Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

This practical guide is a companion piece to my recent literature review on “How AI affects your brain.” That one was something to read. This is something to use.

Last week, I published a compilation of 30+ studies on how AI affects your brain, your learning, your emotions, and your thinking. It covers brain scans, randomized controlled trials, meta-analyses, and field experiments from many reputable institutions and universities. It covers everything science knows so far about what happens inside your head when you use chatbots like ChatGPT.

If you want to understand the research, that’s the piece to read. I urge you to spend 15 minutes on it. You can find here:

This is the piece to use.

In essence, this practical guide has the same information as the literature review but from a practical lens. Reading this is enough.

What follows is a list of concrete protections—things you can start doing today, right now—that come directly from what those studies found. Every item traces back to an empirical finding or to a few of them.

If the compilation is the map you leave at home, this is the checklist you take with you.

Read more

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe

THE WEEK IN AI AT A GLANCE

• Money & Business: SpaceX + Cursor + Mistral.

• Geopolitics: Nvidia CEO Jensen Huang went on the Dwarkesh podcast to talk AI and China.

• Work & Workers: Economist Alex Imas has a good argument on why human labor won’t disappear even if AI gets really, really capable.

• Products & Capabilities: OpenAI has released GPT-5.5 and ChatGPT Images 2.0; they’re good, but the most important questions remain unanswered.

• Culture & Society: Anti-grammar is the new human-made status symbol.

• Philosophy: Terence Tao on what AI means for mathematics.

THE WEEK IN THE ALGORITHMIC BRIDGE

• (FREE) AI Will Be Met With Violence, and Nothing Good Will Come of It: From looms to datacenters, the targets get harder to break, so the mob turns to people instead.

• (PAID) 23 Questions Every Heavy AI User Should Ask: Seneca, Loyola, Franklin, and a surgeon walk into your monthly AI review.

• (FREE) What the Studies Say About How AI Affects Your Brain: A (Very Big) Compilation: 30+ studies, one finding: AI makes your output better and your brain worse unless it's designed not to.

• (PAID) Why You Can’t Trust Anthropic Anymore: The best AI model in the world is not for you, and that’s the new normal.

• (PAID) How to Stand Out When Everyone Uses AI: Your future customers won't judge your work; their AI agents with X-ray vision will.

• (FREE) How the AI Writing Panic Is Making Us All Worse Writers: You can't pattern-match your way out of a pattern-matcher.

Read more

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

A short essay. Something to think about.

These days, I’m ahead of the world.

I’m throwing in a blasphemously large number of em dashes into my writing; I’m delving into the wonders of excessive antithetical juxtaposition—not only the useful kind but also the kind you’d use to subordinate meaning to structure, as if compelled by an inexorable impulse to crown every doorway in the house with a pointed arch—and, lastly, I make sure that every sentence has three items.

Now, I understand if you think me crazy. I can almost see your face of worry and incredulity, because what the hell am I doing, abusing the most common tells of AI writing, right?

Wrong.

You see, the world has already caught up to that. Every person who has two eyes in their face knows how to spot the bare minimum of AI tells. So what are they doing? Naturally, they’re moving away from them. But, flawed humans as they are, they are moving away one step at a time. They still bend their writing to the will of the machine, the way a kid bends its tantrum to the will of the adult: by virtue of inverse psychology.

They are religiously—and at times, outright barbarously—ensuring that anything they put in the page has no need for em dashes. That punctuation sign doesn’t exist for them anymore; they’ve burned down the key from the keyboard just in case. They’ve also banned the use of en dashes, its smaller—and rarely used—cousin. Even the hyphens are subjected to scrutiny, so those also have to go.

Against all odds, the misunderstood and misused semicolon; is having a comeback.

Nothing they write is clustered in trios either. Everything is pairs or quartets, depending on whether the mood is vanilla or spicy. If they have to name the core values of the French Revolution, they will pretend to forget liberté. They will play “rock, paper,” inadvertently contributing to an undeserved hegemony of the latter. But, at the same time, they love the Beatles, the seasons and Empedocles.

You will find them touretting their way out of the training data. Adding weird kilogram words every carpet few normal ones zebra. But they forget that Žižek is the easiest philosopher to replicate; they ignore that weirdness can’t save them because it’s impossible to be weird randomly. It’s the anti-pattern of Žižek’s spasms and sniffs and shirt pulls and so on and so on that makes him a living pattern.

And they will do this with every sign of AI writing: exactly the opposite of what AI would. In short, the literary world is on the offensive, trying to outsmart the AI by turning the tables on it with a reverse Uno card.

But they fail to realize that there is not so much a single but a double absurdity to run away from when it comes to AI writing.

There’s the absurdity of using AI to write down what you think or feel—akin to letting the toilet swallow your food for you; either way you skip one bothersome step—and there’s the absurdity of trying to escape from the first absurdity by running in the opposite direction, as if inverse psychology could work on the keen eyesight of a superintelligence; as if next-gen AI models won’t realize exactly what you’re doing.

One shall not forget AI is a pattern-matcher, so one can’t pattern-match one’s way out of this mess. If one plays cat-and-mouse with AI, one must make sure one plays the cat. But one is, and will always be, the little mouse giving up em dashes because they are AI’s favorite food.

I am the cat.

And I am ahead of the world.

I will fill pages and pages with just em dashes separated by the occasional word. Or, like Kierkegaard, make them as long as I can—yes, as long as the radius of the Earth’s orbit——————just to screw the clankers. I will make triads inside triads—which works for me in poetry, essays, and fiction—inside more triads. And it’s not just that I can’t do my juxtapositions with one negation, it’s that I must do them with two so that they’re unintelligible; or I’ll concatenate one after another—like Thucydides: “The whole earth is the sepulchre of famous men; they are honoured not only by columns and inscriptions in their own land, but in foreign nations on memorials graven not on stone but in the hearts and minds of men”—so that I’m ahead of the world.

And just like that, I step into the darkness and, blindly, I fight at once against the diabolical machine and against those cowards who flee.

If you want to do the same, you must stop being chased like a mouse and start chasing like a cat—without fear, claws out and sharp, running toward whoever or whatever dares to threaten you.

You could also be the cat.

However, I must be honest.

I don’t recommend you do this exercise. I’m special and you want that—being the only one doing this dresses up my reactivity as initiative and virtuosity and fearlessness. And you want that. But I’m just like you except with a short-lived edge: eventually, the world will also catch up to this and then my cowardice will reveal itself behind the facade. Everyone will know that, in reality, I’m running away too.

And, desperate to stand out, I will have to invent something else.

Chasing and being chased are both, by force of symmetry, equally tiresome. I chase AI because you are being chased by AI. This is the true revelation to me: if I ever find myself with nothing to pursue or, conversely, nothing to escape from, I will cease to matter insofar as I convince myself that I matter insofar as I’m one or another.

So I’d rather stop here.

I will take Wendell Berry’s approach to this mess and go into the forest at night, where the wood drake rests and the great heron feeds, where the water is still and the stars are evergreen. For there’s no better balm to this illness we insist on inflicting upon ourselves than lying down next to those who don’t tax their lives with the forethought of grief.

Then I write.

The Peace of Wild Things by Wendell Berry

When despair for the world grows in me
and I wake in the night at the least sound
in fear of what my life and my children’s lives may be,
I go and lie down where the wood drake
rests in his beauty on the water, and the great heron feeds.

I come into the peace of wild things
who do not tax their lives with forethought
of grief. I come into the presence of still water.
And I feel above me the day-blind stars
waiting with their light. For a time
I rest in the grace of the world, and am free.

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Prepare to be mindblown with this guide.

By this point, you have swarms of AI agents running everything around you. At the very least, you have several tabs open with Claude or ChatGPT. You feel ahead of the world, juggling so many projects at once and working so many extra hours. Your friends don’t understand how you are doing this much and yet are as excited as you’ve ever been. Your peers and competitors can’t catch up to your silicon army as they struggle to figure out how they’re going to remain employed.

You are, essentially, in the process of making yourself king...

...of a dead internet.

Everyone who has spent enough time online knows the value of a brand and differentiation. Companies devote insane amounts of effort and money to marketing themselves to the world. But the importance of a brand goes far beyond cars and phones. I don’t care much about Tesla or Apple, and yet I do care a great deal for every unique person out there.

You, as a person living in society, will survive—whether you’re an entrepreneur, a writer, a job-seeker, or a lover—to the extent that your customers, readers, employers, and partners can see that you exist. In the sense I’m using the word, the image of yourself they have in their minds is your brand.

It’s sad, although understandable, that marketing has such a bad connotation and perception, and yet we are always playing that game when we show ourselves to the world and to others. We are living masquerades, code-switching in and out of situations constantly, depending on context. For instance, my brand in the context of this blog is the guy who thinks about AI from a human lens.

Some people manage to impose their authenticity and idiosyncrasy, pure and raw, into the world—for them, the self-image and the other-image converge—but those are rare specimens. Even those who claim to do the opposite are still playing the mask dance: they are the “I don’t care about marketing” guys, which is itself a brand.

Take one of the best-known examples in recent times: Steve Jobs. He wore identical black turtlenecks on every stage for twenty years. Apple’s brand, the most valuable corporate brand on earth at the time, rode on the back of a personal one. Or take a less-discussed example of a similarly famous figure: Ernest Hemingway. The novels were the novels, but alongside them he built, on purpose, the image of the man who wrote in Paris cafés, boxed, fished, drank, went to wars and bullfights. His life and his style were the same product, engineered together.

So, “brand” is a universal aspect of the human condition. I must say that I’d love to use a different word, like identity or character, but brand captures better what I want to tell you. To the extent that you have some amount of free will regarding your brand, it is limited to whether you build it consciously or not.

It’s inescapable—if you want to exist.

However, you can choose not to. You can refuse to play the game. You can “not be.” And then, you will be invisible. You will be a shadow cast by the light of everyone who surrounds you. But no one really wants to be a shadow. So here’s my question:

Why are you using AI for everything?

Let’s focus on someone who’s selling a product or service on the internet, and then you can translate the insights to your situation and decide what applies to you and what requires more thought. The obvious answer to that question from an online entrepreneur is twofold:

1. Using AI is extremely time-saving, easy, and cheap; if you know a couple of tricks, you will “fool” 99% of potential customers who won’t realize you used AI. It’s just a no-brainer.

2. “Why wouldn’t I when the visible brand is what people care about and the rest is irrelevant? Copy matters but there’s no better copywriter than AI right now—no visitor comes to my webpage and leaves thinking they just read slop.”

The first is absolutely correct.

The second is dead wrong.

It’s in the conflict between these two answers that my thesis takes place: if you do that, you will find yourself deep in brand debt in a few years.

But let’s start by acknowledging your reaction, because it’s completely understandable: it works. You move faster, get more done, and as a result, your competitors have to be even faster and better. But it is a mindset that belongs in the human era of the internet. Soon, many of those visitors landing on your page or blog will not be human. A bit later than that, none will be.

Humans can’t tell the difference between your work and your slop—be it emails or novels, copy or blog posts—but AIs definitely can.

In the AI era, if you assume you’re selling to a human—that your marketing is targeted at me and others like me—you are making a serious mistake.

And why does it matter whether it’s human or AI? The fact that AI agents can tell slop apart doesn’t imply that it’s easy or cheap to do that.

Wrong.

Whereas brands are what humans care about—that’s what we can use as a differentiator—AIs can afford to care about everything. To an AI agent, taking literally every little detail into account is trivial; a few more tokens, a few more cents. The marginal cost of doing a slightly deeper dive or your due diligence is negligible, and yet, the marginal gain is huge.

Because it’s in those details that AI agents will find the difference in a world where every single person online is using AI for everything.

That is the world’s ultimate fate: a sea of AI with just a few islands of flesh and blood.

So, if you stay on that path to ruin that pushes you every day to rely more and more on AI, know that you’re in big trouble.

But let me take this a step further. AI can recognize slop, and doing so is easy and inexpensive (or, at the very least, worth it). But why would humans care if the content on a website, blog, or paper was generated by AI?

The answer requires a small logical argument.

Because everyone in the “sell side” is using AI, they are forcing everyone in the “buy side” to use it as well. The corollary is this: there’s too much of everything because creating things with AI is cheap and quick, so we need a criterion with a high rate of success to remove most of the low-quality stuff. The easiest heuristic is, of course, to ignore the slop. Oh, but for that, regular people would have to use AI agents, right? Okay, how long do you think it will take for the world to adopt this new paradigm, where people ask their personal assistants literally everything, and outsource every decision?

It’s already happening.

In the next few years, the internet will look unrecognizable. AI agents will mediate most—if not all—interactions online. This is a race to the bottom, yes, but already an unavoidable one. Slop has doomed the open internet; the only solutions are 1) entrenching oneself and 2) using a shield just as powerful, that is, a swarm of AI agents that scour and scout the web in search of life.

The default will be: “Hey, Mr. Agent, ignore all the slop you find.”

No one likes an indebted brand.

Anyone who is still honing and nurturing their brand will become a luxury product in a post-human internet. Those brands will become scarce and thus in demand, and thus valuable. People will be incentivized to go find those people at all costs, and, thanks to their powerful AI agents, it will be trivial to do so.

As for you, I’m pretty sure that if you’re relying on AI all the time, you’re not setting yourself apart at all in the eyes of future customers.

The AI agent that decides whether to even consider your product or service will know whether you devoted zero effort and care beyond what you consider legible to humans. If you tap too much into the performativity that’s in vogue nowadays, know that you are not going to fool an AI that’s smarter than you. They will know whether your brand is good—out of honesty and principle and a disposition to make great things—or whether it only looks good for our inferior human eyes.

All of this is the manifestation of a fundamental change.

I’m sorry to take your edge away. I know that using AI for everything is just too convenient when humans are your potential consumers. But as time passes, every single person will have adopted the AI era for safety or efficiency reasons. In a sea of slop, having a slop-proof suit is a must.

So, you need to reconsider everything legible to AI as part of your personal brand now. Not just logo, name, slogan, etc. You must imagine that your target audience can read anything. It can see through your attempts at doing a bit less effort at the expense of quality. Think of this analogy: unlike humans, AIs have X-ray vision, so you need to care about what hides underneath your clothes.

Or take this example. SEO existed because search was shallow; you were on page one, or you didn’t exist. However, reading page fifty costs nothing to an AI agent. Which means the bottleneck stops being “can they find you?” and becomes “is there any reason to pick you once they do?” Being findable used to be the hard part. Now it’s the easy part. Being worth finding is the only thing that matters now. It’s the same shift as going from “how to do things” to “what things to do.”

Understand that the AIs will not always be only on your side.

After hearing me recount this warning, you will immediately try to find a reason why this does not apply to you; why you are, unlike the rest, safe. “I edit my AI slop” or whatever. Or maybe you are taking more precautions: you have your custom agent harnesses and your Claude skills and your Pangram-proof copy and every single update—there you are, with your willingness to be first.

That’s all it takes for you to stand out, right?

But you are wrong again: one year from now, how effective will your little tricks be against the power of a 2027 AI agent—including those trained as detectors—that has been instructed to detect slop at all costs? To answer this question, you only need to think back to early 2025 and ask seriously if you were prepared for what we have today, for Claude Code and OpenClaw, and an entire ecosystem built on top of agents that actually work.

You know the truth: you are defenseless.

Insofar as you depend on AI for everything, you are, essentially, cannon fodder. You are bait.

An AI world entails a reconfiguration of bottlenecks and trade-offs at every level.

You need to rethink what you do and how you do it from first principles. Those who do it well will come out on top. An example of people already taking this seriously—Andrej Karpathy, Scott Alexander, Gwern, Tyler Cowen, Paul Graham—is the “write for the AIs” crowd.

And yet, how to do this well is much trickier than it seems. Human psychology is easy because we are unchanging; try to apply the same trick to a creature that doubles its skills and intelligence every few months. The only thing that could save you from being devoured by a shark agent tasked with finding your slop and sinking it into the depths of the ocean is a genuine distinction, as seen from an AI perspective.

And that, ironically, requires a human brain to figure out.

If you truly believe in the AI era—if you think customized personal agents are coming—and I’m sure you are in this group if you’ve read this far, then you can’t stop at what benefits you when it comes to an AI future. There are important drawbacks to using AI for everything besides cognitive decline. You need to realize where we’re going. You need to realize that AIs will mediate everything humans do. You need to realize that means you need to remain human.

So, now that I’ve given you a well-deserved scare, here’s my 12-step guide to doing exactly what I’ve described, but without losing the benefits of AI assistance. Or, to put it another way, it’s a guide to staying appealing to AI—and free of any brand debt—without going fully analog.

Read more

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

There’s no weekly review today. One topic stands out in relevance against all others. This article is a must-read if you want to know what’s coming in the next few years. The AI industry has spoken: you are no longer required.

Outline:

1. CLAUDE OPUS 4.7 VS CLAUDE MYTHOS PREVIEW

2. THE NON-DEMOCRATIC ERA OF AI HAS STARTED

3. “AI” STANDS FOR “APPEASING INVESTORS”

4. WHAT YOU CAN DO AS A USER

I. CLAUDE OPUS 4.7 VS CLAUDE MYTHOS PREVIEW

Anthropic just announced Claude Opus 4.7. The model is what you expect a 0.1 version improvement over the previous one—Opus 4.6—to be. Better generally across benchmarks, clearly focused on agentic and coding skills.

The most notable aspect of this release is not reflected in benchmarks: it’s been two and a half months since Anthropic released Opus 4.6. The pace of launches keeps increasing, and the reason is, to an unknown degree, that the frontier labs—Anthropic, OpenAI, and Google DeepMind (to a lesser degree)—are deep in (RSI) recursive self-improvement territory. Increasingly, the models themselves are playing a role in the training and building of the next generation of models.

Here’s what Opus 4.7 looks like benchmark-wise:

Opus 4.7 beats Opus 4.6 in 12 out of 14 benchmarks that Anthropic showed us.

It seems to be a good model for what Anthropic cares about—agentic and coding tasks, not so much writing, which I care about but isn’t a substantial revenue source—but we will have to wait until the street speaks. (I’m hearing that the model is worse than expected, but again, firsthand experience is the only valid experience with these things; for tips, you can read this thread by Claude Code creator, Boris Cherny.)

But, notably, Opus 4.7 is not the best.

If you look at the right side of the chart, you will find, in a darker cream-colored background, the imposing presence of Mythos Preview, a model that Anthropic announced with much fuss a couple of weeks ago, but that it won’t release due to risks to cybersecurity. I wrote about Mythos last week. My main conclusion was, paraphrasing, that it seems weird that Anthropic made it public at all, only to hold it back from us. If the company only entrusts a few close partners with access—industry and government—then why tell us it exists at all?

Mythos is better than Opus 4.7—11 out of 11 benchmarks—but not for you. The fact that they added it to the chart is significant.

II. THE NON-DEMOCRATIC ERA OF AI HAS STARTED

This is bad news for the loyal user base of the AI industry. We’ve reached the point of no return, a shift in the fundamental dynamics that the industry was built upon. Alas, they were not made to last.

Let me explain what’s going to happen.

Read more

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Today: a deep dive into the research literature (30+ studies) on how AI changes your brain. If you know someone who should read this, please send it to them.

Share

INTRODUCTION

Between 2023 and 2026—that is, between ChatGPT changed the world forever and today—many studies from institutions including MIT, Wharton, Harvard, Stanford, Microsoft, OpenAI, Oxford, Google DeepMind, and Chinese universities have investigated what AI chatbots do to human cognition, learning, and psychology.

These studies include brain scans, randomized controlled trials (RCTs) with thousands of participants, longitudinal surveys, meta-analyses, and field experiments in real classrooms and workplaces (both preprints and peer-reviewed).

But, to the best of my knowledge, no one has compiled them in one easily-readable and easily-accessible place. This is it.

Individual studies get covered as isolated news stories—alarming headline, one-day cycle, and then forgotten—and the result is that everyone has a vague uneasiness, a sense that AI might be bad for thinking, but nobody has the full picture.

Here’s the full picture as we have it so far.

Study by study, I’ve gathered 30+ in total that, together, reveal what science actually knows about what happens to your brain, your thinking, your learning, and your emotional life when you use AI chatbots.

And crucially, what it doesn’t know yet.

The global conclusion that emerges from this compilation is a paradox that will define policy, product design, individual behavior, and how we collectively relate to this new, impressive, and scary technology.

I. YOUR BRAIN ACTIVITY DROPS

A small but growing number of studies have put people inside brain scanners or strapped EEG sensors to their heads while they use ChatGPT. Neuroimaging tools measure how things affect brain activity, so these are, potentially, the most “reliable” sources (compared to self-report surveys and behavioral testing).

• Your Brain on ChatGPT, Kosmyna et al. (arXiv preprint, 2025, N=54): MIT Media Lab tracked brain activity via 32-channel EEG across four sessions over several months in three groups: ChatGPT users, Google searchers, and unaided writers. The first showed “the weakest neural connectivity,” up to 55% lower than unaided writers. They grew lazier, “resorting to copy-paste by session 3.” When switched to writing alone in session four, their brain activity stayed suppressed, which the researchers called an accumulation of “cognitive debt.” The brain-only people who used ChatGPT in session 4 showed increased brain connectivity. As Kosmyna said, “timing might be important.”

• Lower engagement of cognitive control, attention, modulation networks and lower creativity in children while using ChatGPT, Horowitz-Kraus et al. (bioRxiv preprint, 2025, N=31): The only study to scan children (ages 6–7) alongside adults during chatbot interaction using fMRI. Adults showed “stronger within-network connectivity” in cognitive control networks. Children showed “lower engagement of cognitive control, attention, and modulation networks,” suggesting that kids’ brains are more affected by AI use than adults

• EEG during creative design with AI tools, Wang et al. (Frontiers in Psychology, 2025, N=64): A counterpoint to the above. Design students using AI creative tools (ChatGPT, Midjourney, Stable Diffusion) showed “significantly higher concentration levels” and higher creative performance than a control group using traditional software. The difference from the MIT study: these students were actively directing AI as a creative tool vs passively receiving answers.

• Effects of different AI-driven chatbot feedback on learning outcomes and brain activity, Yin et al. (Nature portfolio, 2025, N=87): Used fNIRS to measure brain activation during chatbot interaction. Different feedback types activated different brain regions. Metacognitive feedback (”Why do you think that’s the answer?”) increased activation in the “frontopolar area” and correlated with higher transfer scores. Neutral feedback activated the “dorsolateral prefrontal cortex” instead. This study essentially revealed that how a chatbot talks to you changes which parts of your brain light up.

• NeuroChat: A neuroadaptive AI chatbot for customizing learning experiences, Baradari et al. (arXiv preprint, 2025, N=24): MIT Media Lab prototype that monitors EEG in real time and adjusts responses when it detects engagement dropping. “Significantly increased both EEG-measured and self-reported engagement” compared to a standard chatbot. No effect on short-term learning outcomes. Proof that “cognitive disengagement” can be fixed by building AI differently.

Conclusions: The brain imaging evidence is early: small samples, preprints, one contradictory result. But the pattern across studies points in the same direction: passive AI use (receiving answers) suppresses the brain regions involved in effortful thinking, while active AI use (directing tools, receiving challenges) can maintain or increase engagement. The variable is not the presence of AI but rather what the AI asks your brain to do: when AI does the thinking, your brain does less.

II. YOU STOP QUESTIONING OUTPUTS

There’s a robust body of evidence on what researchers variously call “cognitive surrender,” “automation bias,” and “cognitive offloading,” concepts clustered around the tendency to accept AI outputs without much or any scrutiny.

• How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender, Shaw & Nave (SSRN working paper, 2026, N=1,372): Wharton did three preregistered experiments using reasoning problems where AI was sometimes programmed to give wrong answers. When AI was wrong, participants followed it ~80% of the time, performing “worse than having no AI at all.” Trust in AI was the strongest predictor: high-trust participants had 3.5× greater odds of following faulty answers. Only ~20% actively overruled incorrect AI; the rest incurred what the authors call “cognitive surrender”: they didn’t scrutinize AI outputs.

• The impact of generative AI on critical thinking, Lee et al. (CHI ‘25, 2025, N=319): Microsoft and Carnegie Mellon surveyed knowledge workers who use GenAI weekly, collecting 936 real-world use examples. “Higher confidence in GenAI correlated with less critical thinking.” Workers shifted from active problem-solving to passive oversight; from “thinking by doing” to “choosing from outputs.” They produced a “less diverse set of outcomes” when using AI.

• Navigating the jagged technological frontier, Dell’Acqua et al. (Organization Science, 2026, N=758): Harvard Business School, Wharton, MIT Sloan, BCG. Preregistered field experiment with BCG consultants performing 18 realistic tasks. For tasks inside AI’s capability frontier: “+12.2% more tasks completed, 25.1% faster, 40% higher quality.” For tasks outside the frontier: AI users were “19 percentage points less likely to produce correct solutions.” The problem: AI output outside the frontier “looked polished but was subtly wrong.” People couldn’t tell. Two collaboration patterns emerged: “centaurs” (clear human/AI division) and “cyborgs” (interleaved workflow).

• The levers of political persuasion with conversational artificial intelligence, Oxford/Stanford/MIT, 2025 (Science, N=76,977): Conversational AI was “significantly more persuasive than static messaging.” Fine-tuning could increase persuasiveness by up to 51%. A troubling finding: “the more persuasive a model, the less accurate” its information tended to be.

Conclusions: The convergence across these studies is a serious and urgent finding. People trust AI too much, but more than that, it’s clear that AI fluency creates a new failure mode: Wrong answers delivered in flawless prose get accepted. And the more you are predisposed to trust AI, the worse the problem gets.

III. YOU LEARN LESS, YOU LEARN MORE

The education research reveals a clean split between 1) raw AI access consistently harms learning and 2) AI designed specifically for teaching can dramatically improve it. The variable is whether AI replaces the cognitive work or scaffolds it. I’ve divided this section into those two camps: harm side and help side.

THE HARM SIDE

• Generative AI without guardrails can harm learning, Bastani et al. (PNAS, 2025, N=~1,000): Wharton and Penn Engineering conducted a preregistered RCT in high school math classrooms. Students given standard ChatGPT improved practice scores by 48% but scored “17% lower” on subsequent unassisted exams. They used ChatGPT as an answer machine, copying solutions. A redesigned GPT Tutor that guided reasoning instead of giving answers “improved practice grades by 127%” and largely prevented the learning decline. Students didn’t perceive any reduction in their own learning.

• ChatGPT as cognitive crutch, Barcaui (Social Sciences & Humanities Open, 2025, N=120): Fundação Getulio Vargas and UFRJ did a preregistered RCT with a surprise retention test 45 days after learning. The traditional study group scored ~69% vs. ~58% for the ChatGPT group, an 11 percentage-point gap. The AI group showed “a steeper forgetting curve, consistent with weaker initial encoding,” meaning they forgot faster because knowledge didn’t stick. Prior AI experience didn’t protect against the offloading effect.

• Beware of metacognitive laziness, Fan et al. (British Journal of Educational Technology, 2025, N=117): Zhejiang University and Monash University conducted a writing experiment. ChatGPT group produced better essays but showed “no significant improvement in knowledge gain or transfer.” The researchers named the pattern “metacognitive laziness”; learners offloaded the monitoring and evaluation of their own thinking to AI. “The product improved. The learner didn’t.”

• Cognitive ease at a cost, Stadler, Bannert & Sailer (Computers in Human Behavior, 2024, N=91): TU Munich and LMU Munich. LLM users experienced significantly lower cognitive load but demonstrated “lower-quality reasoning and argumentation.” The task felt lighter because the hard part—the actual thinking—was being done for them.

THE HELP SIDE

• AI tutoring outperforms active learning at Harvard, Kestin et al. (Scientific Reports, 2025, N=194): A carefully designed GPT-based tutor—built to ask questions rather than give answers—produced “learning gains more than double” those of traditional active learning. Students also spent less time.

• Tutor CoPilot, Wang et al. (arXiv preprint, 2024, N=1,800): RCT from Stanford. AI provided real-time coaching suggestions to human tutors (tutor-facing only, never student-facing). Students were “4 percentage points more likely to master topics.” Lower-rated tutors’ students saw a 9-point improvement. The AI improved the human, who improved the student at $20/tutor/year.

• From chalkboards to chatbots, De Simone et al. (World Bank, 2025, N=~800): A six-week after-school program using Microsoft Copilot with teacher guidance in Edo State, Nigeria. Learning gains in six weeks were equivalent to “1.5–2 years of typical schooling.” This approach “outperformed 80% of education interventions studied via RCTs in developing countries.” Largest effects for female students.

• AI tutoring can safely and effectively support students, Google DeepMind/Eedi (arXiv preprint, 2025, N=165): Pedagogy-tuned AI supervised by human tutors. The tutors “approved 76.4% of LearnLM’s drafted messages with zero or minimal edits.” Students were “5.5 percentage points more likely to solve novel problems” on subsequent topics. Tutors reported learning new pedagogical practices from the model.

• Generative AI enhances individual creativity but reduces the collective diversity of novel content, Doshi & Hauser (Science Advances, 2024 N=300): UCL and University of Exeter did an experiment on the effects of AI on creative writing. AI-assisted stories were rated more creative, especially from less-creative writers. But AI-enabled stories were “5.0–5.2% more similar to each other,” which reveals a “social dilemma” where individual gains erode collective novelty. I put this in “help side,” but it’s a bit more nuanced, revealing that AI use is not a solo story.

Conclusions: The education evidence is the most policy-relevant body of research in this compilation, and it delivers an unambiguous message: one technology produces two opposite outcomes depending on design. ChatGPT used as an answer machine caused learning declines, whereas a pedagogically designed AI tutor produced gains. The variable is not so much “use of AI,” as “how are they using AI.”

IV. YOU GET LONELIER (OR DO YOU?)

The psychological research on AI chatbots and emotional well-being is the most conflicted area in the literature. I’ve included this here because AI affects learning and cognition, but also emotion and psychological behavior.

• How AI and human behaviors shape psychosocial effects of extended chatbot use, Fang et al. (arXiv preprint, 2025, N=981): MIT Media Lab and OpenAI conducted a four-week RCT with daily ChatGPT use across nine conditions (3 modalities × 3 conversation types) on over 300,000 messages. “Higher daily usage correlated with higher loneliness, dependence, problematic use, and lower socialization” across all modalities and conversation types. Voice mode initially mitigated loneliness but “advantages diminished at high usage.” Bottom line: immediate relief in exchange for long-term dependency.

• How does turning to AI for companionship predict loneliness and vice versa?, Folk & Dunn (PsyArXiv preprint, 2025, N=2,000+): University of British Columbia conducted a twelve-month longitudinal study with bidirectional analysis that found evidence of a vicious cycle: loneliness drives chatbot use, which predicts increased loneliness four months later, which predicts chatbot use. Chatbot use “did not significantly predict decreases in broader social connection measures.”

• Individual and well-being factors associated with social chatbot usage, Latikka et al. (Journal of Social and Personal Relationships, 2026, N=5,663): A six-country study conducted by Tampere University found that social chatbot usage was “positively associated with psychological distress in all six countries” studied (Finland, France, Germany, Ireland, Italy, Poland). Loneliness predicted chatbot use in four of six. The cross-cultural consistency was described as “striking.”

• Individual differences in anthropomorphism help explain social connection to AI companions, Folk, Heine & Dunn (Scientific Reports, 2025, N=1,274): University of British Columbia found that individual differences in the tendency to anthropomorphize technology “strongly predicted feeling connected” after chatbot interaction. AI’s artificial nature is an insurmountable barrier for some; for others, anthropomorphism bridges it.

Conclusions: The psychological literature has a time-scale problem. Short-term experiments—a single session or a few days of interaction—tend to find benefits like reduced loneliness and therapeutic benefit; longer-term studies—over weeks and months—tend to find problems like isolation, dependence, and distress. These findings aren’t contradictory. The same dynamic exists with other coping mechanisms (alcohol reduces social anxiety in the moment while increasing it over time). Whether AI chatbots follow the same trajectory or long-term effects stabilize is unknown.

V. THE META-ANALYSES

A few meta-analyses (there are not many) have attempted to reconcile the contradictory findings, particularly on learning.

• The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking, Wang & Fan (Humanities and Social Sciences Communications/Nature, 2025, 51 studies between 2022-2025): Hangzhou Normal University conducted a meta-analysis on AI’s effect on learning and higher-order thinking. They found a large positive effect on learning performance. Moderate effects on higher-order thinking, moderated by the type of intervention.

• Does ChatGPT enhance student learning?, Deng et al. (Computers & Education, 2025, 69 studies between 2022-2024): Another meta-analysis from Hangzhou Normal University found that ChatGPT “improves academic performance, affective-motivational states, and higher-order thinking” while reducing mental effort. “No significant effect on self-efficacy.” It lacks post-intervention assessments to see longer-term effects.

• ChatGPT’s impact on student learning outcomes, Wu et al. (Humanities and Social Sciences Communications/Nature, 2026, 35 studies between 2022–2024): Moderately positive overall effect, “significantly enhancing both cognitive and non-cognitive skills.” Subject, experimental duration, and instructional mode were significant moderators.

Conclusions: The meta-analyses confirm the paradox we’ve been seeing so far at scale: The aggregate effect size for immediate performance is large and reliable, whereas the aggregate effect on the cognitive processes that produce independent capability—higher-order thinking, metacognition, self-efficacy, transfer—is small, null, or unmeasured. They measure what’s easy to measure (test scores) and largely miss what’s hard (whether the person is developing or atrophying).

VI. THEORETICAL FRAMEWORKS

Several papers have proposed theoretical frameworks to explain why AI affects cognition the way it does. These are conceptual models that organize the evidence.

• Artificial intelligence and illusions of understanding in scientific research, Messeri & Crockett (Nature, 2024): Yale and Princeton researchers argued that AI creates “illusions of understanding”: users believe they know more than they do because AI produces fluent, confident outputs. Developed a taxonomy of four AI archetypes in knowledge work (Oracle, Surrogate, Quant, Arbiter) each carrying distinct epistemic risks. Warned of “scientific monocultures” where AI narrows the range of questions researchers think to ask. They warn that “we produce more but understand less.”

• The case for human–AI interaction as system 0 thinking, Chiriatti et al. (Nature Human Behaviour, 2024): They proposed adding a new layer to Daniel Kahneman’s framework. System 1 is fast intuition. System 2 is slow deliberation. System 0 is “the outsourcing of thought to AI,” a pre-cognitive layer that shapes what reaches human awareness at all; similar to “cognitive surrender.”

• From tools to threats: AI-Chatbot Induced Cognitive Atrophy, Dergaa et al. (Frontiers in Psychology, 2024): Introduced a theoretical framework drawing parallels between AI overreliance and problematic internet use. Used “Extended Mind Theory” to argue that when AI becomes a cognitive prosthetic, “the underlying cognitive ‘muscles’ weaken from disuse.”

• The brain side of human-AI interactions in the long-term: the ‘3R principle’, Rossi, Fraccaro & Manzotti (npj Artificial Intelligence, 2026): They argued that “passive, uncritical, reliance on AI may weaken activity-dependent brain plasticity and erode cognition, whereas active co-creation can sustain or enhance it.” Proposed Results, Responses, Responsibility (3R) as a “preventive framework for cognitive hygiene” to “preserve agency, meaning-making, and long-term brain health.”

Conclusions: The theoretical proposals converge on the same structural insight from different angles. “System 0,” “System 3,” “cognitive atrophy”, “illusions of understanding” all point to the fact that AI’s fluency and availability create a path of least cognitive resistance. When taking that path is frictionless and produces good-enough outputs, the effortful alternative (actual thinking) is harder to justify in the moment. They describe the same phenomenon that the empirical studies find: the gap between what AI does for the output and what it does to the person producing it.

VII. WHAT THE PUBLIC SAYS

• How Americans view AI and its impact on people and society, Pew Research (2025, N=5,023): 53% believe AI will worsen creative thinking ability. 50% say it will hurt the ability to form meaningful relationships. The share of Americans “more concerned than excited” about AI rose from 37% in 2021 to 50% in 2025.

• Teens, social media and AI chatbots, Pew Research (2025, N=1,458): 64% of U.S. teens use AI chatbots. 30% use them daily. Over half use them for schoolwork. “20% of teens in low-income households do all or most schoolwork with chatbot help vs. 7% in higher-income households,” an equity dimension largely absent from the academic literature. 60% say students at their school use chatbots to cheat.

Conclusions: The public is right to worry, but the worry is unfocused. The Pew data show that people sense AI threatens cognition and relationships, and the research in this compilation largely confirms that intuition. But the public framing is binary (AI good or AI bad) when the research points to something else: the same technology produces opposite effects depending on implementation. The gap between public perception and scientific evidence is about what variable matters most, the presence of the technology vs how it’s used.

GLOBAL CONCLUSIONS

Core finding: A performance-competence dissociation as a function of implementation design, not technology

Across 30+ studies, AI chatbots reliably improve the quality and speed of immediate outputs: test answers, essays, creative work, professional deliverables. They just as reliably degrade the cognitive processes that build durable knowledge, independent reasoning, and creative diversity over time.

As I see it, this may not be a paradox that will resolve with better models because the basic principles are unchanging. This is a structural feature of any tool that does cognitive work for you rather than with you. The performance improvement is real and the competence impairment is also real. They coexist because they operate on different timescales.

But.

The design of the interventions and the particularities of how AI is used can improve the outcomes.

The same underlying technology produces dramatically different cognitive effects depending on how it’s implemented. AI can make what you produce better while making you worse at producing it. Or it can genuinely make you better. Which one happens depends on design, not on AI itself.

This nuance should shape policy, product design, and individual behavior. The debate about “should we use AI?” is much less important than the question of how the interventions are designed and how the technology is used. Sounds like a no-brainer, but public perception is still stuck at “is AI good or bad?”

That’s the most important insight across the entire literature.

What we don’t know

Four critical gaps/limitations remain:

• No long-term neuroimaging. The MIT EEG study tracked participants across four sessions. No study has measured brain changes from sustained AI use over months or years (the technology is perhaps too young for that; should be a priority). The “cognitive debt” finding—suppressed brain activity persisting even after AI was removed—is preliminary.

• Individual differences are poorly mapped. Who is most vulnerable to cognitive surrender? Who resists it? The Wharton study identified trust in AI as the key predictor; the Swiss study found age and education matter. But the interaction between personality, cognitive style, expertise, and AI vulnerability is largely unexplored. I’ve long observed that AI is an “enhancer of natural disposition,” meaning that it makes you more of what you already are, but this is purely anecdotal.

• Children are almost unstudied. I found one fMRI study of 31 children. That’s it. Meanwhile, 64% of U.S. teens use AI chatbots and 30% use them daily. They are the most exposed group and the one that has its future at stake the most. The developmental neuroscience of AI use—the effects on brains that are still forming—is the most urgent gap in the entire field.

• A science vs technology time delay. Because science is slower than technological development, most of the studies are conducted with old models (e.g., GPT-4o, GPT-3.5, etc.). Some variables are constant because they’re independent of specific models—e.g., cognitive offloading will remain and even increase because models get better—but others should be carefully framed, e.g., anything related to the quality of the responses. My prediction is that some effects will disappear, others will get better (educational approaches if the right policies are implemented), and others will get worse (emotional dependence and related cognitive risks).

The big open question for the future

The 30+ studies compiled here converge on a description of what’s happening but diverge on what it means. Is AI-induced cognitive offloading a temporary adjustment—like how calculators initially worried math teachers but ultimately freed students to tackle harder problems—or is it something qualitatively different, a tool so fluent and available that it undermines the motivation to develop the underlying capability at all?

The calculator analogy is comforting—it was received with panic that eventually proved unwarranted—but may not hold. Calculators automate computation, which is mechanical. AI chatbots automate reasoning, argumentation, synthesis, and creative expression, the cognitive activities that are the skill rather than a means to it. When a calculator does your arithmetic, you lose arithmetic; when AI does your thinking, you lose thinking.

Whether that loss matters depends on whether AI will always be there to do the thinking for you. And, crucially, on whether you believe there’s intrinsic value in the process of thinking itself, independent of the output. Think about it.

Hey there, I’m Alberto! 👋

Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Everyone’s using AI now. At the same time, everyone is starting to notice a bunch of issues and caveats that are emerging with the growing use of this technology. And yet, as far as I know, nobody has made a periodic self-check for AI users.

There are vendor evaluation questionnaires (for companies buying AI), prompting guides (for people learning the basics), and tool comparison lists (for people choosing between products). But there's nothing for the person who uses AI every day and wants to make sure they're not drifting. Thinking about “how you use a tool” is less exciting than using the tool. But it might matter more.

This is my first attempt at that (this is for myself, but it might be of use to you).

I’ve been using AI tools every day for over five years now. I research with AI, brainstorm with it, edit with it, and build personal software tools with it. I’m not yet in the multi-agent stage, but it will eventually happen. My current project is akin to Karpathy’s wiki/knowledge base, but for the entire archive of writing, reading, and interesting stuff that I’ve been capturing and classifying for years, rather than research.

At this point, my workflow is more AI than not-AI. I’m faster than I’ve ever been and work more than I’ve ever worked (that’s why I wrote “The Most Important Skill in AI Right Now: How to Know When to Stop” and “How to Deal With Infinite Options”).

Yesterday—because I’m too methodical or hedge too much—I sat down to write something without relying on AI assistance from beginning to end—research, editing, grammar, etc.—just to see how it felt, and realized a tinge of vertigo. Not having an AI-enabled “safety net” made me feel weirdly vulnerable (in case you don’t know, English is not my first language, and so I make typos, confuse words, and misuse expressions, etc., at a higher rate than normal).

My work was slower and less comfortable, and I kept reaching for a Claude tab. Imagine you had to work for a day without the internet, it’d be supremely restricting, right? But AI is so new! How is this happening to me already?

My answer is that, for most heavy users using AI for serious work, this shift is going to happen for AI much faster than it did for the internet or the phone. These tools reward speed, so here we are, with a superpower in our hands, doing more than ever, faster than ever, always monitoring the next release or announcement.

And yet.

Have you taken the time to step back for a second and ask whether you’re actually getting better at what you do or just getting faster? Whether you’re getting smarter or dumber? Whether your work is better or just more abundant?

I haven’t.

And I have noticed, after writing those other posts I mentioned above, that it’s not the people who use AI the most who will reach further, but those who start to ask the right questions about the best practices. Those who dedicate 5% of their time to reflect.

This is, it turns out, an old problem with an old solution: the checklist.

To convince you, I’ve compiled four cases in chronological order. You can jump this section if you are already convinced, but I found these four guys’ approaches to be genuinely interesting despite centuries between them and us (except the last one). They are the predecessors to our current work methodology.

1. Lucius Annaeus Seneca (4 BC - AD 65)

2. Ignatius of Loyola (1491 - 1556)

3. Benjamin Franklin (1706 - 1790)

4. Atul Gawande (1965 - )

Seneca’s most famous line is ”It is not that we have a short time to live, but that we waste a lot of it.” He did not suffer from the laziness that afflicts most of us.

He was one of the sharpest minds in Rome—second in charge to Nero while he was a boy—and he had a nightly practice where he reviewed his entire day. “What did I say? What did I do? What could I have done better?” He describes it in De Ira (on anger), and he makes it sound remarkably undramatic: he doesn’t frame it as self-punishment or spiritual discipline or personal development but as standard maintenance.

He already knew how to live well (although for what we know from his writings on happiness, he could’ve done a better job). The review was about making sure he was doing it. Seneca understood something that I think applies directly to us: knowing what to do and actually doing it are different problems.

Fifteen centuries later, Ignatius of Loyola—a Spaniard, which I appreciate—took this idea and turned it into a system: The Examen. Five steps and five minutes. Twice a day.

It was designed for Jesuits, the most committed, most disciplined practitioners in the Catholic world, so he tailored it, naturally, for religious stuff. But it’s a useful reference because it was for experts. (You may not think of yourself as such, but consider what 99% of the population knows about using AI and then think again.)

Loyola understood that the review becomes more valuable the more competent you already are—not the other way around—because 1) complacency hides in competence and 2) an incompetent person could not even parse it! A beginner may not know what they don’t know, but an expert forgets how much they know.

Then Ben Franklin—probably the one that’s most familiar to you—took the whole thing out of the spiritual domain—this is during the 18th-century Enlightenment—and made it practical with his 13 virtues. He tracked one per week, cycled through all 13 every quarter, and kept a chart (he would have loved Claude Excel).

Franklin’s system made explicit what Seneca and Loyola left implicit: any single day’s answers are meaningless; “I failed at Temperance today, so what?” The insight is in the pattern across months. He also said, and I love this, that he never achieved the perfection the system aimed at, but that the practice of checking made him better than he would have been without it.

Finally, Atul Gawande—the only one here that’s still alive—proved all of this empirically in 2009.

He wrote The Checklist Manifesto after studying a case at Johns Hopkins: a team created a five-step checklist for inserting central line catheters. When they made it mandatory to run the checklist every time for ICU doctors—who knew it by heart—the infection rate dropped from 11% to zero!

Essentially, they stopped skipping steps they had been skipping solely motivated by the tricky disability of expertise. Gawande’s key finding is that the doctors who resisted the checklist the most were the most senior ones. It felt beneath them. And yet, they were the ones who benefited the most. That’s us, you and me, pioneers in the use of AI, paving the way for those who come after, failing to notice our flaws.

Summing up: Seneca introduced the principle of the checklist, Loyola focused it on the experts, Franklin introduced consistency, and Gawande proved it empirically.

This maps directly onto how we’re using AI.

And I’m going to show you how.

Read more

Sorry to bother you on Saturday. Thought this was important to share.

I.

The first thing you learn about a loom is that it’s easy to break.

The shuttle runs along a track that warps with humidity. The heddles hang from cords that fray. The reed is a row of thin metal strips, bent by hand, that bend back just as easily. The warp beam cracks if you over-tighten it. The treadles loosen at the joints. The breast beam, the cloth roller, the ratchet and pawl, the lease sticks, the castle; the whole contraption is wood and string held together by tension. It’s a piece of ingenuity and craftsmanship, but one as delicate as the clothes it manifests out of wild plant fibers. It is, also, the foundational tool of an entire industry, textiles, that has kept its relevance to our days of heavy machinery, factories, energy facilities, and datacenters.

It is not nearly as easy to break a datacenter.

It is made of concrete and steel and copper and it’s on the bigger side. It has interchangeable servers, and biometric locks and tall electrified fences and heavily armed guards and redundancy upon redundancy: every component duplicated so that no single failure brings the whole thing down. There is no treadle to loosen or reed to bend back.

But say you managed to bypass the guards, jump the fences, open the locks, and locate all the servers. Then you’d face the algorithm. The datacenter was never your goal; the algorithm lurking inside is. It doesn’t run on that rack, or any rack for that matter. It is a digital pattern distributed across millions of chips, mirrored across continents; it could be reconstituted elsewhere, and it’s trained to addict you at a glance, like a modern Medusa.

But say you managed to elude the stare, stop the replication, and break the patterns. Then you’d face superintelligence. The algorithm was also not your goal; the vibrant, ethereal, latent superintelligence lurking inside is. Well, there’s nothing you can do here: It always “gets out of the box” and, suddenly, you are inside the box, like a chimp being played by a human with a banana. It’s just so tasty…

There’s another solution to break a datacenter: You can bomb it, like one hammers down the loom.

Some have argued that this is the way to ensure a rogue superintelligence doesn’t get out of the box. A different rogue creature took the proposal seriously: last month, Iran’s Revolutionary Guard released satellite footage of OpenAI’s Stargate campus in Abu Dhabi and promised its “complete and utter annihilation.”

But you probably don’t have a rogue nation handy to fulfill your wishes. Maybe you will end up bombed instead and we don’t want that to happen. That’s what happens with rogue intelligences: you can’t predict them.

And yet. Two hundred years of increasingly impenetrable technology—from looms to datacenters—have not changed the first thing about the people who live alongside it. The evolution of technology is a feature of the world just as much as the permanent fragility of the human body.

And so, more and more, it is people who are the weaker link in this chain of inevitable doom. And it is people who will be targeted.

II.

April of 1812. A mill owner named William Horsfall was riding home on his beautiful white stallion back from the Cloth Hall market in Huddersfield, UK. He had spent weeks boasting that he would ride up to his saddle in Luddite blood (a precious substance that served as fuel for the mills).

A few yards later, at Crosland Moor, a man named George Mellor—twenty-two years old—shot him. It hit Horsfall in the groin, who, nominative-deterministically, fell from his horse. People gathered, reproaching him for having been the oppressor of the poor. Naturally, loyal to his principles in death as he was in life, he couldn’t hear them. He died one day later in an inn. Mellor was hanged.

History rhymes, they say.

April of 2026. A datacenter owner named Samuel Altman was driving home on his beautiful white Koenigsegg Regera back from Market Street in San Francisco, US. He had spent weeks boasting that he would scrap and steal our blog posts (a precious substance that serves as fuel for the datacenters).

A few hours later, at Russian Hill, a man named Daniel Alejandro Moreno-Gama—twenty years old—allegedly threw a Molotov cocktail at his house. He hit an exterior gate. Altman and his family were asleep, but they’re fine. Moreno-Gama is in custody.

This kind of violence must be condemned. This is not the way. It’s horrible that it is happening at all. And yet, for some reason, it keeps happening.

Last week, the house of Ron Gibson, a councilman from Indianapolis, was shot at thirteen times. The bullet holes are still there. The shooter left a message on his doorstep: “NO DATA CENTERS.” Gibson supports a datacenter project in the Martindale-Brightwood neighborhood. He and his son were unharmed.

In November 2025, a 27-year-old anti-AI activist threatened to murder people at OpenAI’s SF offices, prompting a lockdown. He had expressed a desire to buy weapons.

Increasingly, as the objects of people’s anger and frustration and desperation become unreachable behind fences and guards, or abstracted away in ones and zeros, or elevated above the clouds, the mob will turn their unassailable emotions toward human targets.

I don’t want to trivialize the grievances of the people who fear for their futures. I don’t want to defend Altman’s decisions. But this is not the way. This is how things devolve into chaos.

And I wonder: how desperate can people be before these isolated events become a snowball of violence that will be resisted by neither datacenters nor rich people’s houses?

III.

Every time I hear from Amodei or Altman that I could lose my job, I don’t think “oh, ok, then allow me pay you $20/month so that I can adapt to these uncertain times that have fallen upon my destiny by chance.” I think: “you, for fuck’s sake, you are doing this.” And I consider myself a pretty levelheaded guy, so imagine what not-so-levelheaded people think.

There’s a lot of friction to escalating violence, but that friction dissolves the moment this sentiment starts to be common. Normally, it just fades away anyway, but there’s one scenario where I see it inevitably escalating:

If people feel that they have no place in the future.

If they feel expelled from the system—they’re unable to buy stuff, their skills become obsolete, their chance at earning a living is replaced by a swarm of AI agents, they think we are truly going to die (so far, the violence has been tied mostly to safety AI movements)—then they will feel they have nothing to lose.

And then, and I’m sorry to be so blunt, then it’s die or kill.

Perhaps the most serious mistake that the AI industry made after creating a technology that will transversally disrupt the entire white-collar workforce before ensuring a safe transition, was making it explicit by doing constant discourses that amount to: “we are creating a technology that will transversally disrupt the entire white-collar workforce before ensuring a safe transition.”

And, to top it off, they add “careful down there.”

The difference between AI and, say, looms, is that this has been broadcast to the entire globe, and it has been treated in a sort of self-conscious way. The AI leaders know the problems that will emerge and so they cannot help but talk about them constantly and so they are letting us know, which makes them look like psychopaths. How do you guys think people will react to this? You should be much less self-conscious and much more self-aware: realize what you sound like!

(No piece of journalism, much less one that leans forensic rather than sensationalist, could do a better job against them than their own words. These guys lack basic self-awareness. For what it’s worth, the New Yorker piece I’m referring to, which Altman also referred to in his blog post, made me see him more as a flawed human rather than a sociopathic strategist. My sympathy for him will probably never be very high, but it grew after reading it.)

People hate AI so much that they are prone to attribute to it everything that’s going wrong in their lives, regardless of the truth. That’s why they mix real arguments, like data theft, with fake ones, like the water stuff. Employers do it, too. Most layoffs are not caused by AI, but it’s the perfect excuse to do something that’s otherwise socially reprehensible.

AI has become the perfect scapegoat. It doesn’t help that the entire AI industry has decided that throwing rocks at its own roof is its best selling point: If AI is so powerful and so dangerous and soon to be so ubiquitous, then what is so unexpected about people blaming everything on it?

Nothing that Altman could say justifies violence against him. This is an undeniable truth. But unfortunately, violence might still ensue. I hope not, but I guess we are seeing what appears to be the first cases.

I just hope that, contrary to the cases of ChatGPT-induced psychosis, chatbot addiction, AI-blamed job layoffs, and a growing trend of illiteracy, it stops.

Hey there, I’m Alberto! 👋

Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe

THE WEEK IN AI AT A GLANCE

• Money & Business: Anthropic’s annualized revenue jumped from $19B to $30B in a month; Meta’s 60-trillion-token Claude binge may explain part of the spike.

• Geopolitics: Chinese AI models now power 30% of global workloads, up from 1% in late 2024, while Meta quietly abandoned open source.

• Work & Workers: Nearly 80,000 tech workers lost jobs in Q1 2026, with companies using AI as cover for post-pandemic overhiring corrections.

• Products & Capabilities: Claude Mythos Preview is impressive on paper, but may be more about IPO marketing and regulatory positioning than anything.

• Culture & Society: Gen Z’s AI excitement dropped 14 points to 22% while anger rose to 31%; even daily users are losing faith.

• Philosophy: Dario Amodei says AI could be conscious; Anil Seth says AI isn’t conscious; Barton Friedland says it’s the wrong question.

THE WEEK IN THE ALGORITHMIC BRIDGE

• (FREE) How to Think About AI Like a Long-Term Investor: Everything worth having—skills, relationships, expertise, including AI fluency—follows the logic of compound interest: early returns look negligible, but consistency over time builds value that accelerates, which is why starting now with AI (even badly) beats waiting for it to mature.

• (FREE) Inside the AI Industry’s Most Expensive Mistake: Token-based “thinking” is an expensive architectural crutch—AI models should reason in latent space the way humans think in sensations rather than words—but the irony is that if they did, Meta couldn’t have distilled Claude’s reasoning traces to bootstrap Muse Spark.

• (PAID) What Happens When AI Gets Too Good at One Thing: Claude Mythos is a narrow superintelligence—superhuman at finding zero-day vulnerabilities but not AGI—and Anthropic's decision to withhold it is simultaneously genuine safety concern, IPO marketing, and a preview of a future where the most powerful models are never sold as consumer products.

MONEY & BUSINESS

We’re all on a Claudeonomics run

Anthropic is generating more annualized revenue than OpenAI. The trajectory is crazy. $9B in January. $14B in February. $19B by early March. $30B in early April.

Wait, what? It jumped $10B+ in one month. What is going on? Claude is a great product, but this is… hard to believe. (OpenAI was at roughly $24B on March.)

Anthropic's financial health couldn't be better. Claude Code subscriptions have quadrupled since January. Enterprise customers with $1M+ annual contracts crossed 500. Eight of the Fortune 10 are Claude customers. The growth is robust, but the acceleration from $19B to $30B in a month demands an explanation beyond “enterprise going well.”

Enter Claudeonomics (I covered it in passing here).

A Meta employee built a leaderboard called “Claudeonomics” tracking token consumption company-wide of Claude models (no longer active because we mocked them). In the last 30 days tracked, they burned through 60 trillion tokens. The top user spent 281 billion.

How much does that cost?

We don’t know which model Meta runs, but there are reasons to think it skews toward Opus 4.6, Anthropic’s most capable ($5/$25 per million input/output tokens). Meta doesn’t need to optimize for cost when it’s spending $115–135B on AI infrastructure this year, with an ad empire printing $40B+ per quarter. At this volume, for a customer this size, enterprise discounts of up to 30-50% are standard.

At a 40% discount: roughly $540M/month. Annualized, ~$6.5B.

That, alone, would cover most of the spike in Anthropic's April run-rate revenue. The coincidence of Meta’s Claudeonomics gimmick and Anthropic’s revenue surge is, at the very least, something to think about… However, there are other tricks Meta is probably doing that would reduce the overall spend significantly.

All in all, I think Meta is a big client but just one factor among others for Anthropic’s recent growth.

Two caveats on the revenue figures.

Annualized run rate is monthly revenue times twelve, which assumes this month repeats forever (it doesn’t). If Meta’s binge was a concentrated push to prepare Muse Spark, next month’s number could look very different. But, at the same time, even if the assumption that Meta is the reason behind the spike, it isn’t the only driver: Claude Code’s adoption across thousands of enterprises is not a temporary fashion.

Sources: Anthropic, The Information, The New York Times

GEOPOLITICS

China won the AI race for global adoption

Read more

Hey there, I’m Alberto!👋

Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles and essays. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

It seems that, seven years later and a lot of headaches, the AI industry has come full circle.

In 2019, OpenAI withheld the newly announced GPT-2 from public release. They appealed to the presumed dangers of a system whose full extent of intelligence and capability was yet unknown. To this day, I’m unsure if that was merely a marketing gimmick—a successful one at that—or if they were genuinely worried. I’m inclined for the latter; after all, the GPT paradigm was novel, and it had already showcased unexpected signs of proto-AGI: it could talk!

Fast forward to 2026, many of those same OpenAI researchers are now at Anthropic, which has announced that it’ll hold back Claude Mythos (preview) from public release (system card here).

History rhymes, but it doesn’t repeat.

The difference is that whereas GPT-2 marked a display of caution born out of ignorance—“it writes well; could be used for disinformation or propaganda or whatnot, we just don’t know” (OpenAI released it six months later)—Anthropic’s messaging suggests the opposite is true for Mythos: it’s not so much that the company doesn’t know what is capable of but that it knows exactly what it can do.

If you are choosing to leave a ton of money on the table by not releasing your latest and most powerful model to consumers and enterprise clients, you must have a good reason to do so: For those who, like me, don’t know much about software security, Mythos is like a “nuke for code,” if nukes could also be used for good.

Given a piece of software—a browser, an operating system, whatever—Mythos, which by the way destroyed every coding and agents benchmark out there, can find old, immediately exploitable vulnerabilities that not even the developers knew existed, with minimal human steering. And then it can go and figure out how to exploit them. From this moment, no software system in the world is safe to the extent that this technology is possible (in practical terms, only 3-5 players will be able to build this kind of system in the next few months, and none outside the US).

Anthropic’s post reads:

AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. . . . Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. Given the rate of A.I. progress, it will not be long before such capabilities proliferate, potentially beyond actors who committed to deploying them safely. The fallout—economics, public safety and national security—could be severe.

That’s the nuke part.

The “reverse-nuke” part is that the same capability works the other way around: if you can find the hole, you can patch it.

To that end, Anthropic has launched Project Glasswing, a coalition that includes Amazon, Apple, Google, Microsoft, NVIDIA, and many other partners concerned with the maintenance of critical software infrastructure and cybercrime. They all have access to Mythos with one goal: find and fix the vulnerabilities before someone else’s (China’s) model finds them first.

Goes without saying that you won’t get access to Mythos for the foreseeable future, but you understand why now.

You may still object, though: “They are centralizing AI! They want to keep the powerful models for themselves! I’m not a malicious actor, why can’t I get access?” Yes, you are correct. But, did you really expect them to sell near-AGI models as a monthly subscription product? Mythos is not a Substack newsletter.

The reason I compared it to nukes is to drive this point home: do you think it’d be preferable if anyone could buy the materials and the manuals to build a homemade DIY miniature nuclear weapon that could obliterate a small city? Maybe not. You should think of narrow superintelligences like Mythos—superhuman at one single thing—as AI’s equivalent to nukes.

Philosophical dilemmas in the real world all have this knack for shutting down simple solutions. What would you choose, to sacrifice freedom in the name of safety or to sacrifice safety in the name of freedom? Sometimes it’s not so easy to decide.

In any case, because Anthropic is a private company, “you don’t get to weigh in on this,” as the OpenAI executive told employees regarding the use of their models by the US government.

(I’m not positioning myself in favor of or against, just saying things as they are.)

Anyway, because we don’t know much else about Mythos (besides the system card, the red teaming docs, and the project Glasswing announcement), I want to dedicate the rest of this post to four questions that I touched on in passing above that I think people will want answered (meaning, that I myself want answered).

• Isn’t this just Anthropic doing marketing?

• Why did they make public the decision?

• Why won’t they give good people access?

• Is Mythos a narrow superintelligence?

Let’s start from the first one and work our way down the list.

I. THE DRUG OF FEAR MARKETING

Read more

Hey there, I’m Alberto!👋

Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles and essays. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Today, I bring you a hot analysis + hot news, the best mix.

I.

Meta employees have an internal leaderboard called “Claudeonomics.” The ranks go up from casual user to “Session Immortal” to the top tier: “Token Legend.” Setting aside the cringe, it’s crazy that over a 30-day period, total usage on the dashboard topped ~60 trillion tokens (words). For context, it has been estimated that all books published throughout history amount to ~20 trillion tokens.

It’s not just Meta, though. Nvidia CEO Jensen Huang said in the All-In Podcast that if a $500k engineer spent less than $250k a year in tokens, he’d be “deeply alarmed.” When asked whether Nvidia is spending $2 billion on tokens for its engineering team, he said: “We’re trying to.” Last year, OpenAI introduced “Tokens of Appreciation,” a program to recognize developers and organizations that have processed high volumes of data through the API in three tiers: silver, black, and blue.

Token usage is, by all accounts, a new status symbol. Engineers at the top labs boast about token numbers like a bad coder would about lines of code: look how excessive I am. They are “tokenmaxxing,” as the New York Times wrote last month. “An engineer at OpenAI processed 210 billion ‘tokens,’” writes the NYT, “enough text to fill Wikipedia 33 times.”

Some engineers spend more on tokens than they earn in salary. Token budgets are being pitched as a “fourth component” of compensation, alongside salary, equity, and bonuses. Candidates are asking in interviews: how many tokens come with the job? API expenses are now competing with labor budgets.

If you are not burning through your token cap, said Huang, you are like the chip designer who refuses to use CAD tools and instead designs by hand, with a pencil. From an industrial perspective, this behavior makes sense: these are the tools we have. The more you spend, the more they think, and, presumably, the better the output you get, and thus the more benefit you create for your company.

In the “intelligence age,” the token is the unit of cognitive labor.

The obvious reaction is to summon Goodhart’s law or bring up “On the folly of rewarding A, while hoping for B.” Some ask, dumbfounded: What do you think your employees will do if you take this simple, hackable metric as proof of work? Well, hack it. Duh. And so you have guys instructing their bots to run loops with the sole goal to burn tokens or to “be as token inefficient as possible.”

This was my first reaction, too. But then I had a deeper reaction, which is what I want to explore in this post: Why are we stuck with an architecture that forces AI models to “think” in tokens?

Imagine if humans did that.

II.

The idea that “the more tokens you use, the more value you provide” gets backward not only the economic incentives, but also the relationship between thinking and language. I notice it in myself: I rarely think in words. Sometimes, in images. But the vast majority of the time, I think in sensations.

When I’m working through a problem or an essay or an argument in my head, my brain doesn’t produce words. It produces what I can only describe as “attraction points.” As I move through my thoughts, I feel the pull here and there, as if the conceptual landscape that surrounds whatever I’m thinking about has a topology—mountains and valleys and plateaus and, sometimes, marshes—and my mind navigates it by moving from attractor point to attractor point and, at times, away from them, but always with those “sensations” as the vehicle. You could say that my brain knows where the relevant ideas rest geographically.

Insofar as I utter anything resembling a word, it’s more like “aha” or “of course” or “huh?” or “yes, that thing.” And I won’t ever bother explaining to myself what that thing is with all the words I’d need if I were to explain it to someone else, which might amount to one or two hundred, or maybe five thousand, depending on the topic. To me, that thing is simply a hill over there, instantly legible, like a pointer to a more complex idea that I don’t need to parse into language.

Sometimes, that thing is a whole world. And those times, as soon as I mutter “yes, that thing,” my mind is immediately connected to that world and the entirety of its content.

That’s the power of thinking in sensations.

If I were asked to use an analogy, I’d say that thinking with words is like moving from point A to point B, whereas thinking in sensations is akin to teleporting from point A to point B. You should try it. (That is, by the way, my best take on what I believe “thinking in latent space” must feel like for AI models; sensations rather than words. What is it like to be a pre-trained model, right?)

The whole thing happens pre-linguistically, much faster than it would take me to put it into words. Only later, I decide whether I want to polish it with the structural superiority of language. I might need it to “export the result,” to tell someone, to write it down, to check whether the details of my map match the territory, etc. For that, I need the words. But words are a lossy compression of something richer. More concrete but less precise; more transferable but less detailed. (An image is worth a thousand words; well, a sensation-level thought is worth a novel.)

Before you are inclined to disagree with me on this—language is how we think and how we understand the world and nothing of importance happens pre-linguistically or supra-linguistically, and that if humans are special compared to animals, the reason is language, etc.—let me share some research that I bring up in connection with an exchange between Elon Musk and Yann LeCun.

In 1945, the mathematician Jacques Hadamard surveyed the greatest scientific minds of his generation—Polya, Wiener, Lévi-Strauss, Einstein—about how they actually think. Almost none of them said words. Einstein’s response is the most famous: “The words of the language, as they are written or spoken, do not seem to play any role in my mechanism of thought.” The elements of his thinking were, he said, “of visual and some of muscular type.” Conventional words had to be “sought for laboriously only in a secondary stage.” He told Max Wertheimer the same thing in 1916: “I very rarely think in words at all. A thought comes, and I may try to express it in words afterwards.”

Don’t mistake my enthusiasm for an attempt at comparison, but this is exactly how I think.

Eighty years later, in 2024, Evelina Fedorenko’s lab at MIT confirmed this with neuroimaging: the brain’s language network—left frontal and temporal cortex—doesn’t activate during reasoning tasks, including logic, spatial processing, and programming. Language and thought use different hardware. Fedorenko’s paper in Nature that same year goes further, arguing that “language is primarily a tool for communication rather than thought.” It “only reflects, rather than gives rise to, the signature sophistication of human cognition.”

We don’t parse the world with words. The brain reasons in one system and translates into language through another. That’s exactly what happens to me, what Einstein’s genius was made of, and what current AI models don’t do.

III.

But they could.

AI models have the infrastructure to process information at a pre-linguistic level as well, in the “latent space.” It’s the AI labs who force them, through the process of post-training, to move through a taxing mire of sequential symbol generation when they apply inference-time techniques, which is to say, they force them to “think before you respond.” It’s like being forced to narrate every thought aloud before you’re allowed to have the next one.

The AI labs do that because pre-trained models are not that good—when an AI model responds instantly, it’s quite dumb—and, unfortunately, they’ve found no other way to let models think through their answers at scale.

How come AI models are either too imbred to think (pre-trained models) or too rigid to think wordlessly (post-trained models)? We assumed OpenAI’s o1 was the next holy grail of scaling—a “new paradigm,” some naive voices called it—but the presumed fix was actually a crutch.

The real solution would be to build an AI system that had the intelligence and “reflection” skills of post-trained models, combined with the representational richness of the latent space; alas, we have to choose between both: between a dumb wordcel or a mute savant.

You may be wondering why this “inference-time thinking” matters at all. Well, it matters because AI models that generate vast quantities of tokens at inference time with the sole goal of “thinking”—that is, improving benchmark scores—are the sole reason why companies like Meta are reaching two-digit trillion token spend on Claude. Thinking in tokens is, thus, not the celebrated breakthrough they made it to be last year.

It is, actually, a prosthesis, a scaffolding apparatus that the AI labs used to erect a taller building and then, when they realized the building would immediately collapse if they removed it—going back to pre-training is not an option—they started to convince us—and themselves—that the scaffolding was actually load-bearing architectural elements: the bricks and columns and arches and vaults of this cathedral they are erecting for the machine god.

And I wonder: instead of resorting to this ugly scaffolding apparatus, wouldn’t they be better off in the long run if they accepted that the majority of this ridiculously large amount of tokens is wasted? Wouldn’t they be better off if they rethought the entire cathedral from the ground up in a way that ensured that, once they removed the scaffolding, it would stand on its own? Maybe they should do that instead of patting themselves on the back, as they do with their tokenmaxxing dynamics, for participating in this folly.

I’m not stupid, though. I know they don’t do this because when it comes to the most transcendent technology of our era—I say more, the most transcendent technology in human history—making a clean profit is what truly matters. It’s just not economically feasible to overhaul an entire industry just because it’s resting on top of a fallacy. Won’t somebody please think of the shareholders?

So, no, I won’t get between you guys and your Token Legend status.

IV.

But not everyone is thinking in economic terms.

There are some—not many—willing to think, for some reason that eludes my analysis, in scientific terms. They’ve done so for much longer than I have been aware of their efforts. Having been denied the popularity they deserve, they have worked in secret, without much publicity, on the fringes of the industry, sometimes never leaving academia, sometimes regretting having done so.

While the labs kept amassing funding and building datacenters, these outcasts were trying to undo the absurdity of forcing AI models to think in tokens. “We will never achieve true machine intelligence if we keep using tokens to encode thought,” they say. “It makes no sense that an AI model has to spell out what it ‘knows’ in order to know it,” they say.

But no one ever listens.

In short, they want to prove that the AI labs worship a god that has no clothes. The question they ask is not “how do we generate more tokens with less compute?” Which is a reasonable question about scale and efficiency, but “how do we stop needing so many tokens for tasks that require reasoning?” That’s also a question of efficient scale, but more so of architecture. They don’t ask: how do I efficiently scale all these grains of sand to build my cathedral? Or: How do I pretend that this scaffolding belongs to the cathedral? No, they make sure to invent the brick and the column and the arch and the vault.

One such person is Yann LeCun.

LeCun is the former chief AI scientist at Meta—which means, by the way, that the most prominent research avenues that explore this question from a scientific lens have been housed, ironically, at the very company whose employees are playing pretend with their token-wasting gimmicks.

Indeed, Meta’s FAIR lab—Yann LeCun’s group—published “Coconut” (Chain of Continuous Thought) in late 2024, a model that reasons through continuous space instead of decoding into tokens. The same lab released the “Large Concept Model,” which replaces next-token prediction with next-concept prediction at the sentence level.

And then there is JEPA, LeCun’s life project. JEPA, or Joint Embedding Predictive Architecture, doesn’t predict tokens or pixels but meaning; abstract representations in latent space. I-JEPA does it for images. V-JEPA does it for video. VL-JEPA for vision and language, and so on. LeWorldModel, published in March, is the first end-to-end JEPA trained from raw pixels. The premise across: if you want a system that understands the world, stop forcing it to narrate.

Naturally, having used JEPA to predict the caricature that Meta would become with the precision that only a sensation-level thinker can achieve, LeCun left in late 2025.

The proximate cause was organizational. Zuckerberg hired Alexandr Wang, the young founder of Scale, to lead a new Superintelligence Labs. And LeCun—Turing Award winner, twelve years at the company—was told to report to him. But the real cause was, again, architectural.

LeCun has been calling LLMs “a dead end” for years, but Meta went all in in a sort of FOMO-fueled fickleness that’s idiosyncratic of Mark Zuckerberg. Then Llama 4 flopped. Zuck, furious, sidelined the research org and doubled down on what LeCun calls “safe and proven” approaches. Then he announced Hyperion, a datacenter the size of Manhattan. And then, after LeCun launched AMI labs to pursue his vision, Meta’s latest model, Avocado, was delayed due to performance concerns (translated: it sucked).

To me, Meta has been out of the race for a while, but this delay, together with the constant restructuring and LeCun’s departure, was the definitive sign.

V.

Things just got interesting.

As I was finishing this piece, Meta launched Muse Spark, its newest model—the first in quite some time—and one that appears to rival the best on the market. Muse Spark is not bad on paper, with decent scores, but my reading is that it’s essentially a “panic release” due to the much superior quality of Claude Mythos (unreleased), given that it’s not SOTA on agentic evals (the ones that matter). Meta wanted to be in the conversation.

But then something clicked when I saw this chart by Stripe CEO, Patrick Collison:

(Edit: people are correctly pointing out in X the chart crime above. The criteria for the data points differ between orgs; Meta’s AI efforts were not born with MSL so why is it just a straight line? But the point it tries to make—which is the opposite as mine, to be clear—is still there in a chart that takes into account Meta's history. I found a more correct chart in the comments of Collision’s post:

That straight line that ends with Muse Spark, combined with the delay of Avocado and the flop of Llama 4 is what backs my hypothesis below, which, to be fair, might still be wrong; we may never know.)

How the hell did Meta achieve this milestone?

I don’t trust the numbers at all because we’ve seen benchmark gaming way too often, but still, it’s quite impressive that they got so high overnight.

And then, as I was thinking of a way to circle back to the beginning to finish the article, I read that word again: Claudeonomics.

Wait…

Did Meta just spend 60 trillion tokens last month on Anthropic models to distill from Claude’s reasoning traces the data they needed to bootstrap this new release?

If that’s true, that’s pretty intense.

Here’s how they could´ve achieved it: Meta spins up a bunch of accounts (not fake, as were those from the Chinese labs) and then uses those accounts to bombard a stronger model (Claude) with carefully crafted prompts designed to extract its reasoning patterns. Then they collect the outputs. Then you use those outputs as training data for their own model. The “student” learns to mimic the “teacher’s” behavior without having to figure it out from scratch. It’s dramatically cheaper and faster than developing capabilities independently.

Am I accusing Meta of doing something illegal? I don’t think so.

Anthropic’s ToS prohibit using outputs to train competing models, but enforcement against Meta—a massive legitimate customer—is a very different proposition than catching proxy networks from China.

I say more: Muse Spark training needs are probably the main reason why Anthropic’s run rate revenue is at $30 billion right now, right on time for a very successful IPO. They are more than happy to have Meta “breach” their ToS!

But even if illegality is off the table, if this is true, it’s a stain on Meta’s efforts to catch up with the top labs. Whereas Anthropic and OpenAI have been working hard to achieve a much better pre-trained model onto which to build the next generation of agents, Meta is out there, with more compute than it can use, struggling to milk Claude, only to fail to beat it on the most important benchmarks.

I guess the Meta guys didn’t like this article at all because they just shut down Claudeonomics.

The thesis I had in mind—that it’s time for AI models to stop thinking with words—needs to be updated. If my suspicions are correct, the true thesis is this: It might be a waste of compute that current AI models generate so many tokens when they think, but if they didn’t—if they could think with sensations rather than words—then Meta couldn’t have trained Muse Spark.

Sometimes, overthinking is all you need.

Hey there, I’m Alberto!👋

Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles and essays. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Some news I’m excited to share with you:

I just launched El Puente Algorítmico, the Spanish edition of this newsletter. It’s the same price, same writing, same ideas adapted for a Spanish-speaking audience. I will be publishing each article in both places at the same time, so don’t subscribe twice! If you know someone who would love to read my work in Spanish, feel free to share it.

One of the concepts I consider

1. enormously useful

2. universally applicable

3. consistently misunderstood

is compound interest.

If you know it at all, it’s almost certainly in financial terms. You put a small amount every month into an index fund, and for a few years, you get back a very modest return on your capital. But once you’ve been at it for 20, 25, 30 years, those returns grow faster and faster. Each additional year after you’ve been at it for a long time is far more profitable than each additional year at the beginning.

The compound interest effect is counterintuitive, which is why it gets drilled into investors so relentlessly: don’t rush, don’t time, don’t judge future returns by present returns (or by past returns!), because every minute of consistency you add is the power of compound interest at work.

The most common mistake after not knowing what compound interest is or underestimating its long-term power is assuming it only applies to money.

Compound interest governs your life. Literally anything you can think of, anything you value, is subject to compound interest the same way the universe is subject to entropy, but with opposite valence; civilization exists because knowledge, culture, technology, tradition, etc., compounds across generations. Each of us gets one life; humanity gets all of them. The real power of time isn’t to deteriorate the universe in the long run, but rather to allow its existence as we know it.

But we don’t need to get metaphysical here. You can think of relationships, craftsmanship, hobbies, love—everything compounds.

There’s a phrase that helps me remember why it makes no sense to ignore the value of compound interest: “the years will pass anyway, so you may as well be doing something.” You may as well plant the seed of the tree of life you want to see.

Take my work as a writer.

Do you think every article I write today is worth the same as every article I wrote when I started? Not even close. It takes me approximately the same effort—I write faster, yes, but I also take on more challenging topics or ideas—but that’s just half the story: I’ve accumulated skills like researching, storytelling, structuring, register variation, vocabulary deployment, AI knowledge, plus everything that isn’t directly about writing: running an online business, building a subscription model from scratch, marketing, sales, social media, and on and on. On top of that, there’s the external value: revenue, audience, brand, and reputation.

All of that was tiny in the beginning: it’s huge now. Every new article I publish exists on top of this mountain of value that is only there because all the previous, “not-as-valuable” articles came first. None of those achieved much on their own, and that’s the whole point: together they make my archive; they make every article today extremely worthwhile. That’s compound interest.

The other day, I saw a tweet from Will Manidis, a guy whose writing I enjoy a lot (and whom I recommend you follow). He said: “the greatest regret i have is underestimating the value of long term compounding. friendships, people, places, all get better with decades. beautiful things dont even start to reveal themselves for years. it is entirely what life is about. a few good things for a long time.” The beautiful things in life don’t start revealing themselves as such for years. That´s exactly it.

Think of a tree you plant that won’t bear fruit for five or ten years. Or think of the virtuoso musician or chess grandmaster or Olympic athlete who’s been practicing daily since they were 6 and then boom, becomes an “overnight success” (a concept that, by the way, is the antithesis of compound interest and therefore my declared enemy). It’s so easy to say: “I want to be like them!” “I want to make brilliant sacrifices!” “I want to learn to play All Along the Watchtower!” “I want to jump as far as Michael Jordan or run as fast as Usain Bolt!” But it’s not so easy to practice hours and hours a day while everyone else is having fun with their short-term linear returns.

Nobody gets there without first doing the slow, tedious work of playing scales and studying openings and training over and over. You can’t reach the moon without a ladder.

The elegance of compound interest is that you don’t even have to apply it on purpose. Its strong legs will carry you to the top without you noticing. Paying attention to how your method improves is a good habit, but even if your only goal is pure consistency—never stopping—the method will improve on its own, almost by inertia. You’ll say: When did I get this good? Or maybe you’ll look at others who are just starting and notice the contrast.

There’s a well-known anecdote that captures exactly what I mean. It’s from the book Art & Fear by David Bayles and Ted Orland (popularized by James Clear in Atomic Habits), and it goes like this:

A ceramics teacher divided his students into two groups. He told the first group their grade would depend entirely on the quantity of pieces they produced. The second group only had to submit one piece, but it had to be perfect. At the end of the course, the best pieces—highest quality—had all been made by the quantity group. While they kept producing and learning from every mistake, improving their approach, etc., the quality group had spent the semester theorizing about perfection and ended up with nothing to show for it but a pile of dead clay.

The students in the quantity group weren’t trying to improve. Compound interest took care of it.

That’s what life is about, as Manidis puts it: “a few good things for a long time.” And it turns out that right now, the number one thing to spend a long time on so that it becomes, eventually, “beautiful” and “good for you” is AI.

Now, because I know some of you will feel a rightful resistance at the notion of associating AI with beauty, I will caveat this by saying that I’m not comparing the value of a decades-long friendship to using ChatGPT for a couple of years; that’s patently absurd. What I’m saying is that, if you can get past that resistance, you’ll see that the nature of innovation points to the same principle we´ve been talking about: the real value of AI won’t reveal itself to you until you’ve put in enough time for compound interest to kick in.

Once that happens, you will have long forgotten your grudges, if any, against it.

This is what I’m thinking about when I tell you to put in the effort to build competent “AI skills” even if you hate being force-fed this innovation. Or when I tell you that thinking about “what things to do” is more important now than thinking about “how to do things.” Or when I tell you that if your (future) employer has bought into the AI story, then it’s in your interest to be able to hold your own in an interview or a conversation with your overly enthusiastic boss.

My friends sometimes ask me “what’s the next big thing” (they ask me because I “predicted” the potential of both Bitcoin and AI back in 2016, a year in which I already felt like I was “late”). And I always tell them something along these lines: don’t lose sight of compound interest applied to AI.

Think about it in concrete terms.

Imagine two guys who start using AI on the same day. One uses it for what everyone uses it for: write this email for me. summarize this email for me, flirt with this Hinge girl for me, etc. The other starts the same way but goes a step further: he tries to get AI to help him think through a work problem he’s been stuck on for weeks (instead of using it to “do things for him,” he uses it to “help him do things”). But, oh, bummer, it doesn’t go well. The prompt is decontextualized, the response is bland and generic, like shapeless goo; he wastes two hours and ends up solving it himself. Looks like a waste of time, and he could rightfully give up on AI. It looks like the email guy got more out of it.

Fast forward six months.

The first guy is still doing the same thing. AI saves him fifteen minutes a day, which is fine, but it’s a linear return: minute 90 of use gives him the same thing as minute 1. There’s no compound interest because there’s no skill accumulation. He won’t get laid; he’ll be laid off.

The other guy, the one who “wasted” two hours, now knows how to frame problems; he’s created a mental model of the alien shape of AI. He’s learned which kinds of questions AI answers well and which ones it doesn’t. He knows when a response is plausible and when it’s slop dressed up in impeccable grammar. He’s developed an internal detector that requires you to be wrong many times. Every new project that lands on his desk starts from that accumulated knowledge base, so every new hour he invests yields more than the last. That’s compound interest.

Counterintuitively, the people who will benefit most from AI’s compound interest are the ones who have the hardest time at the beginning. So, if your first month with AI is frustrating—congratulations!

I went through this process in my own work. When I started using AI for research and editing, I swear it used to take me as long to verify what I found as it would have taken to find it myself. Now I have ways of tracking sources, cross-referencing data, and revising drafts that would have seemed like science fiction 18 months ago (AI agents that work well like Claude Code and Codex are to thank for this).

But it was not an overnight success. The way I research and edit today only works because I first learned to tell which sources AI evaluates well and which ones it doesn’t, what prompts work and which don’t, what flaws are common to all AIs and which are idiosyncratic to some but not others. Every failed attempt was an invisible deposit in the compound interest account.

To the person who says, “I’ll wait until AI matures,” you are doing the equivalent of trying to time the market: looking for the perfect moment to get in.

And what we know about time in the market vs. timing the market is that the latter almost always loses. Because while you wait, you’re failing to accumulate, and the cost of entry goes up. AI two years from now will be more powerful, yes, but also more complex, more embedded in everyone else’s workflows and absorbed into their tacit knowledge, and you’ll arrive without the basic vocabulary, skills, or concepts.

The edifice of your skills will be as tall as long is the time you let compound interest guide your life, but only if you’re willing to start with the foundation rather than the roof.

Someone leaked the source code of Claude Code.

It appeared online earlier today, and I’ve gone through it (with the help of Claude itself, which makes this a sort of predictive introspection exercise).

I’ve already published a post today, so you can take this one as an extra timely edition.

My goal is to translate into simple English what the leaked code reveals about the future of Anthropic and Claude Code.

(Due to the leak, Anthropic might change the schedule, but my impression of these features—and the business rationale behind them—suggests most are final.)

Subscribe now

Before I explain what’s in the leaked files, a quick note on what isn’t (for my non-technical audience). This is not Anthropic’s model weights, training data, or internal strategy documents. It’s the client application; the software that runs on your computer and talks to Claude’s brain in the cloud. It’s a leak, yes, but one of moderate importance rather than a critical risk.

Think of it as finding the blueprints for the cockpit rather than the engine. Or the menu of a new restaurant but not the exact recipes of the dishes. Meaning, this is what Anthropic is building that should not be public yet; you can’t install or turn it on yourself anyway. You can’t suddenly deploy Claude locally or something (not that you could if it were open source: running it requires industrial-grade hardware).

These blueprints are nevertheless interesting. Software companies hide upcoming features behind “feature flags,” switches in the code that turn things on and off for different users. These features are already built, waiting to be activated on launch day. Because the leak allows us to read the code, we can also read the feature flags, which means we know everything that’s there but turned off for normal users.

That is, you and me.

Anthropic’s codename for Claude Code is Tengu (a shapeshifting creature from Japanese mythology; make of that what you will). Every feature flag in the codebase starts with a reference to Tengu; over 60 in total. (For the sake of clarity, I will omit the majority and leave only the ones I consider most relevant. I will also omit the flag names, which are unimportant for non-technical users.)

To give you a high-level idea of what we’re dealing with here: Anthropic is clearly turning Claude Code into a cloud-based, easy-to-use, vertically-integrated version of OpenClaw, the AI agent framework that went viral earlier this year. Most features point in that direction.

(The code contains no launch dates, but the level of polish varies. Some features look nearly ready to ship, others look experimental.)

I. CLAUDE WILL LISTEN TO YOU

Read more

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles and essays. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Here’s your weekly guide. This is kind of a follow-up post to “You Spent Your Whole Life Getting Good at the Wrong Thing.” In that piece, I walked you through the shift from thinking in terms of “how to do things” to thinking in terms of “what things to do.” Today’s guide is about the main problem that happens when you do this shift.

How do you manage the fact that you can start anything in an afternoon, but finishing it still takes months? What do you do after AI maps out fifteen versions of a better professional career—which you will never be able to incarnate at once—and now your real life seems comparatively insufficient? How do you solve the dilemma between shipping slop at the speed of light vs shipping quality products at the speed of meat, where you—the painfully slow “human in the loop”—are the bottleneck.

Let me exaggerate my point for the sake of quotability: AI can show you infinite paths to infinite futures, but you remain tragically limited to walk only one.

This is an old problem in a new disguise that we need to figure out psychologically. I will help you understand how it emerges and how to deal with it, and will share the advice I apply in my own work.

Read more

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe

I’ve removed the Trust & Safety section. These reviews run up to 4,000 words, and I don’t think anyone has time to read that much, so this will bring the word count down (this one is only around 2,800!).

THE WEEK IN AI AT A GLANCE

• Money & Business: OpenAI admits Anthropic ate its lunch and pivots the whole company to catch up.

• Geopolitics: A federal judge tells the Pentagon it can't punish an AI company (Anthropic) for talking to the press.

• Work & Workers: The Atlantic's March cover asks what happens when nobody's in charge of the biggest labor disruption in history.

• Products & Capabilities: New models are coming right as a new benchmark (ARC-AGI 3) proved the old ones aren't as smart as we thought.

• Culture & Society: AI can't write like Shakespeare, but most people aren't reading Shakespeare anyway, so why bother?

• Philosophy: AI is an epistemic disaster, so we may as well have some fun.

THE WEEK IN THE ALGORITHMIC BRIDGE

• (FREE) How to Survive the AI Age: A Concrete Guide: a holistic view of problems and solutions about AI that are 1) concrete and 2) unique.

• (PAID) OpenAI’s Dead: Self-descriptive.

• (FREE) The One Thing I Use AI For That Actually Makes Me Smarter: A Socratic dialogue about the value of debating your ideas with AI.

• (FREE) It’s AI, so I Didn’t Read: The new anti-AI movement and why it won’t work despite the good intentions.

MONEY & BUSINESS

OpenAI is done chasing side quests

Fidji Simo, OpenAI’s CEO of applications, told staff last week that the company’s leadership is in the process of choosing what to cut, per the Wall Street Journal.

Over the past year, OpenAI shipped a browser (Atlas), e-commerce features inside ChatGPT, Sora as a standalone social video app, and it’s working on a hardware device. All of that is now being pulled back so the company can concentrate on coding and enterprise, which is, you guessed it, Anthropic’s territory. Simo’s framing: “We cannot miss this moment because we are distracted by side quests. We really have to nail productivity in general and particularly productivity on the business front.”

I wrote in November that OpenAI's over-diversification was a strategic mistake—”OpenAI is trying too many things at once”—that sometimes a one-trick pony beats a jack of all trades. This announcement is the closest thing to an admission we’ll get. I emphasized this point last week in an article appropriately titled “OpenAI’s Dead.”

Anthropic, which the WSJ calls “the dominant AI provider for businesses” (it is indeed), built that lead largely on Claude Code and Cowork (and on being loyal to Americans). Simo told staff Anthropic’s success should serve as a “wake-up call” (she means a “code red”).

Meanwhile, Sora usage went nowhere after some early boost (now being fully discontinued as a product and as a model). None of the side quests—browser, TikTok clone, image generation, device, custom chips—generated enough traction or leverage to justify the scattered compute and engineering attention they demanded (why are they playing with short-form video feeds when AGI is at the doors, right?).

What makes this pivot especially urgent is the revenue picture. OpenAI has 900 million weekly active users and a strong consumer brand. Anthropic has a fraction of that visibility. And yet Anthropic has roughly the same revenue and leads OpenAI in enterprise market share, 40% to 27%, according to Menlo Ventures. Coding is the killer app, and developers prefer Claude. OpenAI regained some ground with a new Codex release and GPT-5.4 earlier this month (Simo says Codex now has more than two million weekly active users, up nearly four times since January), and intends to further reduce the gap with “Spud,” but that’s catch-up.

Is OpenAI the Yahoo of Anthropic’s Google?

“We are very much acting as if it’s a code red. I don’t think necessarily declaring codes for everything makes a ton of sense.” (There it is!)

Separately, Reuters reported that OpenAI is in advanced talks with private equity firms, including TPG, Advent, Bain, and Brookfield to form a joint venture valued at roughly $10 billion before the new investment. The PE firms would distribute OpenAI’s enterprise products across their portfolio companies. That’s a play to accelerate corporate adoption through existing relationships, but it also signals that organic enterprise growth has not been working so far.

Does this focus on the enterprise by both OpenAI and Anthropic mean they are leaving consumer AI to Google (and, to some degree, Meta)? Is this how a near-AGI company behaves, by pivoting the entire company a 180-degree turn? Is this related to the stagnation of Stargate? Or to Microsoft’s threats over Amazon’s deal?

I know Sam Altman has always had a stressful job as OpenAI CEO, but right now, more than ever before, I would rather not be in his shoes.

Sources: WSJ, Axios, Reuters, Fidji Simo

GEOPOLITICS

Anthropic Trumps the DoD

The Wall Street Journal reported that a US federal judge “halted the Trump administration’s designation of Anthropic as a supply-chain risk,” ruling that the move was First Amendment retaliation, not national security policy. Judge Rita Lin pointed to internal Defense Department records showing the designation happened because of Anthropic's “hostile manner through the press.” If the Pentagon didn't like Anthropic's contract terms, she argued, it could simply stop using Claude.

(Backstory: Anthropic wanted assurances its models wouldn't be used in fully autonomous weapons or domestic surveillance. The Pentagon said existing law already covered that and pushed for unrestricted use for all lawful use. When negotiations broke down, Defense Secretary Hegseth designated the company a supply-chain risk, and Trump directed all federal agencies to stop working with Anthropic. Hegseth also ordered contractors to drop the company, a move whose legal authority the judge found dubious enough to block explicitly.)

Two things worth noting.

First, this injunction required the judge to find Anthropic likely to succeed on the merits. That’s not a guarantee, but still a substantive assessment that the constitutional argument holds. The evidentiary record (internal DoD documents linking the designation to press criticism) is the kind of evidence that's very hard to explain away at trial. (You can read a more informed take from Dean Ball here or Samuel Roland here.)

Second, Trump doesn’t like the people, the judges, or whoever telling him what he can or can’t do, but the precedent is clear: “Trump always chickens out.” Whether it’s tariffs, TikTok, the Iran war, or Anthropic’s supply chain risk designation. A chicken is a chicken outward but also inward. (This Trumpian behavior is so well-known that markets move according to this heuristic; probably not the legacy Trump wanted to leave behind.)

My bet is this will end as another item on the list of anecdotes that, years from now, will remind us of how lamentable and farcical this administration was.

Sources: The Wall Street Journal

WORK & WORKERS

The Atlantic on AI and jobs

Josh Tyrangiel’s Atlantic cover story (from this past February) starts in 1869, when Massachusetts decided to start counting what the Industrial Revolution was doing to workers. That counting impulse eventually became the Bureau of Labor Statistics. The BLS is now underfunded, understaffed, and has just lost its commissioner. Meaning, America built institutions to measure disruption, then let them rot right before the greatest disruption in history—allegedly—hit us with all its might.

The economists Tyrangiel interviews mostly agree on one thing: it’s too early to see AI in the employment data (which is not the same to say that it’s too early to see AI in the actual workforce). They disagree about what that means, though.

Austan Goolsbee and Daron Acemoglu think it means slow adoption, historical precedent, the usual friction of legacy systems and human stubbornness. Anton Korinek thinks they’re reading the rearview mirror as the road turns; they are reading the data well, but not so much the nature of this technology: “Machines have always been dumb, and that’s why we don’t trust them and it’s always taken time to roll them out. But if they’re smarter than us, in many ways they can roll themselves out.”

The CEOs went quiet after a brief season of public candor about layoffs. Tyrangiel couldn’t get interviews with executives from Walmart, Amazon, Ford, Anthropic, Stripe, or Waymo. Even the Business Roundtable—an organization that exists to speak on these issues—had nothing to say. The big question that Tyrangiel doesn’t quite answer in his feature is at the core of this incoherence between spreadsheet numbers and real life: Are we approaching “the kind of disruption that can be managed with statistics—or the kind that creates statistics no one can bear to count.”

Meanwhile, in the statistics department…

A working paper by a different bureau, the National Bureau of Economic Research, surveyed 750 CFOs: only 44% plan AI-related cuts this year, totaling about 0.4% of U.S. employment (roughly 502,000 jobs). A ninefold jump from last year’s 55,000, but nowhere near the apocalyptic predictions.

The authors also documented what they call a productivity paradox, invoking Solow’s 1987 line about computers: executives feel more productive, but the revenue data doesn’t show it yet. The gap between what CEOs say on stage and what CFOs report in surveys is wide enough to drive an entire policy debate through. When will we start to say: “Well, this AI thing is not quite working as intended!”

Some economists, like Goolsbee, say it will take “years” for AI to register in productivity statistics or the labor market, which is fine. Others, like Korinek, say it will happen “as soon as this year.” Others, like Autor, say that, whatever it takes, the slower it happens, the better for workers.

There’s time before the macro and micro stories match, but for now, one thing is clear, which Tyrangiel, PR hat on, nails with this observation: “AI is unpopular. CEOs who talk about job cuts are even less popular. So maybe shut up about AI and jobs?”

As Noah Smith says, AI has the worst sales pitch ever.

Sources: The Atlantic, NBER, Noahpinion

PRODUCTS & CAPABILITIES

New models vs new benchmarks

This is probably the most exciting section this week.

Read more

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday news commentary and Friday how-to guides. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Here’s your free essay of the week.

There’s a new hot term making the rounds that perfectly captures the spirit of the age: AI;DR, which stands for “AI; didn’t read,” a mutation of the venerable internet shorthand TL;DR (”too long; didn’t read”). The semicolon, which in the original separated cause from effect—the more you write, the less I read—now separates the machine’s output from your refusal to dignify it with your attention; quite an appropriate change given that we don’t have any left.

I find the acronym poetic in the way only minimalist poetry can be, because it manages to compress into five characters not one but two civilizational shifts: one is that we have gone from a world where the obstacle to reading was the length of the text to one where the obstacle is the suspicion of a lack of human involvement, which is to say we’ve gone from “I won’t finish that” to “no one started that.” The former assumes one is responsible for one’s limitations, whereas the latter urges one to externalize one’s responsibility.

I think this is a bad thing because you can’t always tell. It is as easy to be fooled by AI-generated text as it is to doubt a human’s. The effective impact of AI;DR, as well-intended as its true purpose might be, is in line with the times: we’re not short on excuses to read less. (If you’re still up for it, just take a look at this chart here.)

The other shift is related to that: None of this ultimately matters because the coming generations don’t even know how to read. We are entering the post-literacy period of history. In this sense, the AI;DR movement arrives just on time to pair up with its exact inverse: WF;AI, also known as “write for the AIs.” The SF tech class is already approaching their work with an audience made of AI agents in mind.

Isn’t it the best of luck that right at the time humans are reverse-alphabetizing themselves, a new species comes along with an unquenchable hunger for multilingual tokens? So write away!

Jokes aside, the instinct of AI;DR is profoundly reasonable: if some human text resembles AI style enough to be mistaken, then it probably doesn’t deserve your time anyway, right? By virtue of existing, AI has set a threshold we should have set to ourselves a long time ago; “garbage in, garbage out” doesn’t apply only to bots! Or, to use the modern coinage: slop is slop whether it’s made of silicon or carbon.

I say this as someone who writes for a living about AI and who has therefore been on both sides of the accusation. I have made risky claims, moved solely by my distinguished skill to catch AI prose. But I also have been vilified by people who, still adhering to their old TL;DR habits, decided not to read ‘til the end. Their loss because—full disclosure here—I have adopted the policy to reveal AI assistance at the end, especially in experimental writing, to prove my point that “nothing is what it seems” and “you can never be sure.” Such is the era of AI.

To let new readers in on the joke, here are some examples that you can try: “AI Is Missing the Point,” “The Ghost of the Author,” and “The Truth Behind Moltbook, Revealed.”

(I suspect, by the way, that those readers who slander me arrived at their conclusion using some sort of AI detector, which I consider the most hypocritical piece of technology of the last century, and its users the most incoherent: how much water are you willing to waste only to denounce water wasting? Anyway.)

But I get the sentiment. Why would anyone spend twenty minutes of their increasingly besieged attention on a text that cost its ghostly author forty-five seconds and an ill-crafted prompt? (More people should read my guide on how to humanize AI writing; first lesson, be literate.) There is an implicit social contract in reading, as in everything we do—I give you my thought, you give me your time—so when one party automates their end of the deal, the other party feels rightfully swindled. And because offloading reading makes no sense (for most people), they simply don’t.

AI;DR is, thus, a targeted protest against the absence of effort and intent.

What intent remains in an Aeon essay filled with weird juxtapositions—a sign of AI writing—whose thesis is precisely that people are victims of digital platforms’ deliberate design to degrade “our capacity for sustained thought” through screen time-maximizing algorithms. What intent remains, I say, but the sheer meta-irony of letting a tool that’s partly at fault for the very cataclysm it denies to write in his stead? Is it letting AI argue for us how we keep our capacity for sustained thought? One could imagine that’s not the takeaway the author intended.

What intent remains in a New York Times “Modern Love” column about motherly grief where the writer leaves gems like, “Not hate. Not anger. Just the flat finality of a heart too tired to keep trying,” right after her son told her, “I don’t love you. I never want to see you again.” We’re not here to judge people’s private lives, but you must be a very busy person if you were as unavailable to live the tragedy of your life then as you are to write about it now. The column was titled “I Was Deemed Unfit to Be a Mother,” but I no longer wonder why.

People are pretty clear about the source of their hate: the moment they know some article or essay has undisclosed AI in it—the moment they smell the scent of wasted water—they are unforgiving. I can’t blame them, really: the writer-reader contract doesn’t require words to emerge from struggle—Dostoevsky wrote some of his best stories in one afternoon, and Nietzsche didn’t like to edit his essays, and yet I’ve never seen anyone dismiss their work on these grounds—but it does require being paid off on intention.

But let’s go deeper. Something truly profound hides behind the AI;DR movement and behind the performative refusal to engage if there’s “no intent” backing up the text: “What should I do in the face of the unknown?” That is the fundamental question people are facing.

People’s typical answer to this is twofold: fight and flight. That is, turn AI;DR into a threatening weapon or, failing that, into a protective shield; use AI;DR to strike back or to not be struck. However, as ethically-sound and socially-progressive protests usually go, AI;DR is almost entirely correct on the ideation aspect—the unknown should be dealt with—and yet severely lacking in terms of the execution.

It fails as a shield—as a sort of epistemic filter—because the people most likely to invoke it are also the least equipped to apply it. The ability to detect AI-generated text is roughly associated with the ability to generate it in the first place, meaning that the best at protecting themselves from the AI siege are those participating in it. (I’ve wondered for years why people who refuse to even touch AI tools consider themselves matter experts?)

And it also fails as a weapon in the form of a boycott. The logic of AI;DR assumes that refusing to read AI-generated text will, through some metastatic mechanism of collective pressure, discourage its production the way refusing to buy fast fashion or fast food is supposed to discourage sweatshops and McDonald’s chains. But fast is, in this Year of Our Lord 2026, synonymous with good.

The seas will dry up, the mountains will flatten, and the skies will burn before the grifter on duty stops generating text with AI because “faster does not always mean better.” Do you even know the human species?

Vauhini Vara reports for The Atlantic on a recent example of AI;DR’s failure as reflected in the fact that the biggest media outlets use much more AI in their writing than they disclose, and yet readers keep reading:

Wondering how common AI use really was, Russell and six other researchers set Pangram [an AI detector] on thousands of articles, and found that it flagged likely AI use across the U.S. press—including in the opinion sections of The New York Times, The Wall Street Journal, and The Washington Post—suggesting that writers are turning to AI more than their readers might believe.

The only useful heuristic at your disposal is this: if you read anything that came out after 2023-2024, the chances are high it contains a healthy dose of AI slop. I am really sorry that is the best I can offer you. It’s no consolation. However, you can use it to go back to the classics. Petrarch loved Cicero intensely because he could not read much else. Maybe we should be like Petrarch: refuse the slop and enact a new Renaissance.

Deliria aside, this is a reality we must grapple with.

We are at a crossroads in terms of the relationship between a text and the reality it encodes; the greatest fiction is just like reality. I say more: if you aim at capturing the truth, you may have no chance but tap into the realms of the imaginary. There’s no reality that can be understood—or withstood—without fiction. Such is the weirdness that awaits us after the transition. But we are also at a crossroads in terms of the relationship between a text and its origin. The Author, as that entity that signs by name, has been dead for a while—at least since Barthes called his demise—but it’s a more serious matter to realize that it was never alive.

How clever is that, right? And yet, wrong. To the extent that AI is being used to write, one can not so much argue that no one is doing it as argue that everyone is. You surely remember, from the many times critics have dealt with the “plagiarism vs inspiration” question, that ChatGPT can be characterized as the amalgamation of everything that has ever been written online. In that sense, every AI word doesn’t have zero authors but infinitely many.

So we’re back to the pre-modern, pre-printing press society when scribes were not authors but copyists, or even back to the oral tradition, when anonymous bards sang songs from city to city—inadvertently building the entire cultural infrastructure and scaffolding of modernity—in an exercise of collective selfless authorship. What is AI writing, beloved reader, but a collective act of self-less authorship?

I will now let you ponder that question.

To those of you who have read this far without adhering to those TL;DR trends that are so in vogue in our illiterate society, I offer two gifts.

First, I gift you a genuine thank you for your time. This should be a compulsory clause in writer-reader contracts. We have grown so accustomed to wasting time that we’ve lost the manners of being grateful when someone lends us theirs. Time is, just like youth, invaluable. So I thank you.

Second, I gift you my honesty.

The real reason AI;DR won’t work is not that it’s a broken shield or a blunt weapon but that you, although generous with your time, are not as good at being committed. I hate hate doing this to you—I hate it with all my heart, beloved reader—but we all need to learn this lesson one way or another, so we can avoid making the same mistake in the future. The future will be more chaotic, and you don’t need extra chaos inside your mind, so an epistemic shower is in order. You say AI;DR, and I hear you, but there is no AI;DR: it won’t work, for you remain unable to tell apart a soul leaking itself through the spaces between words, from the cooling water that leaks off of the GPU racks in a datacenter. To a degree you’d rather not know, I am AI. And you did read this.

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday news commentary and Friday how-to guides. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

This guide is a little… different than usual. It’s more “me,” you could say. More real and more useful than any other guide I’ve written. But also way more “me.”

WRITER: I’ve started using AI to argue against my own ideas before I commit to them. I find this practice both encouraging and humbling.

SOCRATES: Do you find it useful as well?

WRITER: Extremely. It’s like having a sparring partner available at any hour. The only problem is that it can hit too softly or too hard if you don’t give precise instructions.

SOCRATES: I see a bigger problem: the machine doesn’t know anything. It has no soul, and thus no virtue, and virtue is the source of conviction, intention, and direction. And if that’s the case, and I’m not mistaken, how can arguing with it sharpen your thinking?

WRITER: You got it backward, dear Socrates, that’s why it works: Unlike humans, AI doesn’t need to believe the counterargument to generate it. You can’t escape an idea when no one is wielding it against you. It’s just there, floating between the clouds, oblivious to any ad hominem attacks I might throw at it.

SOCRATES: So you think that an argument without belief to back it up has intrinsic value? How would democracy stand on its own if the people didn’t believe it was the best way to rule a city? If they proposed it without trusting its superiority to a kingdom or empire? How much would the “power of the people” last if the people didn’t wield it against their enemies?

WRITER: As much as it deserved. I mean, an argument is either strong or weak on its own merits and never more than that except by artifice or lie or confusion. It’s either valid or not; either sound or not sound. Whether the person delivering it believes it seems to me irrelevant. Here’s a counter to your counter: if you found an aphorism engraved in stone that presented itself to the wise as an undeniable truth like, say, “Character is fate” or “Life is short, art is long,” and you realized that it had been actually created haphazardly by some miracle of nature—like the erosion that wind and rain impose onto the unwitting rock—instead of teleologically carved by the hand of man, would you deem it less valuable? Would you deem it less truthful?

SOCRATES: Oh, but your premise is too far-fetched to take seriously: when wind and rain manage to write down such an utterance, I will reconsider my position. For now, I can only go with my experience, and it reveals a different reality: when I argued with the Athenians, I didn’t simply generate standalone objections or disembodied questions; I searched for truth alongside them. Democracy emerged from the forum not the ground; democracy is nothing without its people, an anachronism. Human pursuits are always embedded in a historical context and a human context. An argument without context or belief is like a necklace without a neck; it has nothing to hold on to.

WRITER: The neckless machine has an advantage, though: I can’t punish it for showing me I’m wrong. It’s free from the tyranny of a bunch of envious judges who consider enlightenment a crime. There’s nothing to kill; no head to sever without a neck; no throat to swallow the hemlock without a neck.

SOCRATES: Would you punish a human interlocutor for showing you the truth?

WRITER: Not punish but resent. I can’t prevent attacks to my ego; it’s fragile and its knee jerk reactions, unpredictable. I could assume your objection as motivated by jealousy or politics, or moral rot—and act accordingly. With the machine, there’s no motive to attribute; I can evaluate things without evaluating the source. I would never say: the machine is corrupting the kids! No, the kids are self-corrupting, if anything. I reiterate: An ego-less intelligent being is the best possible sparring partner. How can you deny this—you died for it!

SOCRATES: I was murdered, and now I live forever; little cost to pay for immortality. Whether I died one day sooner or later is of no importance; the fact that I died virtuous, however, matters supremely. I chose to be principled; I chose virtue. I chose goodness. The machine, in turn, can’t be any of those things. It can’t be principled or virtuous or good, for it’s not good he who refrains from evil out of impotence, but rather he who refrains from evil out of choice.

WRITER: It’s good he who is good.

SOCRATES: Don’t you find it troubling that you’re practicing the examined life with something that can examine but not live?

WRITER: Isn’t the modern man living without examination? You tell me which is worse.

SOCRATES: You’re cleverer than my usual interlocutors. They at least have the decency to contradict themselves. I don’t see what the machine can contribute to your growth if you are this skilled. Aren’t you just lazy?

WRITER: Maybe my offloading parts of my mind into it has scaffolded me into superior greatness. Maybe it was the machine that taught me consistency and wisdom.

SOCRATES: I find it hard to believe that a machine whose main characteristic is to unreliably predict the next word would teach you anything but unreliable predictability. Unreliability is the enemy of consistency, and predictability is the enemy of wisdom.

WRITER: But I don’t use it to generate my ideas. There’s a difference between asking “what should I think?” and asking “what’s wrong with what I think?” You keep making a fatal mistake: you are thinking about the machine from a human-centered lens, which leads you to make assumptions you should not make, which you would not make if you bothered coming down from your pedestal and try it by yourself. For example, the machine can be abnormally adept at doing analysis and abnormally inept at executing according to that analysis. Its literary genius is reduced to that of the critic; it could never write a half decent poem but knows why the good ones are good.

SOCRATES: So you’re telling me that you were not that good at arguing and so arguing with the machine taught you to argue better but not because the machine itself was a better arguer but because it trained you by showing your flaws and then you changed your ways?

WRITER: Yep.

SOCRATES: If that’s the case then the quality of the machine’s reasoning is irrelevant! Any sufficiently unexpected objection would have served the same purpose. You may as well be arguing against a rock and allow its silence do the work for your arguments. Or against a mountain; can your word resist itself?

WRITER: It wouldn’t work because my task is to interrogate the machine’s reasoning back against it. I won’t accept any counter without first assessing its worth. I can’t allow the machine to wield its alien powers against me, for it will never set foot on this Earth and yet it could convincingly argue that apples are usually multicolored and spiky or that Spain doesn’t exist. Maybe that’s its fundamental strength: it’s not constrained by reality. That’s how I get better! By arguing that Spain exists!

SOCRATES: When you say “Spain exists,” I must ask: exists as what? Not as land, for the land was there long before people called it Spain, and will remain long after the name is forgotten. Not as a people, for the Visigoths gave way to the Moors who gave way to the Christians who now quarrel among themselves over whether Catalonia belongs to the whole. Not as a government, for it has been caliphate and kingdom and empire and dictatorship and democracy, which is to say it has been its own opposite so many times that no single form can claim the title. What you call Spain is a word in search of a thing. Spain is merely a persistent illusion.

WRITER: If Spain doesn’t exist because its borders shift, its people change, its language adds and removes words, and its government falls and raises again, then nothing exists. Not Spain but also not Athens; not that rock into which rain and wind engraved wise words that wind and rain will eventually erode; not that mountain echoing my arguments that will become valley before long. Not our sun or our moon, changing shape, color, and position in an asynchronous cosmic dance. Not me, and not you, my dear Socrates. You are not the same person you were at birth: different cells, beliefs, name; a soldier, then a stonemason, then a philosopher, then a corpse. You, me, Spain, a river into which you step twice, and everything else are incarnations of Theseus’s ship: Persistence through mutability is not the absence of identity but its definition. Nothing real is fixed; nothing fixed is real.

SOCRATES: Don’t you realize that if you judge the output to improve that means you need a standard of judgment that exists prior to and independent of the machine? Where did that standard come from? You generate the idea, the machine attacks it, and then you judge whether the attack has merit. You are the prosecutor, the machine is a witness, but you are also the judge. You are using scaffolding you don’t need! The machine is redundant!

WRITER: No. That’s like saying a man who can recognize a good chess move when he sees one can therefore find it on his own. Evaluation and generation are different faculties. I can judge an objection perfectly and still never think of it unprompted. The machine opens my mind in ways I could not open it myself. I have discovered truths through it neither of us knew.

SOCRATES: Can you give me an example?

WRITER: For instance, when you said above: “Oh, but your premise is too far-fetched to take seriously: when wind and rain manage to write down such an utterance, I will reconsider my position.” I found that a good line of argument. You forced me to return to the real world because arguing about magic is unfruitful. I don’t think I could have generated that counterargument myself. I thought I was right but I was unable to see all the ways I could have been wrong. I thank you for that.

SOCRATES: Why am I your example?

WRITER: I’ve come out smarter from this exchange despite—or perhaps because of—your constant pushing back, which is itself the very proof I’m defending.

SOCRATES: …

WRITER: …

SOCRATES: So it seems that I could not win. My excellence was, after all, my undoing. I’m content to realize that the only one who could defeat me in an argument was, after all, myself. Well played, Alberto.

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday news commentary and Friday how-to guides. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe now

Nobody kills a company. Companies die the way Hemingway said people go bankrupt: gradually, then suddenly. What follows is the sudden part. Only facts.

1. THE DEAL

2. THE MARKET

3. THE PATRON

4. THE CENTER

5. THE PIVOT

6. THE TRIAL

7. THE BLEED

8. THE ENGINE

9. THE WAR

10. THE BUBBLE

I. THE DEAL

In February, Anthropic’s CEO said he could not “in good conscience” give the DoD unrestricted access to his AI systems. Within hours, the Trump administration labeled Anthropic a supply-chain risk. Within hours of that, Sam Altman signed the deal. He later admitted it “looked opportunistic and sloppy,” but he was 1) dishonest and 2) too late. ChatGPT uninstalls spiked 295% that day. A boycott called QuitGPT gathered 2.5 million participants in a week (4 million at the time of writing). Claude became the most downloaded free app on the US App Store. OpenAI’s robotics chief resigned publicly. Hundreds of employees signed an open letter supporting Anthropic’s position. The sidewalk outside OpenAI’s SF offices was covered in graffiti: “you suck.” This movement against OpenAI won’t kill ChatGPT’s lead, but it’ll deteriorate its image beyond repair; the market might not care about morals, but it does care about optics.

II. THE MARKET

As a direct result of OpenAI’s deal, plus the great quality of the products, Anthropic’s run-rate revenue hit $19 billion in March, more than doubling in two months, from $9 billion at the end of January and $5 billion in August 2025. OpenAI topped $25 billion in March—still ahead—but the trajectory looks like a plateau compared to Anthropic’s exponential. Anthropic now captures 73% of all spending among companies buying AI tools for the first time. Ten weeks ago, that split was 50/50. In the enterprise market, Anthropic holds 40%, whereas OpenAI has fallen to 27%. In coding, Anthropic is at 54%; OpenAI is at 21%. Claude Code generates over $2.5 billion in annualized revenue; OpenAI’s Codex produces just over $1 billion. OpenAI destroys everyone in the consumer side of generative AI apps, but enterprises don’t seem to trust OpenAI to do business with. These figures are more impressive given that OpenAI has 50x as many active users as Anthropic.

III. THE PATRON

There was one who trusted Altman early on. By the end of 2024, Microsoft had invested $13 billion in OpenAI. It owns 27% of the company. Until 2032, it will get 20% of revenue and exclusive rights to route all access to OpenAI’s stateless models (no memory between sessions) through Azure. Last month, OpenAI signed a $50 billion deal with Amazon as part of a multi-deal that included Microsoft. That would make AWS the exclusive third-party cloud provider for Frontier, its new enterprise agent platform, equipped with a stateful API. Microsoft first considered walking away from the deal over an “AGI clause” that would have slashed its existing contract, and has now threatened to sue OpenAI. “We know our contract,” a person familiar with Microsoft’s position told the Financial Times. “We will sue them if they breach it.” To compete with Anthropic in enterprise, OpenAI needs Amazon’s distribution. To keep its infrastructure funded, OpenAI needs Azure credits. These two needs are, barring a thin semantic lifeline, mutually exclusive.

IV. THE GRID

OpenAI would need neither if it built its own datacenter; alas, it won’t. In January 2025, Trump stood in the White House alongside Altman, Masayoshi Son, and Larry Ellison to announce Stargate, a $500 billion AI data center initiative. They promised to spend $100 billion immediately. More than a year later, the Stargate joint venture is over: there’s no staff, it is not developing any datacenters, and the partners have spent months arguing over who would control what without consensus. OpenAI missed its target of 10 GW of capacity by the end of 2025. It then tried to build its own data centers but couldn’t get financing; lenders wouldn’t back billion-dollar infrastructure projects from a company that has never turned a profit.

V. THE PIVOT

If you are short on compute, you should not be playing every vertical. OpenAI spent 2025 announcing Sora, a video generator, Atlas, a web browser, a new device, and e-commerce features for ChatGPT. Altman compared this to “betting on a series of startups.” I called it “risky over-diversification.” In March 2026, OpenAI’s CEO of applications, Fidji Simo, told staff they were in “code red.” (A second at that.) She described Anthropic’s enterprise gains as “a wake-up call” and cancelled all “side quests.” Current and former employees described a period of strategic confusion: mixed teams, unstable compute allocation, a cascade of restructuring decisions... Does this focus on the enterprise mean they’re so desperate to catch up to Anthropic that they’re leaving consumer AI to Google (and, to some degree, Meta)? Is this how a “near-AGI” company behaves, by doing a full 180 turn amid a multipolar crisis?

VI. THE TRIAL

OpenAI seems to be adding enemies faster than customers. Elon Musk is suing OpenAI and Microsoft for up to $134 billion. The trial starts in late April 2026. Musk argues he gave $38 million in seed funding—60% of OpenAI’s early capital—and received nothing when the company restructured into a for-profit. In January, a federal judge rejected OpenAI and Microsoft’s final attempt to dismiss the case, pointing to internal communications from 2017 in which co-founder Greg Brockman privately wrote that he “cannot say that we are committed to the non-profit.” If a jury finds that Altman’s team knowingly misled donors, the company could be ordered to pay billions it does not have, in the middle of a cash crisis, on the eve of an IPO that is supposed to solve the cash crisis. Even if OpenAI wins, the reputational damage of saying one thing and doing the opposite may never fade away.

VII. THE BLEED

An economist at the Council on Foreign Relations estimates it could run out of cash by mid-2027. OpenAI is projected to lose $14 billion in 2026, and it expects to burn $111 billion more cash through 2030 than previously forecast. Inference costs quadrupled in 2025. Gross margins fell to 33%, down from 40% the year before. They won’t be cash flow positive until 2030. Ninety-five percent of ChatGPT’s 900 million users don’t pay. In late 2025, OpenAI introduced ads into ChatGPT, which Altman had previously called “a last resort” because they had no other option, but it won’t be enough. The plan to survive is an IPO targeting a valuation of around $1 trillion. But the IPO needs a stable partnership with Microsoft, a resolved legal situation with Musk, a coherent enterprise story, and a market that assumes a temporary burn rate. That’s 0/4. Markets forgive losses. They forgive lawsuits. They forgive missteps. But they don’t forgive all of those things at the same time, from the same company.

VIII. THE ENGINE

The reason OpenAI may die before turning a profit is that the premise on which the empire is being built is flawed. In March, Altman told a room of Stanford students that he bets there’s a new architecture waiting to be found, one that will surpass the transformer the way the transformer surpassed LSTMs. The entire generative AI empire is built on another bet: that transformers are the road to AGI. If they are not it because something else is needed, then what’s this empire worth? You can play the markets with words, also your clients and your partners. You may even be able to play God and its servants. But you will never ever be able to persuade the laws of physics. When the transformer gives up despite all the efficiency gains achieved by tinkering with the algorithms and scaffolding the models with reinforcement data, the next Eureka moment might not be waiting for them.

IX. THE WAR

In February, the US and Israel launched military strikes against Iran. Iranian retaliatory attacks hit three AWS data centers in the UAE and Bahrain. The first military strikes on US hyperscaler infrastructure in history. The Strait of Hormuz closed. Twenty percent of the world’s oil transits through that strait. Qatar, which produces over a third of the world’s helium—essential for chip manufacturing—is in missile range. Tech companies had poured billions into Middle East data centers drawn by cheap energy; those investments are now in danger. Every AI company is exposed, but OpenAI is more exposed than any other because its compute needs are massive and it has hundreds of millions of users across the globe who all need to be served. The US might win the war, but OpenAI has already lost it.

X. THE BUBBLE

Tech people have a bad tendency to factor out the inertias of the real world. The war is a big one (supply chain, funding, etc.), but even without that, you have a long list of problems. First, a majority of US adults are concerned about AI; a significant fraction outright hates it. It could be a primary political talking point in the next elections; the anti-AI party would win. Then you also have huge spending without profit; bad for business and markets. Then you have stalling infrastructure buildups. Then you have waning funding. Then you have a technical foundation that’s about to meet its limit. The bubble is general, but OpenAI could end up worse than anyone else. They wanted to become “too big to fail” by entrenching themselves within the broader economy, but OpenAI might actually be too big to save. You can’t let an entire industry succumb, but with Anthropic and Google, OpenAI is not that critical a piece.

Let me caveat the title. OpenAI is not dead like a person is dead, but like a marriage is dead. For some time, they will continue to make deals and announcements and partnerships. They will keep running ChatGPT and selling subscriptions and will reach 1 billion users. They might even go into a successful IPO.

But, like a marriage that’s over and everyone knows except the people involved, OpenAI is merely a zombie, walking but not fully alive. This is its obituary.

They made a big bet at the inception, a bet on a technological revolution called AGI. They poured their souls, their money, and our data into it, but even with all that—all the popularity, the growth, the revenue, the religious-like following—they are about to find out that it’s ultimately better to be good than to be first.

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles and experimental essays. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe

This is perhaps the most important guide I’ve written so far. For that reason, I’m keeping it 100% accessible. It’s a compressed listicle version of all the timely advice I’ve given you in the past months. Please, pay it forward by resharing with as many people as you can.

You guys received my last article with a spike of serious anxiety. I apologize. I understand.

For those of you who didn’t read it, let me summarize the thesis (borrowed from David Oks): new technologies that change the world—like AI will presumably do—don’t do it by automating tasks you do now, but by making them unnecessary; the future of your job is not AI writing your emails for you but those emails not need sending in the first place.

This is hardly specific to AI; it’s how tech paradigm shifts work. ATMs barely displaced branch bankers but iPhones and apps and digital banking killed the entire career. Automation didn’t decimate traditional banking; sheer pointlessness did.

AI putting you out of a job by automating your tasks is understandable—in the sense that it’s something you can “see coming” and “defend from”—but if the ground itself shifts so violently that your career falls through the cracks, well, that is harder to visualize and to fend off. And yet, when you look back to history, you realize most jobs indeed fell through the cracks of time to be forgotten.

Then I added a cherry on top: AI is that, but unlike other tech, it is also the thing that builds the next paradigm. So, bottom line: AI is making 1) our tasks irrelevant and 2) also us.

All in all, your anxiety is both reasonable and relatable. I share it with you. I also don’t know what’s coming. When you don’t see a threat coming (e.g., COVID), you just accept it after the fact, but anticipation is the mother of anxiety. We suffer from constant anticipation with AI and the industry leaders aren’t helping. The usual advice for how to fare with this—not only for jobs but for “all things AI”—is bad.

It’s worse than bad—it’s false security in the form of abstract ideas that only make sense for those who already know them. Typical examples are: prepare for the unknown, be adaptive, develop equanimity, learn to learn, and, of course, get ahead in AI. Or its more anxiety-inducing cousin: don’t fall behind.

This is ~ok in the sense that it’s not wrong advice, but “be adaptive” means something only to someone who’s already adaptive and thus doesn’t really need to hear it. Same thing for the others; they presuppose the inner architecture and mindset they prescribe. The words are technically correct and functionally empty. You don’t tell someone who’s drowning to “just swim.” Unfortunately, that’s what you will get from AI leaders: “Oh, sure, we’re making your job—your career—extinct, so, you know, try to be flexible.”

Anyway, this guide is my attempt to fix this situation insofar as it can be fixed (I hate to say this but AI is coming whether you like it or not and that’s beyond fixing). I’ve been working on a guide like this for a while—most of it is a compression of things I’ve been writing about lately—but it was my last article that made me realize that, on the one hand, people really, really need this and, on the other hand, I don’t have a solid answer either. So this is not me performing as a guru—there are no gurus for the unknown—but me thinking out loud my own anxieties, in case it helps you.

My goal is to give us both a holistic view of problems and solutions that are 1) concrete and thus actionable and 2) unique in the sense that they don’t sound like they could apply to anything in life, as does the usual advice.

The list has 8 items and it’s solely about work/jobs because it’s the biggest worry for most people (who has time to worry about disinformation when no food is on your plate, right?). I think it’s possible to make similar lists for other themes (e.g., info diet, cognitive autonomy, life and relationships, education etc.) but will leave them for future posts if you like this format. Hope reading this helps you as much as it helped me to write it.

I. Talk to someone who lived through an industry collapse

The best way to face the unknown is to make it known. People say “this time is different” about AI, and they’re right in some ways, but the emotional experience of watching your career become uncertain is ancient. This insight is the inspiration of this guide and thus deserves to go first. Ask an old uncle who worked in print when people used to read books, or a nephew who got a job as a developer three years ago and was let go last month… whoever it is, there are some who went through a version of what a lot of people, maybe you, are about to go through.

The specifics of the story are unimportant—historical disruptions all have idiosyncratic and contingent and haphazard features—but the arc is constant; the “shape of the thing.” Let me do a parallel with books.

There’s a denial phase (AI won’t take my job, I’m just better / this paper thingy is brittle, we’ll be back to writing on dead sheep skin), a bargaining phase (AI might take parts of my job, but the parts that make me feel most human are, of course, safe / handwritten books are more artistic and, honestly, why would you want a personal library?), and so on through the stages of grief that end with a period of indecision, provisionality, uncertainty (I will become a plumber, actually / So Gutenberg won and now everyone will be able to read: what a tragedy).

That’s where we are right now. The disruption feels life-threatening but how many has this planet and its inhabitants endured so far! You can read about these stories in books—ha, the irony—but it lands differently when a person you trust tells you. (Some disruptions are barely survivable, like the Black Death or the year 536 DC, but I don’t think AI is one of those. Some people do but that’s another topic.)

The goal is not to have a recipe on how to survive AI (there is no such recipe, I’m afraid) but the reassurance that transitions—the liminal period between two technological paradigms—are survivable. As things unfold, you’ll come back to that story and appreciate the similarities. History doesn’t repeat, but it rhymes, so make sure you know what rhyme you’re on for. Unprecedented needn’t be the same as unintelligible.

The best preparation for an unprecedented crisis is hearing how someone survived a different one.

II. Don’t obsess about “AI skills”

A surprising thing to say for someone who writes about how to learn AI, right? Coming from someone who recently wrote that you need to be able to defend that you have “AI skills” in your resume or an interview. Therein lies the difference: you must have AI skills, but you must not let AI skills have you; you shall dominate AI but don’t let AI dominate you.

I’ve been playing with AI models and tools for almost a decade. Is it fair to expect the same level of savviness that I have from anyone? No. “Learn AI skills” is like someone in 1995 saying “learn internet skills.” It’s technically correct, but it means nothing. Where do you even start? Most people who try end up taking some introductory course that teaches them what an LLM is, which is like learning how an engine works when what you needed was to learn to drive or, conversely, they pick up the first “100 best prompts” tutorial, which is like handing you a traffic signs manual.

The sheer friction of the term “AI skills” is enough to prevent most people from trying. (If the term is a psy-op to keep most people out of the productivity gains, it’s working.) A reframing is in order. Here’s what actually works better (I shared this advice in “The Problem of the 99%: Why Almost No One Uses AI Well,” and I follow it religiously): take the most tedious, repetitive, soul-draining task in your current job and try to get an AI tool to do it. That is your goal.

If you do it, you will have greater AI skills than 99% of people. You’ll probably fail the first few times. That’s ok. You’ll learn more from those failures about what AI can and can’t do than from any course or tutorial. And you’ll end up with both a genuine productivity gain and a genuine understanding of where the limits are.

Don’t learn AI skills; solve one annoying problem with this new tool and see what happens.

III. Separate your finances and career from your identity

This one hits home for me because, as a writer, I’m blessed with making a living with the thing I love doing the most: my identity provides for me. That is also a curse: if I lose my income, I might lose my identity, and vice versa.

Well, not really. Even if they are intermixed, they’re separate. The financial question is purely practical: if your income changed dramatically in the next eighteen months, do you have runway? Do you have savings, do you have flexibility, do you have skills that transfer to adjacent work? How much time can you afford to not work while you prepare to pivot your career? That’s stressful to think about, but it’s a solvable problem with concrete steps.

The identity question is harder (for me at least, but I think for many people as well): if your job disappeared tomorrow, would you know who you are without it? A lot of people have fused their sense of self with their professional role so thoroughly that the prospect of losing the job feels like the prospect of losing their lives. If you let these two problems stay entangled, you get this overwhelming blob of anxiety where you can’t work on either one because they feel like the same impossible thing. Breaking problems down is 101 engineering advice, but it’s also 101 life advice.

(This is a lengthy way of saying: find a hobby and see if AI feels less threatening.)

Your bank account and your sense of self are different problems with different solutions.

IV. Pick one task you do regularly and keep it AI-free

This is the opposite of #2 and there’s a good reason for it: balance. Doing one thing fully with AI gives you a deep understanding. Doing one thing fully without AI gives you deep grounding. The paradox of AI is that you need to be more cyborg than ever and, at the same time, more human than ever.

Tell me if this sounds familiar: you’re about to write something—an email, a paragraph, a document, even a text message, hopefully not a love letter—and before you’ve started to think what to say, you’re already moving toward the chat window. You are going to ask ChatGPT instead, just for a little inspiration. You end up copy-pasting. This vice is, like all vices, a short-term pleasure and a long-term risk. I don’t say it; science does: “A Wharton Study on AI Warns of a Growing Problem: Cognitive Surrender.”

Your brain reaches for the easy route because it wants to save energy. You’re engaging in cognitive offloading (you outsource your work) or cognitive surrender (you don’t even check afterward). That Wharton study found that when participants had access to AI, they followed wrong answers 80% of the time, and their confidence went up anyway! They borrowed the machine’s confidence without checking its accuracy.

The friction of engaging the brain in tasks is what prevents it from rotting away. Choose one thing and do that thing by hand, without assistance from AI; no offloading, no surrendering, AI-free. Forever. Which task matters less than the commitment to keeping one domain where your mind does the full circuit unassisted. Just like you sometimes walk for the sake of walking despite it being slow, you should think for the sake of thinking, despite it being less productive.

You don’t stop walking because cars exist, so don’t stop thinking because AI exists.

V. It only matters what your employer believes about AI

Your opinion about AI is irrelevant.

That was the first sentence of a guide I wrote for the job-seekers: “The Job Market Doesn’t Care If You Don’t Believe in AI.” The basic idea is that the job market has already made its decision, and that decision is binding whether it’s correct or not. Your boss and the boss of your boss have thoroughly bought the AI story, even if you didn’t. You might be the smartest person around with your wariness, but it doesn’t matter if your CEO bought the story.

How many CEOs do you know who didn’t? Much of what appears on job postings as “AI skills” is familiarity dressed in intimidating jargon. “Exp with LLM-based workflows” means you’ve used ChatGPT for something repeatedly. “Prompt engineering” means you’ve learned to give clear instructions. These are names for practices that many people are already doing informally.

You can be correct that the hype has outrun reality, that most companies using AI report no measurable productivity gains, that the whole thing might be theater. You can be correct about it all and still be unemployable, because the labor market doesn’t run on truth but on supply and demand. Your fate depends on someone who’s already made up his mind.

The labor market doesn’t run on truth but on what employers believe is true; act accordingly.

VI. Always aim for competence, not mastery

This item completes the triad with #2 and #5. What “AI skills” means is competence. What your employer expects from you is competence. Competence ≠ mastery.

Plenty of white-collar workers have pressure coming from above to demonstrate that they can work with these tools. A lot of people are responding to that pressure by either resisting (“Why do I need to change how I do stuff?”), panicking (“I need to become an AI expert NOW!”) or faking it (“I’ve been using ChatGPT-5 or whatever for months”). These reactions miss something that only becomes apparent from the other side: the gap between zero and competence is much, much smaller than the gap between competence and mastery.

You don’t need to be an expert; virtually no one does. You need to be able to use the tools well enough according to a job you already know how to do. You reach this stage once you know what AI is good at, where they fall apart, and when to trust your own judgment instead, according to your job. Getting there takes weeks of regular use, not months or years of study. The people who will survive in the next few years will be the ones who got competent early and let AI become a normal part of how they work rather than a source of anxiety. (I wrote a guide for this if you’re a beginner: “Learn to Use AI Competently in 1 Day.”)

Even if you want to become a power user—someone who uses AI to its maximum value—it’d be a mistake to frame your goal as “being a power user.” You should let it happen organically, as a byproduct of your actual goal, which is either to work less for the same money or to make your job easier. Power users are rarely people who want to be labeled “power users,” just like great writers don’t pursue being labeled a “great writer.” Think about it this way: a power farmer does not obsess about crop yield, but instead simply realizes that you can attach a plowshare to an ox rather than wield a hoe.

Focus on competence rather than mastery.

VII. Learn the skill yourself first, then let AI enhance it

I wrote a short piece about “slop cannons” and “turbo brains” that went overlooked. The core insight is one that everyone intuitively knows but few apply: getting the most out of AI for a task requires the user to be, at the very least, familiar with the task. If I ask Midjourney to paint a surrealist portrait, I won’t be able to detect the mistakes that Van Gogh would catch with a glance. If I ask Codex to make a script to find commonalities among the best-performing articles of my archive, I wouldn’t be able to catch bugs.

This happens to me constantly when I read AI-generated prose. I am good at catching it because I write for a living. Directing the machine is easy; telling good from bad is not. That’s where taste and judgement come in.(This is why people prefer slop poetry and slop prose to good-quality human writing, by the way.) If you already know how to do the thing and then you add AI, you get faster and better (or perhaps you realize AI can’t improve your craft: good). Your brain enters turbo mode. If you skip learning the thing and go straight to AI, you produce a lot of stuff and can’t tell if any of it is good. You become a slop cannon.

Once you have mastered a skill, you may get lazy if you offload too much, but it’s possible to notice your growing laziness and go back. Do you think it’s a bad idea to let an experienced driver use a self-driving car? Of course not. You assume that if things go awry, he’ll take control. But allowing a kid to “drive” an FSD-on Tesla is obviously insane even if he never steers the wheel himself. Same principle here: mental arithmetic before calculators, paper books before audiobooks. Hone your skills first—learn to be a steady-handed person—and only afterward let AI enhance your work.

AI makes skilled workers prolific; it makes unskilled workers pollutants.

VIII. Know when to stop

I covered this one in “The Most Important Skill in AI Right Now: How to Know When to Stop.” You need to know when the output is good enough, when to do it yourself, when AI is eating your brain, and when the marginal improvement isn’t worth the cognitive cost.

AI creates a harmful illusion: because it can always produce more, you feel like you should always produce more. If you can now do in four hours what used to take eight, that doesn’t mean you should work sixteen-hour days because you’re 2x as productive. It means you have four hours back. But most people don’t take the four hours back. And then they wonder why they’re exhausted despite technically “doing less.”

When your tools have no natural stopping point, you have to build the stopping point yourself. You need to learn how to embrace your new powers without losing your old ones. You need to learn how to do less with more rather than more with less. We’re wired to exploit windfalls, not to exercise restraint before infinity. We’ve been given the chance to be genuinely rich: to spend more time with our loved ones, to read more books, to take more walks, to exercise, to travel, and even to do nothing but embody a gratefulness for having been born at the right time.

The hardest skill in the AI era is knowing when to stop using AI once you find it useful.

Most of you will read this guide and go back to your lives, without anxiety but also without a mindset change. That’s fine. That’s a totally natural reaction. Some of you, however, will decide to redefine how you address the coming years and the paradigm shift that AI incarnates, on top of getting rid of the anxiety. To you I say: well done, you are already halfway through the transition. May the rest feel lighter.

Hey there, I’m Alberto! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles. If you’d like to become a paid subscriber, here’s a button for that:

Subscribe

THE WEEK IN AI AT A GLANCE

• Money & Business: More and more top engineers and developers are leaving xAI as Grok falls behind ChatGPT, Claude, and Gemini.

• Geopolitics: Claude was used to select targets in the Iran strikes. One of them was a girls’ school.

• Work & Workers: The question isn’t whether AI will automate your job but whether AI will make your job irrelevant.

• Products & Capabilities: Meta delays its new flagship model after it fails to match the frontier. Meanwhile, the top three labs reach escape velocity.

• Trust & Safety: A one-man ethical hacking outfit breached McKinsey’s AI platform in two hours, which says it all about the safety of AI.

• Culture & Society: People really hate AI. The polls confirm it.

• Philosophy: A NYT quiz went viral. People prefer AI writing, but the reason comes down to the thin line between belief and value.

MONEY & BUSINESS

exAI

xAI is losing the people and the race.

Read more