One recent German court ruling concluded that:

“…transforming someone else’s photograph into a comic-style image using AI does not automatically constitute copyright infringement.”

The key word is automatically.

That doesn’t mean every AI-generated image is lawful. It means courts are increasingly analyzing these cases based on the specific facts, rather than assuming infringement simply because AI was involved.

If this reasoning eventually influences U.S. courts, it could have important implications. It may suggest that using someone else’s image as a reference input for a generative AI model could be permissible in some circumstances… particularly when the resulting work is found to be sufficiently transformative.

That said, I think there are still areas where copyright concerns remain much stronger.

For example, one trend I’ve been watching involves using depth-map video to capture a dancer’s movements and transfer that performance onto an AI-generated character. (Check out my YouTube channel where I cover a weekly AI News Hour.) In my view, that “performance transfer” raises a different set of copyright questions. A choreographed or recorded performance fixed in a tangible medium may itself be protected, and reproducing that expressive performance through AI could implicate copyright in ways that differ from simply referencing a still image.

The larger lesson is that copyright law isn’t producing simple, universal rules for AI.

Courts continue to evaluate these disputes case by case, weighing factors such as transformation, substantial similarity, the nature of the copyrighted work, and market effects. That’s likely to remain true for the foreseeable future.

As always, these cases deserve careful reading. The headlines often suggest sweeping victories or defeats, while the actual opinions are usually much narrower.

Article:

https://petapixel.com/2026/07/22/dog-photographer-loses-copyright-case-over-ai-generated-comic-version-of-her-image/

And I still stand on the fact that training an AI model is “quintessentially transformative”. There’s absolutely nothing that references the original copyrighted works from the process of training an AI model. This was a fantastic ruling by Judge Alsup.

Subscribe

Share

So... on July 4th, I spent the day reading the United States Constitution and all twenty-seven amendments... closely, section by section. I expected to be reminded about institutional design, Congress, the presidency, federalism, voting rights, and the courts... and I did. But I came away with a larger question.

Apparently the Constitution is a rather old document... it was signed on September 17, 1787 (who knew?!?)... and it was written for a world of horses, handwritten letters, and human bureaucracies. Fast forwards to Anno Intelligentiae of today, and what happens when government gains access to advanced, agentic artificial intelligence... systems capable of researching, planning, coordinating, monitoring, and acting across complex systems?

AI could help governments solve persistent problems... improving healthcare, modernizing infrastructure, simplifying regulation, reducing fraud, and anticipating policy consequences. But it could also make government far more capable before it becomes wiser. That possibility makes constitutional limits more important, not less.

The Machinery Is Old. The Principles Are Not.

Many constitutional mechanisms reflect eighteenth-century constraints... the Electoral College, the former indirect election of senators (later changed by the 17th Amendment), slow transitions, and physical notions of privacy. But the underlying principles remain durable. The Framers assumed power would accumulate, institutions would protect themselves, and even well-intentioned actors could abuse authority.

Their solution was structural... divide power, create checks and balances, and protect individual rights. AI changes the capacity of institutions, but not the core problem the Constitution addresses... how to empower government without allowing it to dominate the people.

(Side Note: I still believe that we now have a better understanding of the mathematics behind the Electoral College. And I feel like we can do better. Someday I’d love to see us pass a amendment that makes voting a “ranked choice”.)

What Happens When Government Becomes Competent?

I’d argue that government failure stems from complexity rather than bad intent. Modern governance involves coordinating vast systems, data, and competing obligations. Humans struggle at that scale.

Agentic AI could expand what government can do. Policymakers could simulate policy outcomes before acting. Citizens could use AI agents to navigate services. Agencies could continuously audit spending, detect fraud, and identify regulatory inefficiencies. Scientific knowledge could be synthesized across domains in real time.

Existing frameworks from GAO and NIST focus on governing AI safely, but the deeper question is what governance looks like when administrative intelligence is abundant.

The Most Dangerous AI May Be a Perfectly Obedient One

The primary risk may not be rogue AI, but highly capable systems executing government directives with precision. A government able to monitor populations, analyze behavior, predict opposition, and automate enforcement at scale could be extraordinarily powerful.

The danger is not loss of control, but concentration of control. Constitutional safeguards... especially separation of powers... become more critical when capability increases.

The Fourth Amendment in an Age of Inference

The Fourth Amendment protects against unreasonable searches. Historically, that meant physical intrusion. Today, it may mean something else.

Modern data... location, communications, purchases, biometrics... can be aggregated to reconstruct a person’s life. Even if each data point is collected legally, AI can infer patterns, beliefs, and future behavior. At some point, inference itself may function as a search.

This raises a critical question... when does aggregation become constitutionally significant? The answer may define privacy in the AI era.

Algorithmic Due Process

The Fifth and Fourteenth Amendments guarantee due process. But what does that mean when decisions are made or influenced by AI?

If an AI system denies benefits, flags fraud, or triggers enforcement, individuals must be able to challenge the outcome. They need to understand the reasoning, access the evidence, and identify who is accountable.

A principle emerges... no sovereign action without an accountable human authority. AI may assist, but responsibility cannot be delegated away.

The Delegation Problem

The Constitution assigns authority to institutions, not machines. Agentic AI introduces a new layer of delegation... from Congress to agencies, from agencies to systems, and from systems to autonomous processes.

At what point does delegation become abdication?

For example, if Congress says, “Here is the law, here are the goals, here are the limits... use your expertise and tools to implement it.” That’s very different from Congress saying, “You decide what the policy should be, what values matter, whose interests should win, and then enforce your own decision.”

The nondelegation doctrine, though historically weak, becomes more relevant. Key decisions must remain tied to accountable actors.

(And as much as I wanna give up my Critical Thinking skills… and completely and totally abdicate all of my effort to AI… I just don’t see that happening.)

The Constitution Has a Private-Power Gap

The Constitution primarily constrains government, but many powerful AI systems will be privately controlled. These systems may shape employment, credit, information, and access to essential services.

When private platforms exercise power that resembles governance, the distinction between public and private authority becomes blurred. This gap may become one of the central governance challenges of the AI era.

And at the same time, we’ve got lots of SOTA LLMs available to us. This is why Open Source/Open Weights models are so fundamentally critical to everyone’s success.

AI Could Strengthen Federalism

AI may also decentralize capability. Smaller jurisdictions could gain access to advanced analytical tools, reducing reliance on centralized expertise.

One of the practical arguments for centralization has always been capability. The federal government has more money, more experts, more data, and more administrative infrastructure. If AI reduces that capability gap, states and local governments may be able to handle more problems effectively themselves.

This could strengthen federalism by enabling states and local governments to act as more effective laboratories of democracy.

(Side Note: This could benefit regional Innovation Ecosystems, which I’ll discuss in a future post.)

Intelligence Does Not Resolve Politics

AI can improve prediction, but it cannot resolve value conflicts. Questions of morality, fairness, liberty, and distribution remain inherently political.

Even optimal policies may lack legitimacy without consent. Democratic participation remains essential, regardless of technological capability.

From Episodic to Continuous Government

Government today operates episodically... laws are passed, then evaluated years later. AI could enable continuous monitoring and adjustment, allowing policies to evolve based on real-time feedback.

This could improve governance, but it also raises questions about oversight and control.

(Side note: This could be incredibly exciting to explore for a Case Discussion. But I’m concerned that the speed could have negative consequences.)

Friction Is/Was Sometimes a Feature

Traditional inefficiencies limited government power. AI removes those constraints. Automated enforcement systems could operate at massive scale, making power more efficient and potentially more intrusive.

Safeguards that once emerged from human limitations must now be designed intentionally.

I’ve often argued that when things change quickly, people die. Therefore, friction in government is good. It means that things can’t change quickly, And it gives people, businesses, institutions, and ecosystems time to adjust to change.

What’s Missing?

I think there are several issues that deserve more explicit thought:

• Legitimacy and consent: How do citizens meaningfully participate in AI-mediated governance? (Long term... I’m thinking brain implants. #Borg)

• Transparency standards: What level of explainability is required for public systems?

• Security risks: How do we protect AI-enabled government from misuse, hacking, or manipulation? (I think this is a perpetual cat and mouse of Mutually Assured Destruction.)

• International dynamics: How do constitutional principles hold up in a world where other governments deploy AI differently? (”Bomb first, ask questions later” is not a viable solution.)

• Institutional adaptation: Which branches or agencies are best positioned to oversee AI governance? (Oh! Oh! Pick me! I’d be an excellent Galactic Emperor!)

The Constitution’s New Role

AI may allow governments to optimize nearly everything. The Constitution defines what cannot be optimized away... privacy, due process, freedom of expression, and limits on power.

The central challenge is not building smarter government, but ensuring that increased capability does not erode liberty. The question is no longer whether AI will transform governance, but whether constitutional principles will remain strong enough to guide that transformation.

Regardless of your political perspective, I think the Constitution is on the right track.

“More Than Just a New Tool in the Toolkit: VFX Executives Take Stock of AI’s Impact”

I think they are getting a lot right… and I also think they are missing something fundamental.

The panel keeps returning to the idea that AI is “just another tool” in the VFX toolkit. I understand why. VFX has always absorbed new tools. Optical compositing became digital compositing. Matte painting became digital matte painting. Workflows expanded to include rotoscoping, tracking, simulation, rendering, previs, and virtual production. From inside the current VFX business model, the framing makes sense.

AI accelerates the artist. It helps generate assets, support look development, clean up shots, enhance images, assist with lighting, and create environments. The human artist still guides the work, refines the output, composites the final shot, and protects quality, taste, continuity, and control.

That is mostly true today, but it is not the deeper truth.

AI is not merely entering the VFX pipeline as a tool. AI is becoming the pipeline.

The panel is looking at AI through the lens of linear automation. (This is a “job shop” business model, btw.) A human performs a task, and AI helps the human do that task faster. The workflow remains fundamentally the same, but the task is accelerated. That’s Augmentation Effects.

The next phase is not just augmentation. The next phase is Agentic AI, and that changes the business model. That’s Transformation Effects.

An Agentic AI pipeline does not simply generate a background element and hand it to a human compositor. An AI compositing agent can communicate with a lighting agent, a camera agent can reference the director’s notes, a continuity agent can check the previous cut, a brand agent can evaluate strategy, and a legal agent can verify rights, likeness, licensing, and provenance. The system can generate variations, evaluate outputs, incorporate feedback, and loop thousands of times.

Not because a human manually executes each step, but because the workflow itself has become computational.

This is the shift from AI as a tool to AI as an operating system. Once you see that shift, the “just another tool” narrative starts to break down.

The panel is right that AI is not yet a one-click solution, but that is the wrong benchmark. The future of AI in VFX is not one-click. It is one intent. A human defines the desired outcome, and the AI system decomposes the work, assigns tasks to agents, uses software, evaluates outputs, incorporates feedback, and iterates until the result meets the specification.

That is a very different world!!

The VFX executives also lean heavily on the idea that human taste, curation, and judgment remain the ‘ultimate moat’. I genuinelyyyy understand that perspective. Their jobs and businesses depend on it. And in a philosophical sense, they may even be right. Human intent will likely remain central to art, story, advertising, film, strategy, and meaning.

But the “taste and curation” moat is shrinking. #Moneyball

Taste is not magic. It is partly pattern recognition, cultural memory, exposure, feedback, and an understanding of what has worked before and what audiences respond to now. AI is exceptionally good at pattern recognition.

An ‘agentic creative system’ can be trained on cinematic history, composition theory, camera language, advertising history, brand guidelines, engagement data, sales data, customer feedback, and real-time client notes. This does not mean AI has a soul. It means AI can simulate taste at scale.

That matters because AI can generate options, test variations, compare references, detect inconsistencies, predict audience response, and optimize toward a creative goal faster than any human studio can.

This is where the VFX industry may be making the same mistake many industries make during disruption. They assume the future version of the technology will preserve the current shape of the work. It rarely does.

New technologies do not just make existing processes faster. They reorganize them entirely. The camera did not simply make portrait painting faster. Digital editing did not simply accelerate film cutting. Streaming did not merely improve television distribution. Each innovation reshaped business models, labor structures, economics, and culture.

AI will do the same to VFX, advertising, design, filmmaking, education, consulting, and any (all?) fields built on expensive human cognition packaged into billable hours.

The panel seems to imagine a future of empowered, nimble teams of highly skilled humans doing higher-level creative thinking with better tools. That is one possible future, and it is certainly the comforting one. In that version of disruption, repetitive work disappears and humans move upmarket. They become more strategic, more creative, more valuable, and more essential.

That can happen, but Clayton Christensen would likely urge caution. Disruption usually begins where incumbents feel comfortable dismissing it. Early outputs are lower quality. Early customers are less demanding. The work is cheaper, faster, and “not good enough” for the high end. Then the technology improves, the workflows mature, and the economics become impossible to ignore.

Eventually, what once looked like a toy begins to consume the middle of the market.

If a single director, creative director, strategist, brand manager… or a kid in High School… can use an Agentic AI team to do the work of a 50-person VFX house in an afternoon, the business model of mid-tier VFX and post-production studios faces an existential contraction. Not because artists lack skill, but because the unit economics change.

The industry keeps asking whether AI will replace artists. That may be the wrong question. The better question is, “what happens when the work of the studio becomes software?”

What happens when the pipeline becomes autonomous? What happens when brands build their own internal AI advertising teams? What happens when a creative director can generate, test, revise, localize, and deploy hundreds of campaign variations without a traditional production army?

What happens when execution is no longer scarce, and intent becomes the limiting factor?

This is why brands should seriously consider building their own Agentic AI advertising and content teams in-house. (And, I can teach you… but I’d have to charge.) Not just hiring AI vendors. Not just giving tools to marketing departments. Not just asking agencies to reduce costs. I mean constructing the actual operating system for creative production.

In that world, the agency of the future may not be an external company. It may (should?) be a software architecture inside the brand.

Human creatives do not disappear, but the labor pyramid collapses. The old model required large teams because execution was scarce. The new model requires fewer people because execution becomes abundant. As a result, judgment changes shape. Value shifts from “Can you make the thing?” to “Can you define the right thing, evaluate it, improve it, and decide what it means?”

That is a different skill set, a different organization, and a different industry.

So yes, AI is a tool. But calling it “just another tool” is like calling electricity “just another candle.” It misses the system-level transformation.

The AI operating system of today allows humans to define intent and specifications while AI performs more of the work. The AI operating system of tomorrow will allow humans to define purpose, constraints, values, strategy, and desired outcomes while agents execute the pipeline.

That is not a new tool in the toolkit.

That is the toolkit waking up.

The concept is rooted in Outside-In Open Innovation.

(And for those of you who know me… I can talk about Open Innovation and AI for freakin’ HOURS, so <cough> I’ll do my best to shorten this.)

Generative AI has fundamentally challenged the traditional homework assignment model. For decades, most classroom assessments followed a familiar pattern: an instructor provided a “prompt”, a student produced a single artifact, and that artifact was evaluated. Whether the assignment took the form of an essay, business plan, design proposal, campaign concept, lesson plan, painting, illustration, sculpture, animation, or case analysis, the structure was largely the same. The student’s work was represented by one final submission.

That model is becoming increasingly difficult to sustain. Students can now use AI tools to generate essays, outlines, marketing campaigns, startup concepts, case analyses, lesson plans, code, and visual directions in seconds. As a result, many educators consider this plagiarism or cheating. There’s a much deeper challenge than simply preventing cheating. The more important question is how assignments should evolve when competent first drafts, plausible arguments, and polished surface-level or final outputs have become inexpensive and widely accessible.

Remember, in the age of Anno Intelligentie, the AI does the work.

The problem is not only that students can use AI to produce work. The deeper problem is that many traditional assignments have always measured output more visibly than thinking. Generative AI exposes that weakness. It makes first drafts cheap. It makes average work cheap. It makes “good enough” cheap. If the educational value of an assignment resides primarily in the production of a single artifact, then AI forces us to ask whether we were measuring the right thing in the first place. (Hint: We’re NOT!)

This leads to a central pedagogical question: how do we design assignments that continue to require meaningful intellectual engagement, cognitive effort, and critical thinking in a world where production is abundant?

One possible answer is what I am calling a “Curated Judgment Method”.

At its core, Curated Judgment shifts the focus of learning from producing a single answer to evaluating, selecting, synthesizing, and defending among many possible answers. Rather than asking students to use AI to generate one response, instructors ask students to generate a large field of possibilities and then engage in the higher-order cognitive work of determining which ideas are most valuable, why they are valuable, what they are missing, and how they might be improved.

In this model, the key educational challenge is no longer simply whether students can produce an answer. Instead, students must demonstrate their ability to define quality, compare alternatives, identify weaknesses, synthesize stronger solutions, and justify their decisions. These are precisely the kinds of skills that become more important, not less, in an AI-enabled world.

The Open Innovation Analogy

This approach draws inspiration from Henry Chesbrough’s theory of open innovation, which argues that organizations create more value when they look beyond their own internal boundaries and learn to leverage ideas, knowledge, and contributions from broader ecosystems. In traditional closed innovation, the organization relies primarily on internal expertise. In open innovation, the organization creates mechanisms for sourcing, evaluating, adapting, and integrating ideas from outside the firm.

A similar shift is now possible in education. The traditional classroom often treats students as isolated producers. Each student receives a prompt, works independently, and submits one answer. But an open innovation classroom can place students inside a much larger ecosystem of ideas. Students are not merely asked to create. They are asked to evaluate, compare, refine, combine, and defend.

This logic is visible in challenge prizes, Kaggle competitions, XPRIZE-style contests, crowdsourced design initiatives, and open calls for solutions. The value of these systems does not come only from generating a large number of submissions. The value comes from the funnel. Many possible solutions enter. A smaller number survive. The strongest are evaluated, combined, refined, tested, and defended.

That same logic can be applied to the AI classroom. Generative AI can (should?) create the initial field of possible solutions. Students then become the judges, curators, editors, synthesizers, critics, and defenders. AI does not replace student thinking. AI creates the field in which student thinking becomes visible.

AI as the Synthetic Crowd

This approach also connects to work on crowdsourcing and collective intelligence, including Ann Majchrzak and Arvind Malhotra’s argument that organizations can create value by enabling broad participation, recombination, and collaborative problem-solving. Crowdsourcing is not simply about collecting many ideas. It is about designing a process through which ideas can be evaluated, improved, combined, and transformed into useful solutions.

Generative AI creates a strange new version of the classroom crowd. It can produce a large set of possibilities quickly. Some will be weak. Some will be obvious. Some will be repetitive. Some will be wrong. Some will be surprisingly good. That unevenness is not a problem. It is the learning material. (We’re trying to mine the gold in the outliers of the distribution.)

The student’s job is not to accept the AI output. The student’s job is to judge the field.

This is where AI becomes pedagogically interesting. Not as an answer machine, but as a possibility machine. The value of the assignment is not in asking AI for one polished response. The value is in asking students to confront abundance. What happens when there are one hundred possible answers? How do students know which answer is better? What criteria do they use? What do they reject? What do they combine? What do they improve? What are they prepared to defend?

Those are not easy questions. Those are the questions that create learning.

(And the students need to come up with the rubric for judging.)

The Basic Method

The Curated Judgment (aka Open Innovation AI Pedagogy) framework can be implemented through a structured sequence. The specific details can vary by discipline and grade level, but the core process remains the same: generate many possibilities, define quality, evaluate the field, narrow the set, synthesize a stronger final version, and defend the reasoning.

1. Define the Challenge

The instructor begins by presenting a meaningful problem, question, case, design challenge, or project. For example, students might be asked to develop an argument about a public policy issue, design a campaign for a new product, create a business model for a startup, develop a lesson plan for a difficult concept, propose solutions to an environmental challenge, or analyze a case and recommend a strategic decision.

The challenge should be open enough to generate variety, but structured enough to allow comparison. That balance matters. If the prompt is too narrow, the AI-generated outputs will look too similar. If the prompt is too broad, the outputs may be too diffuse to evaluate meaningfully. The instructor’s first task is to design a challenge that produces a useful field of variation.

2. Generate the Solution Field

Students then use AI to generate a large number of possible responses rather than a single answer. Depending on the assignment, this might mean generating 25 thesis statements, 50 campaign concepts, 100 startup ideas, 100 essay outlines, 100 design directions, 100 policy proposals, or 100 case recommendations.

The objective is not to produce a collection of uniformly excellent outputs. In fact, many of the responses will be mediocre, redundant, derivative, flawed, or inaccurate. That is precisely what creates the learning opportunity. Students need to see the difference between weak, average, good, and excellent work. They need to experience the abundance of plausible responses and then learn how to evaluate that abundance.

The abundance becomes the raw material for analysis.

3. Build the Rubric

Before evaluating the generated outputs, students must define what quality means. This step is critical because it requires students to articulate what distinguishes a strong response from a weak one. Depending on the discipline and assignment, criteria might include originality, accuracy, evidence, feasibility, ethical awareness, audience alignment, strategic insight, clarity, creativity, emotional resonance, implementation cost, or disciplinary rigor.

This changes the student’s relationship to the assignment. The student is no longer simply asking, “What did AI produce?” The student is asking, “What would make this response good?” That question forces students to externalize standards that are often left implicit. It makes quality discussable. It makes judgment visible.

It also transforms the rubric from something the instructor hands down into something students must understand, build, apply, and defend.

4. Evaluate the Field

Students then apply their rubric to the complete set of AI-generated outputs. During this stage, they identify patterns, eliminate weak responses, recognize repetition, detect inaccuracies or hallucinations, and compare stronger alternatives against weaker ones. Through this process, students learn to distinguish between fluent language and substantive thinking.

That distinction is essential. AI is very good at producing text that sounds confident. It is good at creating the appearance of structure, completeness, and authority. But a fluent answer is not necessarily a strong answer. A polished paragraph is not necessarily an insightful paragraph. A plausible strategy is not necessarily a workable strategy.

Students need practice seeing that difference. This stage teaches discernment, and discernment may become one of the most important human skills in an AI-rich environment.

5. Narrow to the Top Ten

After evaluating the full set of responses, students narrow the field to the ten strongest candidates. This selection process must be justified through explicit reasoning rather than intuition alone. Students should explain why particular responses were retained, why others were discarded, which criteria proved most important, and what strengths or weaknesses emerged across the solution set.

They should also identify recurring assumptions, areas of convergence, and points of divergence. Which ideas appeared repeatedly? Which ideas seemed promising at first but failed under scrutiny? Which ideas seemed strange at first but contained an unexpected insight? What did AI do well? What did AI miss?

This stage transforms evaluation into a deliberate and reflective learning activity. Students are not merely selecting favorites. They are practicing reasoned judgment.

6. Synthesize the Final Version

Students then create a final product by synthesizing insights from the strongest responses. The goal is not simply to select the best AI-generated output. The objective is to produce something superior to any individual response generated by AI alone.

Students combine strengths, correct weaknesses, incorporate evidence, improve organization, refine arguments, adjust tone, add context, and contribute their own insight. The final artifact should demonstrate the value of human judgment and synthesis. It should be stronger because the student evaluated the field and made choices.

This is the human value-add. The student is not being rewarded for outsourcing the assignment to AI. The student is being evaluated on whether they can transform AI-generated abundance into higher-quality work.

7. Document the Process

Students should submit documentation that makes their reasoning visible. This may take the form of a reflection, appendix, portfolio, or process report that includes the original prompt, the generation strategy, the evaluation rubric, the assessment process, the top selections, the final synthesis, key revisions, rejected ideas, lessons learned, and reflections on AI’s strengths and limitations.

This matters because the assignment is no longer just the final artifact. The assignment is the reasoning process. If only the final essay, proposal, design, or recommendation is graded, AI becomes a shortcut. If students must show the prompt, rubric, evaluation process, selection logic, synthesis decisions, and reflection, the shortcut becomes much less useful.

The process is the pedagogy.

8. Present and Defend

Finally, students present both their final solution and the reasoning that led to it. The presentation should not focus only on the finished product. Students should explain what AI generated, what patterns emerged, what was eliminated, what survived, which criteria mattered most, what they changed, and why the final version is stronger than the alternatives.

Class discussion can then explore the reasoning behind the work. Why were particular criteria prioritized? What assumptions influenced the evaluation process? What alternatives may have been dismissed too quickly? What trade-offs were involved? How might critics challenge the final solution? Would the proposal succeed in a real-world context?

This stage requires students to take ownership of their judgments and defend them publicly. Not the AI. The student.

A Collaborative Variation: The Classroom as Innovation Community

The model becomes even more powerful when students work collectively rather than individually. Each student can generate a set of AI-assisted solutions. Those solutions can be pooled into a shared repository. Teams can create a common rubric, evaluate one another’s submissions, identify the strongest ideas, and synthesize hybrid solutions by combining elements from multiple responses.

This approach transforms the classroom into an innovation community rather than a collection of isolated learners. Students learn not only how to evaluate AI-generated ideas but also how to build on the contributions of others. They experience firsthand that innovation often emerges from recombination rather than individual genius.

The educational objective shifts from “my answer” to “our best answer.”

That is a profound shift, and it mirrors how much real work actually happens. Rarely does a finished strategy, design, policy, product, or venture emerge from one person thinking alone in a room. More often, it emerges from a messy process of generating, comparing, arguing, revising, combining, and defending.

The classroom should teach that.

Why This Matters

This method addresses several challenges at once. First, it acknowledges the reality that students will increasingly use AI tools in academic and professional settings. Rather than resisting that reality, it incorporates AI into the learning process in a structured and purposeful way.

(I believe we should REQUIRE the use of AI in the classroom.)

Second, it raises expectations. Students are not rewarded for submitting average AI-generated work. Instead, they are evaluated on their ability to identify, improve, and defend work that exceeds the AI baseline.

Third, it emphasizes judgment and taste. As AI systems become more capable of generating content, the ability to evaluate and select among alternatives becomes increasingly valuable. Fourth, it makes thinking visible. Students must demonstrate how they moved from a large set of possibilities to a final recommendation or solution.

Finally, it reframes AI from an answer-generating tool into a learning environment. The educational value lies not in the outputs themselves, but in the student’s ability to navigate, evaluate, and improve those outputs.

The Role of Human Effort

A common criticism of AI-assisted assignments is that they reduce student effort. This concern is valid when assignments are poorly designed. If a student can paste a prompt into an AI tool, receive a polished response, submit it, and get credit, then the assignment has failed.

But Curated Judgment does not eliminate effort. It redirects effort. The intellectual work shifts away from producing every sentence from scratch and toward evaluating alternatives, defining standards of quality, identifying weaknesses, synthesizing ideas, collaborating with peers, and defending decisions.

The rigor remains. It simply resides in different aspects of the learning process. We should not preserve effort merely because the old assignment required it. We should preserve effort where it produces learning.

Example 1: The Hundred Essay Method

Consider a traditional essay assignment. Instead of asking students to write a single essay from the outset, students could use AI to generate 100 possible thesis-driven responses to a prompt. Students would then develop a rubric for evaluating quality, apply that rubric to the responses, narrow the field to the ten strongest candidates, analyze common strengths and weaknesses, and synthesize a final essay that improves upon all of them.

Finally, they would present and defend both the essay and the reasoning behind their choices. In this model, students function not only as writers but also as evaluators, editors, researchers, synthesizers, and advocates. That is a harder job. It is also a better assignment.

Example 2: The Hundred Startup Ideas Method

A similar approach can be applied in entrepreneurship education. Students generate one hundred startup concepts within a defined market and evaluate them using criteria such as customer pain, feasibility, market size, defensibility, revenue potential, ethical considerations, and implementation cost.

After narrowing the field to the strongest candidates, students identify patterns, synthesize a stronger venture concept, and present it to the class. Discussion then focuses on assumptions, competitive dynamics, customer behavior, pricing strategies, and execution challenges.

The educational value lies not in generating ideas, but in recognizing which ideas merit further development and why. That is entrepreneurship.

Example 3: The Hundred Design Directions Method

This method can also be applied in creative education. Students generate one hundred design directions for a campaign, product, logo, interface, poster, or experience. They define the criteria for quality, sort the outputs into categories, eliminate derivative or impractical directions, select the strongest concepts, and synthesize a final creative direction.

They then present the design rationale and defend the trade-offs. The student learns that creativity is not merely generation. Creativity is selection, taste, judgment, refinement, and point of view. AI can generate options. The student must develop taste.

The Bigger Shift

Traditional assignments are often built around scarcity: one student, one prompt, one submission, and one grade. AI enables a different model based on abundance. Students can explore numerous possibilities, compare alternatives, examine failures, and synthesize insights from a much larger solution space.

In this environment, student value is no longer measured solely by the final artifact. It becomes visible through the decisions, evaluations, and judgments that shaped that artifact. The old classroom asked students to produce work under conditions of scarcity. The new classroom can ask students to think under conditions of abundance.

That may become one of the defining skills of the AI era.

The New Assignment

Generative AI has made average production inexpensive. That reality does not diminish the importance of education. It changes where educational value resides.

The future of teaching may depend less on asking students to produce a single answer and more on helping them navigate an abundance of possible answers. AI can generate countless alternatives, but it cannot relieve students of the responsibility to determine which ideas are strongest, what is missing, what should be combined, what should be rejected, and what conclusions they are ultimately prepared to defend.

That is the central premise of the Curated Judgement Pedagogy.

The assignment is no longer simply to produce an answer. The assignment is to transform abundance into judgment.

References

Bowen, J. A., & Watson, C. E. (2024). Teaching with AI: A Practical Guide to a New Era of Human Learning. Johns Hopkins University Press.

Chesbrough, H. W. (2003). Open Innovation: The New Imperative for Creating and Profiting from Technology. Harvard Business Press.

Chesbrough, H. W. (2020). Open Innovation Results: Going Beyond the Hype and Getting Down to Business. Oxford University Press.

Majchrzak, A., & Malhotra, A. (2019). Unleashing the Crowd: Collaborative Solutions to Wicked Business and Societal Problems. Palgrave Macmillan.

Notes:

• In case you’re wondering, I’ve been a faculty member at The School of Visual Arts for 30 years. I currently teach “AI for Creatives: Strategies for Success”, and I previously taught Computer Systems in the MFA Computer Arts Department. And, I’m an Entrepreneur in Residence at Cal Lutheran University.

• I am NOT sure how this pedagogy should apply to K-12 Education. I think AI in the K-9 age group should focus more in the approach that Salman Khan has built with KhanMigo. K10-12, on average, has more mature students, and they may have the critical thinking skills to engage with this methodology. I’m just an uncle… and I have no direct K-12 teaching experience and this is NOT my area of expertise… so… IDK. Here’s a new NBER paper on AI Integration in K-12.

This is a MUST READ article from Anthropic on how they’re using Claude to automate business analytics and data science.

But the most important point is not that Claude can write SQL. That part’s easy <cough>. The hard part is knowing which table to use, which definition of “active user” the business actually means, which dashboard is canonical, which metric is stale, which filter everyone forgets, and whether the answer is safe enough to forward to leadership.

That’s the real work.

Anthropic’s post makes something very clear: the future of analytics is not about giving everyone a chatbot connected to Snowflake. That’s how you manufacture confident nonsense at enterprise scale.

The future is a governed analytics harness: semantic layers, canonical datasets, agent skills, evals, CI, provenance footers, adversarial review, and human-owned definitions. In other words... today… the model is not the moat. The harness is. (For now.)

Anthropic says their internal data science team now automates 95% of business analytics queries using Claude, with roughly 95% aggregate accuracy, and around 99% accuracy in some domains.

Nice. Very nice.

But the real insight is this: when business analytics fails, it’s usually not because the AI can’t write SQL, R, or Python. It’s because the AI doesn’t understand the business.

Because “revenue” is not always revenue. “Active user” is not always active user. “Retention” is not always retention. Every company has a swamp of definitions, exceptions, stale dashboards, undocumented assumptions, legacy tables, and institutional scar tissue.

Welcome to enterprise data.
Please wipe your feet.

That’s why this is not really a post about SQL. It’s a post about analytical trust. And that means the analyst’s job is changing.

The analyst is no longer just the person pulling numbers or whispering to the data science team. The analyst becomes the architect of the system that makes AI-generated analysis trustworthy.

That means the new skill stack is not just SQL. It’s metric governance, semantic layer design, agent instruction design, evaluation engineering, and provenance design. Can you define the metric? Can you route the question through the canonical data model? Can you encode the judgment of a senior analyst into reusable agent skills? Can you test the agent’s answers before executives rely on them?

That is the work now.

The analyst does not disappear. The analyst gets promoted into the higher level architecture.

(Side note: This is an example of the effects of “Disruptive Innovation”.)

The Skills Are the Interesting Part

Anthropic’s skills are where this gets really interesting. Their pattern has three parts: knowledge skills, runbook-style skills, and reference docs written for LLM retrieval.

Knowledge skills route the agent to the right domain references, semantic layer, tables, joins, exclusions, and gotchas. This is the map that tells Claude where the reliable knowledge lives.

Runbook-style skills encode the workflow of a senior analyst: clarify the question, find the right source, use the semantic layer first, run the query, review assumptions, validate the result, and produce the final answer. This is not just knowledge. This is procedure.

Reference docs written for LLM retrieval turn messy human knowledge into agent-readable operating instructions: scope, exclusions, required filters, “when to use / when not to use,” routing triggers, and known traps.

This is the part I find most fascinating. We are watching institutional tacit knowledge get converted into agent-readable operating procedures.

The stuff that used to live in someone’s head now has to become explicit: “Don’t use that dashboard.” “That table is deprecated.” “Finance uses a different version of revenue.” “Ask Sarah before sending that number to leadership.”

All of that messy human knowledge now becomes part of the harness.

That is the real transformation.

We’re able to codify the tacit knowledge in an organization!

A Few Caveats

The 95% automation / 95% accuracy claim is impressive, but still under-specified. How complex were the queries? What counts as accuracy? How many answers needed human correction? And how much of this depends on Anthropic being Anthropic?

The jump from 21% without skills to 95%+ with skills is also striking. But that proves the point: the gains are not just coming from the model. They are coming from the operating environment around the model: the skills, documentation, governance, CI, semantic layer, and maintenance discipline.

Many companies do not have that level of canonical data hygiene.

And this quietly undercuts a lot of “just connect AI to your database” vendor hype. Access alone is not enough. Your enterprise chatbot does not become smart because you connected it to a swamp. It becomes dangerous.

The Big Takeaway

Is this going to be generalizable? Maybe. Highly likely.
Is this a state machine? It feels like it.
Is this a decision tree? It can be.
Is this deterministic? Yes, similar to an expert system.
Is this fragile. Yeah… everything is fragile. _(But…???)
Is this gonna burn lots of tokens? Yes. By design.
Is this better than fine tuning a model? Fine tuning models is a waste.

Anthropic’s article is not really about Claude writing SQL. It is about the new enterprise analytics architecture: governed data models + semantic layers + agent skills + evals + provenance + adversarial review + human ownership.

That is the new stack.
_{(Well… for now. I give it a few days/weeks and something else will replace it.)}

The analyst’s job is not going away. It is moving up the value chain from pulling numbers to designing trust, and from answering questions to governing how questions get answered.

So yes...

Claude can write SQL.

Big freakin’ deal.

The real question is:
Can it understand revenue?

The Dumb New Metric

Tokenmaxxing is what happens when companies confuse AI activity with AI productivity. Michael Burry said that tokenmaxxing is a, “crazy, rushed, temporary phase.1” I completely agree.

It’s not about using more tokens. It’s about better thinking, better agents, better outcomes.

Businesses are treating tokens like they’re some “proof of innovation”, but tokens are not productivity. They are input costs. The real question is not “How much AI are we using?” It’s “How much value are we creating per unit of AI cognition?”

The Wrong Scorecard

And whenever we try to measure something… we need to be mindful of Goodhart’s Law (“when a measure becomes a target, it ceases to be a good measure”).

“Pretty sure 50% of internal token spend is completely useless, but right now it’s hard to know which 50%.” -Akshat Bubna 2

This sounds like it was inspired by the attributed quote from John Wanamaker, “50 percent of my advertising budget works, I just don’t know which half.”

When token usage becomes the target, the organization doesn’t become more intelligent. It just becomes more expensive.

The Better Metric

Tokenomics is the discipline of maximizing value per token in human-agent workflows.

Tokenmaxxing is what happens when organizations confuse AI activity with AI productivity. Tokenomics is the discipline of converting AI tokens into measurable business value.

Agentic AI Problem

Agentic AI makes this problem worse because agents don’t just answer questions. They plan, search, reason, call tools, retry, reflect, and loop. An inefficient agent is not a chatbot with bad habits. It is a tiny autonomous cost center.

The Efficiency Counter-Signal

At the exact same moment companies are learning how easy it is to waste tokens, researchers are showing us how much efficiency is still trapped inside the system.

Last night during my office hours, I mentioned two important research papers that make AI orders of magnitude more efficient.

DiffusionBlocks

The first is DiffusionBlocks from the freakin’ geniuses at Sakana.ai.3 Their algo potentially democratizes AI training. Right now, training frontier AI models requires thousands of specialized GPUs and tens of millions of dollars, largely because the models are too massive to fit into standard hardware memory. By reducing the training memory by a factor of the number of blocks used, DiffusionBlocks opens the door for individual researchers, students, and smaller startups to train or fine-tune powerful, large-scale AI models on much humbler hardware. Technically, they discovered that data running through a neural network is surprisingly similar to a ‘diffusion model’ (the same algo behind Stable Diffusion, that’s making all the kewl AI images these days.) A diffusion model starts with random noise and gradually clears it up step-by-step until it’s a perfect picture. DiffusionBlocks treats each block of the neural network as one of those “clearing up” steps. Because this is mathematically grounded, the blocks can train independently without losing coordination. BRILLIANT!

OSCAR

The second paper is called OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization4. (This gets me uncomfortably excited.)

When you chat with an LLM over a long conversation, it stores a history of the interaction in what’s called a Key-Value (KV) cache. This cache grows linearly with the length of the prompt and the number of users, quickly eating up expensive GPU memory and slowing down text generation.

Engineers want to compress this KV cache down to ultra-low precision. Specifically to a 2-bit (INT2), which would theoretically reduce the memory footprint by about 8x compared to standard 16-bit floating point (BF16). However, standard 2-bit compression usually destroys model accuracy because LLM activations contain extreme “outlier” values in certain channels, and this compression loses those outliers.

OSCAR’s big leap is shifting the goalpost from raw data reconstruction to downstream attention impact. Instead of trying to perfectly preserve the exact original numbers in the cache, OSCAR asks: “What does the model’s attention mechanism actually care about?” It runs a lightweight, one-time calibration pass to calculate how the model focuses its downstream attention, and it basically finds the mathematical directions that the model reads most strongly.

By aligning the BF16 to INT2 compression with the model’s actual mathematical intent, OSCAR achieves staggering system gains without losing the model’s mind. They maintain a 1.4-point gap with the uncompressed performance. (That’s insanely close!) And this cuts the KV cache memory usage by 8x.

Signals Decoded

The next phase of AI will not be about who uses the most tokens. It will be about who creates the most value per token.

What this means is that we’re seeing MASSIVE algorithmic gains, at the same time we’re seeing increased demand for tokens, at the same time we’re seeing massive investment in AI data centers.

And we’re moving towards an “unlimited intelligence” stage where edge devices (like laptops and smartphones) will run the equivalent of SOTA models that require cloud compute.

I think we’re gonna see an inflection point in late 2027 or early 2028.

Not because companies learned how to use more AI.

Because they learned how to waste less of it.

The next phase of AI will not be measured in tokens consumed. It will be measured in value created.

Tokenmaxxing is AI theater. Tokenomics is AI strategy.

And honestly… I get it.

Entry-level jobs are being decimated. The career ladder they were promised is starting to look like it’s missing the first few rungs. The future feels unfair, confusing, and hostile.

So yes… I understand the anger.

But if I were speaking to those graduates, I would have read the room first. I would have acknowledged the anti-AI sentiment. I would have told them their fear is rational.

And then I would have shown them a pathway forward. Because we need strong leadership right now.

We are living in Anno Intelligentiae… the Age of Intelligence

And I believe this anti-AI mindset is going to hurt recent graduates far more than they realize.

Not because their concerns are wrong. But because their strategy is wrong.

And, to be blunt, I blame the professors and university administrators who “banned AI” in their classrooms. That story taught students that using AI is cheating. That AI is something to avoid. That the moral position is refusal. That the professional position is human-authored purity.

And that was exactly the wrong lesson to teach.

Those professors should have been removed from the classroom three years ago because they were not preparing students for the world as it already existed... three years ago!

And now we have 2026 graduates entering the workforce completely unprepared for the world we actually live in. They don’t have AI skills. They see AI only as a threat. They do not see the massive opportunities sitting right in front of them.

Shame.

And honestly… shame on me for not screaming louder.

I mean… I tried.

I really did.
I raised my hand to ask the uncomfortable questions. I spoke at conferences. I sent emails. I taught my classes. I gave workshops. I hosted Live Session discussions.

I’ve been trying to be a leader here.
I’ve been trying to show the pathway forward.

And unfortunately, we’re exactly where I said we would be if we refused to change.

The cost of cognition and creativity is basically $20/month right now. Or, you can run an open weights LLM on your laptop with LMstudio or Ollama for free. That changes everything.

The good news is that humanity is NOT running out of problems to solve any time soon. Which means the future of work is not simply “jobs.” The future of work is Entrepreneurship. That’s the north star.

Not because everyone needs to become a startup founder. But because everyone needs to learn how to identify problems, create value, use tools, build systems, tell stories, serve communities, and bring something new into the world.

We need to master “The Art of the Possible.”

That is the real AI literacy.

And right now, too many schools are still preparing students for a world that no longer exists.

And I predict that we’re gonna get “Unlimited Intelligence”… like unlimited data/voice/texting plans… some time around 2027-2028.

When that happens, the question will no longer be: “Do you know how to use AI?”

The question will be: “What can you do with unlimited intelligence?”

That’s the conversation we should be having with students! And that’s the future they should be preparing for.

But Unlimited Intelligence is for an upcoming article… so… PLEASE SUBSCRIBE!

And… please subscribe to my YouTube Channel!
I host weekly Office Hours where we discuss the latest in AI News.

Open Disruption YouTube

And check out what GPT Image 2 generated as thumbnails for this post!

Not by changing or retraining the model itself... but by building systems around it.

Those systems can preserve memory across runs, score what went wrong in prior agent trajectories, rewrite prompts, rewrite Python code, update rules, swap tools, and even change which agents get used for which tasks. In other words, the learning is increasingly happening in the harness around the model... not only inside the model.

When an agent fails, we can inspect the failure, trace where the chain broke, formulate a better rule, and feed that lesson back into the system. That might mean changing a prompt, updating a retrieval step, modifying a tool, or replacing one agent with another better suited to the task. In multi-agent RAG systems, this can look like optimizing not just the words, but the topology... who does what, in what order, with which context, under which constraints.

This is not the same thing as updating the weights and biases of the base LLM. It is something more modular, more legible, and in some cases more practical. We are building externalized forms of learning... memory layers, experience libraries, rule sets, prompt evolution, code edits, orchestration policies, and agent teams that adapt over time.

Here’s an actual example:
A research agent cites the wrong source. Instead of retraining the model, the agent team updates the retrieval logic, adds a verification step, rewrites the agent prompt, and stores the failure pattern in memory so the system is less likely to repeat it.

One recent paper describes exactly this move in multi-agent RAG: improving system performance by evolving orchestration strategies and role-specific prompts, while keeping the underlying model frozen1. Another shows that major gains in agent memory did not come primarily from tuning knobs, but from fixing bugs, changing architecture, and improving prompts through repeated autonomous experimentation.2

Of course, all of this is still constrained by the underlying model’s capabilities. The harness can only stretch so far. They still inherit the reasoning limits of the base model, they can compound hallucinations if badly designed, they can overfit to local heuristics, and they can become brittle if the memory layer is noisy or if the evaluation loop is weak.

And the base models are improving quickly... which means the leverage of these outer systems is improving too.

The real shift is not that the model learns after deployment. It’s that the system around the model can now accumulate experience, encode lessons, learn human subjective ‘taste’, and adapt over time.

This also connects to the broader push toward Continual Learning. Some of that work focuses on updating model parameters over time without catastrophic forgetting, and realtime alternatives to back propagation. But another branch, and to me the more immediately (and cost effective) important one, focuses on test-time adaptation, modular memory, and externalized learning systems that can accumulate experience after deployment without retraining the entire model. The frontier is no longer just how to make the model learn... but how to make the system learn safely, selectively, and continuously around the model.345

Why does this technical shift matter?

The scarce resource is no longer only access to a powerful model. It’s now also:

• the ability to instrument and observe failure.

• the ability to capture experience and taste.

• the ability to encode lessons into prompts, tools, workflows, and memory

• the ability to continually improve system performance faster than competitors.

That’s where the power shift becomes visible. In other words... the moat is not just the model. It’s also the adaptive layer around it.

What worries me is how few people understand any of this.

Most people I speak with, and many I teach (here and here), still do not know what a token is. Or an embedding. Or a context window. Or RAG. Or a vector database. Some have never even touched a markdown file. (And some people I know still actively resist AI, ban it, refuse to learn it, and truly believe that if they ‘stick together’, that they will win. I worry that most of them will suffer.)

They do not know where these systems break, where they succeed, or how quickly they are improving. That ignorance is not trivial. It’s economic.

Knowing what lies inside and outside of the Jagged Frontier… the uneven capability boundaries of LLMs… is increasingly tied to the value of knowledge work.6 The people who understand these systems, their limits and boundary conditions, and the layers forming around them are gaining an exponential advantage over those who don’t.

Some people are learning how to build compounding cognitive infrastructure. Others are still learning vocabulary.

And the organizations and business leaders that learn how to externalize, preserve, and operationalize machine experience will outcompete those still treating AI as a simple chatbot.

A new architecture of power.

I still hold a positive vision of ‘post-labor economics’... of abundance, dignity, and broader human flourishing.

But I worry that most people are functionally unprepared for Anno Intelligentiae.

And if that does not change... the more likely future is not liberation, but dependency. The danger is not just that a few firms own the models. It is that they may also own the memory, the tooling, the orchestration, the feedback loops, and the adaptive systems that make those models actually useful.

Not abundance for all, but serfdom inside someone else’s kingdom.

Folks… please pay attention!

Please check out my YouTube channel!
Each week I cover the latest news in AI.

It’s always been a bit sloppy. And now it may be dangerous.

Because in the age of AI… the Age of Intelligence… Anno Intelligentiae… the difference between a fine artist and a designer is no longer just a cultural distinction... it may become an economic fault line.

We’ve trained generations of creatives to believe these categories overlap so much that the distinction barely matters. Both make visual work. Both use craft. Both have taste. Both shape culture.

But that misses the point.

The real difference is not what they make. It is how they create value... and why.

It is the logic of value creation underneath the work.

Designers create value primarily by solving external problems.

Fine artists create value primarily by advancing internal vision.

And once you see that... a lot of what is about to happen becomes easier to understand.

The 2x2 that clarified the whole thing

I originally thought about the difference as a simple single dimension spectrum.

On one side, self-expression. On the other, commercial or industry orientation.

That seemed useful... until HBS Professor Carliss Baldwin said, “I think this is empirically wrong”. She said that there’s actually a 2x2 that better describes this.

She is right. The one-dimensional model is too simple. A 2x2 reveals far more.1

Once you put commercial or industry orientation on the x-axis and self-expression on the y-axis, the fog starts to lift.

You get four quadrants:

Amateur or Hobbyist
Low self-expression, low commercial orientation.

Designer
High commercial orientation, lower self-expression.

Fine Artist
High self-expression, lower commercial orientation.

Hybrid or Star
High self-expression, high commercial orientation.

Simple framework... but it exposes something people in the creative world often do not want to say out loud.

Designers and fine artists are not doing the same job

A designer’s job is not mainly to express themselves.

A designer’s job is to make something work. On average, they solve problems for an industry and are part of a value chain.

They solve for users, markets, brands, systems, constraints, production realities, business goals, conversion targets, manufacturability, legibility, usability, and scalability. (And many more things, I’m sure!)

A designer creates value by responding effectively to an external brief.
(Sometimes they’ll create that brief too, and I kinda place this into the Entrepreneur or Product Manager category. I need to think about this more.)

A fine artist does something else.

A fine artist creates value by advancing an internal vision.

Their work is tied to authorship, perspective, identity, interpretation, meaning, symbolism, and cultural signal. The work matters because it comes from them... not because it optimizes a funnel, improves usability, or hits a KPI.

That doesn’t make one superior to the other.

But it does make them fundamentally different.

And pretending otherwise has allowed a lot of confusion to accumulate.

Here’s the part that will likely annoy people

A lot of designers want to be treated like artists... without giving up the commercial logic of design. They want the prestige of authorship, the mystique of self-expression, the status of being a visionary.

But they are still working inside someone else’s system, solving someone else’s problem, often for someone else’s business model.

That is not an insult.

That is just a different value proposition.

And on the other side...

A lot of fine artists want to claim they are above the market... while quietly wanting followers, collectors, validation, sales, brand collaborations, and institutional legitimacy. #StarvingArtist

So both sides perform a kind of hypocrisy.

Designers often romanticize artistic freedom.

Artists often romanticize commercial independence.

And both sides sometimes avoid admitting what game they are actually playing.

There’s another layer here... the Artist Tax, or Passion Tax that creatives pay for the opportunity to work on fun projects. Business leaders know that they can pay creatives less money for their labor because the pool of creatives looking for good projects is larger than the pool of good projects. Creatives, on average, have a lower “Willingness to Sell” than many other professions. This is also one of the reasons why ‘Outside-In Open Innovation’ works. (Please read “Better, Simpler Strategy: A Value-Based Guide to Exceptional Performance” by Felix Oberholzer-Gee.)

Why AI makes this distinction much more brutal

This is where the conversation gets real.

AI does not care about our cultural confusion.

AI is disproportionately good at tasks where success can be defined by constraints.

Give it a brief. Give it examples. Give it a style. Give it a target customer. Give it performance criteria. Give it brand rules. Give it a workflow. Give it a SPEC.md.

That is exactly the kind of environment where design labor begins to get squeezed.

Not all design... but a lot of it.

Especially the layers of design that are repetitive, derivative, production-oriented, optimization-driven, or heavily constrained by existing systems. (And by my calculation, too many of the courses offered at SVA are teaching skills that are within the path of the AI Tidal Wave.)

That should make designers very nervous.

Because much of the economic value of design has historically come from being the human who could translate constraints into competent output.

And “competent output under constraints” is exactly where AI is improving fastest.

Meanwhile, fine art is exposed differently.

The fine art market is not just buying the artifact. It is buying authorship, provenance, scarcity, story, status, identity, and cultural position. It is buying the relationship between the object and the human who made it.

That does not mean fine artists are safe.

It means the rules of disruption are different.

AI can generate images.

It cannot so easily generate you... your biography, your position, your social signal, your lived experience, your symbolic place inside culture.

At least not yet... and not in the same way.

That makes fine art less immediately vulnerable to replacement, even if it is still vulnerable to flooding, dilution, and aesthetic confusion.

Oh… and I can’t forget to mention that even the Fine Art Industry has a long history of “gatekeepers” and market forces. (See this, and this. I often wonder how many Fine Artists actually understand the business that they are in? Have they even read these reports? Do they even know they exist?)

The uncomfortable irony

For years, many people treated design as the practical path and fine art as the risky one.

But in an AI-saturated world, that assumption may flip in strange ways.

The more your value depends on solving clear external problems under known constraints, the more you may be standing directly in the blast radius.

The more your value depends on irreducible authorship, cultural signal, and singular point of view, the more defensible your position may become.

That is a brutal irony.

The person trained to be “market relevant” may end up more replaceable than the person trained to be distinctive.

Read that again.

Because I think a lot of schools are still educating creative people for a world that has already disappeared.

Want proof? If you are looking for a Master’s Degree in UI/UX… and you think you’re gonna get a job designing user interfaces after you graduate, then you need to take a look at Google Lab’s “Stitch”. Because this ‘beta proof of concept tool’ already replaces between 25% to 95% of the UI/UX designer’s tasks.

The fantasy of the hybrid

Of course, everyone wants to imagine they live in the top-right quadrant... the Hybrid or Star.

High self-expression, high commercial orientation.

The person whose voice is unique and scalable. The creator whose taste becomes a market. The artist who becomes a brand. The designer who becomes a cultural force.

This is the dream.

It is also rare.

Painfully rare.

Most people are not hybrids. They are aspirants.

And that matters, because false self-diagnosis is one of the biggest career risks in the creative economy.

If you think you are being rewarded for your originality when you are actually being paid for your utility... you will misread the threat.

If you think you are an artist when the market sees you as a service provider... you will make bad strategic decisions.

If you think you are above commerce when your whole career depends on attention markets... you are lying to yourself.

That’s called the “Outlier Economy”... and it’s really hard to get there.

So what is the actual question?

Not “Which one is better?”

That’s a lazy question.

The real question is:

What kind of value do you create... and how exposed is that value to systems that can now generate competent creative work at scale?

That is the question every designer should be asking.

That is the question every fine artist should be asking.

That is the question every art school, design school, agency, gallery, and creative founder should be asking.

Because the old prestige games are breaking down.

And once AI can produce infinite amounts of polished, plausible, on-brief creative output, the premium shifts.

Toward taste.

Toward judgment.

Toward strategy.

Toward identity.

Toward trust.

Toward cultural positioning.

Toward work that feels difficult to substitute, even when images themselves are easy to generate.

The bottom line

Fine artists are not designers.

Designers are not fine artists.

They may share tools, mediums, skills, techniques, and talent.

But they do not always share the same economic logic.

One creates value primarily through utility. The other creates value primarily through authorship.

And in the age of AI, that distinction is no longer academic.

It may determine who gets commoditized first... and who doesn’t.

Moving Forward: A Proposal

If this diagnosis is even partially correct, then designers, creatives, and artists need a new agenda. Not one built on nostalgia, denial, or purity politics, and not one built on the fantasy that the old creative economy is somehow coming back. We need an agenda built around leverage, ownership, and adaptation.

First, creatives need to embrace AI as leverage, not betrayal. AI is no longer a purity test. The question is not whether these tools will be used. They will be. The real question is whether creatives will use Generative AI, cognitive automation, and Agentic AI to expand their own agency... or whether others will use those same tools to reduce creative labor to a cheaper input.

Second, creatives need to move from labor to ownership. Using AI to become a faster worker is not enough. If all we do is become more efficient inside someone else’s business model, then we may simply help accelerate our own commoditization. The real goal should be to own more of the upside... products, intellectual property, audiences, systems, brands, and businesses.

Third, creatives need to move up the stack. As execution gets cheaper, the premium shifts upward, toward taste, judgment, problem selection, direction, synthesis, and trust. AI can generate options and automate parts of production, but it is still much weaker at deciding what is worth making, why it matters, who it is for, and how it fits into a larger system. Creatives need to position themselves where value is hardest to automate.

Fourth, creatives need to become more entrepreneurial. The ability to stare at a blank canvas and imagine a future state is entrepreneurial at its core. Creatives are already trained in ambiguity, iteration, reframing, and possibility. Those are not just artistic skills... they are builder skills. In the age of AI, the blank canvas is no longer just a painting or a layout... it needs to be a company. It needs to become a product, a platform, a brand, a service, or some new engine of value creation and value capture.

Fifth, creatives need to build collective power. Individual adaptation is not enough. If people respond one by one, many will simply become more efficient, more isolated, and more replaceable. Creatives need new organizations, networks, and alliances that help them share knowledge, increase leverage, and negotiate the terms of participation in an AI economy.

And sixth, we need a new educational agenda. Art schools and design schools should not be preparing students for shrinking roles inside legacy workflows. They should be teaching AI fluency, entrepreneurship, business models, systems thinking, and the economics of creative work. They MUST require students to use AI in every class. And teachers who ban the use of AI are no longer qualified to teach in the classroom because they are no longer preparing students for the world we currently live in. In the future, it will not be enough to leave school with a portfolio. More students will need the ability to build ventures.

That’s the proposal.

Creatives must stop thinking only as ‘makers for hire’ and start thinking more like builders, owners, and founders. In the age of AI, the most important thing a designer, artist, or creative may ever design is not just an artifact... it may be a company.

I assume you’ve already read this:

Citrini Research

THE 2028 GLOBAL INTELLIGENCE CRISIS

Preface…

Read more

6 months ago · 5289 likes · 82 comments · Citrini and Alap Shah

Worth repeating… it’s explicitly a scenario, not a prediction… but it models the “abundant intelligence” downside in a way most people still haven’t internalized.

(And somehow… that triggered a market selloff. I’ve been talking about this for 3+ years, and it triggers a selloff now?!?! IDK anymore.)

This is just the beginning of the transformation.

Welcome to “Anno Intelligentiae”, the Age of Intelligence!

Constraint Zero… compute, energy, grid, cooling… and WATER

When intelligence becomes abundant, the moats are physical constraints and bottlenecks.

Compute and Energy are the major constraints.
Invest in compute, energy, chips, cooling, grid capacity… and WATER! (See below)

Everything else becomes negotiable, copyable, or automated.

Also… we have a global compute single point of failure. Taiwan risk is not a tweet, it’s a supply chain reality. That’s not a prediction, it’s a dependency graph.

“Build your own power”! Hyperscalers MUST move from buying electrons to building generation (gas, nuclear, microgrids) because grid queues and turbine supply become chokepoints and are risky. (See Trump SOTU)

Democratization Effects

Democratization Effects means capability goes abundant, and business models built on scarcity get repriced toward zero. Your domain expertise is no longer as valuable as it once was. It’s still valuable… but the value is decreasing.

“Democratization Effects” destroy entire businesses and industries.

Intermediaries get hollowed out first. (anything that exists to help humans compare, choose, schedule, fill forms, route tickets, etc.)

Be someplace else! Preferably a platform… or HALO (see below) that has a revenue generating complimentary that allows for the loss leader.

Labor and productivity shocks show up where work is digital, subjective, and lightly regulated. Advertising and Marketing can now be fully automated. Those who don’t see this are at risk.

Whiplash Effects

Whiplash Effects means hype, panic, layoffs, regulation talk, market rotation, consolidation… then it dissipates… then it repeats.

“Whiplash Effects” always happen and always dissipate… like ripples in a pond.

These cycles repeat: hype, layoffs, regulation talk, market rotation, consolidation

Watching #Klarna gives you a “second mover advantage”. Dont be Klarna.
Smart CEOs wont make the same mistake twice. #HuntForRedOctober
(They will make new/different mistakes.)

Reuters basically described this mood swing as an “AI doom bubble” moment… which is just another way of saying, whiplash.

SaaS is dead

(Micro-glossary) “SaaS is dead” means seat-based pricing and “ARR as religion” breaks once agents replicate workflows.

SaaS is dead. (Ok,fine... “traditional SaaS is dead”, but do you really think any non-traditional SaaS is going to survive without a clever moat? Nope! SaaS is still dead.)

Revenue compression. ‘ARR as religion’ breaks when agents replicate workflows.
Transform your business model or die.

Outcome-based work is interesting. Token, task, and outcome billing replaces seats. You’ll see agent fleets priced by consumption, throughput, guarantees, and SLA, not per-user licensing.

Big Consulting is dead… they need to pivot to implementation with some ‘skin in the game’. (Imagine ‘Consulting Implementation’ that bills on outcomes and has a piece of the action. If the customer does well, then they do well.)

A2A… Apps to Agents

(Micro-glossary) A2A means agents coordinate directly across systems, and humans stop being the routing layer. Google is explicitly pushing an “Agent2Agent” interoperability protocol framing.

A2A transforms how people and businesses create and capture value. From “Apps to Agents”, app loyalty collapses. If you have an ‘App’, then build the A2A RIGHT NOW.

Agents kill funnels!
Discovery is A2A!

Agent marketplaces become a distribution war. Not “app stores,” more like “capability stores” with trust scores, proofs, and liability wrappers.

(And yes… the protocol stack soup is real… MCP tool, A2A, AP2, UCP, A2UI, etc. That’s not just product… it’s also risk.)

Jagged Frontier

(Micro-glossary) Jagged Frontier means AI is uneven… it crushes some tasks and faceplants on others that look similar. The term is popularized in the HBS work on “jagged technological frontier.”

(Depending on Industry and Jagged Frontier, Agents outperform 95% of all humans.)

This is why “HITL everywhere” is dumb… and “full autonomy everywhere” is also dumb. The frontier is jagged, not smooth.

HITL

(Micro-glossary) HITL means humans stay in the loop for accuracy, safety, accountability, auditability… not because humans are faster.

HITL = Accountability, and is mostly for the C-Suite.
(HITL answers the question, “Who goes to prison when the AI hurts someone?”)

Signatories, escalation points, and liability owners… especially at the top.
“human sign-off layers,” are for accountability and risk, not for productivity.

Insurance becomes an AI Agent feature. Products bundle “who pays when the agent screws up,” and that becomes a differentiator, especially in regulated workflows.

GaaS

GaaS means Governance as a Service… a productized layer for permissions, audit trails, monitoring, and “safe, compliant autonomy.”

‘System of Record’ and ‘governance’ are still critically important. (#Watching carefully!)
Truth, provenance, permissions, audit trails… all go up in importance and value.

GaaS is an HUGE opportunity: “Governance as a Service” becomes a product.
“Safe, compliant autonomy” becomes an enterprise buying center.
(governance, identity, permissions, audit, safety)

If you can’t produce audit trails, risk assessments, and oversight artifacts on demand, you can’t deploy at scale.

Agent Identity Governance is desperately needed. Agents need identities, permissions, key management, revocation, and monitoring… like employees, but at machine scale.

“Agent Managers” are going to be as short lived as “Prompt Engineers”. (Think ‘prompt optimization’ but for Agents)
^ Must read article, btw!

Continuous Learning… and tacit knowledge

Continuous Learning algos (ie Parametric CL) have a viable working proxy… the file system!
(And… your data exhaust becomes more valuable.)

We can now codify the ‘tacit knowledge’ in an AI-first Organization, and act on it.

“Code must not be written by humans. Code must not be reviewed by humans.” (StrongDM)

Brand, trust, and substitution

Brand loyalty collapses as ‘longitudinal trust’ in your personal Agent grows.

There’s a “Coke vs Pepsi” debate in here, however. Consumer preference still matters! Trademarks have value.

The more substitutes in your industry, the greater the risk.

Security… nothing is secure

Security! Nothing is secure.

This is about risk and liability.
(Talk to a lawyer, and accountant, and an Entrepreneur and you’ll get 3 different recommendations.)

Agent incidents will look like real-world fraud, not “model mistakes.” Tool misuse, privilege escalation, prompt injections, goal hijacking, memory poisoning, cascading failures… the failures move from “wrong answers” to “unauthorized actions.”

(I’m also wondering about the MCP tool/A2A/AP2/UCP/A2UI/etc protocol connections risk. More to come on that.)

Bayesian risk vs Knightian uncertainty

Bayesian risk is modelable uncertainty… Knightian uncertainty is “unknown unknowns,” not susceptible to clean measurement, and that distinction matters in decision-making.

Pay close attention to Agents that can operate within Bayesian risk and Knightian uncertainty.

Agent behaviors in “unknown unknowns” scenarios that can reduce risk in unmeasurable and uninsurable situations can unlock massive potential.

HALO

HALO means Heavy Assets, Low Obsolescence… a defensive moat when investors fear software gets atomized.

And I love this concept of “HALO”: Heavy Assets, Low Obsolescence

Investors fear “software gets atomized”… HALO is a possible defensive moat.
(I wonder if there’s gonna be an investment rotation?)

Water… we are water bankrupt!

Bankrupt! We’re past salvation.
There’s only one long-term path forward: Ocean desalination.

Water constraints on AI will drive more NIMBY backlash than rising electricity costs.

GLOBAL WATER BANKRUPTCY
United Nations University Institute for Water, Environment and Health
Living Beyond Our Hydrological Means in the Post-Crisis Era

We need a lot more energy.
We should be building solar everywhere.
We should use that energy for every industry, AI, and to desalinate the oceans.
The US is lacking in energy compared to China.

(And what about Regulation? This is a can of worms for another post.)

The minimum viable team collapses

The “minimum viable team” within an organization collapses. HR struggles to adapt.

Tiny teams ship what used to require departments, and the new scarcity becomes taste, domain judgment, and risk ownership.

(Where this gets fun is that we can now encode “taste”, “domain judgment”, and “risk ownership” into an AI agent.)

Five moves (for founders, operators, investors)

1. If you sell software… assume revenue compression, then design the moat.

2. If you ship an app… build A2A, or accept irrelevance.

3. If you deploy agents… treat governance as product, not policy.

4. If you operate in uncertainty… build escalation paths and decision traces (HITL isn’t optional, it just moves). Codify Tacit Knowledge at every stage possible!

5. If you invest… follow constraints (compute, energy, grid, cooling, water), then follow HALO.

And while I have your attention, please subscribe to my YouTube channel!
I post a weekly news hour on AI.

If you’ve ever seen p < 0.05 treated like a divine proclamation handed down from Mount Sinai, congratulations… you’ve witnessed one of the most successful rituals in modern science.

Today, I want to talk about what p-values actually are, why they’re so easy to misunderstand, and how c-values offer a more intuitive way to think about evidence. And if you haven’t heard of a c-value… keep reading. Because in the real world, decisions don’t wait for statistical significance.

Share

A Very Quick Refresher: Hypothesis Testing

Hypothesis testing is basically a structured way to ask:

“Is this result real… or did randomness just prank me?”

We start with two competing ideas:

• Null Hypothesis (Ho): nothing interesting is happening

• Alternative Hypothesis (Ha): something interesting is happening

Example:

• Ho: This new study method does not improve test scores

• Ha: It does improve test scores

We collect data… run a test… and compute a p-value.

Professor Mike Parzen @HBAP taught us, “If the P is low, the null must go.” That means we “reject the Null Hypothesis”. If the p-value is high… or higher than 0.05… then we “fail to reject the Null Hypothesis.”

What a p-value Actually Is (Not What People Think)

Let’s say your statistical test returns:

p = 0.03

This does not mean:
“There’s a 3% chance the null hypothesis is true.”
“There’s a 97% chance our method works.”
“We are 97% smart and 3% unlucky.”

What it actually means:
If the null hypothesis were true, results this extreme would occur about 3% of the time.

That’s it.
It’s a statement about data assuming the null… not about the truth of the null.

Don’t hate me for this. But… the formal definition of a p-value:
“P values are the probability of observing a sample statistic that is at least as extreme as your sample statistic when you assume that the null hypothesis is true.”

Yeah… one of my biggest fears is trying to fully understand that. Good luck with that.

(one of my heroes) said, “Does the Evidence we collect make the Ho look Ridiculous? Yes, or No?” This is genius.

Why p-values Feel So… Unsatisfying

Here’s the problem…

Humans don’t naturally think in terms of:
“How surprising is this data if nothing is happening?”

We think in terms of:
“Given this evidence… should I believe the null?”

But p-values don’t answer that question.

They answer the backward version.

It’s like asking:
“If I weren’t hungry, how surprising would this pizza be?”

Not helpful.

The Hidden Distance Meter Inside Every p-value

Here’s the part I feel most textbooks dont explain well.

Before we ever get a p-value, we calculate a test statistic.

That statistic measures how far the data are from what the null predicts.

Any stats folks reading this likely know what I’m referring to:
z-score = distance in standard deviations
t-statistic = standardized distance
chi-square = squared distance

Then we convert that distance into a probability… the p-value.

So while a p-value is a probability… it’s really a probability translation of distance from the null.

Small p-value = the data is far from null hypothesis
Large p-value = the data is close to null hyptohesis

I like to think of it as a “null discomfort index.”

Enter the C-Value: A Confidence-Based Credibility Score

Now we meet the new understudy that never got Broadway billing: the c-value.

Back in 2023 an incredible paper “Confidently Comparing Estimates with the c-value” was published in the Journal of the American Statistical Association1. The authors, Brian L. Trippe, Sameer Deshpande, and Tamara Broderick are nothing short of genius. I think this may be one of the most important developments in Statistics in years.

The “c-value” is short for credible value, it asks the question people think the p-value answers:

How credible is the null hypothesis given the data?

Not “How surprising is the data if the null is true?”
But “How believable is the null after seeing the data?”

Subtle shift… huge difference.

Technically, the c-value measures the smallest confidence level at which two estimates would still be considered compatible.

Think of it like this:
p-value = how extreme is the data under the null
c-value = how confidently can we say two quantities differ

A Courtroom Analogy… because Statistics is basically “The Law” for numbers

And when I say “The Law”, you should imagine this spoken in a Judge Dredd voice.2
(Ok.. focus, Todd.)

So… imagine a trial.

• p-value asks:
  If the defendant were innocent, how surprising is this evidence?

• c-value asks:
  Given this evidence, how credible is the defendant’s innocence?

One evaluates the evidence under innocence.
The other evaluates innocence under the evidence.

My brain prefers the second.

Side Bar: How the Law Actually Reasons About Evidence

In criminal law, the standard is:
Proof beyond a reasonable doubt

That’s not a p-value threshold. It is a judgment about belief in guilt after considering the evidence. That’s much closer to Bayesian reasoning. (That’s coming!)

The Direction of Legal Reasoning

A jury is not asked:
“If the defendant were innocent, how surprising is this evidence?”

They are asked:
“Given the evidence, are we convinced the defendant is guilty beyond a reasonable doubt?”

That’s evidence → belief.
Not hypothesis → probability of data.

Courts don’t ask how surprising the evidence would be if the defendant were innocent, they ask whether the evidence makes innocence believable, which is why legal reasoning feels more Bayesian than frequentist. Plus, reducing guilt to a probability would create false precision. And lawyers would argue about whether “beyond a reasonable doubt” means 90%, 95%, or 99.9%… forever.

The prosecution must show evidence that would be very unlikely if the defendant were innocent. In other words, evidence must be inconsistent with innocence, and alternative explanations must be implausible. This resembles the idea of rejecting a null hypothesis. But the court never assigns a numerical probability threshold. There is no legal equivalent of p < 0.05.

Why This Matters (a.k.a. The Replication Crisis Has Entered the Room)

The obsession with p < 0.05 has led to tons of problems. “Significant” results that don’t replicate. Trivial effects declared important. Researchers treating 0.049 like victory and 0.051 like failure. Entire careers built on decimal dust.

The real issue isn’t that p-values are “wrong.” It’s that binary threshold thinking is fragile. When 0.049 and 0.051 produce opposite headlines, the system is incentivizing noise.

(And don’t get me started on p-hacking researchers who use an Alpha of 0.1 and 90% Confidence Intervals to show an effect when they KNOW that they wouldn’t be able to publish anything at the 0.05 Alpha.)

C-values encourage a more nuanced view. Evidence is graded, not binary. Credibility is contextual. Uncertainty is acknowledged.

In other words… science behaves more like science.

A Student-Friendly Mental Model

If your brain starts melting, use this:
p-value = how weird the data would be if nothing is happening
c-value = how confident we are that something is happening

or… how about this one? (Thank you ChatGPT for this nugget!)
p-value asks: “Is this surprising if nothing is happening?”
c-value asks: “How confident are we that nothing isn’t happening?”

That’s the whole game.

A Simple Example

You test whether a new app improves focus. (Oh look! A Squirrel!)

Result: p = 0.02

Classic interpretation:
“Statistically significant improvement in focus.”

More honest interpretation:
“If the app had no effect, results this extreme would occur about 2% of the time.”

C-value framing:
“The ‘no effect’ null hypothesis has low credibility given the data.”

Notice how the last version avoids overconfidence… and sounds like something a grown-up scientist would say.

In case this is still confusing… hold onto this:

A p-value asks how surprising the data are if nothing is happening.
A c-value asks how believable “nothing is happening” is after seeing the data.

Why I’m Talkin’ About This…

Because the future of research isn’t about memorizing thresholds… it’s about understanding evidence. And because somewhere, right now, a spreadsheet is about to produce a p-value that someone will misinterpret with absolute confidence.

Let’s make sure it’s not you.

(I’ve got more coming on this soon… so please stick around. I’m still learning this myself. There are some limitations on c-values, but they’re powerful, This is the beginning of a new tool in Statistics. Share. Like. Subscribe. Feed the Algo!)

From a workforce comparative advantage standpoint… people who know how to automate any human task with an Agentic pipeline are the only jobs that organizations should hire. Most (all?) other roles within an organization should be used to provide information and knowledge to the Agentic pipeline builders.

That’s an example of a #Tushman Ambidexterity “Explore” business unit.

Replacement Effects are in full swing.

This is not about ignoring human value in a business. This is about replacing the non-human value in the business. What’s uncomfortable is that there are lots of jobs that people do that can now be considered non-human value.

In a recent HBAP | STARTUPS discussion I said (referring to Rory Sutherland’s doorman fallacy), “It’s important that we Do NOT replace MEANING with FUNCTION! Dont replace a doorman with an automatic door! The doorman does a LOT more than open and close doors!” Human Value is NOT going away… it’s just going to shift DRASTICALLY away from the middle and lower layers of an organization. The C-suite will control a ‘swarm’ of AI Agents with very few human touch points.

This is called “Zero touch”, and it maximizes trust in the system.

If you’re a knowledge worker, you need to pay attention. The number of knowledge worker tasks that are outside the “Jagged Frontier” are rapidly dwindling.

This is why I’ve been saying that Entrepreneurship is the future of work!

Workforce replacement is not malpractice at all. This is Milton Friedman 100%; “maximize shareholder value”. This is about the pure pursuit of profit within the framework of capitalism. Businesses are ALWAYS incentivized to adopt automation and technology due to the cost savings and efficiency gains. This is nothing new. These are the droids you’re looking for!

It’s not… in any way shape or form… that “humans are bad for business”! (Humans are good for business!) It’s the “comparative advantage” that humans provide over Agentic AI is what we’re talking about… and where business leaders need to focus.

Consider Clayton Christensen’s “Disruptive Innovation”. New tech can render traditional business models obsolete. AI-driven business models and operating models are disrupting conventional human-centric approaches, pushing businesses to adapt or risk obsolescence. (Hint: Don’t be an Integrated Steel Company! And… something about hard drives.)

What’s CRITICAL to understand about “Disruptive Innovation” is that it occurs from the bottom up. Not everything is a disruptive innovation.

Here’s how the “SaaS is Dead” reality will come into play. SaaS customers can build their own solutions:

All business leaders need to understand a simple truth: if we are not exploring how Agentic AI can be deployed within our organizations right now, then our competitors are. We have access to technology that can perform better, faster, cheaper and safer than your current operating model offers, and ignoring this simple fact is the CEO’s malpractice.

To be clear, the primary driver behind the acceleration of Replacement Effects is profit motive.

Andrew Yang Newsletter

The End of the Office

I write this filled with sadness…

Read more

7 months ago · 102 likes · 4 comments · Andrew Yang

Now, for those of you who know me… I’ve been saying for 4+ years now that we’re likely entering a Post-Labor Economy. This means that people will no longer be able to exchange their time, knowledge, or labor for money… and we will need a new social contract.

The reality is that the most likely scenario is that we will be surfs in someone else’s kingdom. (Elon… if you’re reading this… I’m available to help grow your kingdom.)

I would add to Andrew Yang’s article.

It’s currently possible to replace entire teams of people with AI Agents. The primary bottleneck is ‘systems integration’… and of course compute and energy, but this applies at the macroeconomic level.

“Trust” in Agentic AI will come with longitudinal data, and C-Suite HITL oversight. (HITL = accountability. “Human in the Loop”… is about accountability. This answers the question, “who goes to prison when the AI hurts somebody?” It’s not about your taste, creativity, or subjective opinion. It’s a legal and liability construct.)

If you’re in the C-Suite of an incumbent company… PLEASE Go build a #Tushman “Exploration Business Unit” RIGHT NOW. The mission is to replace the entire Core of your business with AI. (If you dont know #Tushman, go read “Lead and Disrupt”, “Corporate Explorer”, and “Corporate Explorer Fieldbook”. There’s a quiz on Friday.)

If you’re NOT in the C-Suite, then you better get onto that “exploration team” RIGHT NOW, because that’s the only place to stay employed at your company.

The next thing you MUST do is focus on “Modularity”. (If you don’t know what modularity is, go read “Design Rules, Volume 1: The Power of Modularity” and “Design Rules, Volume 2: How Technology Shapes Organizations” by Carliss Baldwin. There’s a quiz on Monday.)

Modularity tells you where to run your AI experiments. Look for everything within the organization that has well-defined inputs and outputs. This will give you a place to start deploying AI Agents.

“Sovereignty Effects” (another chapter in my upcoming “AI Effects” book) where AI systems don’t merely replace workers… they become sovereign decision making ‘entities’ within organizations… are now in the early stages of deployment.

“Zero Touch” systems that maximize trust are here. This is the transfer of cognitive sovereignty from humans to machines. And that’s a civilizational shift.

Can you say, “Synthetic Cognitive Sovereignty Architecture”?

It captures a growing frustration that generative AI is flooding the world with low-quality output… generic images, shallow writing, synthetic noise. A firehose of mediocrity.

And the critics aren’t wrong.

We are drowning in average. (Statistically speaking, we’ve always been drowning in average. That’s the damn definition of average… it’s the mean of a distribution. But dont get me started on Statistics.)

And focusing on AI slop is like judging the future of electricity by early flickering lightbulbs. It mistakes a transitional phase for the destination. (Don’t do this!)

Because at the very same moment we are complaining about low-quality AI output, something far more important is happening underneath the surface.

We are quietly crossing the threshold from AI Slop to AI Gold.

And that shift maps perfectly onto a deeper pattern I’ve been tracking for years:

Augmentation Effects
Replacement Effects
Transformation Effects
Acceleration Effects

Augmentation: When AI Still Needed Us

The first wave of generative AI was clearly augmentative.

AI assisted humans. Designers used it for inspiration. Writers used it for drafts. Coders used it for snippets. Marketers used it for ideation.

Humans remained central. AI was a tool.

This is when most people first encountered “AI slop.”

Early models produced output that was fluent but generic, confident but shallow, impressive at a glance and disappointing up close. The human still had to clean it up.

AI was an intern… fast, tireless, and unreliable.

Replacement: When AI Stops Being an Assistant

We are now entering the replacement phase.

Here, AI no longer supports tasks. It substitutes for entire roles.

Not because the output is perfect, but because it is fast enough, cheap enough, and good enough.

This is why creative fields feel the shock first. Designers, writers, illustrators, marketers, musicians.

Not because their work is least valuable… but because it is digital, subjective, and lightly regulated.

Which is why creatives are not overreacting.

They are early indicators.

They are the canary in the coal mine.

When AI displaces designers, it is not the end of the story. It is the beginning of the story.

Because the same economic logic will later move upstream into law, finance, consulting, medicine, engineering, and management.

Creative work is simply the most exposed layer of the cognitive economy.

Transformation: When Work Itself Is Redesigned

This is the phase almost no one is ready for. (Except for the early adopters!)

Transformation is not about doing the same jobs faster.

It is about eliminating the old structure of work entirely. It means completely redesigning and rearchitecting your business practices and procedures from the ground up. (And, designing them so humans are not needed to do those tasks.)

If you haven’t already completed your “Digital Transformation” then you’re in trouble. It means that you’re not ready for this, and the risk is that your competitors are gonna eat you alive.

And this is where AI Gold truly emerges.

The Rise of “Zero Touch Agentic Systems”

We are now witnessing the rise of Zero Touch Agentic Systems.

These are not chatbots waiting for instructions.

They are autonomous operators.

Systems that can define objectives, break them into sub-goals, determine success metrics, assemble digital or physical resources, execute multi-step workflows, monitor outcomes, and self-correct in real time.

All without a Human-In-The-Loop. Except at the C-suite. Why? Because HITL = Accountability. Someone has to go to prison when the AI hurts people.

No approvals. No supervision. No micromanagement.

This is not automation.

This is delegation.

This is the moment AI stops being software and starts behaving like an organization. That’s happening RIGHT NOW! (And most folks are not ready.)

Why AI Slop and AI Gold Exist at the Same Time

This is the paradox confusing everyone.

We see low-quality AI content everywhere and assume intelligence is degrading.

In reality, we are watching a bifurcation.

At the bottom, AI Slop: cheap, mass-produced, low-signal output.

At the top, AI Gold: high-reliability, high-autonomy cognitive systems driving real-world outcomes.

The public sees memes and bad images.

Enterprises are quietly building autonomous intelligence pipelines.

Two completely different realities.

Side note: I discussed this in a previous post. This can be explained by the Disruptive Innovation Model from Clayton Christensen, where incumbents move upmarket in response to downmarket innovations.

From ‘Content Machines’ to ‘Cognitive Machines’

Old generative AI produced artifacts… documents, images, videos, code.

AI Gold systems produce decisions.

They do not just generate a marketing plan. They analyze markets, design strategy, allocate budgets, launch campaigns, run experiments, and optimize continuously.

The output is no longer content.

The output is executed strategy.

This is the transformation phase.

And it is far more disruptive than people realize.

The New Divide

The future will not be divided between humans and AI.

It will be divided between those who consume AI Slop and those who command AI Gold.

The valuable skill is no longer prompt writing. (I’m not really sure if “prompting” was ever a valuable skill beyond the general ability to have a conversation with another person?)

The valuable skill is designing missions, architecting feedback loops, governing autonomous systems, and auditing machine decision-making. This is often found (only?) in the C-suite.

We are not training prompt engineers. Prompt engineering is no longer a job. We’ve already automated that!

We are training AI conductors.

Why the Canary Matters

Creative professionals feel the panic first because they are closest to the blast radius.

They are not wrong about what they see.

They are simply early.

What is happening to designers today will happen to knowledge workers tomorrow.

But there is a deeper structural reason creatives are hit first, and it has nothing to do with talent. It has to do with measurement.

Creative work lives in the world of subjective metrics.

What makes a design “good”?
What makes writing “compelling”?
What makes an image “powerful”?

There is no single answer.

Unlike engineering, finance, or medicine, creative output is rarely judged by hard thresholds. There is no pass–fail test for originality. No universal benchmark for taste. No equation for beauty.

Success is contextual, cultural, emotional, and interpretive.

Which makes it fragile in the age of AI.

When output quality is subjective, the bar for replacement is not perfection.

It is “good enough.”

And “good enough” is a moving target that depends on the viewer, the budget, the deadline, and the use case.

This is why creative industries are uniquely exposed. Because when performance cannot be cleanly quantified, substitution becomes easy.

Subjective metrics collapse faster than objective ones under automation!

If a logo only needs to be “acceptable,” if copy only needs to be “usable,” if an image only needs to be “on brand,” then the system does not need to be great.

It only needs to be cheap, fast, and plausible.

This is the same logic that will eventually reach other white-collar fields.

The moment knowledge work becomes evaluated by soft metrics rather than hard thresholds, it becomes vulnerable to AI substitution.

Which is why creatives are not just emotional early adopters of fear.

They are structural early adopters of disruption.

They are the canary in the coal mine for all subjective-output professions.

AI Slop is not the end state.

It is the visible turbulence of an entire system reorganizing itself!

The Real Economic Shift

Zero Touch systems do not merely raise productivity.

They collapse the marginal cost of complex cognition.

When strategy, analysis, coordination, and execution become autonomous and scalable, entire organizational structures become optional.

This is not another software wave.

This is a new economic substrate.

And those who control AI Gold systems will capture disproportionate power… just as platform owners did in the last cycle.

AI Slop Is a Symptom, Not the Disease

We are not witnessing the decay of intelligence.

We are witnessing its industrialization!

Slop appears when a technology becomes cheap before it becomes refined.

Gold appears when refinement catches up.

It always does.

The Only Question That Matters

The real question is not:

Will AI replace us?

The real question is:

Will we be operating AI Slop… or commanding AI Gold?

Because the distance between those two worlds is about to become freakin’ enormous.

And that distance will define the next decade of work, power, creative control, and humanity.

If you wanted to build something in software, you needed someone who could remember syntax, understand (arcane?) frameworks, fight compilers, and debug cryptic errors at 2 a.m.

That bottleneck is gone.

AI-assisted programming has quietly… and now very loudly… eliminated the need for advanced, artisanal coding skills to build functional software and tools.

Let’s be precise about what changed…

Coding is no longer the scarce resource

Today, an individual with basic programming literacy, decent problem framing, and access to AI-assisted coding tools, can build things that once required full teams. And they can do it in record time!

CRUD apps. APIs. Dashboards. Scripts. Automations. Internal tools. MVPs. Glue code. Data pipelines. Frontends. Backends. Tests. Documentation.

The act of writing code… the mechanical labor… is being automated.

And that’s not a future prediction.
That’s just… Tuesday.

But software developers are NOT obsolete

(they’ve been unbundled!)

Here’s the critical distinction most conversations miss:

We don’t need fewer software developers.
We need fewer people whose value was primarily typing code. (for now.)

What we still desperately need are people who understand how systems work. How components interact, where failure modes live, how data flows, how state is managed, how to architect pipelines instead of files, how to secure systems that will be attacked, how to verify, test, and reason about correctness.

In other words…

The job isn’t “writing code.”
The job is designing systems that don’t break when reality shows up.

AI is incredible at generating code. It is terrible at owning consequences. Businesses and People cannot blame the AI for making mistakes (yet).

That’s why we still need Humans in the Loop (HITL)… for now.

HITL = Accountability.

HITL answers the question, “Who goes to prison when the AI hurts someone?”

Architecture > syntax

Always has been… now it’s unavoidable

AI-assisted coding can write a function, scaffold a service, refactor a module, generate a UI, translate between languages, and a whole bunch of things I’m forgetting.

What it cannot reliably do (yet) is to decide what should exist, reason across long-lived systems (I might be wrong on this one, btw), understand organizational constraints, anticipate scaling, compliance, or security tradeoffs, and perfectly design for humans instead of demos (Note: this ‘empathy’ loop is rapidly closing. I give it 4-6 months from today.).

This is where experienced developers still matter… arguably more than ever.

The center of gravity has moved up the stack!

From syntax to structure.
From implementation to orchestration.
From code to cognition.

AI-assisted programming = a low-end disruption (not a sustaining one)

This is the exact same “skill relocation” I’ve been discussing for years. This is “Replacement Effects”. And… for the Business Theory academics out there… this should remind you of Clayton Christensen’s “Disruptive Innovation Model”. This “upmarket” movement is not new. From Mini Mills to Hard Drives to Kodak to Blockbuster.

AI-assisted programming did NOT start by being “better” than senior engineers.
It started by being good enough for people who couldn’t previously build software at all. Early AI coding tools were inaccurate, verbose, brittle, untrusted, and… quite frankly… mocked by senior professionals. This is exactly how disruptive entrants look at the beginning.

Whenever you hear a senior developer saying that AI coding sucks… they are moving ‘upmarket’.

The Christensen Test: If a new technology is dismissed by experts as inferior, first adopted by non-experts or low-end users, dramatically cheaper and easier to use, and initially serves jobs that incumbents consider unattractive or unprofitable… yet improves rapidly over time… then it is not a sustaining innovation. It is a disruptive one! AI-assisted programming passes this test cleanly: it began as “toy” code for amateurs… line-by-line code autocomplete… was mocked for quality and reliability, required little formal training, and targeted work that professional developers had already over-engineered. As its performance accelerates, it moves upmarket, not by replacing expertise, but by redefining which expertise still matters.

And I’ve got two words for you: “Ralph Wiggum”!

And now… the uncomfortable part

Designers are in real trouble

This is the part no one wants to say out loud.

For a huge portion of design work…especially UI/UX production work… the bad news is that moat is collapsing. (There’s good news, though. Hang in there!)

It is now trivial for an AI-assisted developer to screengrab a UI, drop it into an AI tool, ask for a few changes, replicate it in working code.

Tools like Google Stitch make this even more explicit!
design → prompt → deployable interface with the code!

What used to require a designer, Figma, Adobe illustrator, handoff, iterations, misunderstandings, pixel debates… now takes minutes to complete.

And, NOT because design doesn’t matter… design matters! (Just look at Apple’s latest Liquid UI that is arguably the worst UI blunder in the company’s history.)

In the Software Universe… we’re focussing on programmers at the moment… much of what we called “Design” was really just UI assembly.

And assembly is exactly what AI is good at.

The Good News is that Designers, on average, have a Superpower. We can fill a blank canvas and imagine a future that doesn’t exist today. It’s about Entrepreneurship and finding solutions to problems. This requires a new mindset that you can learn!

The hard truth.

AI doesn’t eliminate taste or judgment. (Yet!)

It eliminates production friction. It does the actual ‘work’.

And the professions most exposed are the ones whose value was speed, execution, and surface-level differentiation.

Designers who focused on visual systems, interaction thinking, product strategy, human behavior, narrative and meaning… they will adapt.

Designers whose primary output was screens, layouts, component variants… are about to experience what developers felt when Stack Overflow showed up… only faster.

What actually matters now? Entrepreneurship!

The future of work is Entrepreneurship!

The new value stack looks something like this:

• Problem/Solution framing.

• System design and architecture design.

• Constraints awareness.

• Human judgment and Taste (for now.)

• Ethics (for now).

• Responsibility and Ownership

(I feel like I’m missing a few hundred bullet points here… but this is a good start.)

Yes, AI is a force multiplier. But it still needs someone who understands what should be multiplied. (For now.)

Final thought

We didn’t eliminate the need for software developers.

We only eliminated an excuse to confuse typing with thinking.

And the same reckoning is coming… hard… for design.

The question isn’t whether AI-assisted creation changes your role.

The question is:

Were you hired to produce artifacts… or to make decisions that actually matter?

Professions organized around artifact production are at risk.
Professions organized around decision-making under constraints are not.

AI-assisted programming isn’t necessarily a better version of software development. It’s a cheaper, simpler, more accessible version… and that’s exactly why it wins. The real disruption isn’t happening at the top of the market. It’s happening where work has already been over-engineered, over-specialized, and over-priced.

Final Final thought

Did you notice how I keep using the words “yet” and “for now” above?
That’s because those skills and decisions are not going to be as everlasting as folks realize.

Do you wanna know what’s coming next?

It’s called “Zero Touch”!

Ok. It’s not dead… I just think “Design Thinking” needs to change.

For the last 30 years, Design Thinking has been one of the most influential frameworks in creative, business, and innovation culture. It gave us permission to put people at the center of problem‑solving. It turned empathy into a design tool. It taught us that building, testing, and iterating mattered more than getting things “perfect” the first time.

But here’s the thing…

In the Age of AI… the Age of Intelligence… Anno Intelligentiae… design thinking is no longer enough.

Why?

Because design thinking assumes that the “user” is human.

It assumes the interface is something a person touches, sees, or experiences. It assumes the outcome is about improving human life directly. It assumes cognition lives entirely on one side of the screen.

That assumption no longer holds.

The Problem with Human-Centered Design

Human-centered design (HCD) is beautiful… and it’s also limited.

It over‑indexes on empathy for people while ignoring other forces in the system… ie machines, ecosystems, feedback loops, and unintended consequences.

It truly (arguably?) assumes a clean, linear process:

Empathize → Define → Ideate → Prototype → Test.
(side note: It may or may not have taken me a while to find the “→” character!)

Real systems don’t behave that way. They’re messy, recursive, and nonlinear.

HCD works well at the product or service level, but it struggles at the system and ecosystem level… especially when intelligence itself becomes distributed. (Today, intelligence is distributed!)

Most importantly:

HCD ignores the fact that AI is now a user.

AI as a User

Designers are already designing for AIs… even if they don’t call it that.

Search engine optimization (SEO) is designing for Googlebot, not for people.

Algorithmic trading is designing for other algorithms, not human traders.

Recommendation engines (Spotify, Netflix, TikTok) optimize for machine learning feedback loops as much as (or more than?) human satisfaction.

And now, Generative User Interfaces (GenUI) will mean that designers are designing for AI first, humans second… or sometimes… not at all. (For more info: Read this report from Flybridge from 2024! And check out the Live Session with Jeff Bussgang at HBAP|STARTUPS from last year!)

Designing for AI.
That’s the breakpoint.
That’s why design thinking, as we know it, is dead.

Intelligence-Centered Design (ICD)

So, what comes next?

Intelligence-Centered Design (ICD) is what’s next!

ICD is a framework for designing systems where the “users” include humans and artificial intelligences.

It doesn’t replace Human‑Centered Design. It subsumes it.

HCD becomes a subset of ICD. Humans remain essential. But now, designers also have to account for the machine’s “needs”. The data, interpretability, constraints, biases, APIs.

It’s not about empathy in the emotional sense… it’s about system empathy. It’s about understanding how different intelligences (human and machine) perceive, interpret, and act.

Karim Lakhani says it perfectly:

We’re just taking into consideration the needs of our… soon to be… AI overlords.

Comparing the Two

Human-Centered Design vs Intelligence-Centered Design

• Human‑Centered Design focuses on empathy for individual people. Intelligence‑Centered Design expands empathy to entire systems that include humans and artificial intelligences.

• Human‑Centered Design emphasizes interfaces built for human senses. Intelligence‑Centered Design emphasizes interfaces that support cognition, data exchange, and machine interpretation.

• In Human‑Centered Design, success is measured through experiences. With Intelligence‑Centered Design, success is measured through alignment across multiple agents within a system.

• Human‑Centered Design relies on iterative prototyping cycles;. Intelligence‑Centered Design requires continuous co‑evolution between humans and AI systems.

The ICD Loop

ICD isn’t a total rejection of design thinking… it’s a major expansion.

The loop still matters, but the audience changes:

1. Observe: Humans needs and AI needs (data formats, context, interpretability).

2. Define:Multi-agent goals (human desires + machine constraints).

3. Ideate: Co-creation between humans and AIs.

4. Prototype: Dual-use outputs: usable by humans, legible to machines.

5. Test: With people, with models, and within ecosystems.

Designers now have two questions to ask:
How does a human interpret this?
How does a machine interpret this?

Anticipating the Pushback

I’ve already heard the critiques.

(The “AI Slop!” screams of designers and copy writers howling at the moon… recognizing that their entire careers are being destroyed. Relax, folks… you’ve got superpowers. You’re just not using them. You need to change your mindset. It’s time to wake up or be replaced. It’s 1/2/2026. You’ve got 3-6 months max to figure this out.)

“AI isn’t a user… it doesn’t feel.”
True. (For now.) But it consumes, interprets, and acts. Ignoring that is bad design.

“This de-centers humans!”
Humans remain the value-setters. ICD expands the frame… it doesn’t erase people.

“Isn’t this just hype?”
No. It’s scope! Design Thinking was scoped for human-only contexts. ICD is scoped for multi-agent and intelligent systems.

“Won’t this just produce AI slop?”
Bad design produces slop. We’ve had bad design forever. ICD ensures values, context, and ethics stay embedded.

Why This Matters for Designers

If you’re a designer today… your users are no longer just people.
They’re people and AIs.

And sometimes… only AIs.

That means:

• You must design for AI as a customer.

• You must design with AI as a collaborator.

• You must accept that some systems will never have a human‑facing interface at all.

This is disruptive.
It’s uncomfortable.
And it’s already happening.

The Call to Action

Design thinking isn’t gone.
It’s just too small.
In the Age of AI, the future of design is Intelligence-Centered Design.
Designers of the future won’t just design for people.
They’ll design for intelligences.
We’re in Anno Intelligentiae!

And statistically speaking… all predictions are wrong... I’m curious about how wrong.

I think AGI is already here. This stuff is smarter than the majority of humans. It’s just not better than everyone… yet. AI’s impact will be felt across all sectors in 2026, transforming just about every value chain, business process, and industry. Early adopters have a massive competitive advantage.

Models:

Be on the lookout for new model architectures in 2026… likely coming out of China… that provide an e^x improvement in quality, speed, efficiency, and safety. Don’t be surprised if this model runs on a laptop or even a smartphone. (And for all of my investor buddies who are looking for the ‘something that could trigger the AI investment bubble to burst’… this is my favorite indicator. If this e^x algo happens, then we no longer need as many data center build-outs as we’re investing in. Boom.)

Transformers are not the future. The recently announced T5Gemma 2 represents an evolution of the encoder-decoder Transformer architecture. It’s designed to replace or serve as a more efficient alternative to decoder-only models (the architecture used by most current LLMs like GPT) in specific scenarios. There are lots of possible replacements for Transformers… Mamba, RWKV, Large Language Diffusion Models, and Titans are all super attractive.

“Continuous Learning” is the Artificial Super Intelligence (ASI) spark that changes everything. We’re sooo close to this I can taste it. This is where we see AI models that learn on the fly and update continuously instead of having to be retrained in massive batches. Today’s large models are mostly trained once on a static dataset. (This isnt completely accurate… I’m referring to the ‘pre-training’.) In the very near future, we’re going to see more self-supervised learning and real-time model adaptations. This is basically how humans and animals learn incrementally, where systems can ingest new data and new experiences in real-time to improve their performance. Today we have LLMs that have amnesia… click ‘new conversation’, and they catastrophically forget. We need models that can continuously learn and grow. Back-propagation will retire.

“World Models” will continue to gain traction. If you don’t know what World Models are, think of them as a “mental model” of the world we live in. We need to teach AI the ‘bility to understand physics and how the universe physically works. World models, like Genie 3 from Google and Fei-Fei Li’s World Labs are examples. For more info, read this blog post, this website, and this definition from Nvidia.

Mixture-of-Experts will start to scale. Mixture-of-Experts (MoE) architectures make AL models more efficient. MoE routes queries through many small sub-models (“experts”) rather than one giant model. I like to think of this as how the human brain has multiple cortices. We have these ensembles of specialized expert models all working together.

Enhanced Reasoning via Chain-of-Thought and Neuro-Symbolic Systems. Today’s AI often struggles with complex reasoning or multi-step logic. In 2026 new architectures will solve this limitation. One trend is training models with chain-of-thought reasoning, where the AI is encouraged to generate a “thought chain” of intermediate reasoning steps before giving a final answer. (My mom always said, “Todd… think before you speak”.) CoT, along with new architectures that internally self-reflect or run subroutines, will make AI’s reasoning more transparent and reliable. Next year, we’re gonna see models with “Common sense”!

Human Alignment will start to show promise. Some folks are terrified that AI is gonna wipe out humans, and their “P(doom)”… the probability that AI will destroy all humanity… is 100%. I feel bad for these folks because they are not looking at all of the good things that AI can offer. Sorry, Eliezer, “If anyone builds it, Everyone Dies” is not the future. The more likely scenario is that everyone lives. (However… we may be serf’s in Elon’s Kingdom, and we’ll all live happily.)

And beyond 2026, I predict that humans will continue to incorporate technology into our bodies. One day, homo sapiens will undergo a speciation event… the evolutionary process by which populations of the same species diverge genetically and morphologically until they become distinct species. A new sub-species, homo Technoensis will arise, and shortly afterwards… the ‘alignment’ problem will not be a problem anymore.

Chips! In 2026 China is about to pull ahead. READ THIS. I think we will see the first high-end AI chips coming out of China in the first half of 2026, and production ramping up by the end of 2026, early 2027. We’re also going to see new chip architectures such as Neuromorphic and photonic chips. Companies like Extropic, Mythic, Analog Bits, Lightmatter, Celestial (just acquired!)… are gonna be… popular.

(Here’s a crazy prediction: I think 2026 will bring a new stealth chip architecture to market thats gonna prove an e^x improvement in either speed, energy, or cost over Nvidia’s GPU-based chips. Nvidia will buy this company for between $250-$350 Billion. China will steal the architecture and flood the market with inexpensive versions of this design that NVIDIA and TSMC can’t compete against.)

AI Disruption Across the Creative Industries

I’ve been saying this for 3+ years now. If you’re a ‘creative’ or a ‘designer’… you need to take my SVA course.

Film and Animation: Studios will adopt AI tools to generate visual effects, de-age actors, or animate characters, and even create rough cuts from scripts. Independent filmmakers will create full length features with AI. Scriptwriters are already co-writing with language models, using AI to draft scenes and suggest plot ideas. In post-production, AI editing tools will automate tedious tasks (like color grading or continuity checks), letting human editors focus on creative storytelling.

Netflix creates a full GenAI movie or series. Okay, this prediction is a little bit of a stretch for 2026, but I’ve got my fingers crossed. Look… the reality is that Netflix gets me. They know my profile… they recommend the best stuff. I think they could build out a GenAI Agentic Content Pipeline that can create a movie or series that is perfectly designed just for me. See Showrunner for more inspiration.

Deepfakes are going to continue to be a problem. Until legislation either criminalizes it and/or makes it possible to sue for civil damages, the world will continue to stress and suffer over this. And we still need folks to be able to express their First Amendment free speech rights at the same time… soooo… idk. Something about consent being important???

“Fast Content” advertising becomes mainstream. Advertisers will finally figure out… what I teach in my course… that the RIGHT (only?) way to make GenAI advertisements is to make millions of ads. DO NOT to make a single spot that gets angry “AI Slop” backlash. McDonald’s. Coke. If you’re in the advertising industry, and you’re not ALREADY using AI, then you’re gonna be in a lot of trouble in 2026.

Are you an Advertising creative who’s client is asking for 1,000 banner ad variants targeted to different demographics? An AI can pump those out, each with tailored imagery and copy, and following brand guidelines. I’m not sure why a brand needs an advertising agency beyond 2026? Generative AI will handle a huge portion of content creation for ads: copywriting, design, and video ad production can be fully automated. In fact, an IDC forecast expects roughly 30%-50% of marketing tasks and costs can be automated by generative AI by 2027.

If there’s any hope for ad agencies, it’s that they all embrace the fact that AI can output far more content for their clients at lower cost, and potentially grabbing market share from slower traditional ‘human-only’ agencies. However, this ubiquity of AI-generated content will also spark an increased premium on truly novel, human-crafted design in certain niches (much like how handmade artisan goods gained prestige in the face of mass production). No amount of industry consolidation can fix this.

Ad Agencies become Ambidextrous Organizations. Kodak failed because the C-Suite couldn’t embrace digital photography. If the CEOs of Ad Agencies haven’t read “Lead and Disrupt”, “Corporate Explorer”, and “Corporate Explorer Fieldbook” #Tushman… then they’re in TROUBLE. They need to hire me to help them build a separate “business unit”… immediately… whose mission is to replace the “core” of the business with Agentic AI pipelines that automate advertising. (And if you’re a “creative” who’s not already in that top secret skunkworks exploratory business unit… then you’re in trouble too. You need to do EVERYTHING you possibly can to get onto that team ASAP, because that’s the only future of your organization.)

Copyright becomes less important to brands. I still hear that brands believe that having a copyright on their Ads is critical. Let it go. Copyright is not important anymore. In 2026, brands will finally figure this out. And brands still have their TRADEMARKS! (Trademarks are important.)

SCOTUS and/or Congress may need to weigh in on Copyright next year, so we’ll hopefully get some resolutions on the pending cases. The Ninth Circuit courts may clarify a few things. And for all those folks out there who are howling at the moon that AI companies “stole our art” without permission… yup. That happened. The bottom line is that within the confines of Copyright law, training an AI model is “Quintessentially transformative”.

“In short, the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative. Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different. If this training process reasonably required making copies within the LLM or otherwise, those copies were engaged in a transformative use.”

ORDER ON 122 FAIR USE. Signed by Judge Alsup. (Filed on 6/23/2025)

And remember… China doesnt care about our IP laws! The United States is not likely to hand over the future of the entire AI Industry in this country over to China. Training AI models is “fair use”.

Fashion: The fashion industry will undergo a seismic AI-driven transformation in how designs are created and marketed. Generative AI models can now produce clothing designs, textile patterns, and even full fashion photoshoots virtually, which threatens to upend traditional workflows. Major brands are already experimenting with digital models (“digital twins”) in place of human models… and this trend will normalize virtual avatars that showcase most products online, with human models reserved for high-profile runway events. I predict that 2026 will see an imminent disruption of creative fashion jobs… many mid-tier fashion photographers, layout artists, even junior designers will find their roles changing or shrinking. And if you think for a second that the Fashion Industry will not go all-in on AI, you’re forgetting that Photoshopped fashion photography IS the default behavior. Fashion designers and art directors will transform into “visual engineers,” creating unique brand-specific AI models and curating the outputs.

Graphic Designers are basically screwed in 2026. 2026 will see the first commercially available “AI Creative Director”. Today, AI is an indispensable co-creator. Design studios and Agencies will continue to use these tools to generate illustrations, concept art, logos, and layouts at a moment’s notice. This means graphic designers can produce hundreds of variations of a concept in minutes, then pick and refine the best. Now, wrap the entire cognitive, creative, and (highly) subjective process that a Creative Director does into an Agentic AI pipeline. In 2026 brands will be able to scale design without it being called “AI Slop”.

3D Modelers in the entertainment industry are already obsolete. If you are a 3D Modeler and you make models for gaming, visual effects, movies, or film… then many of you are about to lose your job. There are dozens of tools that automate this task already. And Microsoft released Trellis.2 last week. However, if you are a CAD designer, and you do “design for manufacturing”, then there’s some good news. Your skills are still gonna be in demand in 2026. The multiple cognitive constraints that these roles face are still ever so slightly outside of the Jagged Frontier of AI.

Character Animators, Motion Graphics, and VFX are completely transforming. If you’re currently in school and are studying to be a Maya animator, or a Cinema 4D animator, or an After Effects animator… those jobs are at market saturation and are basically over sold. The vast majority of those tasks are already automated by early adopters using a well articulated prompt and an Agentic AI pipeline. 2026 is going to see massive changes in the TOOLS that those jobs and roles use. It’s not too late to change course… you’ve got until March 2026 to become an expert.

Gaming: The video game industry in 2026 will be profoundly shaped by AI, altering how games are developed and experienced. Game development studios, despite the risk of ‘slop’, are adopting AI tools en masse… yet still, the Indy Game Awards (foolishly?) rescinded their winner because it used GenAI. I dont think the Indy Game Awards is going to be around for very long if they actually think banning the use of GenAI is going to keep Indy Game Developers in business. Looking at history, on average, Luddites tend to lose everything… but the term ‘luddite’.

The Music industry will become saturated with AI music as Musicians compete for airplay. Just go play with Suno for a few hours, and you’ll understand.

Side Note: There’s a difference between Fine Artists and Designers… and Creatives

As the line blurs between human and AI-made visuals, 2026 will continue to see debates around authorship and originality in “art”. I’m constantly hearing screams about “AI Art”. I wanna clarify my position on this.

There’s an important difference between a Fine Artist and a Designer (or a “Creative”), and how AI affects these two types of people.

Fine Art is about self-expression.

Design is about a business value proposition.

So, if you make “art”, I consider you a Fine Artist. And if you get paid for the “creative work” that you do, regardless of industry, I consider you a Designer… of sorts.

Fine Art is not primarily about money or commerce. Fine Art is about self-expression. And the moment that any Fine Artist decides that they wanna make money by selling their artwork, instantaneously… in that very moment… they must focus on their customer and business model. Fine Artists must understand that they have a customer who they hope will buy their artwork. So the question artists must ask is, What problem does their artwork solve for their customers? (And… we don’t teach how to answer this this in art school!) The reality is that AI will NEVER replace human artists! Fine Artists can choose to use AI, or choose NOT use AI… it does NOT freakin’ matter.

Designers, on the other hand, are part of an industry value chain. They’re part of a business model… a cog in an engine. Designers are hired to create beautiful-looking things that solve a specific business problem with a certain number of constraints.

Remember, the primary goal of every business is to generate profit and value for it’s owners and shareholders. Therefore, AI-assisted or AI-automated design and creative workflows will become the norm. The truth is that most organizations cannot afford to risk their futures if their competitors can have an “AI advantage” over them.

And, for now, human Designers will continue to direct the “taste” and final touches that set apart the best works. (These roles will trend to the brand, not the ad agency.)

Below is a slide from my courses, lectures, and keynotes:

This figure presents a conceptual continuum differentiating Fine Artists and Designers by primary “creative orientation”, from self-expression to commercial or industry-driven value creation. The horizontal axis represents creative orientation, while the vertical axis indicates the relative share of “creatives” aligned with each orientation. Fine artists (blue) are concentrated toward self-expression, emphasizing personal vision and intrinsic creative goals. Designers (orange) increasingly occupy the commercial end of the spectrum, where creative work is shaped by external constraints such as client requirements, market demand, and functional outcomes. The dashed vertical line denotes an inflection point at which the dominant Value Proposition shifts from expressive to commercial logic. The model is illustrative rather than empirical, and highlights differences in value orientation rather than fixed or mutually exclusive categories.

And I want to clarify an important concept about the “human-in-the-loop”… the “HITL”… that’s the person with the “subjective taste” who determines what is or is not good enough. Human-in-the-loop is only about accountability. It determines who goes to prison or gets fired when the AI screws up.

So, if you’re worried about the Future of Work and jobs… being the “human-in-the-loop” in 2026 is still a viable role.

China will surpass the US in AI

In 2026, China will remain tied with the United States in certain AI areas, and pull ahead in other AI areas. The key words are ‘remain tied’. Yeah… many folks dont realize that the US is not exactly ‘ahead’ of China anymore. They’re tied with the US.

Research quantity and algorithm innovation coming out of China are super important signals to watch. China is currently producing more new research than the US. And… pay attention to the RATE of China’s research publications! If, all of a sudden, the publication rate slows down or completely stops, that likely means the Chinese government is closely involved, and they don’t want the US to get a hold of their innovations. It means that China is likely pulling much farther ahead, and we’re in for a rude awakening.

China’s total R&D spending (all science and tech) is likely to equal or exceed U.S. levels in 2026. Here comes the crossover! Crossovers are important.

Education and AI.

More Universities will finally lean into AI in 2026. Students will be allowed to use AI for their assignments. And the smart Universities will require their students to use AI.

I’ve said this hard-hitting and controversial statement many times in the past… Professors who ban the use of AI should be fired because they’re not preparing their students for today… forget about the world when students graduate in 1, 2, 3, or 4 years. They’re no longer qualified to teach in the classroom, and they should be removed from their roles immediately. It’s AI or nothing!

<cough>

And to balance out the partial insanity of that extreme perspective… of course I can come up with TONs of reasons why this makes no real sense. We need those Professors! Please don’t fire them! Lets get them to embrace “Anno Intelligentiae”.

Side Note: AI for K-12 is a different story. I have trouble navigating how to use AI in this age group. We’re going from “T-S-I-D-D-A-H-N” to the SATs… and I still think Khan Academy’s Khanmigo has this dialed in with a private tutor for each student.

Privacy and AI:

I think 2026 is going to see some serious privacy battles between The People, governments, platforms, and AI companies.

Don’t blindly enable these commercial ‘AI assistants’ to have access to your data. In 2026 it will become super easy to build and host these features yourself.

Lastly… if you wanna stay on top of the latest AI news… please subscribe to my weekly Office Hours on YouTube.

Subscribe now

But this shift? The one we’re living through with AI? It actually did have a moment.

If this were a 3 act movie… the prolog or “exposition” of this story began in June 2017 with that deceptively chill research paper, “Attention Is All You Need.”

And the ‘inciting incident’… that started on November 30, 2022, when ChatGPT showed up and quickly rearranged the universe.

And now here we are, living in what I like to call Anno Intelligentiae…

“The Intelligence Age.”

It’s a new calendar.
A new mental model.
A new way of seeing ourselves in the world.

B.I. = Before Intelligentiae
That’s everything before ChatGPT’s release on November 30, 2022.

A.I. = Anno Intelligentiae
Everything after ChatGPT’s release on November 30, 2022

By that math:
2017 is “Year 5 BI.”
2022 is “Year 1 AI.”
2025… AND welcome to “Year 3 AI.”

It’s tongue-in-cheek, sure… but it’s also a signal. A way of saying out loud, “Hey, something big is happening… and we should probably pay attention.”

So let’s talk about what this actually means.

The “Oh Sh*t” Moment We Missed

Surprise! Deep Neural Networks and Transformer Models work!

Most eras don’t start with fireworks. But the Generative AI moment felt… different. (An opportunity to some, and an existential threat to others.)

One minute, intelligence was something only humans had.
The next minute, we’re staring at systems that can reason, summarize, imagine, create… and they’re doing it at speeds and scales that make us look like we’re thinking through molasses.

We didn’t build a better tool.
We built a new kind of mind.

Not human… not biological… not conscious… but undeniably intelligent.
And in that moment… our long-standing monopoly on cognition quietly ended.

That’s the real shift of Anno Intelligentiae.
Not the tech… but the hierarchy.

The Collapse of the Cognitive Ladder

In the Industrial Revolution, machines took over our muscles.
And we said… “Fine. We’ll move up the stack. We’ll do the brain work.”

Now the machines are taking the brain work.
And there’s nowhere obvious left to climb.

This revolution isn’t coming for factory jobs… it’s coming for the manager, the analyst, the writer, the designer, the coder, the creative director, the strategist, the researcher.

It’s coming for the people who thought they were safe.

And here’s the twist:

AI doesn’t just automate work…
It automates competence. It democratizes expertise!

Suddenly, the distance between beginner and pro collapses.
The average person can produce work that feels… surprisingly good.

Mediocrity becomes free.
Excellence becomes rare. (rarer??)
And the creative middle… the place most of us live… gets squeezed… hard.

This isn’t just economics… it’s existential.

The Authenticity Avalanche

We’re also entering an era where “truth” is getting slippery.

We used to trust our eyes.
If we saw a photo, a video, a voice recording, we believed it.

Now?
Not so much.

Deepfakes… AI-generated images… perfectly synthesized voices… hallucinated facts delivered with total confidence.

We’re drowning in content that sounds right… looks right… but isn’t necessarily real.

In the Intelligence Age, seeing is just the beginning of the verification process.
And honestly? Most people don’t have the time or energy to verify anything.

That’s a problem.
A big one.

(And the only solution is education. We need more people with intelligence to use AI to be smarter.)

The Promethean Upside

But this era isn’t just dark dislocation and identity confusion.
There’s real fire here. Real possibility.

We get tools that think with us… and sometimes ahead of us.

We get scientific acceleration that would’ve looked like magic a decade ago. (Or was that last week?). We get instant creativity, instant experimentation, instant translation between idea and outcome.

We get personalized everything… learning, care, coaching, creating.

In other words… we gain partners.
Strange partners… yes (maybe?)… but incredibly powerful ones. Especially if you learn to ‘prompt’ and collaborate with them nicely.

If we use this right, the Intelligence Age will be the most transformative period in human history.

If we use it wrong…
Well. Let’s not use it wrong.

(I mean… ya’know… I do love a good Dumpster Fire every once in a while. The heat warms my hands a cold evening.)

The Big Question: Who’s Steering This Thing?

The exciting (scary?) part isn’t AI becoming evil.
It’s AI becoming competent… efficient… effective… and completely indifferent to us.

These systems are being trained right now.
They’re learning from the examples we give them.
They’re absorbing the values we encode, intentionally or accidentally.
They’re getting better exponentially.
They’re global.
They’re biased.
They’re wicked smaht.

So the real work of Anno Intelligentiae isn’t building better models.
It’s building better guardrails… better incentives… better human judgment to pair with all this non-human “AGI is already here” intelligence.

This isn’t a tech problem.
It’s a leadership problem.
A cultural problem.
A wisdom problem.

The Invitation of the Intelligence Age

To name an era is to shape how we understand it.

Anno Intelligentiae is our way of saying…
“This is different. This matters. Pay attention.”

What does it mean to be human when intelligence is everywhere?

What does creativity look like when machines can create too?

How do we maintain truth in an age of infinite fabrication?

And how do we ensure that as machines get smarter… humans don’t get lazier?
(And do’t get me wrong… I don’t wanna work. I wanna be a bum! I wanna hangout and have fun and learn and make art. So, it’s not that I’m lazy… it’s that I’m… <cough> efficient.)

This era may have snuck up on most of us… and it will define us.
We get to shape it… and push it forward… wrestle with it… and ultimately partner with it.

Welcome to the Intelligence Age…
We choose its arrival…
And we get to choose what we do with it.

Other stuff I’m thinking about:

Im working on a new “+@AI” notation for “Transparent Attribution of Artificial Intelligence in Creative and Scholarly Work”. It eliminates the ‘plagiarism’ risk by being transparent. For example, this article was:
+@AI [summarize, edit, ideate: L2 : S] ChatGPT 5.1
(I’ll post more about this soon.)

Measuring Innovation (formulas are coming soon!):

Design Thinking is dead! Welcome to “Intelligence-Centered Design”.
Humans are no longer the focus of what we’re designing for! We need to design for AI as a user, and designers and creatives are not prepared for this. I’m giving a presentation on this next week.

https://deepmind.google/models/gemini/pro/

And… check out ANTIGRAVITY!!!!

https://antigravity.google/

Google has gone from a sleeping giant to the world leader in AI. This requires incredible leadership and talent. Congratulations to everyone at Google!

If you wanna check it out…

https://openrouter.ai/openrouter/horizon-beta

I just gave it a task to build a tool for me. It one-shotted the fully functioning tool from my detailed prompt. And when I compared the tool to the one I built last year… the new tool performs MUCH better than my original.

Fully Automated Development Pipelines

Moving forward, EVERYTHING we build must be grounded in a PRD and highly detailed tech spec. This gets fed into whatever Agentic AI coding pipeline you are using, and you then let it run. You still need a Human In The Loop (HITL) to test, and make sure that it’s secure, and fix the errors in the code. (Don’t worry, the security testing AI Agents are going to help you out too).

This means that your CI/CD processes should be mostly/fully automated by your Agentic AI dev pipeline. The goal is to build an always-on improving feedback loop that is grounded from the PRD/Tech Spec.

Here’s the key takeaway:

Imagine an Agentic AI pipeline that focusses on the high-level functionality. This drives an Agile Workflow that drives the Agentic Coding pipeline, and all of this runs inside a CI/CD feedback loop.

Developer’s Like AI Coding!

And… check out Stack Overflow’s latest 2025 Developer Survey. It’s the definitive report on the state of software development, and it’s got some insight on developer’s AI coding satisfaction.

But… it’s not perfect:

The biggest single frustration, cited by 66% of developers, is dealing with "AI solutions that are almost right, but not quite," which often leads to the second-biggest frustration: "Debugging AI-generated code is more time-consuming" (45%).

This sounds very similar to the senior developer’s 50+ year long frustration with dealing with junior developer’s code.