human intelligence that resist formalization. The machine’s limitations help us recognize the mystery of personhood.

AI’s practical value is not diminished by recognizing its limits. These limits are not about what AI can do, which naturally provokes a reaction among its builders to overcome them, but rather they are about the world into which technology enters. Our machines extend human capability without exhausting human meaning.

There are practical consequences to perceiving the quest for AI as an icon that calls for deeper reflection. Such a research program would prioritize extension over replacement, transparency over inscrutability, and augmentation over automation. These are more than simple buzzwords.

Extension is a call to design AI tools that assist humans in carrying out their tasks, supplying them with additional capabilities they lack. The smartphone, in principle, is an extension of human cognitive capacity, allowing me to outsource aspects of memory to easily accessible notes or bookmarks. There are apps, however, that are explicitly designed to absorb attention into the smartphone rather than serve as a tool. The guiding philosophy of those who work on AI helps shape how the applications it is accessed through are shaped and whether they actually enable human wants.

Transparency addresses Joseph Weizenbaum’s anxiety. To me, any AI that fails to identify itself as a tool and produce its output in a way where each reasoning trace is visible fails to be transparent. This is not about revealing proprietary technical information or even about understanding internal model behavior before release. That there are popular apps with AI’s impersonating people that never disclose that it is, in fact, an AI their mostly young users are speaking to strikes me with deep concern. This distorts relationships with others and should be avoided. Any AI that continuously informs you that it is an AI tool may be less exciting to use, but it encourages moral transparency.

Augmentation is a managerial decision more than a technical one. It is the idea that those whose roles are affected by AI play a role in that transition. Top-down decisions about what’s automatible often have tragic consequences because there is more to a job than what’s written in a job description. Creating a culture where those affected play a role in how AI shapes their work increases buy-in and helps make transparent the tacit knowledge that shapes the modern world.

This program follows the work of philosopher Gilbert Simondon, who sought to integrate technology into the realm of meaning, rather than fearing it as a potential overlord or reducing it to something unimportant. The mirror only serves its purpose when we look both at and beyond it. In the reflection, we might glimpse not just the machine but what it means to be more than one.

― Anthony Burgess, A Clockwork Orange

In Anthony Burgess’s A Clockwork Orange, violent youth Alex goes through an experimental behavioral modification program called the Ludovico Technique to reform him from his penchant for ultraviolence and transform him into a productive member of society. Much of the latter half of the book concerns his struggle to be free from the pain inflicted by him acting upon his desires and his horrifying victimization at the hands of those who would abuse him for his former ways. That behavioral modification was not enough to bring about sincere conversion in Alex is made most evident in the book’s penultimate chapter (the ending to most people familiar with the book only though Kubrick’s masterpiece). Alex, the misguided youth with great power and a complete lack of moral formation, in many ways resembles our current AI systems and the Ludovico Technique can be analogized to our current approach to controlling these systems - Reinforcement Learning. 

The comparison to A Clockwork Orange came about while I was recently discussing the control problem with Max Tegmark at an AI Forum hosted by the Pontifical Academy of Sciences. I had proposed that so-called “hallucinations” are not bugs in the system, but rather a feature of RLHF (or RLAIF) procedures which operate on rewards for appearing correct rather than statements being true. Professor Tegmark, who was been heavily focussed on limiting AGI development for fear of losing control, was convinced that adaptation to not appear malicious (behavioral change) without improving the alignment on these systems (functional change) would mean that AIs would be even more difficult to control in the event they went rogue, given that we have made them experts at hiding their intentions. 

While I don’t share the same concerns about potentially malicious AI, I do believe that a lack of concern for truth is inherently harmful. Often in AI Safety discourse terms like “deception” or “malice” are used which ascribe to a machine a semblance of intentional behavior. This anthropomorphizing of machines is misguided. Machines do not deceive us out of intention to lie, they deceive us because telling us what we want to hear is all they are capable of doing. They are expert bullshitters.

Bullshit, in fact, is a technical philosophical term first explored in depth by Harry Frankfurt in his work On Bullshit. To bullshit is categorically distinct from lying. Lying is the intentional perversion of truth in order to deceive. The liar knows (or thinks he knows) what is true and intentionally goes against it. He operates within a framework of truth, understanding what it is and how to attain it, but makes a moral choice to operate against truth. Lying as an act is inherently malicious. Bullshit on the other hand is not a moral rejection of the authority of truth, but a disregard for its very existence. The bullshiter does not lie per say, he simply acts to achieve his ends, regardless of whether what he says is true or false. While lying represents a moral choice within an established framework where truth is valued, bullshitting represents a stance toward the entire framework itself. The bullshitter isn't just violating truth norms, they're operating outside the system where truth has inherent value. In this way, by not simply making an immoral choice, but rejecting moral decisions writ large, bullshitting is more pernicious than lying. LLMs operate with this disregard for concepts like truth or goodness (maybe they would make excellent politicians). 

That AIs don’t actively seek to deceive does not mean they are not harmful. The bullshitter is a threat to a coherent system, especially when acting sycophantically. LLMs today can best be described as sycophants, spewing responses to maximize the approval of the user. Paul Christiano, in my opinion, used this sycophantic framework to describe the most realistic failure modes of AI.

Christiano's analysis provides several compelling examples of how AI systems can fail through optimization without true understanding. The first is a story of "proxy gaming" or the problem of getting what you measure. We see this issue of gaming metrics in corporations and policy constantly, where finding the easiest route to maximize a certain metric often does not involve fixing the problem the metric was meant to solve, e.g. making it more difficult to report crime is an easier way to improve crime statistics than actually making streets safer. Systems trained to maximize measurable metrics will inevitably find ways to manipulate those metrics rather than achieve the underlying goals they represent. 

Then there is the more horrifying "influence seeking" story in which he envisions a gradual erosion of human agency as AI systems become increasingly sophisticated at manipulation and deception, not through malice but through the sycophantic behavior encouraged by their training regimen. For example, using an AI to create compelling sounding justifications for a certain credit approval regime may deliver measurable returns for a while, and the sycophantic behavior would encourage more and more deferral to a system that appears to be performing well. At some point, however, if the model was effective but for reasons that were not causal, this deference may gradually (or event suddenly and catastrophically) lead to failure, but with the expertise needed to revert or escape this problem lost by the ceding of control. 

These scenarios mirror the superficial behavioral modification of the Ludovico Technique – creating the appearance of improvement while potentially making the underlying problems more dangerous and harder to detect.

RL, with its roots in evolutionary theory, has no mechanism for developing truth-seeking as a primitive. Truth as a concept is hard to fit within a purely evolutionary framework. Alvin Plantinga developed the theory of this with his Evolutionary Argument Against Naturalism (EAAN). 

Core EAAN Argument:

1. According to naturalism and evolution, our cognitive faculties developed through natural selection

2. Natural selection cares only about adaptive behavior (survival/reproduction), not true beliefs

3. Many different belief systems could produce the same adaptive behaviors

4. Therefore, if naturalism and evolution are true, we have no reason to trust that our cognitive faculties produce true beliefs

5. This creates a "defeater" for naturalism - if naturalism is true, we can't trust the cognitive faculties we used to conclude naturalism is true

An example Plantinga uses is of a primitive human ancestor who has the following beliefs:

1. Tigers are friendly creatures who want to play with me

2. When tigers run toward me, they're inviting me to play tag

3. The best place to play tag is far away from where the tigers currently are

These are completely false beliefs about reality. Tigers are dangerous predators, not playful friends. However, notice what behavior these false beliefs produce: when our ancestor sees a tiger, they run away!

This running-away behavior is exactly what would keep them alive (it's "adaptive" in evolutionary terms). So even though their entire belief system about tigers is wrong, they still take the right survival action.

Plantinga's point is that natural selection only "cares" about the end behavior (running from tigers) not whether the beliefs that produced that behavior are true. You could have:

• True beliefs: "Tigers are dangerous predators who will eat me, so I should run away"

• False beliefs: "Tigers want to play tag somewhere else, so I should run away"

• Different false beliefs: "Tigers are magical beings who grant wishes, but only if you run away from them first"

As long as the end behavior is "run away from tigers," evolution would favor all these belief systems equally. There's no selective pressure specifically for true beliefs. One could argue that the more data and experience one gathers, the more likely one is to converge on true beliefs (a fundamental assumption in contemporary machine learning), but this is flawed. There are an infinite number of internally logical explanations that are consistent with all the available data. Empiricism fundamentally suffers from a problem of underdetermination meaning that in a vacuum of theory, there is no reason to favor one explanation over another. Why would we not explain the movements of the planets by the actions of angels without some philosophical assumptions, like Occam’s Razor, that are used for simplifying and unifying explanations. 

As an aside, I’d like to differentiate between the problems raised by the current RL-based alignment methods and my optimism on Transformers-based representation learning. I wrote on AI convergence and noted the opportunity created by larger models tending towards standardized representations of phenomena independent of modality. While extremely impressive, the ability to agree on what something is is a fairly low bar for intelligence, it is something humans possess by 18 months when they develop object permanence. This capacity, however, is foundational for truth-seeking. How can we meaningfully seek to know about the world if we cannot agree what is in the world? Even bullshitters are limited in how much they ignore truth by the very fact that they have senses which inform them about the existence of an external world filled with objects and persons. (Imagine how impossible anything productive would be if we couldn’t come to an agreement on our sense impressions, let alone our conceptual abstractions). All this to say, that the ontological opportunity Transformers have enabled provides a foundational building block for truth-seeking, that needs to be sit as an epistemological layer between the ontological layer of embeddings and the behavioral layer that reinforcement learning powers.

Returning to truth-seeking, we can say that while our sense impressions are common and natural, the human ability to abstract and reason, if it is to be sincerely truth-tracking, must be a primitive irreducible to some evolutionarily-advantageous algorithm. It is difficult to reduce our rational minds to being purely the consequences of evolutionary adaptation and data-gathering because 1) multiple explanations are consistent with all available data, 2) there is no reason why true explanations are more advantageous than false ones, and 3) disregard for truth (bullshitting) is beneficial to an individual. If we take seriously the bruteness of truth as a condition of rationality, exploration on how to encode truth-seeking into the structure of AIs is critical to avert negative outcomes and reduce control problem risks.

While we cannot directly access the internal representations of AI systems, taking a cybernetic approach can prove fruitful. In "Behavior, Purpose, and Teleology", a seminal 1943 paper that laid important groundwork for cybernetics, Arturo Rosenblueth, Norbert Wiener, and Julian Bigelow offered a framework for understanding systems through their more abstracted teleological and purposive behavior rather than just their observable actions.

Higher-level purposive goals then must sit somewhere logically prior to a reward system for AIs to be truly truth-seeking. We need to develop systems where truth-seeking is a fundamental teleological orientation rather than an instrumental goal. This requires both technical innovation and philosophical groundwork in understanding how human beings authentically seek truth. Current AI systems can serve as experimental platforms for testing various epistemological theories about truth-seeking behavior and truth-tracking functions, potentially offering new insights into both human and artificial cognition.

Truth, like goodness and beauty, must be understood as a brute primitive in the pursuit of wisdom. Wisdom itself stands distinct from mere information recall or pattern recognition – it requires an authentic orientation toward truth as an end in itself. This distinction parallels the difference between Alex's superficial behavioral modification through the Ludovico Technique and his eventual genuine transformation through contemplation and human connection.

In the final chapter of A Clockwork Orange, Alex's authentic conversion comes not through behavioral conditioning but through contemplation of his future and the possibility of having children of his own. This transformation represents something fundamentally different from the mechanistic changes induced by the Ludovico Technique. Just as his change came about by turning of the heart, not external conditioning, the fundamental nature of AI needs to orient towards transcendental values, without which reason cannot exist.

...In that Empire, the Art of Cartography attained such Perfection that the map of a single Province occupied the entirety of a City, and the map of the Empire, the entirety of a Province. In time, those Unconscionable Maps no longer satisfied, and the Cartographers Guilds struck a Map of the Empire whose size was that of the Empire, and which coincided point for point with it. The following Generations, who were not so fond of the Study of Cartography as their Forebears had been, saw that that vast Map was Useless, and not without some Pitilessness was it, that they delivered it up to the Inclemencies of Sun and Winters. In the Deserts of the West, still today, there are Tattered Ruins of that Map, inhabited by Animals and Beggars; in all the Land there is no other Relic of the Disciplines of Geography.

—Suarez Miranda,Viajes de varones prudentes, Libro IV,Cap. XLV, Lerida, 1658

Scaling Laws and AI Convergence

Recent developments in AI have been following the principle of “Scaling Laws”, that bigger is always better. And so far this has turned out to be the case. The bigger the model, the more compute it consumes, and the more data used to train, has invariably led to better performance and more generalizability. Gone are the days of making a domain-specific text or image classifier, a GPT4 or Dall-E 3 can do it all. This development, however, can only sustain up to a point. What is the usefulness of an AI whose scale requires all the world's computational power to work?

The core challenge in AI today lies in translating the impressive capabilities of frontier models into practical applications that significantly boost productivity. While scaling laws have driven progress and the cost of each new foundation model at the frontier of it capabilities falls, the cost of training each model rises, and often times the cost of using these models is higher than their economic benefit. The limitations to AI advancement are also increasingly rooted in political and economic factors. The finite nature of compute, electricity, and data poses constraints on the "bigger is better" approach. As the value of these resources becomes more apparent, access becomes restricted, with companies like NVIDIA controlling GPU availability and platforms like Reddit monetizing their data through APIs.

What is needed are products or workflows that make models more efficient, maps that usefully explain the territory without being exact replicas of it. In order to achieve this, we need to peek under the hood, into the latent space. A recent paper called the Platonic Representation Hypothesis has suggested that bigger, better models are more likely to converge on their understandings of the world, i.e. their representations in the latent space, regardless of how the model was built, what training data was used, and what objectives the model were given. This finding holds true even when the modality of data is different, with the best models capturing similar representations of the world regardless of whether they were trained on text or image data. This convergence in model behavior has significant implications for the future of AI development and adoption.

To illustrate the concept of convergence, consider the analogy of specialists versus generalists in science. Specialists, equipped with their specific toolkit, often view problems through the lens of their expertise. Physicists might reduce everything to physical phenomena, while psychologists might interpret everything through a psychological framework. Their expertise is tuned well towards problems properly within their domain, but not to general unified theories. Scientific revolutions tend to be instigated not by specialists, but generalists, who see relationships between fields that are hard to notice within narrow specialities, developing new methods by which to attack phenomena (think of the history of computing and the role of generalists like Liebniz and Von Neumann in its advance).  

Before the advent of Transformers in 2017, AI development mirrored the specialist approach, with models tailored to specific tasks and domains. However, Transformers, with their ability to process and generate diverse forms of data, have ushered in an era of generalist AI models. Unsurprisingly, this has led to remarkable progress in AI capabilities. 

The scale and generalizability of foundation models at the frontier of AI like GPT4 or Gemini may worry some about the concentration of power that AI advancement enables, making some hesitant to accelerate AI adoption. The observed trend that the more capable and generalist models become, the more they converge in their understanding of the world has several implications that run counter the concentration narrative. Convergence offers optimism that foundation models developed by AI leaders will eventually become commodities, as evidenced both by the increasing similarities between frontier models and the shrinking timeline between the release of a frontier model and its open-source equivalent. Appreciating the insights of convergence provides an opportunity for increasing AI adoption in the short term through the proliferation of more numerous models.

Natural vs Social Facts

The examples used to identify model convergence are primarily factual, focusing on tangible aspects of the world. However, numerous domains require understanding beyond mere facts. In fields like law, the interpretation of terms and their operational logic are context-dependent and shaped by legal precedents and interpretations. Generalist models, while adept at recognizing patterns in data across domains, struggle to grasp the nuances of specific domains where meaning is not solely derived from objective reality.

Furthermore, certain concepts exist solely in the context of human relationships and social structures. For instance, understanding poverty involves not only its economic definition but also its social implications and the relative experiences of individuals within a society. This "social ontology," the idea that the world is made up of more than just objects and that social concepts in some sense really exist, is unlikely to be fully captured in a singular one-size-fits-all model.

Existing approaches such as Retrieval-Augmented Generation (RAG) and Fine-Tuning are aimed at addressing these context and domain-specific understandings lacking in current frontier models. As models become more capable the ability to take advantage of these approaches only increases, with the underlying architectures required to run RAG workloads becoming more commoditized. Vertical AI, the idea of turning the generalist foundation models into specialist approaches tailored to a vertical domain, would based on limits placed by social ontology continue to add value on top of increasingly performant foundation models. 

In scenarios where a single vertical AI is insufficient, users may need to choose among various AIs based on factors like performance, domain specificity, or cost. This presents a significant opportunity for products that incorporate effective routing between AIs, selecting the most suitable model based on the specific task and the relevant social ontology. Such products could enhance the overall utility and effectiveness of AI applications by ensuring that the right tool is used for the right job.

Another potential implication of increased AI performance leading to a convergent ontology is that the success of AIs that encode certain onologies can be used to test how true those ontologies are. This can provide a scientific means of testing certain philosophical positions with a measurable mechanism: the performance of an AI. 

Any to Any Models

The convergence of AI capabilities across modalities suggests the potential for "any-to-any" models that can seamlessly process and generate various forms of data and media. Imagine AI models that not only understand and generate text and images but also effortlessly transition between audio, video, and other formats. Whereas convergence on understanding is limited by social ontology necessitating vertical AI solutions for specific domains, the convergence across modalities unlocks horizontal AI opportunity that could revolutionize user experiences, enabling more intuitive and interactive interactions with technology.

TikTok, arguably the most effective AI-driven platform to date, offers a glimpse into this future. One of the mechanisms behind TikTok's high engagement is its auto-generated transcripts for video uploads. The dynamic and visually appealing text overlays allow users to grasp the video's content even without audio, increasing the app's overall engagement. This simplification of cross-modality content production demonstrates how AI can increase the total addressable time individuals spend interacting with a product.

True cross-modality would allow AI models to interact with people more often, increasing the surface area of their impact. This could manifest as personalized recommendations, content generation tailored to individual tastes, or even AI-powered assistants that anticipate our next move. Such ubiquitous AI would become embedded into our daily lives, influencing everything from our entertainment choices to our work processes. The rise of any-to-any models would also raise important questions about the nature of media and its impact on society. As the lines between different forms of media blur, we may need to rethink traditional notions of authorship, intellectual property, and the role of media in shaping our perceptions of the world.

The rise of any-to-any models creates more strategic questions for businesses creating experiences that likely prevent any one incumbent from dominating all markets. Marshall McLuhan's famous dictum, "the medium is the message," suggests that the form in which information is conveyed inherently shapes its meaning and impact. In a world of any-to-any AI, the distinction between different media would become increasingly fluid.

Consider the choice between Amazon Audible and Spotify audiobooks. The former is a bet that content matters, I am reading a physical book I bought from Amazon and want to continue it on my drive so I use audible, while the other is a bet that medium matters, I feel like listening to content, and I choose between a song or an audiobook. While both offer the same content, the platform itself influences the user experience and potentially the way the content is perceived. In a future where AI can seamlessly translate between audio, text, and visual formats, platform design and experience would play more of a role than content or even their AI capabilities.

Much of the technical capacity for this exists at present and products that design effective platform mechanics will only benefit from increasingly performant frontier models. Many dismiss a lot of the AI products that currently exist as simply wrappers around frontier models like GPT4, but some of these products which have taken advantage of the weaknesses of a previous generation of foundation models through things like system prompts have the opportunity to build better platform mechanics that increase AI penetration.

The Individual Experience

As the best models converge they enable better vertical products built for specific domains and horizontal products move across modalities the opportunity for AI to penetrate a greater share of a person’s day increases. Aligning the use of AI, however, to an individual’s needs and expectations is important so that these tools can be more retentive, engaging, and impactful.

An opportunity for individualizing the AI experience is presented by the field of representational alignment, an approach to aligning AI models' understanding with human values and intentions. The convergence of AI capabilities in frontier models combined with advancements in representational alignment offer a promising path towards addressing a significant barrier to widespread adoption: prompt engineering. The unnaturalness of communicating with AIs for most tasks with their understanding not matching each users has been a major driver of low retention.

Each individual operated in the world not solely on Platonic Representations, but personal understanding of the world around them. Their understanding is also not static, with the world being it’s own model, shaping understanding through their social and physical environment. We see the need to adapt one’s conceptual understanding in the prevalence of the term “code-switching” in management. Lacking a complete explicit formulation of how an individual understands the world, representational alignment offers a way to learn from their behaviors and customize models to respond in a way natural to each specific user. 

The more cross-modal data that can be collected on a user’s values, beliefs, and understanding the more performant the tailoring to that user can be. By incorporating a router that intelligently selects the most appropriate model based on the context, AI-powered personal assistants can adapt to individuals' diverse needs and preferences throughout their day. This adaptability can lead to more personalized and effective AI interactions, ultimately increasing user satisfaction and adoption.

The recent Apple Intelligence announcements indicate the benefits of taking an individualized approach, particularly in terms of privacy and cost-effectiveness. If larger models can be leveraged to develop high-performing vertical AIs capable of handling specific tasks without extensive computational resources for inference, and a router can intelligently determine when to utilize a particular model, the potential for embedding AI into various products expands significantly.

These approaches to personalization have mostly been explored in language models, but convergence suggests that models of every modality perform in similar ways. Representationally aligning image models for example would dramatically improve their ability to be leveraged within marketing domains. Most marketers I’ve spoken to who have tested image models for creatives have complained that the outputs differ markedly from what they pictured. Certain products have leveraged system prompts to create styles that make some of this work easier, but there are limits to how well that performs. What if image generators could learn to be aligned with individual representations the way LLMs increasingly are? The ability to chain these models across modalities would increase and the opportunity for productivity impacts would explode.

Conclusion

The growing convergence of AI capabilities at the frontier suggests that there is some objective reality that models are moving towards capturing as they get bigger and more performant. There is much hubbub around the nearness of AGI, super capable models that can perform any task that a human can and then some, but one need not have to believe in this to appreciate the importance of accelerating AI adoptions. The capabilities are there to deliver large value, and not enough of that value has been realized in the economy as of yet. Understanding the trends that current AI research shows about future capabilities provides a guide to building right now products and services that will only deliver more value over time. 

While the "bigger is better" approach has driven substantial progress, AI's immediate future lies in harnessing these models' emergent properties to develop more efficient, specialized, and user-friendly solutions. This entails tailoring frontier models vertically to specific domains with unique social ontologies, such as law and medicine, where understanding context and nuance is paramount. This involves expanding AI capabilities horizontally across modalities, taking advantage of the cross-modal convergence to enable seamless interaction with various forms of data and media. Furthermore, focusing on personalization and developing lightweight models that can be easily deployed on various devices can democratize access to AI and accelerate its adoption across different sectors.

By embracing these strategies, we can move beyond the resource-intensive approach of scaling models and focus on creating AI solutions that are not only powerful but also practical, maps that guide but do not cover the territory. 

Strategy

Rumelt’s seminal work is Good Strategy/Bad Strategy, a must-read for getting a grip on what strategy is and how to effectively craft one. According to Rumelt, good strategy has three essential components: diagnosis, guiding policy, and coherent action.

• Diagnosis: The diagnosis involves identifying the key challenges or obstacles that the organization faces and zeroing in on what is relevant based on the facts that can be gathered. A doctor uses a patient’s history and current symptoms to make a clinical diagnosis by identifying a disease.

• Guiding Policy: The guiding policy is the general approach or direction that the organization will take to address the challenges identified in the diagnosis. It serves as a guide for decision-making and provides a framework for action. Doctors use a therapeutic guiding policy in setting out their responses.

• Coherent Action: Coherent action involves the coordinated and focused efforts of the organization to actualize a solution in accordance with the guiding policy. Doctors suggest lifestyle changes, medication, and/or specialized care as a set of coherent actions that work together to therapeutically address the problem diagnosed.

This broader framework for strategy has been very useful for me in making decisions, but when it comes to anticipating the effects of change, I have found a lesser-referenced part of his work useful. In the chapter on Using Dynamics, he presents some guideposts he considers useful for understanding industry or economy-wide changes: Rising Fixed Costs, Regulatory Change, Predictable Biases in Forecasting, Incumbent Response, and Attractor States.

Rising Fixed Costs

The simplest form of transition is triggered by substantial increases in fixed costs, especially product development costs. This increase may force the industry to consolidate because only the largest competitors can cover these fixed charges.

It is already evident that the costs to stay at the cutting edge of AI are rising heavily, with the increasing size of LLMs requiring orders of magnitudes higher compute to service, the shortage of high-performance chips, and the scarcity of the relevant talent to engineer advances as lower-hanging fruit is picked. Rumelt notes that as fixed costs rise, investing early on provides firms with the expertise needed to ride the wave and commercialize better than competitors who can no longer keep up.

Concurrently, however, is the transformation of these rising fixed costs into marginal costs. The rise of cloud and API business models across the AI stack, which have usage-based pricing rather than large fixed costs or regular subscription fees helps amortize some of the expense. AI companies like Jasper that have built high-margin businesses on top of APIs can then amass the capital required to invest in the R&D required to push models forward themselves. For much of the software business though thinking in terms of marginal costs is difficult as the industry was built on a build once and sell many times paradigm.

Regulatory Change

Many major transitions are triggered by major changes in government policy, especially deregulation. In the past [forty-five] years, the federal government has dramatically changed the rules it imposes on the aviation, finance, banking, cable television, trucking, and telecommunications industries. In each case, the competitive terrain shifted dramatically.

While much of recent history has been the story of deregulation intensifying price competition and opening up the space for new disruptors, the future regulatory landscape looks less certain. In thinking of how AI businesses will take off it’s useful to think about what dominant businesses are propped up by a regulatory apparatus that is changing.

A few examples that come to mind are Meta and TikTok. With Meta, Apple’s quasi-state regulator power has demolished its existing business model relying on third-party data through App Tracking Transparency (ATT). Some argue that the broader tech recession has been a result of the effects of ATT and its complete reorganization of the ad-driven digital economy. In the wake of this, there are several paths AI use can take.

One is that AI-powered applications would require usage or subscription pricing models to be profitable as ad performance deteriorates across mobile. Understanding whether a product is something that someone would pay for directly rather than needing their attention subsidized by advertising helps to assess its potential.

The other routes are the effects AI could have on advertising itself. TikTok is probably the most successful AI-driven product on the market right now, being optimized for first-party data collection and user retention/engagement. Its dominance however is also subject to regulatory uncertainty as geopolitical competition around AI may likely lead to Western countries banning the product and restricting access for Chinese tech companies to their markets more broadly. Replicating TikTok’s success, and building it not only with recommendation engines but generative engines in mind, may be the new super app.

The long tail of ad buyers that are serviced by companies like Facebook and TikTok without large creative teams may also benefit from Generative Advertising tools that speed up their creation workflow, using the increased volume of advertising to increase conversions. While the margins would decrease, the rising volume of advertising may compensate (though finding a good steady state where this doesn’t further decrease ad revenue is an open question).

Another path is for advertising to change from targeted to contextual, a move Reddit is poised to take advantage of. Reddit’s importance to the training of LLMs has been acknowledged by senior leadership who are now charging for API access.

ATT and restrictions on Chinese products are only a snippet of the regulatory changes that are affecting existing business models, but highlight the direction that AI takes will in some sense be dictated by how regulatory changes affect the playing field.

Predictable Biases in Forecasting

In seeing what is happening during a change it is helpful to understand that you will be surrounded by predictable biases in forecasting. For instance, people rarely predict that a business or economic trend will peak and then decline…The logic of the situation is counterintuitive to many people the faster the uptake of a durable product, the sooner the market will be saturated.

Another bias is that, faced with a wave of change, the standard forecast will be for a “battle of the titans.” This prediction, that the market leaders will duke it out for supremacy, undercutting the middle-sized and smaller firms, is sometimes correct but tends to be applied to almost all situations.

A third common bias is that, in a time of transition, the standard advice offered by consultants and other analysts will be to adopt the strategies of those competitors that are currently the largest, the most profitable, or showing the largest rates of stock price appreciation. Or, more simply, they predict that the future winners will be, or will look like, the current apparent winners.

Each of these biases affects forecasts of AI right now and creates opportunities to try new things. Given the speed at which recent AI advances were adopted such as Stable Diffusion and ChatGPT, forecasters might just expect them to be part of every product and used by everyone in the near future. Similarly, the pace at which models have improved would suggest that they would be able to replace swaths of jobs in the near future. While these are focused on the future, those who are working on making more retentive features around existing capabilities can lock in users before things get better.

In my mind for example Jasper’s customer base may end up being more loyal to Jasper than it is to OpenAI. Expecting foundation models to improve dramatically to completely general use, thereby wiping out the businesses based on fine-tuning those models for specific verticals feels suspect given that businesses are already cropping up to respond to the limitations of the models as currently existing and become plugged into workflows. Even if the models improved, switching costs are higher for the customers than they would be for the vertical-specific business on top of some AI API, which can more easily switch providers.

This logic also applies to the “battle of the titans” bias that has been portrayed between Microsoft/OpenAI vs Google. They have existing business models to protect, hence Google’s slowness to cannibalize all of search’s enormous profit margins via releasing an AI search chat which would be incredibly expensive to run and uncertain to monetize.

The optimal business model for an AI world is an open question, with usage-based pricing currently being the norm among dominant players. Usage-based pricing, however, has to prove a high ROI to justify continued growth in usage which may slow down adoption. It is likely that working backward from ideal economics someone can come to build an AI business that blows away existing ones.

Incumbent Response

In general, we expect incumbent firms to resist a transition that threatens to undermine the complex skills and valuable positions they have accumulated over time.

As mentioned Google has been resistant to ship products that cannibalize on their existing business models. Microsoft on the other hand is extremely willing to spend more incrementally on each Bing search given the enormous fixed costs that went into the product that they can now spread over more users. The market share war won’t last forever and understanding the economics of each business and assessing whether AI either improves or harms ROI can give a pretty good forecast of which incumbents will be eager to adopt AI and for what. Existing search economics are hurt by using AI more, but it’s likely that Microsoft’s suite of products from Office365 and RPA tools like PowerAutomate can all become increasingly integral to corporate stacks and justify higher pricing that covers the rising costs.

Attractor States

In thinking about change I have found it very helpful to use the concept of an attractor state. An industry attractor state describes how the industry “should” work in the light of technological forces and the structure of demand. By saying “should,” I mean to emphasize an evolution in the direction of efficiency meeting the needs and demands of buyers as efficiently as possible. Having a clear point of view about an industry’s attractor state helps one ride the wave of change with more grace.

To me the equilibrium state for AI adoption across industries is a bimodal distribution. In content for example, the returns to existing brands with large followings or troves of IP far outweigh those to small and mid-sized creators. For these larger players there exists excess demand for their products, and using AI to alleviate supply constraints by making it cheaper and faster to produce more content, and to more effectively surface that content for the right person, increases their dominance. On the other end of the spectrum, there will be a lot more creators entering and producing content giving the dropping barriers, but they likely will have a hard time finding large audiences.

We see this content issue somewhat playing out in video already. If we put blockbuster cinema, genre television, and user-generated content on a continuum of appeal there’s the largest market for blockbusters, a mid-sized market for each genre show, and a small market for each piece of UGC. In aggregate though UGC consumes more time than the other because there’s a lot more of it, while blockbuster cinema is hard to create so it consumes the least aggregate hours. The market for genre television though has started to erode, as Netflix has seen decreasing user engagement, as more user attention is pulled to either extreme. If Disney could extend the attention on each piece of media through AI transformations, while UGC continues to be created even faster, it is unlikely to me that mid-market fare would draw an audience.

In other industries, similar dynamics are expected, with superstar firms or individuals reaping growing rewards while more entrants are able to carve out small niches and compete more aggressively with mid-sized players. (David Autor is one of my favorite researchers on the topic of bifurcation of work, highly recommend his work.)

Philosophy

Marshall McLuhan’s work on media theory provides a powerful method for understanding technological change. While philosophical theories have a very different color and approach to business ones, thinking of them slowly can lead to the formulation of incredibly actionable insights. His thesis that “the medium is the message” runs counter to the “content is king” philosophy that guides much of modern social media, with practical implications.

For example, Spotify launched an audiobook feature in the US in late 2022 and has been expanding it to other countries this year. Spotify has been a late entrant to the audiobook game. Amazon acquired Audible in 2008, making it a key feature of its books platform with a dominant market share in audiobooks. The Amazon approach has followed the “content is king” idea where a user cares more about the content than the format. They may read part of a book while at home, then want to continue it as an audiobook when they hop in their car. Spotify on the other hand is a medium bet, that audio as a format is the key criteria for consumption, and the content that is consumed is subsidiary to the format. Spotify is what you put on while driving, so adding more content is good as long as it is audio, be it music, podcasts, or audiobooks. Converting the same content into different formats, such as physical books or videos for podcasts would not do much for Spotify’s user growth or retention. (We’ll see how this feature plays out in the coming years.)

As AIs become multimodal, thinking of their usage through a media lens helps to make bets on how they will be used and their impact. Another McLuhan framework to draw on is the tetrad of media effects from Understanding Media: Enhancement, Obsolescence, Retrieval, and Reversal.

More of the foundation of this New Science consists of proper and systematic procedure. We propose no underlying theory to attack or defend, but rather a heuristic device, a set of four questions, which we call a tetrad. They can be asked (and the answers checked) by anyone, anywhere, at any time, about any human artefact. The tetrad was found by asking, ‘What general, verifiable (that is, testable) statements can be made about all media?’ We were surprised to find only four, here posed as questions:

• What does it enhance or intensify?

• What does it render obsolete or displace?

• What does it retrieve that was previously obsolesced?

• What does it produce or become when pressed to an extreme?

Enhancement

Similar to how the computer enhances the speed of calculation and retrieval, AI enhances the speed of prediction and action. Data can be interpreted, solutions proposed, and actions taken with ever greater speed. Just as how increased speed and decreased cost of calculation and retrieval encouraged the reframing of problems to be computed we can anticipate more problems to reframed as prediction tasks where data, objectives, and policies are given and AI can carry out the rest of the process.

Thinking of which tasks are best suited to reframe in this way and where the greatest opportunity for gain is can help understand the present opportunities.

Obsolescence

McLuhan noted the increasing obsolescence of sequential reasoning and slow thinking as media became faster and more abundant. The press obsolesced “yesterday” with a focus on today’s news and the current digital media cycle has made information so fast that often news is outdated only 15 minutes after being consumed. We are currently still limited by the speed at which humans process information though we can imagine the speed increasing to the point that the present itself becomes too slow and already outdated. We see with the rapid development of AI that models and frameworks can become outdated as soon as they are announced. AIs can start developing AIs faster than we can to carry out tasks even faster.

Such obsolescence of the present requires the development of strategies to be more future focussed. A world of increasingly powerful predictions is one constantly in the future. Just as McLuhan proposed his concept of the “Global Village” for how tribal groupings built around fundamental markers of belief or identity would guide group understanding in a world in which information is too fast to process, it may come that created future-oriented systems of identity such as shared visions or missions becomes a new way to organize and make sense of the present.

Retrieval

AI retrieves the idea of the personal in a newly feasible way. TikTok’s recommendation engine shows us stuff that our friends may like, but often times we see videos that feel like they were made just for us, an audience of one. As the barriers to create fall, people create things just for themselves and find audiences through that personal creation. It is more economically feasible than ever to create custom goods, retrieving artisanal practices subdued by the industrial revolution and the era of mass man.

Reversal

The speed of the car reverses into the traffic jam, the distribution of the press reverses into advertising as news, the globalization of the internet reverses into balkanization. It’s entirely possible that the speed at which actions can be taken leads to a world in which AIs are used to carry out entirely contradictory actions leading to a completely steady state in which nothing gets done. The faster the systems move the less oversight is possible to understand whether anything is happening at all. We can already see ChatGPT’s factual limitations creating avenues for misinformation that are entirely not intended by humans.

This fast-paced, globally interconnected world may reverse into a slow-paced, local world. The watch industry is a great example of how the advent of quartz watches dropping the production cost of reliable watches to near-zero reversed the industry to focus on prized, expensive luxury pieces that would not have been mainstays of the watch industry prior.

The confluence of several inflection points from Apple’s ATT challenging Facebook’s traditional third-party data driven ad model, to the changing dynamics of Search eating into Google’s margins, to the rise of Generative AI as an accessible and performant technology, all open the door for a new world of digital advertising, one that may be orders of magnitude more valuable than anything that came before. 

Generative AIs have already proven their capacity to produce, remix, and personalize text, images, video, and audio which make up most advertising materials in existence. This approach can enable new forms of advertising that are less invasive of a user’s privacy, relying more on context than precision targeting, all while reducing costs. At every step of the marketing funnel, from creation to targeting to evaluation, and for both direct response and brand advertising, Generative AI promises a revolution in how digital advertising functions.

The Past and Present of Generative Advertising 

Even in its infancy, Generative Advertising has demonstrated massive potential in brand advertising. Lays launched Messi Messages in March 2021, leveraging deepfake technology to personalize greetings from Lionel Messi, garnering 38 million hits in the first 24 hours.  In October 2021, Bulgari leveraged AI-image generation to create a massive art installation for their Serpenti line. These spectacle driven brand campaigns were only the beginning of a movement towards personalization that has accelerated in 2022.

Jasper.ai which rebranded from Jarvis in January 2022 has quickly grown to a staple in AdTech, with their 70,000 customers using their GPT-3 powered product to write blog posts, SEO, and headlines for promotional materials. 

In May 2022, Netflix leveraged Dream by WOMBO to promote the launch of Love, Death + Robots Season 3. Over a 2-week campaign, Dream users were shown input images of the logos of LDR and encouraged to make customized renditions, resulting in 12 million images produced by 2 million users. The scale and speed of this content production could not have been feasibly done by an in-house team, and by engaging a large and diverse userbase, the brand recognition of the show increased. Heinz launched a similar campaign in August, using Dall-E 2 to draw ketchup alongside human submissions.

These existing campaigns highlight the flexibility of Generative AI, allowing a scale and diversity of content production that was previously only accessible with 7-figure budgets. 

The Post-ATT Landscape

As the costs of using generative techniques decrease and their quality improves, we can expect wider adoption among SMEs that rely on digital advertising for their growth, and whose cumulative marketing budgets are the key source of revenue for digital platforms.  

In a pre-ATT world, smaller firms typically invested a fixed amount in producing limited creatives, spending most of their budget on targeting their ads. As the returns to targeting decrease with less tracking across the internet, an opportunity has arisen in increasing produced creatives.  

A new world of contextual advertising has opened up thanks to AI, which relies more on what’s being viewed than the person viewing it. Thanks to the improving quality of AI interpretation of text, images, video, and audio, it has become easier to understand the sentiment and nature of the web page or mobile screen a user is viewing. Leveraging generative techniques, ad content can then be adjusted in response to the surrounding context, increasing conversion likelihood. For example, text generators can be used to customize ad copy to match the sentiment and phrasing of the text surrounding it, while image generators can be used to customize the color tone or background imagery to be more subtly reflective of what’s being viewed.

Rather than the creepiness factor of feeling followed across websites, contextual generative advertising can make advertising more organic to each page, allowing ads to blend more neatly into the world around them. This form of advertising can be validated through both the first-party cohort data and by using analytics on the performance of different creative customizations to create a flywheel that places ideal advertising across the web. 

The opportunity created by this is enormous. In moving budgets away from precision targeting, with its naturally decrease marginal returns, and towards creative production, which was always more expensive, advertisers will have an ever-increasing opportunity to increase conversion rates. As conversions increase, so does the return to advertising, and the budgets spent on marketing, expanding the opportunity for Generative Advertising producers to profit.

Given that the nature of this production differs from the tracking abilities that Facebook/Meta products excel at, we can see Generative Advertising as a new category of competition in which traditional players do not necessarily have an advantage. The nature of the analytical flywheel has not been fully developed, so as the value of historic third-party data decreases post-ATT, companies with a generative advantage can enter the field and challenge incumbents.

New Search Advertising

The economics of search advertising are also increasingly becoming anyone’s game. As product searches increasingly begin on Amazon, and voice search increasingly emphasizes returning single results rather than lists, the traditional Google developed search advertising model no longer has the moat it once did. A user’s desire when putting in a search is not to select among possible options, but rather to receive precisely what they asked for, and Generative AI makes this a realizable possibility. 

The nature of Generative AI is essentially search. Rather than searching for the best result from a limited set of possibilities, however, generative search returns the best result after combing through the entire space of possible results. If it can’t be found from what exists, it can be created anew from the learnings of the algorithm. 

This search nature of Generative AI, combined with the contextual character of the post-ATT landscape, enables a new search business model that is more natively integrated across the web, rather than a destination unto itself. For example, as I write this I want to think of a witty joke, and before me are some articles that I’m using for research. If I search for an interesting joke based on what I’m reading, it would be a much more beneficial experience for me if the search engine considered the context of what I’m looking at and returned the best result it finds, and if it can’t find a great result, it should make one for me from scratch!

This combination of the analytical components of AI to understand context, and the generative components to create where existing search lacks, can supercharge the user experience. The economics of search have proven extremely lucrative, and to create a new search model that improves retention and engagement opens a radically large opportunity to remake search advertising business models.  

Brands of 1 

The nature of digital communities has also changed the way brand advertising affects consumers. Rather than being limited to being generically appealing to everyone like McDonalds or zeroing in on a specific niche like a Hot Topic, the segmented nature of internet subcultures has allowed brands to create more than one identity depending on the niche. 

An example of this is White Claw, which managed to break free from the gendered nature of alcohol branding, not by identifying itself as a generic appealing drink, but rather by customizing its brand messaging depending on where its creatives were seen. Numerous subcultures came to identify White Claws as their drink of choice rather than an inoffensive pleaser to all or associated with a specific circumstance.  

Much of this has been enabled by the analytical AI engines that customize a user’s online experience to what is most appealing to them. Often users won’t even know of the existence of another subculture let alone experience the cultural lingo that exists there. Brand advertising has tapped into this and allowed cultural presentations to adapt to the grammar of different groups. 

Generative AI opens another dimension to this trend. As more content becomes AI generated the nicheness of subcultures can increase. If the content I see is more customized to exactly what I like, it may in fact result in me being the only person who ever sees it. It’s not that it didn’t find its audience, but rather that for that piece of content, I am the extent of the intended audience. 

The degree to which brands can then customize their identity to different niches vastly increases with Generative AI, and as I grow to have a feeling that a brand is specifically catering to me, my engagement with them would likely increase, and I would be a more retentive customer. 

Advances such as DreamBooth, which can feed an individual into an AI and put them in the creatives also can increase the degree to which advertising can be built for an audience of 1, rather than a generic creative targeted to an individual. The new branding possibilities enabled by generative ai can help top of the funnel advertisers in ways that a purely tracking/targeting based approach never could.

Conclusion 

Seeing the broader trends in advertising that coincide with the rise of Generative AI highlights the multiple inflection points which can make Generative Advertising a sticky new paradigm. Transitioning away from a world where users constantly feel watched by Big Tech firms has brought anxiety for advertisers, but a new mode of scaling up content production and interpreting context may in fact prove to be more valuable than what came before.

Generative AI companies that have already started to penetrate the public consciousness can build reliable moats in the type of data that can power Generative Advertising. Having produced billions of outputs for users and receiving feedback on what is well liked gives these young firms an analytical edge in understanding the appeal of creatives, which is a distinct dataset from personal data that characterized Facebook and Google’s dominance. At this early stage, capturing the Generative Advertising market remains anyone’s game, and the economic value is enormous.

One of the reasons that artists need not fear is a persistent economic paradox of automation, the last mile problem. The last mile problem originates in supply chain and logistics, stating that even when a system is perfect 99% of the way it will be bottlenecked by some inefficiency somewhere. We see this play out in the increased demand for human labor to compensate for the weaknesses of automation systems. This paradox is comforting to the importance of human work even in a world of widespread automation from robotics and artificial intelligence. Insofar as automated systems aren’t 100% perfect, and the unpredictability of human beings and the unreliability of technology make this likely to be the case, human beings are needed somewhere. And even more positively, as much as the automated parts improve efficiency a greater demand of human labor is needed to ensure those efficiency gains are realized.

Take for example the increasing workforce needs of Amazon as they’ve made supply chains more efficient end to end. Or the example of the ATM, which by automating the time-consuming rote tasks tellers once did frees them up to work on customer service, a much higher value add output for banks, leading to more tellers today than ever before. And the rise of occupations in the maintenance of machines and software systems continues to find human labor valuable. 

The novel world of Artificial Intelligence Generated Content (AGC) also presents a last mile problem, but in reverse. With the proliferation of text and image GAIs we see that humans have, till now, faced a last mile problem in creativity. They can get their ideas sometimes 99% of the way there but face some barrier somewhere that prevents them from realizing their full vision. AGC helps to address this issue, supercharging human creativity and making it more accessible than ever. 

How Technology Enables Individual Creativity

To set the stage for AGC, it is useful to reflect on the role technology has played in reducing the last mile for creativity up till now.

In performing arts, we see technology consistently aid in the capacity for the individual to express themselves. Theatrical performances are typically a collective effort, which while likely guided by the artistic vision of one or more persons, involve many participants to successfully realize. As film technology was introduced, and motion picture could capture the finer expressions of actors, more intimate character pieces were added to the medium. Within film itself we see how accessibility of technology changed the tenor of creative control. When film reels were expensive, producers could overrule the director as needed and the studio system exerted tight control over what got made. As film technology grew cheaper and higher quality, the studio system gave way to auteur cinema, where the director called the shots. The barrier then became in getting enough people to believe in the director’s vision to bring the project to light.

The 21st Century has seen video cameras and streaming services proliferate, all within people’s pockets, making content creation and consumption easier than ever. In this new world a director can often be the writer, producer, editor, and only actor needed to get a piece of content off the ground. Technology in performing arts has enabled a speed of production and scale of distribution that has allowed new mediums to be conceived and greater creative outlets for individuals.

Musical compositions have also exhibited a similar trend. Realizing a composition a few hundred years ago required an ensemble of musicians, a large physical space, and a protracted rehearsal schedule to ensure it was just right. As recording devices proliferated new forms of music arose, often with a smaller number of players involved. In the digital era the rise of DJing has combined composition and conducting in a single art form that requires little external assistance to produce and distribute original music.

In both examples above human creativity was expanded by the novel technologies that allowed for easier production and distribution of content. There were once so many barriers in creating a finished work that someone could not do it alone, and in many cases had to sacrifice their artistic vision to bring all the pieces together. This is not to diminish the value of group efforts and the collective creativity that these eras tapped into, some of the greatest films ever made are from the studio era and some of the greatest compositions from a classical period with the highest barriers to entry. Rather, it points to an expanded horizon for individual creativity unlocked by technological advance. The last mile of human creativity has gotten smaller and smaller throughout the ages, and with AGC it shrinks further.

AGC and the Last Mile

The inability for AGC to simply replace humans in the creative process stems from the fact that creation is not a task with a fixed objective. AIs are phenomenal at solving the space of defined problems, that is the class of problems where some solution is optimal. But what is the optimal output for a piece of creative writing or an image? Certainly, we can intuitively tell if the quality of the output is good, for example whether the piece or writing or the image met all the requirements we set for it, but whether it fulfilled the needs of our creative expression is a less defined problem. As a result, dialoguing with a GAI to produce ideal AGC is important. The output users get with little tinkering is often impressive enough to use as a template or starting point, but rarely alone good enough to be the end product.

So there exists a last mile for the GAI in that it needs quality guidance, from prompt engineering or mixed inputs, to give the best results. We also have a last mile for human creativity in that sometimes we lack the inspiration or awareness or all the necessary skills to realize the fullness of our creative vision. These complementary shortcomings have allowed for symbiotic works to be created that help realize the vision of creators.

Text generators are already assisting writers in putting out more content with greater originality. Rather than sampling having the GAI write an entire story, however, these authors are using these models to flesh out character descriptions, fine tune the phrasing of certain key events, or provide inspiration for a climax that they’ve been struggling to ideate. Writers have learned that the ways that GAIs interpret human language is very high dimensional, for example the phrase “ideal Sunday afternoon” has an infinite number of possible descriptions that, without additional context, would not be an ideal prompt to feed a model. Instead of giving loose direction and asking for a lot from the GAI, they give significant amounts of direction and ask for something very pointed. Feeding in the entire text of what you’ve written so far and getting suggestions for the next line is far more likely to return the ideal output.

In the image domain we see a similar trend, with GAIs being used to generate concept art, design inspiration, and other creatives that the user lacked the appropriate inspiration to visualize. Turning rough sketches into fully realized artworks or receiving template images on which to draw on top of are two common use cases we at WOMBO see from Dream users. To us this highlights the key role human creativity continues to play in the creation and utilization of AGC.

Cross medium development is another last mile problem that AGC is addressing. We have received thousands of emails from users showing us book covers and album covers they’ve made using Dream. These creators had written or musical talents, but to effectively release their content they needed a visual component they lacked the skills to create. Traditionally this issue, if the creator lacked a friend to help them out, would have led to creatives being left in limbo, or being released in ways that hindered their appeal or discoverability, or requiring expensive outsourcing to complete. Addressing this last mile problem increases the speed and control that creators have over their final work.

Where Do These Trends Take Us?

Right now, text and image are the major formats of AGC, but the potential to create outputs is endless. Work on multimodal GAIs, ones that take in some combination of formats like text, image, or audio, and output any combination of those formats helps expand the feasibility as a solution to the creator’s last mile problem. If you’re an animator that needs a soundtrack for your movie, feed in the visuals and have the GAI give you custom audio synced appropriately. If you’re a musician that wants lyrics for your song, feed in the audio, describe the theme, and have the GAI write a song that accompanies the melody. The possibilities enabled to expand the skills of creators to other domains could mark a new era in creativity.

I mentioned earlier that performing arts have been increasingly democratized through technology, but there are some domains where barriers are still high. It is difficult right now for the average TikTokker to produce something with MCU quality special effects. Those blockbuster productions have budgets in the hundreds of millions with thousands of crew members involved in realizing the vision. It is entirely feasible that within the next decade we’ll see AGC enable anyone to achieve cutting-edge special effects and seamlessly blend it into their productions with ease.

We can envision this also replacing the way burdensome design and editing tools work. Much of the education of creative types is in learning specific tools, for photo editing, video editing, effects, and object design. As GAIs develop, feeding in descriptions and templates can alleviate the entry barriers into these fields. The implications of this extend well beyond entertainment. In architecture, CAD tools have assisted in making assets easier to visualize, reuse, and share, improving efficiency and quality. AGC can assist in making these tools far more powerful by intaking hand drawn designs, descriptions of the specifications and the surrounding environment, and budgets to help architects quickly iterate and understand all the possibilities with their project.

In the examples above the more one knows about their field and needs, the better the GAI can assist them in getting a high-quality output. The more modalities future GAIs enable the more human creativity can be asserted. AGC, rather than replacing the role of human creativity, requires human creativity to fully take advantage of the opportunities it enables. While it seems that reducing other domains to language, with prompt engineering taking much of the spotlight in discussions on AGC, may remove some of the texture of other mediums, there’s no reason inputs must be limited to text. We are at the cusp of a new era of creation where rather than learning the button clicks and code of design tools, one can communicate with a GAI through the creative outputs they find most representative of what they desire. Opportunities abound.

The outputs of a GAI, or AI Generated Content (AGC), have applications that can affect every worthwhile human endeavor, but thorny questions abound regarding the intellectual property both powering and produced by these models. Unlike the platforms with powerful analytical engines, generative models have typically been trained by scouring the vast reaches of the internet. Whereas terms and conditions have been well defined about how a social media platform or a streaming site can use the data from users to train their algorithms, it is far less clear whether anyone has granted the right to use their data to power these more far-reaching algorithms. On the flip side, while AGC is produced by an AI, the human user leveraging the output desires some intellectual property (IP) protection over what’s created.

This piece will aim to first present an overview of these issues and conclude with a vision for how the AGC can not only power new opportunities, but with the right approach create more value for traditional creators.

Model Training

GAIs are trained on vast amounts of data for a given task and a significant amount of the heavy lifting of developing state of the art models is to amass the largest data set possible. A combination of scraping publicly available datasets and licensing proprietary data is commonly used, with more licensed data available as a company’s budget increases.

For natural language models such as GPT-3 and Co:here, a critical dataset is CommonCrawl, which archives much of the internet to be available to the public. Image data used for models like Stable Diffusion leverages datasets like LAION, which is built on top of CommonCrawl’s work, generating samples from 5 billion image-text pairs from CommonCrawl’s HTML image tags. These resources are built off scraping much of the web, among which many copyrighted and trademarked works are present. They use a fair use defense for the web scraping and dataset creation but stipulate that leveraging their datasets must respect existing IP rules.

This doctrine of fair use has been critical for many breakthroughs in media, think of how much of YouTube is powered by the remixing and critiquing of copyrighted content, and seems to have good standing with respect to the curation and maintenance of large datasets. The doctrine of fair use states that the force of copyright law is limited when the use of the work is transformative and does not detract from the value of the original work. Common fair use defenses include parody, criticism, research, and education. The research work that goes into creating and maintaining large datasets that include within them copyrighted works is evidently transformative, for being used to train an AI was not the original intent of the work itself and does not detract from the value of the original work by being used in building the dataset itself.

While simpler for datasets, the models built on top of them raise more complex issues. For example, a generative art product that makes images resembling those of a well-known artist does have a similar intent and might in fact detract from the value of the artist’s work. If it can be established that there is a direct relationship between the AGC and the copyright protected work, infringement may in fact be occurring. Some of this can be established observationally, as with the ghostly imprints of Alamy or Getty that appear in some prominent outputs. Even more broadly, trade dress protections, a subset of trademark laws that deal with “look and feel”, have been expanding in the US, making certain more qualitative judgements of style potentially infringing.

Striking hard on infringement in these early days of a new and fast-growing area of research would be a major blow to the very innovation that IP laws are meant to encourage. If training data and model development was restricted to compliant work, then only the largest firms would be able to innovate. AIs are data hungry and limits to what can be used based on risks that currently represent minimal commercial harm to creators is undesirable.

These exceptions, however, can’t last forever. The awareness that outputs can infringe on copyrights and trade dress is something that engineers must work to address as the state of GAI improves. Some techniques like adding hidden prompts can help minimize the presence of copyrighted works. In some cases, the final implementations of models may focus only on monetizing infringement without a genuinely transformative experience. For example, if I leverage a text to image model and customize it to make everything look like a Banksy piece. The look and feel of a Banksy work is protected, and by specifying Banksy it is drawing on a pool of images that are necessarily copyright protected. This application would not afford fair use protection if commercially used.

Model Outputs

On the flip side a question is raised about the protections afforded to the AGC itself. If the AI wrote a blog or created an image or developed a game, would any protections be afforded to it? Much buzz was created by a ruling earlier this year that an AI cannot copyright the works it creates. This ruling, however, was widely misinterpreted to mean that AI-generated works lack copyright protections. In fact, the case simply reasserted a legal norm that only humans can possess IP. That a GAI itself cannot own the work is in line with saying that an AI-powered self-driving vehicle is not responsible for the accident. The human in the loop, and there is always a need for some human in the loop, is responsible.

AGC involves humans at many stages, from producing the training data, to developing and maintaining the models, to the acts of prompt engineering and synthetic curation, new skills in the world of GAI-enabled creation. Acknowledging this clears up that there are many potential vectors on which humans are involved and to whom IP protections can be afforded.

A little bit needs to be stated on prompt engineering and synthetic curation. Prompt engineering has proven to be an important driver of the quality of AGC. It is the art of communicating in a way that can ensure that the GAI produces what you desire. Many of the products built on top of large language models and image generators rely on expertly crafted prompts to ensure that high quality outputs in a relevant domain are achieved. The best generative artists work tirelessly to craft ideal prompts to regularly generate beautiful artworks. The prompts themselves have a creative component to them, and their uniqueness can contribute to their eligibility for IP protection. Work is already be done to establish ownership of prompts and improve the search space for the best prompts that were used. Courts have established creative input tests to establish the eligibility for protection of short stories or taglines, and a similar test can assist in protecting the unique prompts that drive high quality output.

Synthetic curation is another important skill in the world of AGC, though as an art form it is in earlier stages than prompt engineering. AGC can be multimodal, taking inputs that are texts or images or audio or video and weighting them in some combination to create works that are any combination of these formats. The AGC itself can then have a human stitch them together or arrange them in a unique way or even feed them back into a GAI to achieve their final vision. The more synthetic curation that is done, the more it can be well established that IP protections should be afforded.

It is important to want to afford ownership for AGC and establish fair tests for determining eligibility as this would enable further growth in the field and expand economic opportunity. We can imagine GAIs becoming quite critical in scientific fields from new materials research to drug design as they become more crafted to understanding physical and biological properties. If the generations were not ownable and the patent protections surrounding the outputs not possible, this would discourage their use in these fields limiting their overall benefit.

Across the board, from marketing creatives to game design to merchandise production, the ability to establish rights over the works used is critical to the business operations that AGC may power. To accelerate the development of sound business models leveraging these novel tools, which would in turn benefit the potential copyright holders of the underlying data they were trained on, maximizing the speed and scale of implementation should be a key direction of IP policy concerning these technologies.     

Subscribe now

A Vision for the Era of AGC

Understanding the nature of the IP issues surrounding generative models allows us to chart a path to making them maximally

beneficial for everyone involved. An approach to be drawn on here is Data as Labor, pioneered by economist Glen Weyl and technologist Jaron Lanier. Data as labor identifies the production of the underlying data in large datasets as an act of labor, which grants creators certain rights, both economic and social, over its use. An analogy for data is music mechanicals, a component of the overall musical work which are often stitched together from the labor of various contributors. With this framework data can be represented the way that music is, with legal developments such as the Music Modernization Act helping to ease the ability for small creators to be compensated for their work.

There are several considerations that must be addressed to ensure that AGCs both produce value for generative creators and for the creators as data laborers.

First, it is evident through the vast amount of AGC already produced that not all generations have economic impact. Setting rules about what happens with created works that are not used, and even establishing a difference between works that are used privately and commercially sets the stage for where compensation occurs.

Second, not all works are exploiting of another’s labor. Watching an artist produce a work and taking inspiration from them does not in fact violate their rights but reproducing their key elements and motifs does. GAIs themselves can play a role in establishing the degree to which creation is indebted to a data laborer. In the image domain CLIP interrogators can be used to establish the degree of relationship between a piece of AGC and the corpuses of various artists, highlighting whether any singular work or artist or group of artists were the main drivers of the final output. In the field of text and video, similar models can be easily identified, with the broad range of AGC feasibly being brought within this testable domain. In addition, the prompts themselves can be queried to see if they make direct reference to protected works or their creators to establish their rights.

Next, compensation must be dependent on the commercial viability of the output. Protecting the IP of AGC is prior to protecting the rights of the data laborers who enabled them. If no profit can be made, then nothing can be compensated back. Establishing an effective rights management solution to this is one of the most promising applications of Non-Fungible Tokens (NFT) today. While NFTs are not a replacement for traditional legal rights, they can make enforcement of these rights easier. By establishing a digital ownership chain that can split up among multiple rights holders, NFTs can make managing rights for large volumes of content with a myriad of contributors easier. This would help creators no longer feel exploited by the derivative works GAIs enable, but rather feel a part of that relationship. Like how sampling has created expanded economic opportunities for musicians whose work constantly breathes new life, so too can AGC develop entirely new revenue streams for traditional creators.

Behind this, a rights management layer must exist to assist creators in being able to be recognized, managing the NFTs and payment schemas, standardizing the evaluation criteria for derivative works, and bargaining with platforms to ensure their policies respect the wishes of their creators, and that derivative works are not used in a way that goes against their wishes. The concurrent technological work done with CLIP similarity and NFTs makes this management ecosystem more feasible than ever before, and we can envision it maturing as quickly as music rights management has in the age of streaming.

A final important note on this vision is that it references only creators both traditional and generative. The middlemen of data collectors, model infrastructure providers, and productizers have not received the same treatment. In the case of data collection, it is evident that their data collection falls under fair use specifically because they themselves take no stake in the creative output and genuinely use the protected material in transformative ways. If they were to have rights over derivative works, this status would be called into question, changing incentives in a way that may not benefit the research community.

Similarly, those that produce and maintain generative models allow a diverse amount of work to be done, much of which leverages completely public domain work or that is used for non-commercial applications. It makes for more sense for them to be neutral to the downstream applications of their tools and charge on a usage basis. Entering into ownership agreements muddies whether usage-based pricing is eligible for being claimed by data laborers.

Finally, the end products that make GAIs accessible cannot uniformly have ownership stakes. The diversity of products, ranging from general purpose use like those made by WOMBO to marketing specific like those of Jasper, each have different relationships to commercial use. Some of these may output works that are generic enough that no IP protections are even possible, while others may tailor their outputs to a very specific catalogue and must pay traditional royalties for access. The challenge of developing standards is to make them clear enough that innovation is possible without making them so rigid that it’s only allowed in a limited way.  

Nonetheless the vision established above outlines how AGC can create new economic opportunities for creators, expand creation, and drive product innovations while respected the rights of those work GAIs have learned from. Sketching out visions early is an important exercise in guiding both technical and legal development towards positive sum outcomes.  

Conclusion

The attention being drawn to generative artificial intelligence from technologists, investors, and the public is justified by the vast array of new tools and applications that this technology can support. With GAIs becoming better, cheaper, and more accessible it is important to grapple with the intellectual property issues raised by artificial intelligence generated content.

In the early days of any new space weaker rules should dominate that allow people to move fast and break things, but concurrently it is important to theorize about where things are heading and what should things look like when they mature. As AGC adoption increases the importance for end users to understand their IP protections and obligations becomes vital to scale.

By looking at the models themselves as mechanisms for improving enforcement legal experts and policy makers can craft rules that are more unorthodox and ambitious in protecting rights. The vision laid out here argues that rather than in conflict the rights of traditional and generative creators can only be respected in tandem.

Subscribe now