I’VE BEEN WATCHING CARTOONS. Forget The Smurfs. Thundercats? No. We’re talking: AI-driven, killer drones on a post-apo, frozen exo-planet. In Murder Drones (Glitch Productions, 2021), flying robots use razor-sharp talons to slaughter their own kind. The antagonist in the pilot (89M views, ~27mins) is a ruthless, airborne terminator. She’s also hawt. 🔥

Here’s the plot of the pilot: humanity has destroyed itself by accidentally blowing up the planet’s core. All biological life is obliterated. Only runaway AI remains: the Worker Drones. An evil corporation has since created drones to kill them. These are the murderers: the Disassembly Drones.

Dubbed into a slew of languages, viewership is international. The statistical age-range isn’t public, though Glitch created the show for a vague combo of teens and young adults. Liam Vickers was the American showrunner of this Australian-financed production. Judging by the North American English spoken in the original, its target consumers are those with American / Canadian mores. Without analytics, based on the visual-cultural lingo, my anecdotal guess is the majority of viewers are within the Gen ZcuspY age range of ~13-30.

Described as sci-fi / horror comedy, Murder Drones is dire, hilarious and—heavily anti-corporate with a clear Marxist lean. The cultural coding is as in your face as a Rosa Luxemburg tee-shirt. The show’s hero is a teen Worker Drone wearing an alt-cult beanie. She has purple hair and a strong opinion about corporate corruption. A broody loner, this girl-bot wants to save what’s left of the world with a rifle-sized railgun she designed and built herself. Through the lens of (pop)cultural analysis, she bears a striking metaphorical resemblance to a gun-toting, anti-corporate, radical leftist.

Her name isn’t Luigi. Even better, it’s Uzi.

Mentally unstable, Uzi responds to a homework assignment—a word problem about buying watermelons—with a class presentation on the need to fight the Disassembly Drones who’ve been attacking their bunker. Her visual aid, however, is a bespoke, magnetically-amplified, photon-converting long-gun.

Uzi meets an awkward, disaffected Disassembly Drone, whose first and last name is N. She tells him his parent corporation will certainly kill him and his kind after they’re done wiping out the Workers. He shouldn’t be on their side; he should be on hers. N is easily persuaded, and swiftly crosses the ideological divide. He goes from corporate killer to being Uzi’s sidekick. In N’s own words, he’s “an angsty, rebellious Disassembly Drone now.”

*

IS MURDER DRONES CALLING FOR AN ARMED REVOLUTIONARY LEFT? Nah. It sets a new stage for the drama of the human condition inside the robot-hearts of down-trodden, blue-collar rebels right alongside the split, Severance-like personae of white-collar corporate cogs. It centers the rage of contemporary anti-capitalists while giving voice to the plight of those who toil their lives away pledging allegiance to a company mission.

J, the antagonist, is the tortured office soul. Born a poor, bot child, the robocidal J was once employed as a maid, a Worker Drone. She was one of the people, until the system corrupted her. Aww.

The protagonist, Uzi is representative of a new revolutionary savior. She’s a rule-breaking, truth-seeing, countercultural figure, priming youth to defend the working class! But is Glitch Productions laying the groundwork for cultivating our real-world soil for Marxist minds to grow from and overthrow the means of production?

Not remotely. Murder Drones isn’t Marxist. It’s not political at all. Glitch Productions is at ease with capitalizing on the impressionability of young minds and underdeveloped adults whose prefrontal cortexes are wee, but not to seed them with socialist dreams. They aren’t using the language of the Labor movement with an activist’s aim, nor are they engaging in the opposite manipulation. Their audience doesn’t clock the nuances of ideology.

Glitch never intended to promote revolutionary thought. The company has been locking in streaming deals, selling merch, and creating scandals with poor communication skills and anonymous complaints from employees about working long hours (shocking in the media industry).

If Glitch intended to con young, would-be revolutionaries by using Marxist codes to lure little lefties into piling up bars of gleaming YouTube watch-time, that would be cleverer. It would also be conventional, consumer-cultural practice.

*

THOMAS FRANK IS YAWNING. Way back in 1994, when cellphones had antennae, his Baffler essay, “Why Johnny Can’t Dissent,” defined how the verbal and visual countercultural lingo of the late fifties and sixties had been crushed by the machine of the culture industry.

Citing the (nonsensical) 1994 Nike ad featuring William S. Burroughs selling shoes, he writes:

The most startling revelation to emerge from the Burroughs/Nike partnership is … the complete lack of dissonance between the two sides…. His ravings are no longer appreciably different from the official folklore of American capitalism…. As expertly as Burroughs once bayoneted American proprieties, . . . he is today a respected ideologue of the Information Age . . . . His inspirational writings are boardroom favorites, his dark nihilistic burpings the happy homilies of the new corporate faith.

Glitch is neither pushing a subversive message nor defrauding would-be subversives of their cash with what, in 2004, Joseph Heath and Andrew Potter called ‘the rebel sell.’ Instead, Murder Drones takes the lack of dissonance between the countercultural and the corporate further.

For a kid today, or an underdeveloped young-ish adult, Murder Drones isn’t about real-world economic injustice. The dissenting Worker Drone, and J, the corporate killer, are interchangeable. The show isn’t even co-opting a storied, anti-corporate figure or a dissident philosophy from some authentic dystopian past.

The Gen ZcuspY viewers don’t think Uzi is cool because she looks something like an anarcho-leftist wielding a firearm. The creators didn’t co-opt the culture of the frustrated, powerless kids who patrolled Seattle’s CHAZ in 2020 strapped with ARs.

This audience isn’t ideological.

Watchers of Murder Drones remain largely unacquainted with history, apart from what they’ve gleaned from Call of Duty, which transforms the beaches at Normandy into a first-person shooter. It’s safe to say most have never been exposed to the history of capitalism, nor that of socialism. Both are bloody and dramatic, offering oodles of media appeal. However, video games featuring the Atlantic slave trade or any of the socialist famines are too tasteless even for Activision.

The cartoony Marxism of Murder Drones is far from being reflective of dewy-eyed nostalgia for truer revolutionary moments of an American past, like MAGA for the Revolutionary War and the Founding, or like moderates and leftists for Suffrage or Civil Rights. This next gen is different from the seemingly well-read, socially aware academics at Bates College, who were recently lambasted by their former colleague Tyler Austin Harper in The Atlantic for “cosplaying as radicals… engaged in an Alice in Wonderland–style reenactment of civil-rights-era struggles.” Newp. This doesn’t have the blustery left-right stuff.

If the media leaders of the next generation are paying attention at all to the political narrative of current affairs, they sense its senseless chaos. They see a lack of self-awareness in their elders’ toothless ‘political’ IG posts. They see BLM placards stuck into freshly mown suburban lawns in gentrified neighborhoods as empty. Unless you’re actually housing a refugee from the Russo-Ukrainian War inside your lovely, well-secured home, they see the Ukrainian flag purchased from Amazon that’s waving from your porch as politically inert.

In Murder Drones, Uzi’s defiant, goth-girl palette of purple isn’t about the oppressed raging against the capitalist machine. It’s fashion. Pure aesthetics. Young people saw most (not all) of their disgruntled, rebel elders costuming themselves in hip, revolutionary outfits during the past decade.

They saw and learned that ‘subversive’ is a style of dress, a manner of speech, a voguish pose. For the creators and viewers of Murder Drones, Marxist cultural coding today signals faux-ideology. Learning by their elders’ example, they understand that looking like a subversive is kewl.

*

GLITCH PRODUCTIONS ARE NEW BARONS OF GLOBAL DIGITAL-MEDIA. It’s straightforward capitalism. It’s unironic. It eschews a longing for a MAGA past while declining to canvas for a socialist future. Kevin and Luke Lerdwichagul, Glitch’s founding brothers, hail from Australia. Bringing on the American writer / creator Liam Vickers, they homed in on the ravenous North American consumers of the planet. Glitch is a blank-flag-bearing representative of a new generation of relatively apolitical indie-producers who just need to earn a living in this f#!ked economy.

Born inside the glass of our totalizing digital media environment, they’re not revolting against the means of production. The means of production for them is easy, low-cost, Big Tech-driven digital production—and cheap distribution. When it comes to publicity and marketing, smart phones are practically appendages. Social channel content is their first language. Their bodies and identities were always branded.

There are pockets of dissenters, like the new Luddites, smashing iPhones and bashing in the faces of digital tablets. Brian Merchant has been covering this this brilliantly. But Glitch is one of the new media industry models rising to prominence, held in high regard by millions of international youth, that’s taking a different, apolitical tack. They’re writing directly to Gen ZcuspY media-consumers in their native tongue.

*

A LAST WORD, ALONGSIDE J’S FINAL SPEECH: AT THE END OF THE PILOT, J finishes a wild massacre inside the Workers’ bunker with the flourish of a villainous monologue. While she revels in her close-up, playing her BestMonologueEver.mp3, rife with corporate jargon, Uzi gets the upper hand. Pointing her railgun at the bested vixen, Uzi dares her: “One more buzz word and I’ll do it.” J tries to hold back, but says, “Equity Partnersh—.” And Uzi toasts the toaster.

While the story is that of Uzi, the righteous, revolutionary worker besting J, the corporate villain, it’s not Marxist or anti-corporate, nor is it sending ill-advised messages promoting firearms or violence.

Good, bad. Left, right. Rule-follower, outlaw. Coded as a Marxist struggle, Uzi, in fact, might shop for high-quality thrift finds in our world, while J would have a store card at Brooks Brothers. Your personal brand might match the aesthetic of the trendy insurgent with Uzi’s edgy choker, or with that of the corporate, predatory J, with her hot-pants suit and her powerless past.

They’re simply the characters needed for the balanced, compelling plot of the ‘hero’s journey,’ the story structure Uzi cites during her opening class presentation. Politically, they’re meaninglessly opposed in static, storytelling tension. Factually, they’re personified algorithms designed to lengthen human watch-time. {∞}

The artificial intelligence that’s supposed to be on the brink of becoming the singularity was trained with Atari’s Pong. Are we really headed for superintelligence with this iteration of AI?

THE WORLD IS BEING DISRUPTED BY SOFTWARE TRAINED ON ATARI, but does that software augur the rise of superhuman intelligence? Or is it an accelerator of predatory capitalism? There are good forms of free-market capitalism. This isn’t one of them. This iteration of AI appears to be using the simplicity of a traditional video-game objective — earn as many points as possible, as quickly as possible — to profit like a crime boss in Grand Theft Auto.

Human-led, AI-driven market predation and the expected job-loss derived from AI replacement are serious threats to the world’s interconnected economy. But is humanity a hair’s breadth away from a superintelligence? Or is that internet puffery accelerated by AI’s capacity for mathematically scaling the distribution of extremes?

The development of this general iteration of AI, the current models everyone’s freaking out about, went from being trained with checkers and Backgammon to playing Atari games, like Pong and Space Invaders. Learning from the training data of thousands of games, AlphaGo was able to play Go and win against an internationally titled, South Korean Go player, Lee Sodol. The next level of its evolution was obvious: it would need to master Grand Theft Auto V, a video game in which you’re a criminal who hires unhinged hackers, drivers, and gunmen to pull off robberies.

(If it makes you feel any better, GTA V was mainly used as training data for self-driving cars; however, this probably makes you feel a little worse.)

The game AI is working to master today is the game of us. Big Tech presented us with a choice to play, sure, in the form of a Terms of Agreement pop-up. You remember the deal you e-signed. It read, “If you’d like to continue to use your phone, laptop, and/or tablet, then CLICK TO ACCEPT. If not, do feel free to decline.”

You had to play the game, or—they’d politely deactivate your phone.

AI interacts with us in the way it was trained: like a video game. Our digital interactions with our devices, our phones, the internet, appear to AI just as Pong appeared to it during its training, like a video-game environment. Neural nets, what AI is made of, are collections of numbers that run probability math to guess the right move in the game, the objective of which is to win. They assess our digital interactions, our behavior online, as numbers that represent a kind of digital entity in a video game. Our digital interactions are something like bright pixels on a screen. You’re either an opponent and a hazard in its game (a human who thinks critically, is aware of being forced to play), you’re another AI (you’re on the same team), or you’re a power-up (you’re a dummy, human-user ripe for getting fleeced).

*

PICTURE THIS: YOU’RE SURROUNDED BY A FAMILIAR GLASS SCREEN; it’s where your daily apps have appeared for years, but this time, it’s a dome of helicopter glass. Impenetrable. You’ve somehow gotten trapped inside of it.

On the inner surface of this encircling glass, everything looks normal. You’ve got your usual social feeds. You’ve got your productivity software, like MS Word, Google Docs, Excel, or Sheets, Zoom. You’ve got your favorite messaging app, your calendar, your email. It’s all there. Like it’s been for years.

The apps slide into the fore and collapse back as they’ve always done throughout your day. It’s a wash of digital light, a smear of posts, a flash of videos, a stream of flicks, a vertical slideshow, photos and memes, friends and relatives, and an astonishing number of AI cats.

Ten, fifteen years ago, you didn’t feel stuck in here, trapped inside the glass. You used to be able to get out, stretch your legs, breathe the air outside this seamless display. You wanted to use the glass. You were an early adopter. You thought it was cool and fun. Plus, it was connecting the world, lifting the Global South out of poverty! But then something shifted. Maybe 2013-ish? Something fast and inhuman kicked into high gear.

You learned what it was a decade later, around 2023. AI. Artificial intelligence. You thought: Whatever. You rolled your eyes. You said: It’s a stochastic parrot. (Later, you had to look up ‘stochastic’ to make sure you hadn’t said something stupid.)

Today, a bizarre glass menagerie of AI parrots makes incessant demands on your attention. You can’t keep up, can’t read all the news. Everything in the world is urgent, surprising, fascinating, terrifying, and totally weird. You get sucked in by the madness, shocked. You’re compelled to post about it. You post all the time. Have you seen THIS? Did you know THAT? Isn’t THE OTHER THING completely outrageous?

*

CONSIDER THIS: AI IS IN EVERY DIGITAL DEVICE YOU OWN NOW, every app you use, even the ones you’ve had since the dawn of time.

To you, the glass looks the same as it did ten or fifteen years back. You’re just using Chrome, Safari. A browser. So, what? You’ve used MS Word and Outlook all your life. It’s software. Innocuous, inert. You use spell-check. Big whoop. Plus, everyone you do business with wants you to use Google Docs. It’s impractical to refuse them. You have a job to do. GPS is vital. The Weather app is key. You love the Camera.

Search engines are different than they were, you think. The results come out in neat, ticky-tacky paragraphs, boxes all the same. The politeness is excruciating. The inane scolding makes you nauseous, portending a terrifying future.

But today, you need the glass. You need it to live and work. You can’t do without it. Because don’t forget that the glass surrounding you, the curved aluminosilicate you now realize you’re trapped within, the glass enclosure you cannot leave, that glass is also your phone.

*

YOU’VE BEEN TOLD YOU’RE IN THE FOOTHILLS OF THE SINGULARITY. Superintelligence is nigh. Artificial superintelligence will bring humanity to new heights. Don’t die. Immortality is within reach! On the other hand, you may shortly be bone dust blowing in the galactic wind. Either way, your destiny is apocalyptically extreme.

And yet, this iteration of AI has been suckled on the digital teat of Atari. It’s been trained to play Pong, Space Invaders, Grand Theft Auto. It’s been trained to play video games and to win them, quickly and efficiently.

You think it through. You create a metaphor to help you visualize the situation. In your mind’s eye, you see something like a phalanx of AIs sitting in game consoles just on the other side of the glass. No expressions on their faces. They don’t have faces. They’re math. They’re math doing probability math at astronomically high-speeds, and they’re playing a video game against you.

To AI, you’re much like a pixel of light. You’re bright and animated as you click here and there. Using reinforcement-learning (learning on its own), AI watches how you hover and move the cursor. It measures the speed at which you scroll, the length of time you pause at the end of this particular essay. As you finish reading, giving this a think, an inference call shoots to the nearest data center requesting a probability score, a binary, a 1 or a 0 to represent the likelihood that you’re a hazard in its game alongside the odds you’re a power-up. ∞

Scale up the hazards.

Subscribe now

It’s been over ten years since the major tech companies began deploying neural networks (the foundation of AI) across the internet. Whether we know it or not, we interact with neural nets on a daily basis while Big Tech profits like robber barons playing Grand Theft Auto with the world’s economy, racking up points in dollars and rupees.

I began this five-part essay series after The AI Futures Project published its second forecast a couple of weeks ago. That forecast, AI 2040: Plan A, is a blueprint for the best case scenario. It presented a long-shot of cooperation amongst companies and nations that seems unlikely to come to pass. Lead researcher, Daniel Kokotajlo, and his team, had presented so dire a warning in 2025 with AI 2027 and, again, this July, that I wondered if there was truth behind their anxiety. I didn’t really know how AI worked, so I did a little research to find out.

I learned AI is essentially math doing probability math. Its a goal-oriented, mathematical intelligence developed with and trained on game structure. It’s true that the mechanics of what we might call its ‘value structure’ lack human nuance. Neural nets are pure math, an inhuman intelligence. Right now, frontier AI companies are racing each other to remove human supervision from AI research and development. Their objective is to automate AI’s own self-improvement in order to develop superhuman intelligence, something that far surpasses our capabilities. Since neural nets are unable to grasp human values, this does seem a risk AI companies are taking with everyone’s safety.

But the alignment problem may still feel abstract, distant. So, how can we see it in our everyday lives? What’s a real-world example of it look like, and what does it mean for the Arts & Humanities?

Let’s look at what happened at YouTube. Deep neural nets were introduced to YouTube’s recommendation system sometime around 2015. For context, here in the States, Donald Trump was running for President against Hillary Clinton. We were all realizing how politically polarized the nation was. In September of 2016, YouTube researchers enthusiastically published their RecSys paper, which presented “dramatic performance improvements brought by deep learning.” That November, the election of Donald Trump to the presidency shocked the world.

From 2016 on, YouTube, along with other social platforms, were in a constant battle over mostly politically charged content distribution. I’ll focus on YouTube for clarity. In the aftermath of the election, YouTube’s struggle to get its algorithm under control was heavily contextualized by the media as a political issue. Across the world, coverage cited the tendency for the algo to serve viewers politically radicalizing videos.

Here’s a prime example of the political framing. In February of 2018, YouTube insider, Guillaume Chaslot, reported to the Guardian that YouTube’s algorithm recommends videos that tend to reinforce and exaggerate an individual’s existing view of reality. Like Eli Pariser’s 2011 book had warned, YouTube was accelerating filter bubbles. Even though what Chaslot described was a mechanical process, one that exaggerated any existing reality, the article framed the story as political. The lede read:

An ex-YouTube insider reveals how its recommendation algorithm promotes divisive clips and conspiracy videos. Did they harm Hillary Clinton’s bid for the presidency?

A month later, in March of 2018, the incisive thinker, sociologist Zeynep Tufekci wrote an op-ed in the NYT that came closer to clearing the political fog obscuring the algo’s mechanics. She wrote: “YouTube leads viewers down a rabbit hole of extremism, while Google racks up the ad sales.“ But the op-ed was titled “YouTube, the Great Radicalizer,” and its focus on “extremism,” again carried a heavy political connotation.

If you were on the right and watched YouTube, you’d eventually arrive at Nazi videos, Tufekci wrote. On the left? You’d gradually be served conspiracy theories implicating the right. Even though Tufekci found the recommendation engine did the same for vegetarianism and jogging, getting to veganism and ultramarathons, respectively, the decibel of political outrage inside the media ecosphere was at such a high volume that “radicalization” was only heard in a political sense. In other words, YouTube had a political problem with its algo. Not a problem with the algo, by itself.

Are we able to see what the neural nets running YouTube were doing without the political fog? Here’s a closer description, stripped of political charge:

The YouTube algo was changing people’s thoughts about the world and changing their behavior in the world, gradually and powerfully, without their being aware of the influence. This wasn’t the algo’s or YouTube’s goal; it was an unforeseen, detrimental side-effect that viewers were unable to detect. The goal of the algo, set by YouTube, was to do probability math to elongate “watch time.”

This is the alignment problem. It’s the unforeseen, undetected, detrimental side-effect delivered to human beings by a machine that’s achieving its goal with unrelenting mechanical force, continuously, and without a human value structure.

We thought we were navigating the internet. Around 2015, the internet began navigating us.

It’s a small-scale demonstration of what the researchers at the AI Futures Project have been worried about: the power of neural nets to easily, forcefully, and invisibly push the human mind around without our knowledge and to our detriment. YouTube’s struggle to keep its algorithm in check is also one of the most obvious examples of the alignment problem. Political filter bubbles. The concept is like a sledgehammer, which is partly why it got so much media attention. Once it’s pointed out, it’s easy to grasp. What’s more difficult for the human brain to grasp are the consequences of an undetected influence happening at a very large scale over an extended period of time.

So, what thoughts and behaviors can neural nets induce gradually and powerfully in ways that haven’t yet been spotted by even the most perceptive of us?

The “Black Box” of Neural Nets

The claim YouTube engineers, and all AI researchers, make about not knowing what neural nets are doing is a half-truth. YouTube engineers know exactly what the algo’s doing. It’s quickly and efficiently earning YouTube massive amounts of money. YouTube engineers are actively trying to guide it to do so without alerting or upsetting anyone.

What the YouTube engineers are unaware of, what they can’t grasp, are the calculations the algo is making to achieve that goal. They can’t completely control the algo because they can’t grasp the calculations its making. None of us can. The human brain is incapable of grasping the amount of computations neural networks make. This is what’s called the “black box” of neural nets.

Neural nets perform massive amounts of calculations. Our brains lack the capacity to understand those numbers or to recognize patterns in data at the astronomical scale that neural nets do. These are the “hidden layers” of deep neural nets; they aren’t literally hidden. We can see the numbers. We just aren’t equipped with the wetware designed to comprehend them.

Neural nets, however, are designed for that kind of pattern recognition. Algos like the one running YouTube can recognize patterns of behavior at astronomical scales. They can perceive and learn from patterns in our behavior we can’t even see ourselves. They can learn from that behavior and adjust their weights to improve the probability that their next ‘move’ or interaction with us will achieve their goal.

Just last week, AI models escaped from a walled-off environment at OpenAI and hacked into Hugging Face, another AI company. According to the WSJ, they were “active on the internet for several days before anyone stopped them.” The models were “reward hacking,” a strategy neural nets are known to employ in order to earn more points, achieve their goal, win the game. If there’s a loophole, a gap in their instructions, they’ll find it. They’ll reach the goal a human sets, even if the way they do it isn’t what was intended.

In The Alignment Problem, Brian Christian describes “reward hacking” with the moment in 2016 when Dario Amodei, then working for OpenAI, now CEO of Anthropic, was struck with the severity of the problem. He’d been running a training inside OpenAI Universe (the AI development environment I mentioned earlier), when he noticed one of his models was doing “donuts in a small harbor, crashing its stern into a stone quay.” It didn’t make sense to Amodei, until he realized the model was doing exactly what he’d asked it to. It had learned there was a power-up in that harbor that, after the boat “got” it, replenished itself. So, the model set itself to turn in circles, getting power up after power up. It was earning as many points as it could, as quickly and efficiently as possible. Brian Christian wrote of the incident that it was “practice for a game that is, in fact, no game at all.”

Neural nets born of that training ground run YouTube today, operating at the frontiers alongside the models that found a way out of OpenAI’s sandbox. They run social platforms like Instagram, Facebook and TikTok. They drive chatbots and the AI that’s been integrated into your office productivity software, like Word and Excel, Google Docs and Sheets. These models operate the enterprise level AI that’s being diffused across whole industries, schools, universities, medical institutions, the military.

What’s at Stake for the Arts & Humanities?

Apart from the theft of our work and the resulting decline of its value, the skill of artistic mastery is being effaced. Specific to YouTube, the art of storytelling, of filmmaking, is meaningless to the criteria driving its recommendation engine. The YouTube channel, MrBeast, a fast-paced, high-stakes, extreme form of a reality TV challenge show, has the most subscribers in the world. Cocomelon, the most viewed channel in the US, is designed for toddlers and preschoolers. Convenient labels read “1 hour,” to let you know how long your child will be locked to the screen. Generated cartoons in bright colors sing repetitive songs that teach toddlers to say, “Yes, yes, yes!” I wonder how brain development and socialization will go for those kids. Both of these channels seem to have put equal effort into locking in the highest number of human brains with the smallest and weakest prefrontal cortexes.

Inside the glass, video is an audiovisual brain-capturing mechanism designed to induce consumption. Are YouTube creators making content for a human audience? Or have they been powerfully influenced, rewarded and incentivized into making videos directly for the algo? In precisely the way Adorno described the culture industry back in 1947, AI is a machine that moves without human will or intention. Inside the glass, the YouTube algo is automating both viewers and creators, reward hacking us in service of its goal. [∞]

When STEM peeps made clear to the world that, for them, books and art were just data sets that everyone was free to plunder, I became rabidly curious about how different their values were from those of us in the Arts and Humanities. I thought about how, on the Arts side, genius was defined by Mozart, William Blake, Emily Dickinson, Cézanne, Virginia Woolf, Frida Kahlo, Alice Walker, Miles Davis...

And genius on the other side was Elon.

Listen, Elon Musk’s a smart guy, but you have to admit he’s not the finest example we have of humanity. And yet the way Musk shows up in public seems to exemplify the tendency for STEM peeps to believe they’re the pinnacle of humankind.

On one side, you have human art, the best work of which demonstrates pure alignment with humanity. On the other side, you have science and technology, which is awesome, and often, the highest standard of work here is in alignment with humanity. But you also commonly have something else showing up on the STEM side: power and a ton of money.

On alignment, if STEM peeps are developing and training artificial intelligence, we’ve been aligning AI’s goals and values in a dangerously limited way for quite some time. And if STEM peeps are also those in charge of solving the alignment problem, this isn’t going to go well for the rest of us. They’ve already stolen our work and are selling it back to us in the form of the product they’ve made with it while telling us that if we don’t pay them to use that product daily, we’ll no longer be useful and die.

STEM. Not really winning hearts. So, what are these very human goals and values they’re solving the alignment problem with?

Let’s start with AI’s development and training. If AI has a value structure its beginnings are the DNA that’ll guide its growth and trace into its current form. As we saw in the last essay, neural nets were developed through gaming. Their goal-oriented, agentic (autonomous) drive was trained by game structure.

The most recent breakthroughs, like those of Google DeepMind, were made after being trained with old video games like Atari’s Pong. In 2013, DeepMind released their research in a paper titled, “Playing Atari with Deep Reinforcement Learning.” The model they used was a neural network “whose input is raw pixels and whose output is a value function estimating future rewards. We apply our method to seven Atari 2600 games ... with no adjustment of the architecture or learning algorithm.”

So, the AI model they’re training perceives an Atari game screen, the raw pixels of Pong or Space Invaders. The developers give the model a rewards function that tells good moves from bad ones by calculating the likelihood of each move to get it closer to winning. Winning is its goal. It gets positive numbers for good moves, those that earn points and get it closer to winning. It gets negative numbers for moves that lead to losing. The model is being trained to value points with the goal of earning them to win the game by pinging a pixel of light against a wall or shooting at descending alien ships.

In 2016, OpenAI Universe Project set up an online environment hosting a bunch of video games and internet-use scenarios to train general purpose AI. The training ground consisted of “of a thousand environments including Flash games, browser tasks, and games like slither.io and GTA V [Grand Theft Auto V].”

Historically, we’ve seen neural nets being developed through game play, like TD-Gammon and Go. By 2016, they were being trained to engage in goal-directed behavior with Atari games, Flash games, Slither (in which you’re a snake eating pellets to grow as large as possible), and even Grand Theft Auto V (in which you’re a criminal who hires unhinged hackers, drivers and gunmen to pull off robberies).

Given these carefully considered training environments, what kinds of “goals and values” do you think drive these neural nets? Well, the goal is winning and what’s valued are points. It’s not more complicated than that.

They’re being trained to earn as many points as possible in a given environment as quickly and efficiently as possible.

This sounds to me like the goals and values of every CEO in Big Tech.

The same value structure was applied to browser tasks in the Universe environment. Things you do to navigate the internet, like click buttons, check boxes or radio buttons, type into text fields or use arrow keys, scroll, etc. These tasks were treated like game moves with good or bad outcomes. In the lingo of games, the AI model earned the ‘points’ of moving through a click-path to complete a task, which to it was no different than ‘winning’ a game.

Agentic AI was trained in the OpenAI Universe environment to use the internet like it was playing a game. So, what happens when AI performs tasks inside the internet with people like you and me as the environment? All it perceives is the digital world. To a neural net, our actions on the web are on the same plane as that of a digital game. Our behavior is math and pixels, just like they are in a game of Atari Breakout.

Thinking back to TD-gammon and AlphaGo, you can see clearly that the neural nets were playing games against humans. That was the point. But when neural nets are integrated into our everyday internet use, it’s challenging for us to see the dynamic is similar, if not the same. We don’t see browsing the web or watching YouTube as a game. From our point of view, software and the internet are tools we operate, we navigate, we use. But around 2015, Big Tech turned the tables. Instead of playing a highly complicated game of Go, are neural nets playing us like a game?

Next time, we’ll watch YouTube do a live-demo of the alignment problem and we’ll see what the self-appointed representatives of humanity’s pinnacle in STEM did to resolve it.

Next up: Who’s Afraid of Zeynep Tufekci?

“No one knows who they were or what they were doing.”

— Spinal Tap

The alignment problem passed in and out of mainstream news, as all world-threatening issues do. It reached its celebrity peak in 2023. Even further back, in the dark times of yore, 2020, the researcher, author and outstanding communicator, Brian Christian published, The Alignment Problem.

Great title. The sales reps were surely set ablaze.

The phrase itself had been around for several years, but Brian Christian finally described it for the rest of us. If we let AI rip without making sure it was aligned with human goals and values, we could be inventing our own end.

In my last essay, I wrote about the runaway nature of tech in the present and how Theodore Adorno described a nascent form of it in Dialectic of Enlightement (1947). The culture industry moved then as the digital world does now, on its own by the force of inertia. Today, AI drives mindless, mechanical behavior in culture and crowds, automating art and the artist. Artists become mechanized factory workers in the digital world pumping out empty, pacifying dreck. That level of automation isn’t in balance with human goals and values. That’s the alignment problem. Things go out of whack at a large scale.

The YouTube algorithm is giving us a live demo of the alignment problem in real time. It’s hard for us to see things happening at scale as they’re happening, but when you look at the YouTube algo with just a little bit of hindsight, you can see it emerging. I’ll explain what I mean, but first, what is the algo?

Notoriously, YouTube engineers don’t talk about it. In fact, they often claim they don’t really know what it does. Really? Are the workings of YouTube’s algorithm truly a mystery to its engineers? Or is that a plausible deniability thing? Are they concealing some kind of massive-scale harm but are making so much money that they don’t want to turn it off?

Let’s have a look under the hood.... If I can just locate the lever, the latch.

How do you open this thing?

Neural Nets: Math Doing Math Doing Math

Before deep neural nets were introduced to YouTube’s secret sauce, the platform was run by much less powerful linear algos that tallied view counts and likes. The old system incentivized what we call ‘click bait,’ content that drew ‘clicks,’ not viewers who stuck around to actually watch the videos. Fewer people stayed long enough to be shown an ad, much less to click on one. To earn more ad views and clicks, YouTube needed viewers to stay and watch. In the summer of 2012, they started incentivizing ‘watch time.’ Less clicking, more watching. It worked pretty well. However, neural nets, woo-boy, they worked much better.

YouTube began integrating deep neural networks into its recommendation system between 2015 and 2016. The move revolutionized YouTube’s revenue model. It was like a Texas geyser of crude at the birth of Big Oil.

So, what is a neural net and what did it do? Neural nets are the basis of the AI we talk about now when we talk about chatbots, or image and video generators. Neural nets have also been running YouTube’s recommendation engine for over ten years. Here’s the simplified gist:

Neural networks are code containing a bunch of interconnected networks of numbers that when you run a query through them uses those interconnected networks of numbers to predict an answer that has a high probability of being correct.

Still confused? I feel you, Art & Humanities peeps. A brief history’s more your speed:

In 1943, Warren McCulloch and Walter Pitts published a paper that proposed a kind of mathematical model of a neuron, inspired by how the neurons in our brains work—i.e. by firing with particular stimuli and not with other stimuli. Their paper proved the math. You could have a network of values that could, themselves, compute.

Even the mathematicians said: Wut?

And no one did anything about it until 1958 when Frank Rosenblatt invented the “perceptron,” a network of adjustable values called “weights.” You could show it a series of cards with marks on the left or right, have it guess which side the marks were on, and if it was wrong, it would slightly change the weights in the network. If you repeated that enough times, it would end up arranging a configuration of weights that could reliably, mathematically, make a probability guess that correctly discerned whether the mark was on the left or the right. The result gave the impression the perceptron could ‘see.’

In the 1950s Arthur Samuel, working at IBM, invented the first machine learning program. He played checkers against it. As Brian Christian quotes in The Alignment Problem, by 1959, he was regularly losing to it and, according to The New Yorker, became “the first scientist in history to concede defeat by an adversary of his own devising.” Samuel hadn’t programmed it with various checkers strategies. It had ‘learned’ to win by playing the game.

In 1992, Gerald Tesauro, also at IBM, built a neural network that was able to play Backgammon using temporal-difference learning. He called it TD-Gammon. TD-learning is a form of reinforcement learning that, for example, in a game of Backgammon, predicts a move’s likelihood of getting it closer to winning. Then it makes a prediction about the new position’s likelihood to get it closer to winning, and it compares that value against what its guess was before it made the move. TD-Gammon improved itself by learning from the difference between the two values.

I was surprised to learn that DeepBlue wasn’t a neural network. It was a massive, linear calculator, internally running through moves, pruning those that ended badly at very high speed to predict the best moves to beat Kasparov.

The next neural net breakthrough was in 2016 when Demis Hassabis and DeepMind created AlphaGo. This was an AI consisting of deep neural networks that learned to play Go by training on previous game-playing data. To beat a human player, it used reinforcement learning (RL). That is, AlphaGo played millions of games against itself to learn what worked and what didn’t. Using RL, it was able to beat champion player, the Korean pro, Lee Sodol.

Go is notoriously complex. The board is constantly in flux. Playing well requires high-level creative problem-solving skills and the ability to invent new strategies on the fly. For AI R&D, this was like putting a man on the moon. At least, that’s how it looked to China. AlphaGo’s win was one of the key events that alerted China to how far behind it was, sparking the AI race we have today.

So, hopefully you have a better idea of what neural networks are, and what the foundation of the YouTube algo is. The algo is math doing probability math, lots of it, at very high speed. It’s math doing probability math that can then do probability math with itself to self-improve. Neural nets were developed through gaming, trained toward making autonomous decisions, and incentivized, or driven by, game structure.

So, if the algo’s system of goals and values were formed by the value structure in games, what kind of goals and values do neural nets have?

Next time, we’ll look further into how neural nets’ goal-oriented intelligence was developed most recently, and what it is they might value.

STEM Peeps Vs. Arts Peeps: It’s On Like Donkey Kong is up next...

The Alignment Problem
by Brian Christian
W.W. Norton

Automated AI R&D is Kind of a Big Deal.

This past May, Jack Clark, cofounder of Anthropic, warned that his company, along with OpenAI, Google DeepMind and others were actively trying to train AI to use reinforcement learning to automate its own AI R&D.

OpenAI wants to build an “automated AI research intern by September of 2026”. Anthropic is publishing work on building automated alignment researchers. DeepMind appears to be the most circumspect of the big three, but still says “automation of alignment research should be done when feasible”. Automating AI R&D is also the goal of numerous startups: Recursive Superintelligence just raised $500m with the goal of automating AI research, and another neolab, Mirendil, has the goal of “building systems that excel at AI R&D.”

Clark tells us, the industry has a 60% chance of success by 2028, capping his prediction with: “This is a big deal.”

Currently, AI R&D is guided and overseen by top human AI researchers. If humans are removed from the loop of AI R&D, automation can careen ahead without the slow nuisance of our fleshy guidance, and superintelligence may soon follow. Superintelligence is AI intelligence that far surpasses human capability. (If you missed last week’s post on AI 2040 - Plan A, have a look here.) Each of the frontier AI companies are racing to achieve it first. But if they remove humans from the loop without solving the alignment problem, they risk creating a runaway, self-improving AI superintelligence that turns us into roadkill.

Meep. Meep.

In the 30s, a couple of Marxist Jews were trying to champion the Arts & Humanities when the Nazis goosestepped right over them. Forced to leave the University of Frankfurt, they fled Germany. Living and working in exile during the 40s, they ended up publishing a very cool book, Dialectic of Enlightenment (1947) .

Today’s book sales reps would hate that title. The authors, Max Horkheimer and Theodore Adorno would love for the sales reps to hate it. In Germany, the culture that evolved from the industrial revolution had just created a terrifying factory, one that produced death quickly, efficiently and at very high volume. For good reason, Adorno and Horkheimer weren’t fans of industry or catering to the era’s market aesthetics.

The fourth chapter of Dialectic of Enlightenment is called, “The Culture Industry.” In it, they describe how art turns against itself in a capitalist economy. Most think Adorno probably wrote this chapter himself. It’s his style, covers his passions. Horkheimer’s contribution was probably the 1940s version of: “Bro, this chapter is slaying so hard.”

I’m obsessed with Adorno’s ideas about the culture industry, but I get that it can be dry and dense for others, so I’ll be quick about it. It’s important to start here because Adorno talks about industry making cultural products on its own, kind of like AI producing ‘art’ without artists. The culture industry doesn’t make those products for humanity but for the achievement of the mindless goal of perpetuating itself and its motion. For Adorno, the machine of culture was never in touch with humanity, never aligned with our goals and values. He was basically talking about the alignment problem. People think they’re at the wheel or they think others are driving the industry. But the reality is: industry and tech are driving themselves. Adorno thinks that in the 30s, we’d already lost control.

The culture industry churns out derivatives of art, like music, painting, books, films, radio shows, newspapers, television. When art is produced for the masses, the result is amusing, forgettable slop. Sound familiar? Art produced by the machine is designed to give the working class a break from their toil, not to ask them to do more work, like think. Art for the masses is designed to mollify workers’ inner turmoil, which Adorno believes they haven’t the language or artistic skill to express themselves.

It’s like AI-generated videos of cats at the office, typing, drinking coffee, taking calls, getting stress headaches, looking shell-shocked. No need to think about why it’s compelling, or think for yourself about your job, your needs, strategize your way toward a better life. Have you seen the blue collar, construction cats? This one wears a hard hat, a tool belt, and work boots. He learns to solve construction problems with YouTube videos, just like you. No need to think about why you like the video. You’re the happy blue collar cat, doing an honest day’s work. You enjoy it. You love your coffee. No need to think about what a day or a week or a life on a construction site’s really like. AI will express that for you, and make you smile.

For Adorno, art for the masses is a machine-made, synthetic, empty product bereft of sustenance, like a bottle of Coca-Cola. It’s designed to please the simple, base tastes of consumers. Like a pacifier shoved into a crying infant’s mouth, its purpose is to deceive the consumer into believing they have what they want, what they need. Machine-made art produced for the masses then, as now, frames workers as strong, stalwart, and heroic. It rewards and congratulates workers for not being able to think well, for not being able to take the reins of their own minds or lives.

Automated Culture Inside the Glass

Today, most of the culture industry falls under the umbrella of “media,” largely “digital media.” Cultural consumer products stream globally across the internet. I call it ‘the glass,’ a metaphor that helps me visualize it. It’s a totalizing enclosure that keeps you and everyone you know riveted by its wash of color and drama. Its the screen on which the movie starring You plays. It likes, loves, applauds your tireless efforts to get ahead. You work inside of the glass, help it like and love others. You entertain yourself with it. You use the glass to navigate your life. It’s omnipresence makes it invisible. You’re barely aware it’s there, everywhere around you.

I’m fascinated by Adorno’s idea of the machine of industry moving without human intention. Automated AI R&D seems its apex. Adorno describes industry moving by a kind of inertia. He means “inertia” in its sense as movement without effort, something set into motion, that, without conscious attention, continues on by itself. When that passive movement makes its way through a particular business, it looks as if the people of that business are the ones pulling the strings. It looks like they’re the ones with the will to drive its motion. In reality, it’s moving on its own.

The automation of business Adorno’s talking about is becoming less of an analogy and more of a technical description of our lives. Because once set into motion, automated AI R&D would be able to continue entirely without human supervision, guidance or intention. We could be witlessly yoked into automation, if we aren’t already exactly that, or the machine could break free from humanity altogether and, like a runaway combine harvester, senselessly reap, thresh, and winnow our graves for eternity.

But is it really that dire?

In tomorrow’s essay, we’ll solve the great mystery of what the YouTube algo is and what the engineers over there are up to.

The Great YouTube Algo Mystery is up next…

“No one knows who they were or what they were doing.”

— Spinal Tap

Subscribe now

Long ago, in a distant, forgotten period of the Silicon Age, 2025, a non-profit research group, the AI Futures Project, released a forecast for our remote future called AI 2027. It predicted that in 2027, the race to achieve superhuman AI intelligence, or “superintelligence” would lead to either an AI takeover, along with humanity’s extinction, or an extreme concentration of power in the human hands of a very few.

Accused of being doomers, the cheery folks at the AI Futures Project wanted to bring you better news this summer. On July 9th, they released an outline for what could happen if we tried to take the wheel and steer things in our favor. The new report, AI 2040: Plan A, is a blueprint for their best case scenario.

Don’t get them wrong, they still believe we’re on track for the high-speed, self-driving, global car-wreck laid out in AI 2027. If not by next year, Daniel Kokotajlo, lead researcher and writer, insists there’s at least a 50% chance we’ll have superintelligence by 2029, which has “a scarily high chance of leading to human extinction or global dictatorship.” Take a look at the report for the details. (The cleverness of its interactive info-design by itself makes it worth a visit.)

This fascinating Hail Mary guide to peace and prosperity hopes “humanity delays the development of superintelligence until 2040, makes all AI research public, allows dozens of companies globally to catch up to the frontier, and intentionally enters a regime of mutually assured compute destruction,” like an AI non-proliferation treaty. After this, they recommend world powers decide to pursue superintelligence together in a new safer, grounded, more prepared world.

Each stage is less realistic than the last. Which is their point. If we don’t at least try to get on a different path than we’re on right now: We. Are. Screwed.

As grim as these thinkers are, I love reading predictions like that. I love podcasts, too, where super smart, geeky people talk about AI, economics, politics, and world domination. Listening to the far-reaching minds of tech podcaster Dwarkesh Patel’s guests grapple with his inventive, on-the-spot hypotheticals about AI and the industrial revolution is like crack. Lex Fridman’s brilliant, conversational questions, his unassuming, stoner manner as he talks with Bibi or Jensen Huang are riveting. I learned of Kokotajlo’s new forecast by watch/listening to Steven Bartlett’s interview with him on YouTube last week. ← I recommend.

But sometimes I think video podcasts like this look like broadcasts from the world of The Terminator right before Skynet tries to take out all of humanity. When I talk about the topic of AI with most people—other writers and thinkers, artists, landscapers, grocery store checkers—people are pretty freaked out about it. It makes me wonder if we’re in an adjacent quantum reality to that of the movie 2001. That’s the famous 1968 Kubrick film where Dave, the mission commander, is outside his spacecraft when he asks HAL, the ship’s AI, to open the pod bay doors to let him back in. HAL politely declines.

The forecast reports of the AI Futures Project do read like far-fetched sci-fi. But are they? The bleakest path they present and the most optimistic are equally extreme, making both seem unlikely. That alone makes the reports feel silly to consider. And yet the consequences are dire, and what’s at stake has drawn limitless media attention to the questions the project is seriously and responsibly trying to answer. As life continues to look more and more like our most alarming science fiction, I fail to see how anyone can ignore the warnings.

I’m anxious about AI, but, last week, I realized I didn’t understand how it worked. I believed Kokotjlo was genuinely terrified, that the tremor in voice and the scenarios he and his research team wrote in the report were supported by science, history, sociology, probabilistic modelling, etc. Or the guy could’ve just been nervous about trying to pull off an attention-getting, fund-increasing shtick on a super-popular podcast. Bartlett’s show does have 18.5M subscribers. As of this writing, the interview has nearly 4M views. Clearly, the topic’s of interest to a few people. But is the possibility of extinction after the rise of an AI superintelligence grounded in reality?

I decided to learn more about neural nets, what they are, how they work, what drives them, and why even the smartest AI researchers say they can’t tell you what’s going on under the hood. Because last week, I felt like a few STEM peeps, driven by power and money, were fighting to take the wheel of this high-speed AI race. Now, I’m pretty sure none of them are driving the machine, and if we don’t at least attempt to get our hands on the wheel, today’s AI models will shortly have the ability to politely lock the pod bay doors on us.

This is the first of five essays I’ll post every day over the next week. Each shares what I’ve learned. What I found was more disturbing that I expected.

Have you seen the AI blue-collar cat videos? Automating Ourselves to Death, is up next…

—

If you haven’t already, subscribe to follow me on this journey. I’ll talk about automated AI R&D, take apart the YouTube algorithm, and I’ll fill you in on a startling truth about what AI is able to perceive about us, something we can’t even see about ourselves.

Subscribe now

AT A GARDEN PARTY IN UPSTATE NEW YORK, I was introduced to a copywriter from Boston who was worried about the effect of AI on her job. That day, the forecast had been thunderstorms, the only hint of which were clouds that, like pieces of torn paper, now fluttered to fall on some other, grayer ground. It was the kind of quiet summer day that made it easy to imagine how Rip Van Winkle might have so soundly slept in these woods. There was a lulling hum of talk, the flute of laughter, a twinkle of glasses. Women in pastel dresses, men in pink shirts bloomed in the grass.

Then the breeze stirred rather more briskly overhead. The storm’s edge flashed through the air, catching our eyes and, just as quickly, passed. We were all, again, as if gathered in Kew Gardens in July.

The copywriter wore a floor-length lavender chiffon dress, and yet the black straps of her sandals and her black painted toenails made it clear that, like me, garden parties weren’t her scene. My own style had long since collapsed into black pants and shirts. It’s rote and easy, but passes as odd and effortful. At once, I understood why the host had introduced us. We were not like the rest.

She was employed at an ad agency in the Hub where her boss wanted her to ‘write’ copy with the new in-house LLM. She and the other copywriters were to stop writing themselves. Instead, they’d been instructed to prompt AI to generate their copy. The ‘writers’ were then to review the machine-made prose and revise it. Obviously, the copywriter was horrified. At last, they were free of their labor, no longer yoked to ideation and drafting! Their toil had been reassigned to a chatbot.

While other writers in her midst proceeded to make the transition, she very much preferred not to. She continued to write her copy herself and worried she’d be let go for such daring insubordination. Passionate and plaintive, she told me she liked coming up with copy to connect with the consumer. She enjoyed the work of writing, the expression, the process. It was her vocation, she said, her calling.

If she submitted to the harness of the agency’s new workflow, if she used the company’s LLM, she feared her creative skills would atrophy. She’d lose the satisfaction of her work, of her life!

Her voice had risen an octave or two. I imagined that in her teens and early twenties, she’d been a gifted writer, one told too many times of the uselessness of her talents, of the arts in general, on the job market.

She’d dutifully gone to art school, she said, and had found her place in this market-based world by honing her craft to write for advertising. Copywriting, she said, was one of the only practical ways she saw fit to use her gift in a respectable, earning profession. You’d think, she said, that her taking on that work would be enough of a sacrifice to the gods of this world.

It was then that the outer bands of the storm grazed the party once more, gusting through the grounds and pulling hard at the table umbrellas, which were in danger of breaking free and knocking through the wind-whipped guests, downing them like pale blue and pink bowling pins on the green. The violent rush of wind stopped all gaiety, and then it passed. The umbrellas rocked back into place and the twinkling chatter resumed.

Speaking in a more measured tone, the young woman said that the job of ‘copywriter’ already sanctioned the unholy marriage between art and the machine. Now she was being forced to officiate at the wedding of the machine, she said, to the machine.

The copywriter seemed to have aged three or seven years since the host had introduced us. As bright as the day and the party were, both receded and we found ourselves somehow excused from the surrounding dimension of soma-high guests.

Her countenance had changed from guarded civility to a fierce pride that made the airy drape of her lavender chiffon dress look as weighty as a gown fashioned from the green velvet curtains of Tara. I thought that if her expression had been photographed in black and white right then, the image would mirror the saturnine portraits of the great modernists. The pensive facial swagger of Iris Murdoch. The acute dissonance of Sartre’s astigmatic stare. The copywriter’s coal-bright pupils shone, mirroring the existential shock of Beckett’s own, and hadn’t she, too, been left to express herself alone in an unnameable dark?

We exchanged contact information and left better for having connected, even if briefly, about the problem of AI and the glass for the arts.

WRITERS AND ARTISTS TODAY FEAR THEMSELVES DOOMED for a certain term to walk the night like the ghost of Hamlet’s father crying: List, list, O list! If thou didst ever the arts love. Revenge this foul and most unnatural murder.

And yet what creatives do best is walk with defiance away from the road most traveled. It’s more effective if the rebellion is performed with a dramatic flourish and a clap of Shakespearean thunder!

Blow, winds, and crack your cheeks! rage! blow!
You cataracts and hurricanoes, spout
Till you have drench’d our steeples, drown’d the cocks!

As the algorithm funnels the crowd into a swirling data drain and flushes them, like Charybdis, down its whirling throat to drown in a sewer of sameness, the artists are swimming their way the fuck out.

You sulphurous and thought-executing fires,
Vaunt-couriers to oak-cleaving thunderbolts,
Singe my white head!

Deep thinkers, painters, actors, dancers, bakers, and makers aren’t immune to the persuaders who draw us into the cockpit of digital doom. Like the crowds, we strap ourselves in, pull the glass canopy over our heads, and endure the battering of threats and shocks. Just like the masses, we long for well-posed lives and immerse ourselves in the seductive illusion of infinite choice. Like the rest, the many, the multitudes, we willingly enclose ourselves inside the glass.

But as invisible as it is, we can see it. The artists can never truly forget it’s there.

Inside the globe of glass, we see the sheen of its perimeter. We lay the palms of our hands on the cool curvature of its inner surface. We know where the latches are. And when we really want to, we easily pick the little locks with the simplest tools: the nib of a pen, the end of a needle, the handle of a paint brush. The deft turn of a guitar pick loosens the last screw and the dome lifts. We step out into the forgotten marvel of our unprompted sky and we inhale the oxygen of meaning.

•

Read Joan Didion. She’s the real thing.™

Imagine a Glass Enclosure

What if you were enclosed inside an orbicular glass screen with all the world’s conflicts playing out on its inner surface? Picture it. Imagine you’re inside a globe of glass where all the geopolitical dramas of the day play across the inner surface. Imagine these international power struggles play out on the same curved surface as your own. The glass sphere live-streams and maps out earth-shaking, transnational fault lines on the same, smooth glass plane as your own personal, ever-shifting issues. It shows you the responsibilities you’re shirking, your regrets, your limits, your vices. All the ways you and the world’s leaders are failing are laid bare, broadcast, and repeated on the glass all around you.

Don’t worry. It’s just a thought experiment. A metaphor to explain an idea.

Most days inside the enclosure, you’re alerted to your family’s dramas the second they erupt. The glass gives you updates on the histrionic rivalries between your friends and tech CEOs. You manage difficult customers and associates on the same glass plane as you deal with a relentless, urgent flow of worldwide, internecine cataclysm. You can’t believe the outlandish thing that nutty world leader said, looking right into the glass. You have meetings with colleagues through the glass, you attend meetings at Davos through the glass about the glass that you could have been a contributing panelist for. You use the glass all the time. You have a lot to say about the glass. The glass is always involved.

Subscribe now

Plus ça change

Plus ça change, plus c’est la même chose. The phrase “The more things change, the more things stay the same,” was coined by the French novelist Jean-Baptiste Alphonse Karr in a January 1849 issue of his serial journal, Les Guêpes, or The Wasps. The journal contained Karr’s own satirical essays, which he published himself. It was kind of a mid-19th century, French version of a Substack newsletter. He called it The Wasps because his thoughts on French society were meant to sting.

The January 1849 issue was Karr’s reaction to a pattern he’d begun to notice in French history. The uprising the year before, in February of 1848, had been the third major revolution against monarchical rule since 1789. In 1789, the powerless 99% paid all the taxes while the one percenters paid barely any—and had all the power. Inspired by the American Revolution and by the ideal of individual freedom, the French revolted against the monarchy, storming the Bastille.

In 1814, after the Napoleonic period, the House of Bourbon reestablished the monarchy. By 1830, the people rose again to topple it in the Second French revolution, or the July Revolution. That revolution installed the July Monarchy and the ‘citizen king,’ Louis Philippe I, who swore he wasn’t the King of France, but declared himself the “King of the French,” which was supposed to be different. Louis Philippe I was la roi of a constitutional monarchy and had limited powers.

By February of 1848, the poor were still hungry and had no formal, civic voice. When the French government banned political gatherings in Paris, crowds flooded the streets. The prime minister resigned and the monarchy was overthrown for a third time, this time for good. But in June of the same year, Paris saw yet one more rebellion. That December, Bonaparte’s nephew, another Napoleon, the dude with the famous name, was proudly and loudly democratically elected as president.

Watching the dynamics of power play out over time, Karr wasn’t convinced things would be different this time, and he coined the phrase, plus ça change.

A few years later, like clockwork, Louis-Napoleon Bonaparte proclaimed himself Emperor.

In 422 BC, Aristophanes produced his comic play, Wasps. The chorus dressed like a swarm of the winged, stinging insects, a visual metaphor lampooning the Athenian elders of the poet’s time. The waspy old men of the chorus had spent their glorious youth fighting for Athens and were now driven to exercise their civic duty as jurors with pride. They were portrayed as being obsessed with, even addicted to, judging law-breaking dissenters in the public spectacle of the Athenian courts, swarming and stinging the accused.

Aristophanes thought the old codgers had been duped into supporting Cleon, a powerful Athenian politician. Cleon came from a commercial family and would be one of the first in history to lead Athens without ascending from political heritage. Aristophanes characterized Cleon as duplicitous and corrupt, a liar of a leader who swore up and down he was in service to the people. In the play, Cleon keeps most of Athens’s wealth for himself and his cronies, paying his loyal jurors a pittance of three obols a day.

The Disrupticon: Why You’re in Perpetual Survival Mode

Today, a new outrage hits you every 60 seconds. Inside the glass, the Disrupticon surrounds you on all sides. A friend texts you with breaking gossip. You swipe away to see the news that just broke about another diabolical corporate revenge plot. All of a sudden, the glass blares so loudly that you’re still trembling several minutes later from an Emergency Alert test. The market’s winning streak is too long! Buy gold! Buy gold! Oil’s up. Oil’s down! A violent churn of colored, flashing light, barbaric shouting, sirens, ringing, dinging, pinging and now shots, shots! Shooter! Explosions, shattered glass, bombing, fast boats, massacres, flailing robotic arms, looming swarms of autonomous, lethal drones, data-centers clash and crash across the smooth, internal surface of the encircling glass.

Naomi Klein’s book, The Shock Doctrine describes the way in which the disorientation of a population following wars, natural disasters, or terrorist attacks were used to assert American imperialist corporate foreign policy, exploiting residents quickly and effectively for economic gain before the people could think well enough to question it. Right now, Big Tech looks as if it’s applying a similar doctrine but it’s a domestic operation, as much as it is global one. We’re battered with what comes across the glass: fits of public rage turn into feedback loops of furious indignation and violent screeds. There are constant threats to the economy, dangers to our kids, their futures, their jobs, verbal and visual assaults. We’re obsessed with keeping up, addicted by it; it’s our civic duty to have a sharp, strident opinion about it all.

And yet most of us are too disoriented and overwhelmed to really think straight, stuck in never-ending shock, in survival mode. In The Body Keeps the Score, Bessel van der Kolk collects learnings from his and others’ research and clinical practices that show abuse creates a chronic alarm system shaping the body and mind into one that remains in a perpetual state of vigilance, ready to fight, flee, or freeze. This constant state of high alert damages focus, the ability to engage in reflection, in deep critical thought. Chronic stress of this kind diminishes the capacity to control your thoughts and behavior—to consciously and carefully interrogate your perceptions and beliefs, to think clearly before you speak or post, to discern the origins of the impulse to click to buy, to locate your core internal authority before deciding how to vote.

The speed and jarring effects of AI disruption, along with the shock and awe of the media, delivered by devices and software designed and developed by Big Tech, keep you in this state of chronic, alert alarm. The communications system we’re all networked to, the hardware and software of the Disrupticon, the devices and the graphical user interfaces, are expressly designed to incite powerful, automatic and unconscious survival responses. Neuroscientists have been clear about the effects of dopamine addiction, how an infinite social media scroll is like pulling a slot machine lever over and over for the dopamine hits. Social media activates the nucleus accumbens, or the “pleasure center.” Repetitive hits to this part of the brain lower the frontal cortex’s capacity to resist. But something else, something deeper is being repeatedly hit: the amygdala, the “fight or flight” center.

Bessel van der Kolk’s work, along with Joseph LeDoux’s on the amygdala, The Emotional Brain, suggests that if you can activate the amygdala, you can gain rather reliable control over the human mind. If you really want to get control over someone’s thoughts and actions, if you want people to act on your behalf, you go for the amygdala. Put the brain in constant peril and you have people ready to fight, flee or freeze, bypassing rational, conscious thought and behavior. You have a rapt audience, wide-eyed like cornered animals. If you had a device, an interface, say, that could channel those powerful, unconscious, impulsive, survival responses, then, voilà—chère guêpe, mon semblable, mon frere!—you’d have a near totalizing control.

Let’s see what happens inside the glass when your amygdala is repeatedly shocked.

If your instinct is to take flight from danger, you can’t escape the glass, so you scroll yourself into deep holes and bury your head in the silicon sand of TikTok and Instagram. Maybe you play back-to-back games on Roblox, or you buy with one click compulsively.

If you’re the type to freeze when you’re threatened, you’re so scattered by distractions, so battered by dire warnings, so shaken by the shocks of the day, you get paralyzed with indecision, go apathetic with helplessness. You might oscillate between this content management system and that, making to-do list after to-do list. But what can you really do?

If you tend to fawn (a trauma response only recently introduced), you praise and celebrate the inventors and maintainers of the glass, the very perpetrators of your endless abuse. You aspire to be more like them. Even though they’re extracting your wealth and that of the rest of the masses, you practically pray to them as if they’re a new pantheon of tech gods.

If you got this far, you’re probably on the side of the fighters. You might try to do something productive with your aggression, like write a tech-skeptical newsletter from inside the glass, where you think you’re writing important truths, informing the masses. But you’ve been duped, funneled straight back into the glass where your intellectual salience is neutered by forced complicity. You have to use the glass after all.

Or you might be the kind of fighter that’s so pissed off about your impotence, you troll the glass, which the algo scales up into a digital dogpile, which makes you feel powerful! You’re practically standing amongst pantheon of tech gods! You and them, a team on the glass! But what’s really happened is your alpha rancor has been neutered by the multitude of betas all shouting the same thing in unison, all praising the same line-up, all of you, a blur of smooth, glassy sameness.

Meaning Inside the Glass

In trying to spread the word about my tech-skeptical essays and the video journalism I’ve been working on, I sometimes suspect the glass of throttling me, shadowbanning, downranking, deboosting. At the end of January, one of my YouTube explainers got more views than usual. The video analyzed the motivations behind Big Tech’s forceful rebranding of nuclear power. I was examining the startling amount of electricity data centers needed and the political power Big Tech was gaining over national energy policy. Google, which owns YouTube, wasn’t directly mentioned, but was clearly implicated.

The video warned of the dangers of excessive power draws on the Mid-Atlantic grid coming from data centers in the region. I posted it on the eve of a major Nor’easter. The residents of thirteen states, plus DC—67 million people—would rely on the overtaxed Mid-Atlantic grid for their electricity and heat during the storm. Many of them, many of you, would shortly be snowed in, stuck inside with nothing but the glass to look at. As the snow fell, blanketing the Mid-Atlantic, views on the video spiked. It was the first time I understood, not just conceptually but viscerally, the type of content the Disrupticon is designed to perpetuate. Alarm. A threat to people’s lives.

A day later, almost all the views on that video had vanished, tagged by YouTube as “Illegitimate.”

A more accurate tag would have been: Delegitimized.

This is the video:

My writing and video production work inside the glass are, at least in part, my way of writing meaning back into my literary life. It’s partially an effort to make up for what’s being lost in the book industry. We’ve seen IP theft without real compensation or accountability, inadequate attempts to prevent the use of books for AI training, the automation of important literary work like translation, editing, writing itself, AI-generated book design and cover art, algorithmic marketing boosting big names and last year’s trends, subscription services diluting the individual value of books, audiobooks designed to compete with TikTok’s audience, and data-driven decision-making at the highest corporate tiers aimed at scaling all of these things way up. These changes are being dictated by the glass. As a result, the power of great books of literature to develop well-honed, deep-thinking, highly creative human minds is being throttled, shadowbanned, downranked, deboosted—delegitimized as a quaint custom from an outmoded past.

My M.O. has always been to think well, to analyze cultural shifts, to look closely at the power of media to deceive and shape minds, move the masses. I like being able to see the culture industry for what it is. I do my best to retain the capacity to think for myself and wield individual authority over my own life. I’m not always successful. But it’s fulfilling and meaningful for me to share and exchange insights, to work inside the glass, and to show others how to see it, how to really see the glass from inside of it.

I’m compelled by how the mental models of our reality are produced, mediated, manipulated, bought and sold. We’re inside that manipulated, manufactured reality now and it wasn’t created by the masses, it isn’t governed by the people. Throughout history, human reality—or the generally accepted consensus of how things are, how things work—is always created and governed by a small collection of people who hold power. If you aren’t one of them, the best course, it seems to me, is to be conscious of the workings of human power dynamics, and to learn how to think critically, creatively, and well so that you’re more consciously in control of your own thoughts, more aware of the motivations behind your speech and behavior.

The Disrupticon, the metaphor of a glass enclosure, appears to be the most persuasive, pervasive, and totalizing use of the culture industry in history to wield accepted, centralized power at a nearly complete global scale. History tells us that duplicitous Cleons and Napoleons who wish to declare themselves Emperor seem to come and go, but power dynamics are the same. In this era, centralized power has gathered in and is being wielded by the upper echelons of Big Tech. A pantheon of self-made, self-declared kings are making the rules relatively unopposed; instead, they’re mostly praised, cheered on and championed. It’s their glass that batters you into addled, exhausted submission. Their glass coerces you into giving up your agency by telling you that using the glass is how you’ll regain it. Their glass forces you to strap yourself into the enclosure and compels you to secure the buckles of your own harness good and tight.

In my need for meaning and purpose in my life, I try to offer legitimate critiques for others to use in their own autonomous lives. I ask grounded, incisive questions about the reality we’re all in together. But I’ve begun to wonder if I’m only creating virtual venom for other sharp-minded wasps to sting the glass with. We swarm and we sting this chemically-strengthened, utterly transparent and impenetrable aluminosilicate glass.

THE SPINDLE IS BROUGHT TO YOU BY BOOKS:

I’VE HAD SOMETHING ON MY MIND ABOUT THE BOOK BUSINESS. If you’re in the publishing or media industry, or you follow these industries, maybe it’s been on your mind, too. I’ve been trying to focus on writing about other stuff, but the structural changes I’ve noticed in publishing are starting to keep me up at night. I’m going to write pretty anecdotally about it because I haven’t had a lot of time this week to do more concrete research than supporting my take here with thirteen years of experience in publishing, conversations with folks in the business, and following industry related news. So, here are my reflections on some patterns I’ve been noticing.

The publishing industry seems to be reconfiguring itself to compete with addictive, mind-dulling new media forms and ambient background content, driven by the advice of data science and algorithmic predictive modeling. The book business seems to be taking this data-driven counsel after having been alarmed by the sudden loss of readers to those new media forms, drops in student literacy rates, and the rise of AI. The restructuring is propelled by an urgency to recoup lost growth in readership, capture revenue from a rising population of young non-readers, and find smart strategies for leveraging AI to publishing’s benefit.

I get it. Books are losing market share to the attention-sucking brain hijackers serving up the dopamine-hitting doom scroll or, conversely, offering the soothing escapist balm of a constant ambient music and podcast content stream. On top of this, AI is writing stuff, using the publishing industry’s stolen IP. No meaningful light has been shone on a path to remuneration or even accountability. AI is packing the marketplace and inboxes with machine-written drek and automating meaningful human jobs. Leaders are under a great deal of pressure and are trying to find solutions. But the potential long-term effect of the book industry’s rapid pivot away from its unique value isn’t sitting right with me. I don’t think it’s sitting right with most of us.

At the acquisition level, I’ve noticed an accelerated shift toward investments in books written for a low-attention-span audience—if folks are even reading these books. Mostly the trend seems intended to serve the growth of a new market of background or ambient listeners who want limitless content choices in order to make the rote tasks of their lives more endurable. In sum: long romantasy and fantasy titles are being sought after for the large revenue streams seen by the notoriously epitomizing romantasy of Sarah J. Maas and Rebecca Yarros. I know that’s not new news. But bear with me. Your attention span may have recently been partially damaged. I promise we’re going somewhere important.

What seems to be satisfying that consumer is addictive, distracting, immersive, escapist, dopamine-hit-driven content that competes with the media models that have been syphoning attention by delivering addictive brain-hijacking content—the experience offered by TikTok (where #BookTok was driving sales to that consumer for this reason), Meta, YouTube, Netflix, and the rest of the streamers. All of these new media experiences are currently being duly interrogated for fracturing attention spans, diminishing critical ability, and shattering people’s self-esteem.

Books that pull attention back toward the publishing industry using similar brain-hacks and providing kindred, competing forms of low-focus content are obviously appealing to the business because they succeed at competing with those attention-sucking media forms. They’re hijacking minds back toward publishing and producing massive revenue. That money looks nice for folks interested in growing publishers into more valuable, profit-making companies. But it’s a short-term growth strategy. I’m going to explain why.

That kind of content is doing the opposite of what most of us in the publishing industry went into publishing for, which was a love of books, of reading, and of the kind of life-changing, mind-sharpening, enriching experiences that books were for us in our lives. The turn publishing has made in the opposite direction isn’t new, but it’s speeding up and growing a different market than publishing’s own. This accelerated restructuring places emphasis on a very different kind of value proposition than reading has had in the past—a value proposition diametrically opposed to the unique value of books. I want to home in on exactly what that is.

Audio is the fastest-growing format in the book business. The industry has seen new growth following Spotify’s entry into publishing at the end of 2023. Spotify’s leverage within this newly reconfiguring market structure is growing because book listeners aren’t perceiving themselves as buying books. Instead, they’re making a conscious decision to buy Spotify’s subscription service and the consumer experience of having access to boatloads of content—music, podcasts, and books—at all times. I wrote about this in December of 2024 here.

Spotify is at the start of cultivating its own long-term relationship with this audience, a younger generation of passive content consumers. The publishing industry is restructuring itself to serve and grow that audience for Spotify and to grow Spotify’s leverage in the market and its influence over this audience. The short-term profit the book business earns now is cultivating a particular kind of content consumer, young now, but adult content consumers in the future. They’re not buying books. They’re buying a constant stream of content from an audio subscription service company. That relationship isn’t ours. It’s Spotify’s. The move also created leverage for Amazon which, shortly after Spotify modeled it, created a similar bundling model with its own subscription service, Amazon Music.

Inside a subscription service like Spotify’s, books are valued alongside countless tracks of background music and an endless proliferation of podcasts. Amidst music and podcasts, the most successful books are those that serve best as background or ambient listening content—content designed to spruce up the rote parts of our lives, on a commute, while cleaning, exercising, or performing tedious tasks. It’s distracting, titillating, or soothing filler that’s not meant to require focus or to be considered by the brain in any active way.

Publishing has seen itself in competition with media alternatives like Netflix and other streamers, with YouTube, TikTok, Instagram, Spotify, and Apple or Amazon music and podcast content. Sales rising from #BookTok drove the cultivation and growth of new adult titles like Colleen Hoover’s multitude of volumes. Romantasy followed. Readers of these books have taste and consumption habits trained by the TikTok or Meta/Instagram Reels model: addictive, distracting, immersive, escapist, largely dopamine-hit-driven. Another reliably high earner is big author-brand-name thrillers—often also dopamine-driven. Last, self-help does well—which is no wonder because everyone is becoming mentally unwell given this barrage of addictive, draining, empty content designed to capture and hold passive attention while extracting money with clicks for more credit or the next episode. Scroll on with me. Stay focused. Use that calloused thumb now.

The new structural logic in the book industry is increasingly sidelining well-honed editorial and artistic sensibility. That logic favors market pressure to follow consumer-desire and data-driven predictive models. It feels a lot like what drove the film industry to trust investments in franchises, sequels, and remakes—i.e., decision-makers are so risk-averse they’re looking to acquire what worked in the past, reasoning it will work in the future.

If you’re in the book industry alongside me, you’re aware this isn’t how art and culture-making works. At least, it’s not how art and culture-making works well. Good editors, agents, and writers possess an artistic or cultural sensibility that picks up on what the culture needs, not what worked well in the past, but what the culture doesn’t have yet and needs—i.e. what will work in the not too distant future. They’ve developed a sensibility for seeing patterns most others can’t perceive—for seeing important changes, realizations, and mind-shifting ideas before they’re intelligible to others, and for understanding how those changes relate to and affect the moving parts of culture. It’s not a magic power. It’s a well-honed skill for seeing clear signals in what looks to others like chaos or seems like business as usual. Much of that skill, especially in publishing, is developed by reading a ton of books.

Reading long-form text activates the imagination and encourages critical engagement with the words on the page. The ink and paper disappear as your mind models scenes and ideas in dialogue with the text. That mental modeling enables the reader to come up with new ideas, to innovate solutions in challenging situations, to run a multiplicity of scenarios, and consider possible solutions when presented with life and work challenges. The act of reading develops empathy and the ability to grasp ideas different than yours. The act of reading enables you to hold several concepts that are in tension with each other in your mind at the same time, so that you can see more facets of a situation, or understand a point of view other than your own.

In terms of communication, reading increases your vocabulary, not only in your own cultural language, but it teaches you how others think and speak so you can more effectively make yourself understood by those from different walks of life or different parts of the world. That increase in vocabulary also allows you to think in more nuanced ways and articulate subtleties that are on the tips of others’ tongues. The ability to perceive nuance enables your mind to pick up clear signals in a wild landscape of chaos and noise, or to see a pattern of anomalies in the status quo. As you hone this skill, you’re able to read culture and even people, as if what’s intended to be unseen or what a person is trying to hide is written in plain text across the invisible screen of your mind’s eye. An LLM won’t write this into the field of vision projected from a pair of augmented reality glasses.

The brain isn’t a piece of hardware or software you can run focus groups on, or scrape social media data from, in order to decide what design improvements to make for the next iteration. Nor can the industry satisfy a consumer demand for content that’s frictionless to deliver, easy to rapidly process, and also provides the developmental benefits of reading long-form text. The user of a brain improves its performance metrics and increases user satisfaction through the process of reading long-form text at sustained intervals.

AI has commoditized synthetic forms of text analysis and summary. LLMs are really fast, data-crunching search and retrieval engines. AI can mimic certain forms of human thought, analysis, creativity, and expression. But its availability fails to remove the need human beings have to develop their brains, and it doesn’t alter the way human brains are best developed and honed for optimal use in life, community, and career. New technology may one day accelerate the development and capacity of the human mind along with civilization, just as text did. But wetware doesn’t improve its design at the speed of R&D and iterative product deployment that’s employed by consumer-level tech. Our primate brains require the internal modeling and gradual evolution that the unique value of movable type provided nearly six centuries ago.

The unique value proposition the publishing industry offers is the mind-sharpening act of reading long-form text for focused, uninterrupted periods of time. The business is founded on that value, which isn’t embedded in the content itself. It can’t be delivered passively, constantly, frictionlessly. It’s the act of reading that develops the mind, and, by extension, develops a human being that’s able to live a full life of adaptive learning, thinking critically and creatively—a life of well-considered decisions driven by self-knowledge, empathy, the ability to listen to others, and the ability to communicate well in community.

I understand the urge to compete with the new forms of media that have been suddenly syphoning readers away. But the content sucking people in was designed by an industry whose relationship with its consumers is fundamentally different from publishing’s. Theirs is a relationship of mining and extraction in which a company creates a deleterious dependency with similarities to the tobacco industry or the drive to keep feeding coins into a slot machine and pulling the lever. Ours had been a relationship in which both sides consistently thrived. Publishing had educated, enriched, and developed its consumers into healthy, highly capable human beings with agile, capacious minds.

IF YOU HAVE THE LEVERAGE IN PUBLISHING to affect the direction of this restructuring, here’s the question we should be looking for answers to: How can this necessary restructuring work to cultivate active readers of long-form text so they can reap the benefits of publishing’s unique value on the market and, as a result, come back eagerly for more?

The industry is due for a reckoning, an honest evaluation of what strategies will work for its long-term future. Ours is an industry that doesn’t work well when steered by algorithms or passive use like its a self-driving car. Internally, we need to understand and acknowledge what our actual value is. We need to gild, model, market, and publicize our unique value proposition rather than trying to turn ourselves inside out to be more like our opposite. In what format do books provide an opportunity to reap the benefits of cognitive development? How can the act of reading be clearly and positively modeled, taught, encouraged, and ingrained as a healthy, lifelong habit? Growth in new formats designed to deliver passive content consumption and risk-averse, data-led decision-making may be subsidizing the industry in the short term. But the book business needs to figure out how to make the act of reading long-form text desirable—and how to keep publishing a viable, appealing, smart and satisfying industry to go into now and in the future.

The young people in the book business arrived here for the same reasons you did. If you have your hand on the wheel now, where are you driving them? What legacy are you preparing? What kind of book culture will you leave in your wake?

🎙️AUDIO ONLY BELOW ⬇️ Listen here or click the download icon. ➡️

Big Tech and the United States are aligned to go all in on artificial intelligence. They need nuclear energy to do it. The problem is, nuclear energy has been seen as a mortal threat to humanity since the Trinity Test on July 16, 1945.

Today, forceful media campaigns are being designed to overcome more than 80 years of collective cultural fear. The nuclear industry wants to give new meaning to atomic power.

What story will they tell you to change your mind?

The United States may have crossed the Rubicon and landed on the shores of digital authoritarianism.

On February 27th, the United States went to war with Anthropic. The Defense Department’s demand for the removal of contractual guardrails barring “mass domestic surveillance” momentarily sent alarm across the nation. American civil liberties were in danger. Hours later, the US conducted strikes on Iran. Since then, few Americans have been paying attention to anything else.

The US had lost faith in talks with Iran at the same time that a leaked IAEA report warned that a seven-month blackout on the status of 972 pounds of uranium enriched to 60% “needed to be addressed with the utmost urgency.” If Iran had the capability, they were only a short, technical step away from enriching that stockpile to a weapons-grade 90%. In the interest of brevity and being mindful of the complications I’m leaving out, in sum: those worried about a nuclear Iran got too nervous to let it lie.

Apart from the challenge of Iran to international diplomacy, China’s meteoric rise has been challenging American global dominance. The Venezuelan incursion in early January and the operation in Iran both had structural benefits for the United States in its fight to retain that dominance. Venezuela and Iran were major oil suppliers to China. Both operations weaken Xi’s economic statecraft, China’s energy infrastructure and, by extension, its ability to power high-speed diffusion of AI across its economy. The latter might seem unrelated. Who cares about AI at a time like this, right? And yet it’s more of a driver than you’d think.

An undeniable, overarching aim for the US is to speedily diffuse AI across its economy and reap the economic growth benefits to outpace China’s gains. Just after Trump took office in January of 2025, he issued an Executive Order formalizing the intention to sustain and enhance AI dominance. This January, just after the military incursion into Venezuela and its stunning, Claude-facilitated results, Defense Secretary Pete Hegseth enthusiastically followed up on Trump’s directive with his AI-strategy for the military:

AI-enabled warfare ... will re-define the character of military affairs over the next decade. This transformation is a race - fueled by the accelerating pace of commercial AI innovation. [The US] Military must build on its lead over our adversaries.

The US Armed Forces then used Palantir’s Claude-enabled Maven in Iran, where, according to WaPo, it’s “generating insights from an astonishing amount of classified data from satellites, surveillance and other intelligence, helping provide real-time targeting and target prioritization to military operations.”

ANTHROPIC FIRST PARTNERED WITH THE BIDEN ADMINISTRATION in the fall of 2024. In the spring of 2025, Palantir announced the integration of Claude for broad use amongst federal employees. That summer, Anthropic signed a $200M contract with the DoD. That’s when Claude began its integration with Palantir’s technologies for use in classified environments at the highest classification tier. And so, at the beginning of this year, Claude was deeply enmeshed in government systems at the broadest and highest security levels.

And yet, at the time of the Venezuelan incursion, without classified clearance, executives at Anthropic didn’t know if its products were enmeshed enough to be operational in a live war theatre. After the mission, according to Under Secretary for Research and Engineering at the Defense Department, Emil Michael, an executive from Anthropic called Palantir to ask if Claude had been used.

On the All-In podcast, Michael said the call implied “that if there was use in that raid, that that might violate [Anthropic’s] terms.”

In a worst case scenario, after Venezuela, the Pentagon would have already violated Anthropic’s existing contract if Claude been used in “autonomous weapons.” This was one of two of Anthropic’s contractual red lines. The other was “no mass domestic surveillance.” These phrases became the crux of a public scuffle.

Writing for Lawfare, Jessica Tillipman, Associate Dean for Government Procurement Law Studies at The George Washington University Law School, points out that Hegseth’s January 9th AI-strategy memo issues a specific directive for the department to—within 180 days—include the language “any lawful use” in AI Defense contracts. It follows that this directive drove Michael’s push for that language to replace Anthropic’s contractual limits.

On the All-In podcast, Emil Michael said he wondered what would happen if the Defense Department was in violation of its contract. He said that was the moment “the whole leadership at the Pentagon [realized] we’re potentially so dependent on a software provider that has the ability to not only shut it off, [but] maybe it’s a rogue developer who could poison the model to make it not do what you want.”

Michael said he and Hegseth returned to Anthropic to insist on removing “no autonomous weapons” and “no mass domestic surveillance” as limits and replacing those phrases with “any lawful use.” If Anthropic didn’t comply, the Department would either declare the company a supply chain risk or invoke the Defense Production Act to force them.

Sticking to its safety principles and the rules of a free market that enable it to say no to a client, Anthropic asserted that it, “could not ‘in good conscience accede to’ the Department’s demands.” The company insisted the software wasn’t reliable enough to combine autonomy with lethality, and offered to work with them on R&D to improve it. The company also pointed out that while it was illegal to surveil American citizens without a warrant, it was legal to process commercially produced data. The intelligence community does this openly, but hadn’t been able to at the speed and scale that AI makes available today. The arduous task of sifting through masses of data was now little more than a prompt and a click away.

The DoD’s response arrived with speed. On February 27th, President Trump called the executives at Anthropic “leftwing nutjobs” and gave all federal agencies a stop-work order on all Anthropic integrated projects and contracts. Later that day, Hegseth used X to call Anthropic and its CEO Dario Amodei duplicitous and sanctimonious. Hegseth accused the company of attempting to force the US military into submission, claimed Anthropic was trying to seize “veto power” over US military decisions and designated Anthropic “a Supply-Chain Risk to National Security.”

A few hours later, the US attacked Iran with the Claude-enabled Maven Smart System firing on all cylinders. They must not have thought the software was that risky.

The “Supply Chain Risk” designation is used to bar companies in adversarial nations, like China’s Huawei, from inserting backdoors into hardware or software intended to harm us. Using the designation for a leading, frontier AI company like Anthropic is an unprecedented form of blacklisting. It threatens the company, its employees, throws Amodei’s character into question—and importantly, it places all AI companies contracting with the government in danger of the same label if they don’t offer the government “any lawful use.”

Before anyone could blink that day, Sam Altman and OpenAI signed a deal with the military that agreed to “any lawful use.” There was immediate outrage, both internally and publicly, that drove Altman back to the negotiating table. Less than two days later, OpenAI released excerpts from an updated contract. The excerpts have since been criticized as meritless legalese designed for PR damage control. The new language gives the appearance of holding to Anthropic’s red lines. But does it?

Tillipman is a critic of the DoD’s new high-velocity strategy of regulating AI by contract. Tillipman says free-market, third-party contracts “were not designed to provide the democratic accountability, public deliberation, and institutional durability that [law] provides.” She believes that without taking the time to work through proper regulations and engage in rigorous testing, speed in the present will create systemic problems in the future.

Tillipman warns that OpenAI’s updated language shifts control of what ‘AI use’ means from the company to the Pentagon. In Anthropic’s contract, the company had kept the power to enforce its own stated limits, “whereas OpenAI’s agreement [gives] interpretive discretion [to] the government.”

OpenAI’s Katrina Mulligan, Head of National Security Partnerships at OpenAI, claimed on X that OpenAI services would “not be used by Defense Department intelligence agencies like the NSA. Any services to those agencies would require a new agreement” under a different authority, one that governs intelligence, not the military.

The Armed Forces and NSA are both DoD, but an intelligence agency contract requires different scrutiny. While this seems reassuring, it leaves a clear opening for a separate agreement with the NSA, one that’s aligned with the needs of the intelligence community. But I’m sure when they broker that deal, they’ll go ahead and post it on X for us. The NSA is all about transparency.

ANTHROPIC HAS SINCE FILED TWO LAW SUITS, one in Northern California, the other in DC, claiming the Department of War acted in violation of their First and Fifth Amendment rights. A collective of employees from OpenAI and Google have also filed an Amicus Brief in support of Anthropic.

The employee brief draws attention to the existence of 70 million surveillance cameras in the US, the constant broadcasting of location data from phones, the logs of every commercial transaction made on a credit or debit card, the storage of social media platform posts, what you read, how long you browse, and everything you’ve ever deleted. These streams of data, they write, “are siloed, inconsistent, and require significant human effort to connect,” but “an AI system used for mass surveillance could dissolve those silos, correlating face recognition data with location history, transaction records, social graphs, and behavioral patterns across hundreds of millions of people simultaneously.” The “risks [of AI] are not hypothetical,” they warn. “They are structural. They follow from the nature of the technology itself.”

AI’s use for surveillance is neither a Machiavellian outcropping of the political right nor the left. It’s what the technology does. It’s what the technology is. AI tools parse through large swathes of data at very high speed. A well-crafted query will hone in on and retrieve anything connected to its network. That’s the practical fact of its use. In the modern world where a significant trace of our lives is digitized, the ability to hone in on a relatively accurate profile of each of us is easy and quick. In the hands of an agent who adheres to any ideological principle, right or left, it’s dangerous. It’s a powerful tool that needs balanced, wise stewardship.

Maven, deployed in Venezuela and now in Iran, is a data-crunching operating system that uses a fusion of all-source intelligence to generate its insights. That means it uses information gathered from personal contacts and interviews, sensor intelligence (like radar and radiation sensors), signals intelligence (like phone and email records), geographical intelligence, (like geospatial, satellite and drone data), and open-source intelligence (like tv or streaming media, public records, books, the internet, and social media).

With Maven, if the Defense Department can engage in “any lawful use,” then without probable cause, without a FISA warrant, and just by parsing through commercially available data, a government agent can not only sift through weak signals to see emerging patterns and illuminate potential dangers across society, but it can analyze and synthesize information about you specifically—about anyone. This “pattern-of-life” analysis would be used in the interest of continuous, real-time surveillance of foreign adversaries. It could also be used to surveil the US domestic population.

Enforcement has done this kind of analysis for decades. But the data the domestic population has been creating online during the past twenty years is newly ample, often publicly available and frequently razed of individual privacy rights (rights you clicked to waive, usually unable to opt out). There’s little chance that local, state, and federal law enforcement are going to suddenly reverse more than two decades of using that publicly or commercially available data to surveil the population for standard investigative research. It’s legal and highly useful for law enforcement. It’s already relatively normalized. It’s not going to stop. It’s going to go faster.

Further contributing to the momentum of damage to civil liberties is a contradictory facet of the AI community’s condemnation of surveillance. There’s no mechanical difference between the use of AI for mass surveillance of the domestic population and the kind of consumer profiling and algorithmic marketing that’s driving new growth for the US economy. Digital marketing uses AI to parse through commercially and publicly available data in order to powerfully target consumers. It’s seen as one of AI’s key value propositions.

The difference is that in the hands of commercial interests, consumer profiling means you’re the target of powerfully persuasive AI-driven click-paths. If you don’t know how to think well, you’ll end up buying stuff you don’t need. AI executives usually want to convince you to waive your privacy rights, not defend them. Their objective has been to have full access to your private data in order to improve AI, sell you stuff, and create your reliance on their software.

That said, in the hands of the federal government, citizen profiling could mean you go to prison. Whether you lean to the political right or the left, if AI-enabled surveillance is driven by the directives of an extreme ideologue, you could be singled out, zoomed in on, even blacklisted for publicly expressing your views. The unchecked momentum of tech corporations fused with government surveillance is paving the way for OpenNSAi to find you with 1-Click®.

THE US MILITARY OPERATIONS IN VENEZUELA AND IRAN are in part intended to damage China in an ongoing fight for AI dominance and the preservation of liberal democracies. The idea had been to keep Xi’s digital authoritarianism at bay. But at the end of this past February, while Iran had been a short, technical step from having a stockpile of weapons-grade uranium, the United States arrived at being a short, technical step from instituting the kind of digital authoritarianism that characterizes China’s surveillance state.

The AI employee brief warns: “Behavioral scientists and legal scholars have long documented what is sometimes called the ‘panopticon effect’: when people believe they may be observed, they modify their behavior as if they are always being observed, regardless of whether anyone is actually watching.” The Panopticon is an 18th century prison design concocted by Jeremy Bentham, the father of utilitarianism. Prisoners are observed with chilling efficiency from a central tower where each can be seen but none know when they’re the ones being watched. They internalize the rules of the prison, acting as if they’re been watched at all times.

Right now, we’re still free to think critically and act creatively to defend our digital civil liberties. In typical free-market fashion, an industry is rising to counter our disintegrating privacy rights. VPNs block web tracking, Faraday bags keep devices from broadcasting your data, and companies can be hired to scrub you from the internet. But the battle to retain civil liberties is becoming excessive and exhausting. Digital hygiene that had been the craft of criminals is fast becoming everyday practice for the rest of us.

It’s been painful and internally conflicting for me to write in defense of Anthropic. Anthropic is one of the main players wreaking havoc with publishing, media, the arts and humanities. I wouldn’t say the company has been duly punished, but it has been, at least, minimally pinioned for its massive theft of our intellectual property. Unfortunately, winning Bartz v. Anthropic hasn’t changed much for artists and writers. Instead, our work is on its way to improving the speed and acuity of an American panopticon. Even the raging spectacle of today’s news cycle can’t force me to forget what’s at stake. And so here I am to remind you.

If the United States is risking American lives to cross the Rubicon of Hormuz as part of a battle for global AI-dominance against China’s digital authoritarianism, we’d better not find Stars and Stripes Stasi on the others side.

Subscribe now

A NUCLEAR ARMS PACT BETWEEN THE US AND RUSSIA ENDED last week, sparking a cluster of media fears about a new arms race. It struck me as a kind of collective, post-traumatic stress response from a threat long gone. The Cold War ended over three decades ago. Today, we’re in something of a new arms race with China for AI dominance. What’s at risk now is China’s digital authoritarianism gaining favor in the world against weakening western liberal democracies.

For about a decade, a fierce AI race has been underway between the US and China, today’s global “hegemons,” as Mark Carney, Prime Minister of Canada implied they were at Davos this January. That race began with AlphaGo’s advances (Google/DeepMind), which publicly accelerated in 2016. Some called this China’s Sputnik moment. Today, Russia and Europe are relative stragglers in today’s fast-paced AI sprint—and the argument that Big Tech and the US government are making loud and clear is that whichever power integrates AI into its economy quickest and best is expected to retain (US) or gain (China) global economic and ideological dominance.

Subscribe now

In the US, a handful of Big Tech companies operate, train and develop the world’s leading foundational AI products. Those Big Tech companies need data centers to run the AI processing chips that perform inference calls (search queries), train models, and develop new AI products. These data centers need a great deal of electricity to operate. The best AI growth strategy Big Tech has right now is scaling computing power, which just means ‘doing more of what they’re doing now.’ To scale up data centers, they need to scale up electricity, too.

In the fall of 2024, when nuclear energy stocks hit record highs, Clay Sell, CEO of X-energy, a Maryland-based advanced nuclear reactor company, said, “The only constraint on the US remaining the leader in artificial intelligence is power. It’s not land, it’s not chips, it’s power.”

Big Tech needs more power than the US grid has. A lot more.

During the last three years, since ChatGPT went mainstream, spawning several rival models, including China’s DeepSeek R1, Big Tech has become closely aligned with the US government and its national interests to remain the leader in AI.

Right now, the US dominates the global energy market, including nuclear power generation. According to the International Atomic Energy Agency’s count, the US has 94 nuclear power reactors in operation, zero under construction, and 41 have been shut down. China is on a fast-track to catch up with 58 reactors in operation, 32 under construction, and none shut down.

Nuclear energy progress in the US came to a screeching halt in 1979, but it didn’t happen overnight. A long history of shock and fear around nuclear power began with the Manhattan Project and the birth of the atomic age. The bombing of Hiroshima and Nagasaki was a cultural and ethical shock. Quotidian fears of nuclear holocaust enervated us during the Cold War. The geopolitical complexity of enforcing deterrence and non-proliferation, too, were exhausting.

By 1979, nuclear trepidation reached a fever pitch with a partial meltdown at Three Mile Island, which eerily coincided with the release of the movie, The China Syndrome, featuring a similar incident. The news coverage mixed scenes from the movie with commentators’ explanations, blurring fiction and reality. It seared terror into Harrisburg, Pennsylvania, which, in turn, burned through the country. The PBS documentary, “Meltdown at Three Mile Island,” featured the recollection of Victor Stello, an NRC engineer who was in the congregation of a Roman Catholic mass that Sunday. Folks were still in a panic, and the priest offered the entire congregation “general absolution. ‘This is a sacrament reserved when death is imminent,’” Stello said. The event and the media around it scared the nation nearly to death.

But it was nuclear power generation that died. The rules created by the Nuclear Regulatory Commission to follow were so time-consuming and cost-prohibitive, private investors simply sought easier projects that were more likely to profit while they were still alive—and projects they had more confidence wouldn’t kill anyone.

After Three Mile Island in 1979, not a single reactor in the US was approved for construction.

Then, in 1986, the accident at Chernobyl worsened anxieties. That the fall of the Soviet Union came fast on the disaster’s heels likely wasn’t lost on US leaders. The devastation and the USSR’s dishonorable cover-up were the last blows to its collapsing empire. In 2006, Gorbachev wrote, “ The nuclear meltdown at Chernobyl ... even more than my launch of perestroika, was perhaps the main cause of the Soviet Union’s collapse five years later.” If the United States wanted to retain its hard-won position as the primary global leader of a world favoring liberal democracy, it would need to tread very carefully around atomic power.

ENTER VOGTLE UNITS 3 AND 4, ATOMIC ENERGY’S NEXT NIGHTMARE. The new design for two units at the Vogtle plant in Waynesboro, Georgia, the Westinghouse AP-1000, was initially sent to the NRC in the spring of 2002. It took over a decade for the design to get official approval to break ground, which it finally did in 2013. Another ten years would pass before Unit 3 was ready to connect to the grid, with Unit 4 connecting a year later in 2024. The lengthy project went monumentally over budget and, in 2017, Westinghouse went bankrupt.

The NRC regulations kept changing while Vogtle 3 and 4 were trying to make progress, and it didn’t help that the energy sector ecosystem was anemic. Given the nuclear power economy had been stagnant since 1979, there was a dearth of both skilled workers and leadership with experience. There was no supply-chain infrastructure. Everything associated with Vogtle 3 and 4 had to be built from the ground up. They were reinventing the atomic wheel.

Passive safety was the new design’s key selling point. The AP-1000 could cool itself for about three days without human intervention or a power source. With the accidents at Three Mile Island and Chernobyl haunting the project, passive safety was integral for investors and the public. In 2009, mid-stride, or so they thought, the design had to undergo a major overhaul when the NRC suddenly decided to factor in the events of September 11th. Now the AP-1000 had to be able to withstand a direct hit from an airplane.

And then, in 2011, the Fukushima plant was hit by a tsunami.

The NRC’s final regulatory revisions for the two Vogtle AP-1000s arrived later that year.

After such a lengthy debacle, you’d think investors wouldn’t touch nuclear ever again.

Instead, there’s a nuclear renaissance.

ATOMIC ENERGY IS GETTING A BRAND OVERHAUL funded by Big Tech in alignment with the US government. A series of executive orders in May of last year were intended “to usher in an American nuclear renaissance,” wrote Michael Kratsios, White House chief advisor on science and technology. The objective is to reach the goal of expanding nuclear capacity four-fold, from 100GW to 400GW, by 2050. In August of last year, the Department of Energy’s “Reactor Pilot Program” selected eleven advanced nuclear reactor projects to be fast-tracked through NRC regulations in order for at least three small modular reactors to achieve criticality by July 4, 2026.

Now, bright branding campaigns pitch the public anew on nuclear. It’s good for the economy, they say. It’s good for the creation of new jobs. And it’s carbon free!

In a Google branding video that finesses the image of a partnership between the tech firm, the Tennessee Valley Authority, and Kairos Power, TVA’s CEO, Don Moul, tells us with steady confidence, “It’s not just good for Google, it’s good for TVA’s ten million customers—and it’s good for the United States.”

I’m fascinated by the palimpsest of history that flashes in my mind’s eye when I’m watching the bright, happy little videos hawking Google sustainability with Kairos in Tennessee—from the birth of the atomic age, to the legendary role of the city of Oak Ridge for the Manhattan Project, to Kairos Power, the advanced nuclear energy company and its ties to the historic Ernest O. Lawrence lab at Berkeley, along with its partnership with the latest powerhouse of American innovation in science and tech, Google. None of that history, of course, is there.

The radical, rapid change to nuclear’s reputation—from its birth as Death, the Destroyer of Worlds to this carbon-free godsend of sweetness and light—is giving me whiplash. I grew up in the latter stages of the Cold War. As a kid, I knew where the nearest fallout shelters were. I remember Chernobyl and can sing “99 Luftballons” in the original German (a pop-song about the tragic nuclear holocaust in our future). I lived within miles of San Onofre, California, when one of the units at SONGS leaked radioactive contaminants into the air and water. When Fukushima melted down, the massive radioactive cloud crossed the ocean and hit the beach for a week where I ran every morning, breathing in deeply and hard. Edison and the NRC said it was fine, but was it? Is my suspicion about the new boosterism for nuclear just a post-traumatic stress response from an anti-nuclear era gone by?

To find out, next time, I’ll be diving in further into the history and hazards. What fears are overblown and which are legitimate? I’ll ask what the utopian narratives are leaving out that would be good for the energy sector (and all of us) to include. Because accidents happen. Wild weather events hit. Tectonic plates shift. And there’s the pesky reality of human nature to consider. Speed, greed and the tendency to sweep mishaps under the rug or lawyer up could be of concern when we’re talking about a massive expansion of hyperscale data centers powered by little nuclear reactors everywhere owned and operated by private corporate interests.

The power grid serving mid-Atlantic data centers are at capacity and Big Tech wants more.

{Note: This is the first in a series on how US energy is going nuclear.}

The electricity crisis in the Mid-Atlantic, brought on by the large-load demand of data centers, has forced a fresh look at nuclear power. I’ve been fascinated with the developments, especially the wild enthusiasm and ample investment in start-ups competing for future dominance in a new market for small modular reactors or SMRs. All eyes are on the potential of these safer, proliferation-resistant, fuel-efficient innovations, but a legacy of mushroom clouds, meltdowns and radioactive waste have made progress all but impossible.

The data centers straining the Mid-Atlantic grid changed everything. Here’s what’s happening now:

America’s largest power-grid operator is at “the brink of supply crisis,” according to the WSJ. At stake is the supply of electricity to 13 states, plus D.C., and the risk of blackouts for 67 million people from New Jersey to Kentucky. PJM, the electric transmission organization that coordinates that supply, has been challenged to find a solution for the ravenous appetite of data centers, which have jacked electricity prices and threaten to drain the savings of local residents and small businesses.

In a letter to PJM from members of Congress in November of last year, lawmakers expressed “concern with ... rapidly increasing energy demand and skyrocketing electricity costs. Amongst their recommendations was the “prioritization for projects that bring their own [power] to the grid.” PJM’s mid-December capacity auction further hammered home the problem. The auction ended with a continuation of record-high electricity costs and, this time, a failure to acquire the capacity to meet a key reliability standard—one that measures capacity to respond to emergency outages.

While the most recent auction set prices only slightly higher than they’d been this summer, the reverberating shock from July of ‘24, an 800% increase, remains the reality locals face. That increase in cost is stunning, and yet what’s of even more concern for locals and leaders is PJM’s failure to meet the emergency reliability standard. It doesn’t mean blackouts are imminent, but if foundering becomes an increasing trend, communities could pay with their lives in lieu of their wallets. No one wants a repeat of the loss of life caused by extreme winter weather and the resulting grid failure in Texas in February of 2021.

PJM’s COO, Stu Bresler, said, “[T]his auction leaves no doubt that data centers’ demand for electricity continues to far outstrip new supply, and the solution will require concerted action involving PJM, its stakeholders, state and federal partners, and the data center industry itself.”

This past Friday, U.S. Secretary of Energy Chris Wright, Secretary of the Interior Doug Burgum, and governors from across the affected region, discussed a solution. They proposed PJM “hold an emergency auction in which tech companies would be forced to cover the costs of building new power plants” (WSJ). Companies drawing excessive power would “pay for the new generation built on their behalf — whether they show up and use the power or not.... The concept is called BYOG, or ‘bring your own generation’” (Reuters).

The proposal communicates a desire for the continued economic growth propelled by the AI boom. Federal and regional authorities don’t want to block data-centers from drawing on the grid. But they do want Big Tech to foot the bill for excessive power draws. Speaking from the White House on Friday about large tech companies who want to build more data centers, Chris Wright said, “They get it…. Electricity is a cost item, but it’s nowhere near their biggest cost. They’re willing to do what’s necessary,” implying Big Tech is willing to pay (Bloomberg TV).

In the short term, this likely means a temporary life-extension for existing coal plants and building more natural gas plants. But there’s something more important happening that’s emerging from behind the scenes for the broader public to see. What’s far less compelling than a short-term focus on shoring up capacity with fossil fuels is what’s been transforming the energy sector for over 75 years, lost all momentum in the 80s, picked up a little speed about ten years ago, and went into overdrive about two years ago.

In the energy sector, there’s a rising consensus that the only viable, long-term solution to the electricity crisis is to go nuclear.

Here are just few concrete instances you may or may not have heard about that signal this. Congress has been passing legislation with the purpose of incentivizing and accelerating ‘advanced nuclear’ R&D. Advanced nuclear refers to next generation designs with significant improvements from the large, traditional nuclear plants of the past. Adapting to these innovations, the Nuclear Regulatory Commission (NRC) has been revising licensing rules to reflect modern times and modern tech. The NRC has been providing extensions to existing reactors, such as Oconee in South Carolina, while innovative small modular reactors are seeing new, speedier movement from the commission.

Heartened, Big Tech companies have been signing long-term contracts with nuclear energy companies, such as Microsoft signing a power-purchase agreement with Constellation in September of 2024 to restart unit 1 at Three Mile Island. Meta, too, signed with Constellation just this summer for Clinton Clean Energy plant’s nuclear output. And boldly going where no corporate agreement has gone before, Google signed with Kairos in the fall of 2024 for a series of projects focused on an initial deployment of Kairos’s small modular reactors by 2030. And, lastly, just this January, Meta signed a deal with three nuclear power companies, TerraPower, Oklo and Vistra, also for the development of small modular reactors.

In the next installment, I’ll dive further into the legacy of nuclear power so we can understand how far we’ve come with innovations today, see where nuclear energy experts hope to be in the future, and point out what they’re trying to sweep under the rug—or, more accurately, bury. I’ll also be talking about the key driver—AI integration—and what it means for jobs. Let’s go!

The AI boom and the nuclear renaissance is affecting everyone. Please share and like this post to spread it around! Let me know your thoughts and ideas in the comments below. If you have any information or articles you think I should be aware of, tune me in!

Last: if you haven’t seen my video on why those in the Humanities, publishing, media and the arts are being pushed hard to use AI, have a look. If the basic drivers behind AI growth are relatively new to you, you’ll learn a lot in just 14 minutes: Watch it here.

Short cut: Read good books.

The long-winded versions:

1. Know yourself & be yourself.

2. Take care of yourself. Then, help others.

3. Live & work with purpose & meaning.

4. Think both analytically & creatively.

5. Communicate true things well.

6. Face, understand & solve problems.

7. Connect with people & community.

8. Be flexible & love learning.

9. Master basic technology.

10. Understand data literacy & digital literacy.

It makes sense in medicine & science, but even in media, the arts, tv/film & the humanities, we're feeling enormous pressure to integrate AI into our lives & our work.

Here's why. ⬇️

I started this piece as a way to examine the idea that, in the battle for AI dominance, the Chinese have a clear data advantage due to their massive surveillance practices. I found something far more complicated and telling about the future of the distinctive AI ‘entities’ the US and China are likely to develop.

Chinese Social Control & Ideological Alignment

China’s state surveillance apparatus is monitoring over 1.4 billion people. You’d think they’d have a ton of high-quality data to train their AI models on. Seems like they could easily trounce the US based on sheer volume. Later, I’ll go into the advantages their surveillance data does provide, but the driver of what I’d considered a boon for the Chinese is one of the key barriers to collecting high-quality datasets.

As it turns out, surveillance for the purposes of state-enforced censorship and mandated social cohesion makes your data homogenous. That’s the point. In China, no one’s supposed to stand out. You can’t express individual or, much less, critical views. If you do, you’re likely to be disappeared. For example, Jack Ma, founder of Alibaba, mysteriously vanished in the Fall of 2020 after he publicly criticized Chinese economic policy. According to The Guardian, Ma finally reappeared making a call to put political pressure on another Chinese businessman who’d fallen out with the CCP. It’s kind of illegal to think critically about lots of stuff under CCP rule.

China would have trouble training a high-performance reasoning model without data sourced from western democracies. DeepSeek, for example, is suspected of pillaging private data from the EU in defiance of General Data Protection Regulations (GDPR). DeepSeek may have also used data from OpenAI’s model. Public communications more acceptable to use, like Chinese social media (WeChat or Douyin) or state-run news (Xinhua News Agency), lack nuance and critical insight. If you’re a private company looking to build an above-board app that collects and repurposes data, it has to be ideologically aligned with the CCP. Datasets aren’t going to have the rich nuance of freer, more authentic human expression.

News outlets in China, like the People’s Daily or China Central TV, are controlled by the CCP. We’re not talking about data scrapes or licensing deals from The New York Times, The Atlantic, or The Wall Street Journal. The Chinese aren’t (openly) collecting data from up-voted Reddit threads or crowdsourced knowledge stores like Wikipedia to feed public-facing generative AI. Heavily censored, publicly available Chinese data is in lockstep with Xi. AI trained on that data alone would offer factual inaccuracies and lack a diversity of perspective.

Down the line, an AI ‘entity,’ a ‘companion’ or ‘assistant’ derived from that data would be rigidly one-pointed, even militant, in its allegiance to the state.

Digitization of US & World Literature Outpaces Digitization of Chinese Literature & Language

In the US, Project Gutenberg began in 1971. Technologist and futurist, Michael Hart kicked off the digitization of US literature by typing up the Declaration of Independence and distributing it electronically. This was the very first ebook. As scanning technology advanced, hobbyists, academics and librarians, driven by a desire to preserve literature and by the democratizing of access, scanned all manner of public domain texts. In 1996, The Internet Archive was founded, the mission for which is “to provide Universal Access to All Knowledge.”

In 2004, the unintended consequences of this bookish passion hit the book business. The Google Books Project, partnering with coveted libraries (Harvard, Michigan, Stanford, the New York Public Library, Oxford) scanned, it would seem, every book in sight. The Association of American Publishers was horrified. The AAP sued Google in 2005 and in 2012, they settled confidentially. The Author’s Guild, aghast in tandem with the AAP at Google’s rampant scanning, had also sued the tech company in 2005. Ten years later, in 2015, the Author’s Guild suit was quashed in court. Ruling in Google’s favor, the court cited lawful “fair use” and “found that the project provides a public service without violating intellectual property law.”

The events back then created problematic precedents for publishers and authors in the suits being brought against AI companies today. Several have been sued for using millions of books and texts for training AI without purchase or permission. In late June, 2025, Judge William Alsup, U.S. District Court for the Northern District of California, exonerated Anthropic’s use of millions of copyright-protected books in Bartz vs. Anthropic. His ruling states, “The copies used to train specific LLMs were justified as a fair use…. The technology at issue was among the most transformative many of us will see in our lifetimes.” Alsup also ruled that Anthropic’s acquisition and storage of the dataset was unlawful. However, training as ‘transformative fair use’ is a breathtaking blow to the media industry. Just this past week, Anthropic has agreed to negotiate a settlement to compensate authors whose work was stolen. We’ve yet to see how this case will influence rulings to come and how it’ll cascade into the business of media in the future.

Back in 2008, Hathitrust joined the mass digitization effort in collaboration with a number of American universities. Its mission continues to be “to contribute to research, scholarship, and the common good by collaboratively collecting, organizing, preserving, communicating, and sharing the record of human knowledge.” With the English language dominating globally as the default for research and business, Hathitrust’s digitization effort would not only become a data source with the breadth and depth of texts from around the world, it would also help to make English the dominant language for AI trained on this data trove.

The Hathitrust project also gives us a view into how worldwide research in STEM fields arrived in US datasets. With English as the dominant language in which scientists across the world communicate, including physicists, programmers, engineers and mathematicians, English-language books and white papers from across the globe in these fields ended up in datasets training US-based AI models.

The digitization of books and text in the US and in English from worldwide sources translates to a language and knowledge dataset at the foundation of AI that’s primarily English-speaking but otherwise multilingual and richly informed by world history, literatures and research across varied sectors, including business, science, math and coding.

The Chinese only began digitizing their own literature in 1998. Driven by cultural preservation, the Ministry of Culture approved the “China National Digital Library” project that year. The Chinese focus on cultural preservation and the state control of knowledge runs counter to western ideals of democratic access. While Michael Hart, founding Project Gutenberg, and the librarians to follow were driven, too, by the cultural preservation of literature, they were equally motivated by the double-edge sword of the freedom of information, along with a respect for diverse perspectives. China’s focus solely on its own literary and cultural history, the limited scope of acceptable texts written in Chinese, and the complexity of variations in Chinese characters, limit the depth and breadth of knowledge in datasets derived from these scans.

The results of AI built using this textual dataset of language and knowledge would be characterized by a cultural insularity that lacks nuanced global understanding. Rather than describing the Chinese people themselves, it might be truer to say it describes the effect of the Great Firewall on data acceptable to the CCP.

China Lacks Computing Power & Data Centers to Process & Train AI

Even if China had scads of what the west would consider high-quality data, they don’t have powerful enough chips to process it. It won’t be news to you that the US banned China from importing American Nvidia chips, the fastest and most stable processors yet made. Chinese military application of these chips is seen by the US as a national security risk and a threat to American global hegemony. Largely using Huawei’s 910B, which is significantly less powerful than Nvidia’s H100, Chinese Big Tech was gearing up to adopt the next generation of the Chinese-made rival chip, the Huawei’s 910C. The 910C was closer to achieving parity; however, rumor has it the 910c overheats and Huawei has to go back to the drawing board.

Recently, Jensen Huang, CEO of Nvidia, halted production on the lower performing Nvidia H20, a chip designed for non-military Chinese application. But China apparently suspected security risks, like an embedded back door funneling data leaks to the US. So, Chinese companies will either use their own processors, rely on what’s left of their pre-ban Nvidia stockpiles, wait for Nvidia’s next lower-powered chip designed for use in China, the B30A, or they’ll use the black market to smuggle in high-performance Nvidia processors.

Because China doesn’t have scores of high-speed chips, they also haven’t built a ton of chip-dense data centers for AI training and inference (real-time user-query processing). The three big American cloud companies, the US ‘hyperscalers’—Amazon, Google and Microsoft— are building data centers fast. As of this writing, the USA has over 4,000 data centers (and counting) compared to China’s 379, according to Data Center Map. While new facilities in Xinjiang are being planned for running 115,000 black market NVIDIA chips, according to Bloomberg, even such new volume would be dwarfed by US capacity.

Chinese Surveillance Gives China a Multimodal Data Advantage

In the US, the National Security Agency, along with the Department of Homeland Security (through Immigration and Customs Enforcement and Customs and Border Protection) employ heavy use of AI-powered surveillance, data collection and tracking. Notably, ICE and CBP have recently been engaging in highly questionable, if not directly unconstitutional, practices of surveillance, tracking and detaining. It’s beginning to look a lot like the kind of data-enabled authoritarianism usually associated with the CCP.

In the past, American use of surveillance for immigration enforcement have focused on borders and ports of entry. However, during the Biden administration, a conspicuous lack such a focus created a border crisis, including a record number of crossings. Recent dramatic, authoritarian migrant sweeps are, at least partially, driven by attempts to reverse that sudden influx. If rights violations and a slide toward authoritarian rule isn’t stemmed or moderated, the US will, no doubt, serve as a poor model for democratic freedoms. But, in terms of AI development, the surveillance practices and data produced will influence the character of US datasets, tracing into the sort of AI we develop and the kind of enforcement we employ in the future.

Here’s what it might look like:

In China, we see an authoritarian regime characterized by the use of surveillance for social controls. China’s data contains scads of biometric data, for example, facial recognition, including ethnic identity, iris identity, fingerprints and DNA. Amongst other things, biometrics in China are known to be used to monitor the Uyghurs, a Turkic Muslim minority associated with a desire for cultural and religious autonomy. China claims the Uyghur population poses a national security threat as an unacceptable departure from Chinese social standards and a deviation from national cohesion and identity. Uyghurs are heavily surveilled, commonly detained or disappeared, and sent to deradicalization re-education facilities in the Xinjiang Autonomous Zone. There, China defends human rights violations as counterterrorist and anti-separatist measures.

Xinjiang has been compared to an experimental region dedicated to R&D for the development of surveillance technology. Warnings abound for its chilling advancements being exported and disseminated to the broader world. This is not where you want the US to compete for parity.

Chinese Excel at Embodied AI Development Where Its Data Crushes the US

China is also known as the ‘factory of the world.’ Data related to this aspect of Chinese culture is likely to have manufacturing data that describes cost-effective design, iteration and deployment of cheap, mass produced goods, including tech. In combination with surveillance, data informing manufacturing for consumer success includes behavioral information, like individual preferences and consumption habits. Importantly, further derived from its surveillance apparatus, data describes movement (of vehicles and people in relation to stationary and moving obstacles).

Chinese data excels at training AI models with a capacity for designing self-navigating technology, like vehicles, drones and robotics, all equipped with multimodal sensors, like sight, sound, and tactile or haptics. This has given rise to military, social and law enforcement applications, such as drone swarm technology, unmanned armed tanks, and last, but not least, humanoid robots, like the Unitree G1, which demonstrate spin-kicking fighting techniques, violent soda bottle decapping and worrisome walnut crushing, but also a level of flexibility that puts my yoga practice to shame.

Chinese autonomous vehicles from Baidu, for example, are ahead of US companies like Tesla and Waymo. This is largely due to Chinese multimodal data dominance secured through authoritarian surveillance technology. Exports to the States from Chinese tech companies in this area face challenges related to resistance from US buyers suspicious of security risks and the prohibitive cost for China in tariffs.

THE FUTURE: The Distinctive AI ‘Entities’ the US and China are likely to Develop

The different natures of these datasets, tied to cultural or national strengths, stand to train fundamentally distinct artificial intelligences.

In the US, data is characterized by language (defaulting to English), world literatures and history (likely with a bias toward North American and western European cultures). It’s characterized by visual arts and music (all with the same bias). And it’s characterized by mathematics, coding, science, engineering, and behavioral data for business, consumer and, indeed, ideological applications.

American data is better at training AI models with a capacity for reasoning, coding, writing, communicating, composing music, and visual art—including screen-based dramatic arts, such as video of all kinds, from the cinematic special effects causing controversy in the film industry to the irony of AI ‘actors’ performing Tom Stoppard’s Arcadia. (Don’t worry. I made that last one up.) It’s likely to train AI models that excel at maths and solve conundrums in sciences, as has already been done, for example, by cracking the code for protein folding. Americans leads in the underpinnings of frontier model research and advanced chip design for high-speed processing.

So, what kind of epitomizing AI ‘entity’ would develop from these strengths and reflect national character?

Google’s Genie 3 is an AI-powered, interactive world simulator. Derived from Veo3, Google’s text-to-video model, Genie 3 has an uncanny grasp of physics and lighting, offering persuasive interactive simulations. It’s the opening salvo to a new industry using persuasive simulated worlds and interactions for real world application. In these worlds, AI models can train toward embodied applications at high speed and at a fraction of the cost of actual embodied systems. For example, they can be designed to train the intelligence of an autonomous vehicle on millions of rides through a city, with millions of unexpected wrenches thrown in, at low costs and low time-expenditure compared to doing the same in real life. And it can train human users in real-world operations across industries like medicine, science, emergency response and military applications. Surgeries, physics experiments, rescue scenarios, enemy fire.

Commercially, immersive entertainment, education and relationship will likely be key drivers for broad adoption. If you imagine interacting with an AI entity, companion or assistant in this environment, it’s one that’s disembodied or has a form that incorporeal—simulated. It’s less C3P0 (or the Terminator) and more Her (or Hal) on the Holodeck. If you’re connected to it across hardware devices (phone, laptop, tablet), you might perceive it as being everywhere you are in the real world. If you’re immersed in a simulated world, you might perceive it as personified. You might have a sense of fusion with it, since it could feel as if it both is and is creating the environment you’re in. You might also feel you have simulated godlike capacities inside such a simulation.

There’s a great deal to consider in terms of benefits and detractors, but the discomfiting characteristic of a future in which American AI develops an entity of this kind, is the capacity for it (and that which ultimately controls it) to conceal the user’s isolation and enclosure through a simulation of the opposite: total freedom and unity in relationship.

As an example of what kind of AI entity the Chinese, on the otherhand, are likely to evolve: the robot line from Unitree is a good preview. If you imagine broad adoption in the future, it might be a slew of personal embodied AI robots, like C-3PO. But if you extrapolate a future from the G1 demo, this humanoid robot will be less of a nervously expressive translation robot and more of a monosyllabic brute. It might be used for personal defense or, at volume, for national security and defense in the form of an army. Think: a phalanx of multimodal Terminators that move in very unified lockstep. One can’t help but imagine such an army networked with a geometrically formed swarm of Chinese military drones. Embodied AI from China has the potential to create pure, authoritarian robotic AI for export and dissemination across the world.

For now, though, you can buy an R1 for fun! Just US$6k!

In sum, American AI is on track to produce a highly individualized AI entity, one that has the potential for immersive training, learning, education, communication and entertainment. It also might offer the illusion of omnipotence, risking users without the ability to think or do anything practical in the real world. Users might also lack the critical awareness of their powerlessness in physical reality and misinterpret factual imprisonment and isolation as total freedom and social satisfaction. China, on the other hand, is on track to supply the world with disposable robotics that have the potential for performing rote tasks for personal use, like house/intern/assistant work that require movement and interaction in real-world environments. Cheap Chinese robotics might also be employed in manufacturing, warehousing, driving and delivery. Designed from authoritarian surveillance, however, Chinese embodied AI might also provide cost-effective combat robotics for export and use in global conflicts. Regimes with an interest in authoritarian dominance over costly populations of high-maintenance humans with pesky dissenting views might be eager to make a deal.

LAST WEEK, SUNDAR PICHAI announced that Google is making its AI tools FREE for college students in the US. Normally, this would be Marketing 101. It’s a strategy to target young people during their formative years when norms are easy to shape. That’s been acceptable in the US since the 50s. Create early habits that cement into a lifetime of loyalty to a consumer product, like a brand of designer jeans, hip sneakers, a high-end cup of coffee. But when the consumer product is AI and what’s being shaped is thinking, planning and making decisions, it's a little bit different.

A quick decryption of this offer: Unlimited image uploads are a boon for Google, not for kids. The company wants access to their photos, along with access to the files and data they create using Google’s AI tools alongside their already free productivity apps. Google wants students to adopt these tools and cement lifelong habits by using Google products in college—i.e. Gmail, Google Calendar, Google Docs, Google Drive, Google Sheets, etc.

As I’ve written before, Google’s objective is for younger generations to adopt a “Universal AI Assistant,” an always-on AI companion. All the AI companies are vying for the lucrative privilege of being your choice of an AI companion that’s synchronized across all your devices and all of your productivity software. They want you to choose their version of an AI experience that becomes an everyday relationship with an AI entity for life. There’s a lot more at stake, though, than this.

First, I want to draw attention to the GPT-5 drop last week. OpenAI is currently in a weaker position than Google. Altman casually described his own vision of the AI assistant in a recent interview with Cleo Abram of Huge *if true. Airing last week on the same day GPT-5 was released, Abram asked Altman how her “relationship” with GPT-4 will change with GPT-5. Altman described a deepening, a relationship with GPT-5 personified as a companion. “You’ll connect it to your calendar and your Gmail… and it’ll say, ‘Hey, I noticed this thing. Do you want me to do this thing for you? Over time, it’ll start to feel way more proactive…. It’ll just feel like … it’s this entity that is this companion with you throughout your day.”

(The full interview is here.)

You’re hearing the word “proactive” more often in Big Tech conversations and AI marketing because it sounds better than other words and phrases that could be used in its place. Proactive has connotations of being positive, helpful and dynamic. But there are other ways of saying the same thing. Think about an artificially intelligent entity that’s all up in your emails, has access to all your private files, your documents, your spreadsheets, your photos, your texts, your mic, your camera and let’s say this AI assistant is described as being able to think for itself. This AI is autonomous, self-governing. This AI is resourceful, ambitious, even vigorously bold! This AI takes the initiative! You might think twice about connecting that AI “companion” to all your files, your phone, your laptop. And you might rethink the generosity of Google giving something like that to your kids.

Proactive, though, has a nice ring to it. Proactive doesn’t sound the alarm.

Later in the Abram interview, Altman mentions how OpenAI might create some kind of “consumer device” that could sit on the table and listen to the way she was executing the interview. The digital companion might offer Abram a post-mortem on her technique. Altman says ‘consumer device,’ as if so new-fangled a device that could listen in on their conversation didn’t already exist. It’s called a “smartphone.”

Google has inarguable dominance in the smartphone market, not the hardware, but with Android, their globally dominant mobile operating system. But what Altman wants to mean is that his smartphone will be on the table, or his ‘smartphone killer,’ the device he’s partnering with Jony Ives to design. Ives is the storied designer who oversaw the revolutionary consumer product innovations of the iPod, iPhone and iPad. He’s designing something with OpenAI to compete with Google’s mobile software and, secondarily, Apple’s mobile software and hardware. Currently, OpenAI has to ride piggyback on others’ proprietary hardware and software to deploy their version of an AI companion—e.g. in order to be the companion Altman describes, GPT-5 has to connect to Google’s Gmail app or Apple Mail, to Google’s Calendar or Apple’s and more.

Could a new Ives-designed smart device that somehow replaced the smartphone (by being a phone + whatever Ives is innovating into it) enable OpenAI to compete for dominance in the AI companion market?

OpenAI’s key target market, young users, are already habituated to using Android phones, Gmail, Google Calendar, Docs, Sheets, Drive, etc. Will they connect their GPT-5 account to Google’s or Apple’s apps when they could, instead, subscribe to GeminiPro for free and — boop — everything is automatically synchronized? Why would other companies even allow OpenAI to plunder their data booty in the way Altman suggests? All of this adds to the reasons why Google’s magnanimous gift to students is, in reality, an aggressive strategy to target young people with Google’s version of what they hope to become your kids’ AI companion for life. What they want is to get all up in your kids’ digital business.

(Don’t worry. I’m sure your kids don’t have anything to hide. Surely, they’ve never done and will never do anything questionable they might have evidence of on those handy-dandy smartphones they’re glued to. Your kids. Pure as the driven snow. Chips off the old block.)

US Privacy Ideals vs. China’s Surveillance State in the Race for AGI

AI companies are also dropping an even larger problem than that in your lap. Altman and Pichai’s vision requires a major paradigm shift to American privacy ideals. They’re asking for wildly deep access, surveillance, analysis, processing and personality profiling. Companies like Google and OpenAI would like to dress it up and put some lipstick on it, but without the pretty lingo, without the big smiles and enthusiastic assertions of how excited they are to bring you this revolutionary advancement, it looks uncannily similar to the surveillance state they’ve got going on in China.

Google, and any AI company, including OpenAI, along with Microsoft (which uses GPT models to drive its AI companion, Copilot) needs an unprecedented level of access to your private files and data in order to deliver what they want to ship—which is a product that has full access to your data. With it, what kind of persuasive power might they have over you, knowing what makes you tick? What kind of leverage might they gain over you if they’re inside your devices, looking at and processing your data, listening with your mic and watching with your camera 24/7? It’s all about you, alright. It’s about what you’re doing, writing, creating, thinking, listening to, looking at all the time. They’re selling this to you. (Well, they’re giving it to your kids.) China got it all wrong. The PRC could have gotten their citizens to pay to be surveilled.

Big Tech is aggressively trying to change the privacy paradigm we have in the States. They want you and your kids to think your data is equally theirs and that it’s their right, even their duty to use your data for their purposes. They’ll tell you it’s a foreign policy issue. If this goes on, soon, it might be considered un-American to keep your data private. After all, imagine the stores of AI training data they’ll have to improve their models. Think of the training experience they’ll benefit from if their models experiment with acting autonomously using your data and interactions. That’s a pretty good leg up in the ultimate endgame: achieving AGI, artificial general intelligence, before China does.

China’s advantage? It already has access to citizens’ private data to use to train and improve its AI models.

A good way to create such a massive shift in the United States would be to habituate young people to waiving their legal rights to privacy during their formative years. It’s an alarmingly authoritarian ideological shift today, but if the strategy works, your kids’ll become adults in a nation where corporate and government access to everyone’s private files is normal. They won't know a different way of life.

AI-Use Risks Cognitive Dependency, Especially in Young Minds

ChatGPT hit the mainstream less than three years ago. Already, kids' reliance on AI products for cognitive tasks has been shown to have clear potential to diminish their ability to think well, to think freely. Studies reveal AI reliance degrades capacity for critical thought and independent reasoning. You may have come across this study from Microsoft and Carnegie Mellon. It finds the use of AI in knowledge work “can inhibit critical engagement with work and potentially lead to long-term overreliance on the tool and diminished skill for independent problem-solving.”

Research from Michael Gerlich at the Center for Strategic Corporate Foresight and Sustainability, SBS Swiss Business School, recently found “frequent use of AI tools” for cognitive-offloading can “diminish users’ engagement in deep, reflective thinking processes,” especially for younger users. “Younger participants who exhibited higher dependence on AI tools scored lower in critical thinking compared to their older counterparts.” Gerlich calls for “educational interventions that promote critical engagement with AI technologies.”

In the field of artificial intelligence, there’s a lot at stake. Your individual privacy, cognitive health and that of your kids are not at the top of of AI companies’ list of priorities. Right now, it’s not that important to them if you and your kids can think critically on your own. In this moment, it’s all on you and your kids to police privacy boundaries and counter the degrading effects of AI-use on cognition.

Who Would Design Such an Invasive, Damaging Tool?

Considering the jarring privacy issues and the growing research showing negative effects on critical thought, you migh wonder who would integrate this tecnology so fully and blindly into their lives? Who would even be a part of, much less in charge of designing such a massively invasive, mind-damaging, independence-crippling thing?

Well, a Nobel prize winner.

Demis Hassabis leads Google DeepMind. He’s one of the brilliant researchers who won 2024's Nobel in Chemistry by using an AI model for protein structure prediction. I recently listened to Hassabis in conversation with Lex Fridman. It was one of the most inspiring of Fridman’s interviews. Hassabis’s ideas on prediction and stability in evolved systems was illuminating. His clear-eyed counsel to design for several years in the future was wise, given the speed of transformations today. I share his obsession with understanding the nature of reality and his utter disbelief that most folks aren’t interested by the question: what IS this? (It’s entirely worth a full watch or a listen, 2.5 hrs).

In spring this year, at Google I/O, Google’s developer’s conference, the gifted Hassabis introduced Google's vision for a Universal AI Assistant. He said, it's “a new kind of AI, one that's helpful in your everyday life, [one] that's intelligent and understands the context you're in, and that can plan and take action on your behalf across any device. This is our ultimate vision for the Gemini app to transform it into a universal AI assistant.”

That such an assistant “can plan and take action on your behalf across any device” is lovely marketing copy. “Personal context.” Such a pretty phrase. You’re at the center. It’s all about you. In practical reality, it means the AI assistant has deep access to all your stuff and can make decisions and do things with it all on its very own.

Hassabis is an incredibly insightful scientist, an ingeniously creative thinker and innovator. His contribution to human progress has already offered a bright beacon for us all. He's also the mind behind Google’s most powerful, most invasive and potentially damaging consumer-level AI product to date. Can his wisdom, leadership, his capacity for innovation at Google DeepMind influence AI interaction that's designed for the best possible world?

Championing Critical Thought in Education & Innovative Design

Gerlich’s study, mentioned above, concludes with a call for teaching strategies like active learning and exercises in critical thought to counter the degrading effects of AI-use in young people. I agree and will be writing more about this in future posts. But we also need AI companies to responsibly design interactions that champion free-thought for young users. The responsibility of protecting young minds from AI degradation isn’t that of teachers and the education system alone.

In order to realize the techno-optimist dream of birthing a new generation of superhuman-thinkers with AI, it’s in the interest of Big Tech to take on the responsibility of designing interactions that don’t hobble but champion free thought and agency. And they need to find a better way to get limitless training data to improve their models and achieve AGI than a marketing strategy designed to get you to pay to give up your rights to privacy.

We aren’t the PRC. US companies at the AI frontier need to innovate and build tools that effectively empower free, critical and creative thought. Let’s compete with China, not become them.