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.

QUICK REMINDER:

I’m running a 50% discount offer on annual subscriptions that ends on June 5th June 12th (extended one week). Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

I know I’m publishing a lot, but I had to share this quick “first impressions” post on Claude Mythos 5 and Fable 5.

Anthropic just released Claude Mythos 5 (blog post, system card). It sits at the top of every major benchmark: coding, cybersecurity, reasoning, biology, vision, etc. It is, by the numbers, the best AI model in the world. Better than Mythos Preview. Better than anything OpenAI has released publicly.

And you will not be using it.

What you’ll be using is Claude Fable 5. It’s the same underlying model, but with a layer of safety classifiers on top that block or downgrade its responses in certain areas.

The launch has been covered extensively from the enterprise angle: what it means for agentic coding, drug design, or long-horizon research. The overall read is that the Mythos-class of models sits one quantitative—and qualitative—level above both Opus and GPT-5.X models. But to taste its greater intelligence, you will need to test it on genuinely complex and long tasks, which means, by definition, that you’d need to expend a lot of money. If you’re anything like me, you don’t have two thousand dollars per month to spare, so I’ve decided to make this post for the rest of us.

It’s a quick bulleted list of my first impressions. I’ve drafted them quickly for the sake of timeliness. It’s intended for standard users: the people who like Claude but have normal salaries. What changes for us starting today? Here’s what I think matters.

1. You have two weeks to try it at no extra cost. Fable 5 is included on Pro, Max, etc. paid plans from today through June 22. Then it gets pulled from those plans and becomes a credits-only (pay-as-you-use) model until Anthropic has enough capacity to serve in the standard plans. My advice is to try it as hard as you can in the two-week window because the restoration doesn’t come with a date. Get used to not having normal access to incoming models; this is the new normal.

2. “Best model in the world” means something different depending on what you do with it. This is the most important takeaway IMO. If you’re a developer running loop-based agentic coding tasks across a 50-million-line codebase, the benchmarks are directly relevant to your life and work. Congratulations, these models are a total upgrade. On the opposite end of the spectrum, you have people using Claude to draft emails, brainstorm ideas, summarize documents, or help them write. Fable/Mythos are not for you. Use the second-best model Anthropic gives you access to because it’s good enough. You won’t find a single task for normal white-collar work for which Mythos-class is clearly, measurably above Opus class. Use the free two-week window to check for yourself and then come back with the mere mortals doing mortals’ work, like me. The gains are concentrated at the higher end in terms of complexity and length of task. That’s only going to be truer over time; it’s why I said: “AGI is here, just not evenly distributed.” Longer tasks, harder tasks, deeper knowledge, bigger models—the edge of AI is now relevant exclusively to the AI labs themselves and a few people outside. Even if OpenAI wants to give everyone in the world a “personal AGI,” as they said in yesterday’s blog post (probably a covert response to Anthropic’s blog post about recursive self-improvement), even then, most people don’t really have any use for an AGI that’s also extremely jagged. A jagged superintelligence is a gift for those living in the peaks and irrelevant for us living in the valleys.

3. Token rich, token poor. So we’re entering a phase where the main resource constraint isn’t the model’s intelligence but how many tokens you’re willing to burn. Test-time compute—letting the model think longer, retry, take notes, iterate, etc.—is becoming the primary scaling axis. The model doesn’t get smarter if you ask it the perfect question, but if you let it work on something for hours. However, at an API pricing of $10/million input tokens and $50/million output tokens (that’s twice as much as Opus 4.8), you may not be willing to burn that many tokens. OpenAI researcher Noam Brown says that “empirically, the plateau [on test-time compute] is very far out. Sometimes we may not observe a plateau at all within practical budgets.” That means that models can keep getting better for longer than you can keep solvent. Relatedly, there’s a new trend in Silicon Valley that’s basically test-time computer taken to the extreme: Loop engineering (started by Claude Code creator and OpenClaw creator). You no longer prompt the model directly but design loops so that the agents themselves decide what to do; you don’t intervene. (Ethan Mollick calls it being a “patron” that commissions work rather than a “wizard” that casts spells.) The funny thing is that one loop that goes for a bit longer than it should can empty your bank account and the bank account of your kids, and their kids after that. I’ve read the testimony of one guy who burned through $1 million in one day on Fable 5. So, the loop engineering trend, which goes hand-in-hand with these Mythos-class models, creates an important divide that goes beyond having or not having tasks suitable for them. I call it the “token rich, token poor” divide. (I wrote about this in 2024 in related terms: “AI rich and AI poor.”) The use cases Anthropic highlights for these models—day-long codebase migrations, autonomous genomics research, multi-step drug design—burn through millions of tokens. At those volumes, a serious power user or small team could spend hundreds of thousands of dollars a month. The model is priced below Mythos Preview (less than half, Anthropic says), which makes it a bargain if you’re already spending at enterprise scale, if you’re already token-rich. If your AI budget is a Pro or Max subscription—that’s my case; I’m paying ~$130/month and I’m probably still in the 0.1% of highest paying users—then you should probably be thankful that you don’t have a Mythos-class-worthy task because otherwise instead of saying “I have no use for this model,” you’d be saying the more bitter “I have no money for this model.”

4. The most important safeguard is invisible to the user. According to Anthropic, Fable 5 has four categories of safety classifiers. Three of them—cybersecurity, bio/chem, and distillation (to avoid China from catching up)—are visible. When they trigger, you get a message that says the response is from Opus 4.8 instead (next best model). But Anthropic has added a special safeguard that limits Claude’s effectiveness for requests related to a fourth category: frontier LLM development. Essentially, other AI companies will not be able to improve their AI game by using Anthropic's top model (fair from a business lens). When the classifier triggers, the model is degraded through “prompt modification, steering vectors, or fine-tuning adjustments.” Anthropic estimates that this affects roughly 0.03% of traffic. My guess is they consider Mythos-class models so good that they could help other labs get ahead, not by distilling the reasoning traces as it works with standard distillation, but by using it normally as a user. That’s new.

5. Jevons paradox for AI-enabled projects. This is not exclusive to Mythos-class models, but it will get more obvious and pervasive as models improve. Three testimonies that illustrate what I mean. In a March episode for the Dwarkesh podcast, Terence Tao said (referring to the previous class of models, like Opus 4.8 and GPT-5.5): “I’m definitely noticing that the style in which I do mathematics is changing quite a bit, and the type of things I do. For example, my papers now have a lot more code, a lot more pictures, because it’s so easy to generate these things now. Some plot which would have taken me hours to do, now I can do in minutes.” Andrej Karpathy (now at Anthropic) says about Fable 5: “you can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps (e.g. a full wandb that is hyper-specific just for your project), you can 10X your test suite, auto-optimize code, run giant research projects with custom HTML for the results, anything! ‘Free your mind’ (Matrix ref).” Ethan Mollick made “a piece of software that researchers have needed for years but was never profitable to create” with Fable 5. The idea is: even if AI models don’t add much to your core work (in terms of depth because you’re extremely knowledgeable or because doing so is excessively expensive or some other reason), these models are good enough now that you can do just a lot of high-quality core-adjacent stuff you wouldn’t do otherwise. Things that are not toy things but actually worth the time and cost.

6. You’re not getting two free weeks; they’re getting two free weeks of you. The classifiers were tuned conservatively (several people report they misfire on harmless queries and way too often), and the only way to fix a conservative classifier is to feed it huge volumes of real, benign, human traffic, which is precisely what a free window on every Pro and Max plan gives Anthropic. (Add to that a brand new mandatory 30-day retention on chats; they say they don’t use it as training data for models, but maybe retraining the classifiers doesn’t count.)

7. Anthropic is pulling ahead of the pack. Recursive self-improvement, which is now thrown around like AGI was when GPT-4 came along (that is, way too often), works like getting closer to a black hole. The closer you are, the stronger the pull and thus the faster you go, and thus the closer you are and thus the stronger the pull, and so on until you reach the event horizon (the singularity). At least that’s OpenAI and Anthropic’s bet (Google had a different one; I will say more about this soon). Both OpenAI and Anthropic have been releasing new models in quick succession since January 2026, which made the global AI race a one-on-one race. If OpenAI doesn’t release something big soon, this will be essentially a race of Anthropic against itself. OpenAI said in yesterday’s blog post that they estimate RSI for March 2028 (Anthropic first said late 2028 and then said that their estimates had been conservative and it’s coming earlier than that). The main difference between the two is that OpenAI insists on keeping the human in the loop: they want AI to work “in tandem with our own researchers.” Let’s see what they have to answer the Mythos 5/Fable 5 card or their opinion will matter little.

8. As a writer. Yeah, it sucks. It sucks just like the previous models. Just like OpenAI’s and Google’s models. Writing—not just the creative kind but plain good writing—is beyond near-AGI models, apparently. I guess dealing with language is too much for language models. No, seriously now, I understand that writing isn’t and will never be the main goal for AI labs, but at this point, it’s pretty clear they just can’t do anything about it. I hate AI writing, not because it’s bad but because it’s everywhere (pun intended). Please, make it stop. Being charitable: If the skill of “good writing” or more generally “creativity” is hidden under a pile of post-training, as some AI people claim, and yet they have never trained a model to showcase it, then the reason is probably that the model would otherwise lose all the other things that come with post-training (like smarts). This is evidence (if not totally convincing) that 1) as models get more powerful, the jaggedness doesn’t shrink much, and 2) AI models are, so far, “either/or intelligences.”

9. My fav benchmark. ARC-AGI-3 results are not out yet, and I really, really want to know how well this model does. My bet is <5%. I will be genuinely surprised if it’s anywhere above 10%.

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.

QUICK REMINDER:

I’m running a 50% discount offer on annual subscriptions that ends on June 5th June 12th (extended one week). Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

I. HOW TO DEAL WITH UNCERTAINTY

You don’t know what AI will do to the economy, to your job, or to the way you spend your days. That’s great news for our topic today: how to prepare for a long period of uncertainty.

AI uncertainty is not a temporary condition. It’s not that we’re waiting for more data, and then we’ll know. No, we won’t know. We never do, actually—AI just makes it weirder because it belongs to a class of phenomena that Nassim Taleb calls fat-tailed: the range of possible outcomes is so vast and the extremes so disproportionately impactful that the average outcome—the most likely AI scenarios—tells you almost nothing about what to do besides “proceed as usual.”

Planning for the average future (which is normally a good rule of thumb) doesn’t work here. A world where AI finds itself recursively self-improving by 2028 is too distinct from one where it collapses the economy by late 2026. The mistake we tend to make with fat-tailed risk in Taleb’s sense is always the same: we optimize for the average scenario and get destroyed by reality. Familiar examples of a fat tail are the 2008 financial crisis or COVID. Well, those were already quite bad, and yet, AI is the fattest of all because by definition, anything is in the realm of possibility.

So, how do you prepare for something you can’t predict? Two things. One, you need to know what to focus on today. Two, you need to know whether to focus on something else tomorrow. It’s that simple. Let’s see what that looks like in concrete terms.

II. WHAT TO FOCUS ON TODAY

Taleb’s answer to fat-tailed risk is the barbell strategy. You load both ends and leave the middle empty. One end is “maximum safety”: the things that are valuable regardless of what happens, the evergreen stuff. The other end is “maximum exposure to upside”: cheap bets that pay off handsomely if a particular future materializes, the timely stuff. The middle—moderate effort, moderate risk, moderate reward—is where you die. Don’t hedge.

Applied to the next five years: the safe end of the barbell is deep fundamentals that will be useful even if AI agents take over the entire workforce. Some ideas: Writing clearly, reasoning through ambiguous problems, persuasion, domain understanding, managing people, judgment and taste, etc. This is a non-exhaustive list. (You can show this post to ChatGPT and ask it to complete the list with the context of who you are and what you want from life.) Don’t forget that the intensity of your focus is inconsequential if the thing you focus on is the wrong one.

There are no shortcuts for this end of the barbell. These skills were as valuable in 2005 as they were in 1905 and as they will be in 2105, no matter what AI does. The reason I can be sure of this is that they pertain to the realm of human-to-human relationships. Our world is built on a set of principles—with the human at the very center—that you can govern with the right skills. These skills are the floor of civilization, and they should be yours.

On the other end of your barbell is aggressive AI-native experimentation. This is where I focus most of my practical guides because I take for granted that you’re already working on the “always important” stuff. I’m trying to emphasize that both types of skills are a must if you’re to be prepared for the next 5 years.

What does aggressive AI-native experimentation look like? Among others: familiarize yourself with the basic concepts and tools, use them daily, build complex workflows to challenge your theoretical understanding, ship things in public to gather feedback, develop a hands-on intuition for what models can and can’t do, etc. This is your lottery ticket. They’re cheap to acquire (time and predisposition) and enormous in value if AI reshapes your field.

The dangerous middle is passive familiarity. Here’s what that looks like: you read articles about AI like this one, but don’t do the work. You watch the news and panic. You’re aware of the industry announcements and the drama, but don’t know shit about the technology. This moderate, mostly passive engagement gives you the feeling of keeping up while building zero actual capability. Keep doing that, and you’ll be tragically unprotected. You could also fall into the dangerous middle if you start to consider evergreen skills useless in the AI age, leading you to forgo taste and judgment, domain knowledge, or overall cognitive competence. Don’t do that.

So the barbell is, in practice, a time split. Spend part of your time on skills that survive every scenario (maximum safety, evergreen). Spend another part getting your hands dirty with AI tools in ways that feel excessive relative to what your job currently requires (maximum exposure to upside, timely). Spend no time on the comfortable middle where you learn just enough to have a vague opinion. As Taleb would tell you: comfort under fat-tailed risk is risk of getting crushed by a fat tail.

III. WHETHER TO FOCUS ON SOMETHING ELSE TOMORROW

The barbell tells you what to do today, and that’s good. But what should you do with the information that arrives at you every day? How can you update once the future starts to materialize? How can you weigh more the evergreen if the AI bubble pops? How can you weigh the AI stuff more if job postings start to ask for hands-on agent experience? That’s what the barbell strategy lacks: It doesn’t tell you when you need to rebalance. For that, you need a different tool.

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.

QUICK REMINDER

I’m running a 50% discount offer on annual subs that ends on June 5th June 12th (extended one week). Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

Here’s a quick industry analysis post for your Sunday afternoon. See you tomorrow, Monday, with a fundamental practical guide: How to Prepare for the Next 5 Years.

In July 2025, President Trump described AI as “a beautiful baby that’s born.” He said the government should “grow that baby and let that baby thrive” and warned against stopping it “with politics” or “foolish rules and even stupid rules,” echoing what Vice President Vance had said in February of the same year at an AI summit in Paris: “The AI future is not going to be won by hand-wringing about safety.” Nine months later, in May 2026, The New York Times reported that the Trump administration was considering an executive order to create, among other things, a formal government review process for new AI models before they are released to the public.

The AI baby grew up fast.

Per the NYT, Anthropic, Google, and OpenAI were all briefed on the plans, which implicitly means the rest of the industry was likely excluded. The EO entailed, wrote the NYT, a “stark reversal” of Trump’s position on AI; he “swiftly rolled back a Biden administration regulatory process.” Up to this point, Trump’s administration had been unconditionally non-interventionist.

I guess that, despite his overt disdain for journalists, Trump reads the New York Times because he swiftly let the press know he wouldn’t be signing. In late May, several outlets reported that Trump postponed the executive order because he “didn’t like certain aspects” of it, namely the part in which he would be called out as a regulator. So the White House changed the terms. Now it’s not a formal government review process but a voluntary commitment for review. Trump got what he wanted: if the press calls him a regulator, he can point at them and say, “Fake news!”

He got his alibi, but in practical terms, the EO is the same. Instead of “Submit your model here,” it will be “Wait, why didn’t you submit your model to review? Hiding something, huh?” Same outcome, sneakier mechanism. So, whether he likes it or not, I will still call him a regulator. And thus the same questions can be asked about it: Why did this happen? How did the EO come to be?

The framing is national security in general, with a specific emphasis on cybersecurity risks, but the NYT reported on two other factors that led to these deliberations: First, David Sacks, heavily anti-regulation, left his role as AI czar. Second, the concerns about AI’s effects on jobs, privacy, and mental health are no longer a “woke” anti-AI leftist thing but bipartisan, per Pew Research (and plenty of other sources). The true catalyst, however, was Anthropic’s Mythos Preview model.

I’ve written about Mythos before, but to give you the gist: Anthropic announced this new, powerful AI model that, they argued, “can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” And they added: “The fallout—for economies, public safety, and national security—could be severe.” They held back the model from public release (initially; now that OpenAI has better public models, they plan to release Mythos anyway). Only a few selected partners, including the government, have access for now. The rest of us were left to be awed at Mythos’ evaluation scores, which Anthropic made sure to show on every release since. The White House interpreted this as a preview of what’s coming.

Anthropic agrees. In a May policy paper on the US-China conflict, they call Mythos Preview a “wake-up call” and use it to argue for tighter export controls, anti-distillation legislation, and accelerated government adoption of American AI. The dangerous model became the centerpiece of a policy agenda; Anthropic is a lobbying firm that happens to have a technological branch.

In a June blog post—of a marketing nature but cleverly disguised as a political call to action—they argue that recursive self-improvement (AI that makes itself better autonomously) is coming sooner than expected, and the implications are so large that the industry should perhaps coordinate a temporary pause on AI development, including, of course, those reckless Eastern actors that prefer to work on the open. For years, Anthropic has argued publicly for government oversight of frontier AI.

CEO Dario Amodei has repeatedly warned against open-weight release of powerful models. The most recent example of this strain of cautionary messaging was from co-founder Jack Clark, who wrote an essay in May where he assigned a 60% chance—now an overly conservative number in light of their updated beliefs—that, by the end of 2028, an AI model will be able to “autonomously build its own successor” via recursive self-improvement.

In general, Anthropic’s scaling policy is built around the idea that increasingly powerful models require increasingly strict controls. Controls that only responsible people like themselves—and unlike, say, OpenAI or Meta or open labs—can safely implement. Everything they do follows a coherent plan. The result is that the regulatory framework the Trump administration has signed into law—even after the watering down—looks, in practice, like a generalization of what Anthropic already does. They’ve raced their way into Trump’s ear.

It’s all for safety reasons, sure, but incidentally, Anthropic’s safety posture has always been aligned with its long-term strategic and competitive interests. (OpenAI GPT-5.5 shows similar cyber capabilities, and it was released normally; to this day, without known cybersecurity issues.) I don’t doubt their honesty, but I can easily imagine how happy they must be now that an unreleased Mythos Preview enabled such a nice belief-market fit. As I wrote in April: “[G]ating and publicity are independent choices, and conflating them lets Anthropic have both at once—the ethical high ground of restraint AND the commercial-reputational win.”

To users, they say: I’m sorry, but this is too dangerous for you to have. To investors, they say: We are the best. To the government, they say: You can’t let this happen with unrestricted models. Anthropic has perfected the art of making the bull case and the bear case identical: “this technology is so powerful it might be dangerous” is both a reason to invest and a reason to regulate, and Anthropic benefits either way.

I’m starting to suspect that Anthropic’s whole strategy since the first Pause AI letter failed in March 2023—more than three years ago!—was to become the #1 AI company and from that position of power indirectly threaten the world—by warning us about what the technology they’re building is capable of—with things like cybersecurity, bioweapons, and recursive self-improvement, and thus force the hand of those that can impose their will onto the reckless. That is, the US government.

And so Trump had no choice.

Per the EO, the NSA, in consultation with a handful of officials, will run a classified—that is, opaque—benchmarking process to determine which models qualify as “covered frontier models.” Developers who participate get 30 days of government access before public release and the privilege of being designated “trusted partners.” Developers who don’t participate get a raised eyebrow and some questions in the backrooms. There’s no mention of open-source AI, and the order doesn’t mandate per se, but it effectively creates a two-tier system: either you’re a friend of the government or you aren’t. If the three top labs voluntarily submit to review and earn “trusted” status, every lab that doesn’t is a second-class citizen in the eyes of enterprise buyers, procurement offices, and investors. You don’t need to ban open AI to halt it.

To top it off, the Washington Post reports that Trump “is considering taking a government stake in leading artificial intelligence companies.” Once the US government links its future to the AI industry, it will set a protective incentive structure. Such convenient timing, right when Anthropic is preparing the IPO. Creating a regulatory-financial moat just before it goes public makes it worth even more to investors precisely because almost no one else will clear the bar.

But wait. Don’t they hate one another?

Didn’t the military cut off government use of Anthropic’s technology after the two sides failed to agree on terms? Isn’t Anthropic suing the Trump administration over a $200 million contract dispute? Didn’t Hegseth officially designate Anthropic a “supply chain risk”? Yes. That’s precisely why they’re doing this. Anthropic is simultaneously courting the government and taking it to court. The government is simultaneously designating Anthropic a supply chain risk and welcoming it into its classified network. Trump claims no regulation and days later he signs the executive order.

This is the political version of that popular term used in pro wrestling: kayfabe or “be fake.” We, the unwitting viewers, cheering back and forth, are the actual target. Because what else would remain of an AI industry but an effective oligopoly?

In the policy paper that I mentioned above, Anthropic describes an ideal 2028. They call it “Scenario One,” in which “a small number of AI labs lead at the frontier with the most intelligent, capable, and performant models. All are based in the US.” We know who the small number refers to: Anthropic, Google, and OpenAI. Their pricing power will be absolute. The subsidies on AI tools will be over. When the three top labs are favored over the rest and the government, which is itself their client, gatekeeps the releases, the competitive pressure that keeps AI affordable disappears. AI has entered its non-democratic era.

The industry no longer needs individual users to survive. The “IPO lens” has replaced the “user lens.” You and I are not the engine of their business anymore and will never be again, but the government, enterprises, and investors are. Everything for the people without the people, I guess. In professional wrestling, they pretend to hurt each other. In AI, they court each other at court.

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.

I have an IMPORTANT REMINDER for you.

I’m running a 50% discount offer on annual subs that ends on June 5th (tomorrow). Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

AGI is Silicon Valley’s favorite three-letter acronym, but there’s a rising contender. Earlier this week, Anthropic announced in a blog post that it has confidentially submitted a draft S-1 statement to the Securities and Exchange Commission. Once it’s reviewed, they will have the option to launch an IPO (initial public offering). Their reported annualized revenue run rate is nearing $50B, and they’re on track for the first profitable quarter ever. Pretty good. OpenAI is closely trailing them. Revenue is lower, and there’s no profitability in sight yet—the consumer market is less valuable than the enterprise market—but they’re pretty much equal. Indeed: they’re both bleeding cash.

Bleeding is not a sustainable business model at their size anymore. One can be unprofitable for a while, and no one will bat an eye—looking at you, Uber—but if you require a planet’s worth of energy and cash injections to run, then we will intensely bat our eyes. At some point, either you figure out how to survive or the economy dies. Well, they figured it out by transferring the risk to the economy. They did become “too big to fail.” Now if they die, we die with them. The timing—non-profitability, exorbitant valuations, huge spending commitments without enough revenue, huge CapEx without returns—raises questions. One might think OpenAI and Anthropic are rushing to go public—that the race was never toward AGI but against the clock.

The cynical view, which I don’t hold but Bank of America’s chief strategist Michael Hartnett does, is that this double IPO story—triple if you count SpaceX, which acquired xAI recently—is less a public-market opportunity than a way for early investors to hand years of accumulated risk to the public market. They’re racing, yes, but not one another; they’re racing against us. The IPO is how insiders will exit the AI bubble. “The pricking is the converting of wealth into money,” says Bridgewater founder Ray Dalio. You thought a bubble meant they’d lose? That’s not how it works: if no bubble, they’d do business as usual, getting rich eventually. If bubble, they’d exit, getting rich now. So even a bubble is bad for us: they’re passing AI’s excessive valuation on to pension funds, index funds, and the broader economy.

If the market realizes that the bubble was actually a bubble and not exaggerated anti-hype before the IPO, Anthropic and OpenAI’s chances to cash out at the peak get shattered. If it happens afterward, however, the problems are redistributed to the broader economy. Privatize the gains, socialize the losses, or something to that effect. If the gap between promise and delivery is fundamental rather than temporary, it’s you and me, normal people, who are screwed.

And the gap is becoming too obvious to ignore: enterprises are losing trust. Just this month, Microsoft announced it’s cancelling Claude Code licenses in favor of an internal tool due to unsustainable costs. Uber is “limiting all employees to $1,500 in monthly token spending per AI coding tool.” Starbucks is retiring its AI tool across North America after 9 months due to reliability problems. Even Sam Altman has admitted that the AI costs have become “a huge issue.” We’re about to get inundated with stories like these because the problem is universal: there’s no clear relationship between AI spending and AI returns. Their IPO window isn’t closing because AGI is almost here, but because the story about “AGI almost here” has a fleeting shelf life.

Was this their plan all along? I don’t think so. I think Silicon Valley is echo-chambered enough that these guys’ beliefs about “superintelligence for 2027” are actually genuine. But so what? They are nonetheless discharging the risks onto the rest of us either way—we have all the reason to be mad.

But risk doesn’t mean danger and it’s only fair to acknowledge what happens if AI does work out. In that case, the IPO would extend what is now a private gain into a public gain. That’s how the world prospers, how the tide lifts all boats, even those that didn’t do any work at all. Unfortunately, as neat as it is that story about the wonders of capitalism, it brings me no consolation: I don’t trust that a technology this unreliable—we can say for sure that AI models won’t stop hallucinating—will pass the test of time as the internet or electricity did before.

This is, to me, the core issue here. They win in any case, either as enablers of a technological revolution or as wealthy shareholders. We only win if AI works out. It’s paradoxical to put it this way, but it’s us—you and me, normal people—whose future is intimately tied to the future of AI, not theirs.

But why won’t AI pass the test of time? Figuring innovation out into productivity and economic growth takes time, so why is AI different? To answer this, I will ask another question: what does a non-figure-out-able technology look like? I do know.

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.

I have an IMPORTANT REMINDER for you.

I’m running a 50% discount offer on annual subs that ends on June 5th. Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

Read more

Hey, Alberto here! 👋 I have an IMPORTANT REMINDER for you.

I’m running a 50% discount offer on annual subs that ends on June 12th. Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

AI is like religion. Either you believe it changes everything, or you don’t believe at all. There is no moderate position; nobody believes in AGI “more or less,” just like nobody is “casually religious.” If God exists, the only coherent response is to reorganize your entire life around that fact, as priests do. If you pray sometimes, then you are just an atheist who’s also fearful. When tech companies spend hundreds of billions on capital expenditures to add sparkly AI features to Office, Gmail, and Instagram, I only see fearful atheists—guys who don’t believe in AI but pretend just in case.

In 2026, the four largest cloud and AI infrastructure providers—Amazon, Google, Meta, Microsoft—committed to spending $670 billion on CapEx. Apple, in contrast, spent $12.7 billion on capex last fiscal year and projects $14 billion for 2026, 2% of what its peers are spending. The conventional reading in Silicon Valley is, naturally, that Apple is losing. Siri has been a punchline for years—an internal executive called the delays ugly and embarrassing—and critics say that Apple has not been the same without Steve Jobs. It is falling behind, they say, and moving way too slowly for AI.

I disagree with this portrayal: Apple is the most powerful tech company in the world right now because it’s acting according to what it believes.

Tim Cook—far from an “AI guy”—announced in April 2026 that he would step down as CEO. His successor is John Ternus, a 25-year Apple veteran who runs hardware engineering. Cook could have named someone who would reorient the company around the most hyped technology in a generation. Instead, he picked another non-AI guy. That is what a true non-believer looks like. One who is not only an atheist but fearless about it. If anything, Apple oozes defiant conviction.

In March 2026, Bloomberg reported that Apple is opening Siri to third-party AI models. “Smart move,” I thought. Users will be able to route queries to ChatGPT or Claude directly from Siri. Apple is building a menu that lets you choose which AI powers your assistant rather than get lost in a zero-sum race with fake believers and actual advocates (that would be Anthropic, the only AI lab with a conviction comparable to Apple’s). You do this if you believe AI models and agents are swappable; a commodity that will play out like electricity, not a divine machine.

So Apple’s position is clear. What is everyone else’s?

The same; they just won’t say it.

The AI industry is playing cool. And right now, roleplaying as a fanatic is the cool thing to do (meaning: it is what keeps the revenues coming, shareholders happy, and valuations intact). The moment the wind changes, we will see their true colors. But for now, let’s look at the data to compare.

Mark Zuckerberg can say shit about “personal superintelligence” to attract investors, but he will pour billions on whatever tech is hot, from 3D printing to the metaverse. Meta’s primary consumer-facing product is a chatbot that lives inside WhatsApp. Sundar Pichai, Alphabet CEO, will happily announce Spark, a “24/7 personal agent” for mundane tasks, side-by-side with Omni, an AGI seed. As if they’re the same. Google spent ~$93 billion on AI last year and then embedded Gemini into Gmail, where it suggests slightly worse replies, and into Search, where it makes up responses. Satya Nadella, Microsoft CEO, will invest in OpenAI before anyone else, but then forces its Office clients to stick to Copilot. Elon Musk will build a giant AI datacenter and then rent it at a profit to a competitor. The entire AI ecosystem is made up of entrepreneur types trying to inflate the pie to take a portion before it explodes. Even Sam Altman. Sam Altman doesn’t believe in anything. Or rather, he believes in everything at different times, whatever the public wants to hear, he obliges. And so on.

These guys keep praying to all the gods just in case their own is not the real one.

That’s not what a company that believes “AI will change everything” does. They are bolting AI onto existing products the way they once bolted social features onto existing products. To them, AI is, if lucky, a new dot-com boom. They speak like revolutionaries but ship like bourgeois. They believe what Apple believes but behave like cowards; if they genuinely believed God was watching them, they wouldn’t hedge.

What they’re doing has a more charitable name in theology: Pascal’s Wager. They can’t prove AI will be transformative. But if it is and they don’t invest, they’re dead. Or worse: the laughing stock. However, if it isn’t transformative and they invest, they only lose some cash they will recoup anyway by doubling down on ads. So they tithe. They go to church on Sunday. Then fast for Ramadan, keep Shabbat, leave offerings at the shrine. They even perform hecatombs, and, every few months, bathe in the Ganges.

Apple is willing to let its actions match its actual beliefs. That’s a stance I can respect. The rest are performative sheeple overdrinking from the fountain of fear.

Over the years, I’ve shared many charts that illustrate how AI and the industry are evolving. But I’ve never put them together. This is the bigger picture that standalone news articles miss. When you see them at once, things start to make sense.

I’ve compiled 11 charts that I consider extremely important: the ones that reveal the ugly side of the story. A chart tells you what the industry wants you to see. Enough charts, and you start to see through the industry.

The format is: title, chart, one-liner explanation with source. In 5 minutes of reading time, you will understand more about where the AI industry is heading than from hundreds of pages of corporate reports.

Here’s the outline:

1. TEN COMPANIES ARE CARRYING THE ENTIRE ECONOMY

2. THE CAPABILITIES OF AI HAVE BEEN EXAGGERATED

3. MOST TOKENS GENERATED BY AI ARE WASTED

4. AI RELIABILITY IS BARELY IMPROVING

5. THE AI INDUSTRY RUNS ON CIRCULAR DEALS

6. PEOPLE ARE NOT USING AI TO ITS FULL POTENTIAL

7. AMERICANS DON’T WANT DATACENTERS NEARBY

8. AI USERS BELIEVE AI EVEN WHEN IT’S WRONG

9. WORKERS DON’T THINK AI IS SAVING THEM TIME

10. AI IS KILLING THE JOBS OF JUNIOR DEVELOPERS

11. THE IMPOSSIBLE MATH OF THE AI BOOM

IMPORTANT REMINDER

I’m running a 50% discount offer on annual subs that ends on June 5th. Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

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 now

A practical guide on how to deal with agentic AI vices that degrade your work.

If AI has taught me anything, it is that struggling because you have to do work is better than struggling because you don’t.

We yearn to free ourselves from work right until the moment we do. Then we yearn to bring it back. AI has put an end to that paradox: it removed the struggle to do work and won’t give it back unless you make the effort to backtrack from the habits you’re currently building. Your work doesn’t require you, and you don’t know how to deal with that.

And so tweets like this one below appear on my feed every day.

The sentiment has generalized across software development jobs and more broadly, white-collar jobs. Work has become, like everything else, a casino slop machine. You pull the lever, you get something, you evaluate the situation. You pull the lever again. Etc. You don’t do stuff anymore. You manage and oversee and organize and hover over tasks, not fully detached but not engaged either.

The sheer amount of work that an AI agent—or a swarm of them—can do prevents you from getting in the mud with it. You are not fast enough, can’t go deep enough. And so you hover over. Your work consists of chaining together a ridiculous number of decisions one after another—including the decision to delegate decisions—without feeling the friction of making them. The number of choices you make is daunting, but volume is not your problem.

The problem is that being an evaluator for a living doesn’t feel like being a generator, which is what your brain wants to be. When you generate the work, you have to ask “what should I build?” or “How do these elements connect?” When you evaluate the work, you go, “Is this what I want?” “Does this meet the requirements?”

You’re constantly in a cognitive mode that’s purely consuming a mental model proposed by the AI; you don’t get to build it yourself. How can you consume a resource you don’t replenish?

That question is at the core of the “brain fog” or the “brain fry” testimonials. The more you take on this manager role, the less time you spend honing the practice of “understanding by doing” and so the worse you actually become at managing itself because your mental models increasingly drift away from the actual nature of the tasks you have to manage.

Mastering AI agents is a matter of finding the sweet spot. Just doing keeps you good at doing. Doing plus managing makes you good at both. Just managing makes you good at none.

Something similar has happened to me.

I use AI to help me research, and it’s good at gathering plenty of material from reputable sources in a matter of minutes. That’s great—it was literally impossible without AI. But then I’d realize that I don’t know what to do with it: I have 500 pages of good-quality information, and it means nothing to me because I’m disconnected from the process that brought it to me. I’m lost in the middle of an ocean. Turns out that the output of research is not the goal of research. If I did scour the internet looking for sources, I would make connections that I need. Ideas present themselves before me during the act of actively looking for stuff, not passively receiving stuff. I lose that by offloading the process. The output of research is only worth for me contingent on my putting it together.

These ideas have never been more relevant. They are not exclusive to AI—as it always happens with innovation, we readily jump into it but rarely zoom out of it to take perspective—and far from new.

In her 1983 landmark paper Ironies of Automation, Lisanne Bainbridge observed that the more you automate a process, the more you need, as the human operator, to be skilled (for edge cases, for oversight, etc.), but the less practice you get. Google and Stack Overflow entail partial offloading, yes, but you still need to generate the work yourself. Generative AI takes that away from you.

With AI, the product of your work comes out faster, but your participation thins and thins. Perhaps, if you’re lucky, the product is better, but you, in contrast, are worse. When you start to think in terms of “what” and forget to do the “how,” you eventually don’t know “what” the hell is going on.

The fix is not hard, just requires an intentional mindset change.

What you need to do to keep the benefits of agentic AI—returning entirely to pen and paper isn’t realistic—whilst not surrendering your skills as a human, is to switch between evaluative cognition and generative cognition. Keep doing the thing, start managing the thing, and get better at both.

I think the idea of “desirable difficulty” is the sweet spot you should aim for. It’s not maximizing evaluative skills. It’s not maximizing hands-on skills either. It’s a weighted mix of both that depends on who you are, your skills, your job, etc. Only you can find the right combination.

My general advice is this: don’t be afraid to backtrack from fully relying on agents just because your peers and the industry are telling you to go full “AI manager mode.”

Eventually, they, too, will have to do this, which gives you an important edge if you start now. Eventually, they will realize that it was a mistake. One born out of a desire to work less and work faster. The correct desire, however, should be to work better, which is rarely achieved by either. In the AI era, better work requires you to gain those AI skills without losing your human ones.

Now that I’ve stated the situation, you’re in a position to understand my practical advice below. Seven things you should stop doing right now to become the best “human doer” + “AI manager”:

1. Stop automating the tasks that teach you things

2. Stop rejecting AI output on fuzzy judgment

3. Stop offloading to agents right away

4. Stop splitting your work into “AI” and “human” tasks

5. Stop choosing two birds in the bush

6. Stop using AI to evaluate AI to evaluate AI

7. Stop deferring the understanding to later

IMPORTANT REMINDER

I’m running a 50% discount offer on annual subs that ends on June 5th. Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

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 now

This is going to be the hottest topic for the rest of the year.

INTRODUCTION

Between 2024 and 2026, AI datacenters went from obscure municipal zoning topics to front-page political fights. A large number of polls from Gallup, Pew Research Center, NBC News, the Washington Post, Change Research, Marquette Law School, and others have measured what Americans feel and think about AI, datacenters, AI companies, and the people who run them.

Everyone has a vague sense that Americans are turning against AI—more so than in any other country—but no one has put together the full map. It’s lost in fruitless discussions on X, localized anecdotes and repetitive news stories.

This is the full map.

I’ve gathered every major piece of survey data I could find.

The global conclusion that emerges from this compilation is simple and, as far as I can tell, unprecedented: every available measure—partisan, demographic, regional, longitudinal—points in the same direction.

Here’s the outline:

1. AMERICANS DON’T WANT DATACENTERS NEAR THEM

2. AMERICANS DON’T TRUST AI IN GENERAL

3. AMERICANS DON’T LIKE THE PEOPLE BUILDING AI

4. AI HAS BECOME A BIPARTISAN PROBLEM

5. WHERE THE AI BACKLASH HAS MATERIALIZED

6. WHAT WE DON’T KNOW

7. GLOBAL CONCLUSIONS

When someone asks what we know about the general sentiment against AI, you show them this compilation. When someone argues that anecdotal evidence is not enough, you show them this. When someone says most people actually love this technology, you show them this.

I. AMERICANS DON’T WANT DATACENTERS NEAR THEM

Multiple pollsters have asked versions of this question between 2025 and 2026, with remarkably consistent results. Americans have gone full NIMBY on datacenters.

• Americans Oppose AI Data Centers in Their Area (Gallup, publication date: May 2026, N=1,000 adults): 71% oppose the construction of an AI datacenter in their local area; 48% are strongly opposed. Only about a quarter favor construction, with just 7% strongly in favor. 70% worry “a great deal” (46%) or “a fair amount” (24%) about the environmental impact of AI datacenters. Among opponents, roughly 50% cited excessive resource use (18% water, 18% energy), ~20% cited quality-of-life concerns, and ~20% cited economic worries. Among supporters, two-thirds cited economic benefits, including 55% who specifically mentioned jobs. (Datacenters have surpassed nuclear as the most-opposed local infrastructure.

  

• Data Center Costs Outweigh Benefits (Marquette Law School Poll, October 2025 and February 2026, N=818 Wisconsin registered voters): In October 2025, 55% of Wisconsin voters said the costs of large datacenters outweigh their benefits. By February 2026, that number jumped to 70%. Partisan breakdown: Republican opinion barely moved (53% to 55%), independents surged (55% to 76%), and Democrats exploded (56% to 85%).

• In This U.S. Hot Spot for Data Centers, Voters Have Turned Against Them (Washington Post–Schar School, April 2026, N=1,101 Virginia registered voters): Comfort with a local datacenter collapsed from 69% in 2023 to 35% in 2026. 57% say datacenters negatively affect home energy bills. 59% say they damage the local environment. In Loudoun County, where datacenters directly fund schools and suppress property taxes, 51% believe they are making their tax bills worse.

  

• Opposition to Data Centers Has Grown Sharply Since 2025 (Change Research, April 2026, N=2,702 national registered voters): Local support for datacenter construction fell from 51% support / 26% oppose in March 2025 to 25–26% support / 65% oppose in April 2026, a more than 40-point net swing in a year. For construction near respondents’ own homes, the opposition is ~74%.

  

• Only 44% of Americans Would Welcome a Data Center Nearby (Heatmap News / Embold Research, September 2025, N=3,741 registered voters): Only 44% would welcome a datacenter near them, 42% opposed. Datacenters ranked below solar, wind, natural gas, nuclear, and battery storage in local acceptability. A follow-up Heatmap Pro poll in early 2026 found opposition had deepened further, with net support for datacenters at −8% among suburban/urban Republicans and −20% among rural Republicans.

• Voters Are Turning on AI Data Center Construction (Morning Consult Intelligence, February 2026, N=~2,200 registered voters): Support for a ban on the construction of AI datacenters near respondents’ communities rose from 37% (October 2025) to 41% (November 2025), while opposition to a ban fell from 39% to 36%. In the same period, the share of voters who said AI datacenters are “very or somewhat responsible” for rising household electricity costs grew from 54% to 58%. 45% of voters said datacenters have adverse effects on the price of electricity—the single most-cited concern.

  

Conclusions: Pollsters, methodologies, and geographies converge. The opposition to AI datacenters is a measurable, replicable fact about American attitudes (it doesn’t necessarily replicate in other countries). To me, the most striking number in this section is that datacenters now exceed nuclear plants in local opposition, an infrastructure category with decades of accumulated fear behind it. Just how badly has the industry botched things to generate such strong opposition?

II. AMERICANS DON’T TRUST AI IN GENERAL

The data center backlash sits inside a broader, older pessimism about AI itself. I don’t think this reflects people’s opinions about the usefulness of the technology, but rather the perception that it’s unreliable and easy to misuse (this applies to both individuals and enterprises).

• How Americans View AI and Its Impact on People and Society (Pew Research Center, September 2025, N=5,023 adults): 50% of U.S. adults are more concerned than excited about AI in daily life; only 10% are more excited than concerned. Concern was 37% in 2021; a 13-point rise over four years. Majorities believe AI will worsen Americans’ creativity, ability to form meaningful relationships, decision-making, and problem-solving.

  

• Majority of Voters Say Risks of AI Outweigh the Benefits (NBC News, March 2026, N=1,000 registered voters): 57% say the risks of AI outweigh the benefits; 34% say the opposite. Net AI favorability: −44 among 18–34-year-olds, −41 among women 18–49.

  

• How the US Public and AI Experts View Artificial Intelligence (Pew Research Center, April 2025, N=1,013 experts, N=5,410 adults): Asked whether AI’s impact on the US over the next 20 years will be positive or negative, only 17% of US adults said positive, while 35% said negative. Among AI experts, the picture is inverted: 56% said positive, 15% said negative. The gap between how the public and the people building AI see its future is one of the largest expert-public divides Pew has measured on any technology topic

  

• Americans aren’t sold on AI’s benefits (Politico / Public First, May 2026, N=2,035 adults): 43% say AI’s risks outweigh its benefits; 33% say the opposite. 44% say AI is developing too quickly, a view shared for the most part by 2024 Trump voters (42%) and Harris voters (49%).

  

• Gen Z’s AI Adoption Steady, but Skepticism Climbs (Gallup / Walton Family Foundation / GSV Ventures, April 2026, N=1,572 ages 14–29): Excitement about AI fell 14 points in one year to 22%. Hopefulness fell 9 points to 18%. Anger rose 9 points to 31%. Even among daily AI users, excitement dropped 18 points year over year. 42% of bachelor’s-degree students have reconsidered their major because of AI. Gen Z uses AI the same amount, but likes it much less.

  

Conclusions: The opposition to datacenters doesn’t exist in a vacuum, but within a broader, deepening pessimism about AI itself that has been building since at least 2021-2022—even as usage has grown, particularly among young people—and has not reversed or plateaued. My reading is that datacenters are the physical manifestation of an abstract anxiety toward the technology.

III. AMERICANS DON’T LIKE THE PEOPLE BUILDING AI

The faces of the AI industry are deeply unpopular, and Americans increasingly view the relationship between tech CEOs and political power with suspicion.

• Americans Deeply Critical of Big Tech CEOs (Tech Oversight Project / PPP, June 2025, N=541 registered voters): Every major tech CEO underwater on net favorability—Zuckerberg −41, Pichai −34, Bezos −22, Cook −22, Altman −16. A later polling cycle put Zuckerberg at −59, Pichai −38, Bezos −45, Altman −36. 70% of voters believe Big Tech CEOs are simply trying to appeal to Trump rather than acting on principle.

• How Americans View Elon Musk and Mark Zuckerberg (Pew Research Center, February 2025, N=5,086 adults): 54% unfavorable view of Musk (36% very unfavorable); 67% unfavorable of Zuckerberg (26% very unfavorable). Musk is sharply polarized by party (R +73 favorable, D 85% unfavorable); Zuckerberg is broadly disliked across parties.

  

• Majorities Concerned About Big Tech Influence and Corruption (Navigator Research, February 2025, N=1,000 registered voters): 69% of Americans believe tech CEOs (Musk, Zuckerberg, Bezos) will have significant influence in the Trump administration; 50% call that a “bad thing” vs. 37% “good thing.” 41% worry Trump will look after billionaires instead of the middle class because of this influence.

Conclusions: Both messengers and messaging are part of the problem; there’s a category-level credibility catastrophe. Whether the backlash against AI infrastructure would exist without the CEO’s favorability collapse is unknowable, but the correlation is there: as the industry’s public faces became more politically visible—and more unpopular—the infrastructure they were building became less welcome. The AI industry is being judged on 1) the infrastructure enabling it, 2) the products they sell, and 3) the character and political motivations of the people building them.

IV. AI HAS BECOME A BIPARTISAN PROBLEM

Over the past few years, there’s been a collapse of the partisan gap on AI concerns. The standard analytical frame—Democrats worry about safety, Republicans worry about government overreach—no longer describes the data.

• Americans Oppose AI Data Centers in Their Area (Gallup, publication date: May 2026, N=1,000 adults): Partisan breakdown: 56% of Democrats are strongly opposed vs. 39% of Republicans (independents 48%). Demographic breakdowns: women 55% strongly opposed vs. men 43%. Regionally: Midwest 76% total opposition, South 75%, East 68%, West 63%. No meaningful differences by age, race, education, income, or urbanicity — this is not a coastal or college-educated phenomenon.

  

• Republicans, Democrats Now Equally Concerned About AI in Daily Life (Pew Research Center, November 2025, N=5,023 adults): Republicans and Democrats are now essentially tied on “more concerned than excited” about AI (50% vs. 51%). From 2021 to 2024, Republicans had been consistently more concerned; that gap has closed entirely. Where the parties diverge is trust in regulators: 54% of Republicans trust the US to regulate AI effectively vs. 36% of Democrats, an inversion of the typical pattern that reflects partisan control of the White House

  

• Majorities of Republicans and Democrats Overwhelmingly Favor Regulating AI (University of Maryland Program for Public Consultation, August 2025, N=1,202 adults): Bipartisan supermajorities favor concrete AI regulation. 79% favor requiring AI programs to pass a government test before deployment (R 84%, D 81%). 78% favor allowing government audits of in-use AI systems.

• Voters Are Turning on AI Data Center Construction (Morning Consult Intelligence, October and November 2025, N=~2,200 registered voters): Republicans are the most pro-ban at 47% (up from 43%), Democrats at 41% (up from 34%), independents essentially flat at 32%.

  

Conclusions: Both sides worry about the same thing at nearly identical rates. That both Gallup and Pew find majority opposition across party lines defies the assumption that AI skepticism is a progressive phenomenon. AI has achieved something rare in American politics: it has united voters who agree on almost nothing else. Where they diverge is about who should fix it, not whether it needs fixing. Dario Amodei says that “We’re going to find that ideology will not survive the nature of this technology. The things I’m talking about are gonna become bipartisan and universal because everyone will recognize the necessity of it.” I gotta give him that one: both sides do agree.

V. WHERE THE AI BACKLASH HAS MATERIALIZED

This section describes what the attitudes emerging from the polls have already caused in the ground: canceled projects, organized opposition groups, billions of dollars in disrupted investment, and a legislative wave at both the state and federal level, moving the regulatory posture from incentive competition to oversight and restriction.

• Data Center Watch Report (Data Center Watch, Q2 2025): $156 billion in U.S. datacenter projects were blocked or delayed in 2025 by local opposition, moratoria, or litigation. 188 active opposition groups tracked across the country, 53 of them newly formed in Q2 2025 alone. (The $156B figure includes both blocked and delayed projects; it should be read as a “disruption” total, not a “cancellation” total.)

  

• Prince William County Drops Legal Fight for Digital Gateway (Virginia, April 2026): The Board of Supervisors voted unanimously to withdraw from litigation defending the Prince William Digital Gateway, a 2,100-acre, 37-building, $24.7 billion campus that would have been the largest datacenter corridor in the world. The county had spent $1.72 million in taxpayer funds fighting its own residents in court. Developer Compass Datacenters formally abandoned the project two weeks later. The site bordered Manassas National Battlefield Park.

• Tucson City Council Rejects Project Blue (Arizona, August 2025): A 7-0 unanimous vote killed a 290-acre datacenter campus linked to Amazon Web Services after weeks of public hearings. The developer subsequently bypassed the city by securing county zoning instead, which generated a second wave of opposition. Water consumption in the desert was the central concern.

• Prince George’s County Pauses All Datacenter Development (Maryland, September 2025): County Executive Aisha Braveboy signed a moratorium order after a resident, Taylor Frazier McCollum, posted a petition against a proposed datacenter at the former Landover Mall site. It gathered over 20,000 signatures in two weeks. Governor Wes Moore later signed the Utility RELIEF Act requiring datacenters to pay for their own grid upgrades.

• Data Center Moratorium Bills Are Spreading in 2026 (Good Jobs First, February 2026): Moratorium bills have been introduced in at least 14 states in 2026. More than 300 datacenter-related state bills were filed in 30+ states in the first six weeks of the year alone.

• Artificial Intelligence Data Center Moratorium Act (Sen. Bernie Sanders, March 2025): A federal bill calling for a national pause on datacenter construction. In December 2025, 230+ environmental groups—including Food and Water Watch, Friends of the Earth, and Greenpeace—sent a joint letter to Congress demanding the same.

Conclusions: Billions in disrupted investment, community groups multiplying faster than projects can be approved, and a legislative system that moved from incentive competition between states to oversight and restriction in under two years is not the effect of a poll result but the consequences of real-world action. This issue is not staying on social media.

VI. WHAT WE DON’T KNOW

Four critical gaps remain in the data:

• No baseline to compare. Gallup’s data-center question was asked for the first time in March 2026, so there’s no trend line. The Change Research and Washington Post–Schar comparisons offer before/after snapshots, but there is no systematic annual tracking of data-center sentiment across the same population over time. We don’t know what the starting point used to look like nationally.

• Causation is unmapped. Americans oppose data centers, dislike AI CEOs, and distrust AI broadly, but which is driving which? Did the CEO favorability collapse poison attitudes toward datacenters by repeatedly claiming AI will take everyone’s jobs? Or did local datacenter fights poison attitudes toward CEOs, which were already pretty dismal? Did the broader AI anxiety create the conditions for local opposition, or did local opposition generate the broader anxiety? The correlations are clear but not so much the causal arrows.

• An (un)informed opposition. What do opponents actually know about datacenters? I’ve intentionally left this question out: for instance, Gallup’s open-ended responses show people citing water, energy, pollution, and utility bills, but no poll has tested whether respondents can distinguish accurate from inaccurate claims about datacenter impacts. This matters for predicting whether the backlash is durable and immutable or responsive to new evidence, but the opposition is real anyway: it shapes votes and kills projects whether or not it is factually grounded.

• The election hypothesis. No American election has yet been won or lost on AI. The 2026 midterms will be the first natural experiment on that. Until the votes are counted, the claim by multiple analysts that AI backlash translates into electoral outcomes is just a prediction. Cost of living still dominates voter priorities, and AI’s role may be as an issue underneath economic anxiety rather than above it.

VII. GLOBAL CONCLUSIONS

Core finding: every available measure converges on a collapse of public opinion

Across every pollster—Gallup, Pew, Change Research, the Washington Post, Marquette, Morning Consult, Heatmap, NBC News, Politico, the University of Maryland—the direction is the same: Americans have soured on AI datacenters, on AI itself, and on the people building it.

It happened faster than the backlash against nuclear power, than the backlash against wind farms, or any comparable infrastructure opposition in the modern polling era. Change Research measured a 40-point net swing in a single year. In three years, the Washington Post–Schar poll captured a 34-point collapse in Virginia, the one state that should be most favorable to data centers due to associated benefits. Morning Consult found support for an outright nationwide ban rising 4 points in one month. This is a collapse of public opinion, not a gradual shift.

The speed makes this story urgent; the convergence makes it important.

When every pollster, using different samples, different question wordings, different modes (telephone, online, in-person), and different time points, finds the same thing, you have a consensus. The opposition to AI infrastructure is bipartisan in composition and only varies slightly in intensity by party, gender, and region. It is not limited to abstract attitudes—it has materialized in $156 billion of disrupted projects, 188+ opposition groups, 300+ state bills, and federal legislation—and even the industry’s own political spending strategy treats the AI brand as too toxic to defend on-message.

When they dare defend it, though, the reaction is as homogeneous as the polls suggest. Below is a video of what happens when a commencement speaker talks to university students about AI (featured: Eric Schmidt, former Google CEO; Scott Borchetta, Big Machine Records CEO; Gloria Caulfield, Tavistock Development VP):

Two things worth noting from this.

1. This doesn’t happen in STEM schools (might be the only exception). That contrast aligns with the Pew Research data on the divide between experts and the general population.

2. Students don’t oppose AI from ignorance but experience. This was quite shocking to me: Gallup found that Gen Z high usage has not decreased despite anger having increased. They know what they’re protesting against because they know what AI is doing to them.

It’s booing today, what will be tomorrow? I don’t want to sound alarmist but there’s only one good moment to warn about violence: too soon. As I wrote in “AI Will Be Met With Violence, and Nothing Good Will Come of It”:

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.

This is America screaming they are being left behind and they don’t want to be left behind.

But why has the sentiment changed so much so rapidly? Why have the poll results on datacenters and AI degraded so intensely over the past year? Why now?

The underlying dynamic is obscured by the intensity of the backlash but my general reading is this: the sentiment was always there but it only became legible after it reached a certain critical mass. The snowball needs to get big before it starts rolling down the mountain’s side. Once it does, though, it rapidly turns into an unstoppable avalanche.

We’re living through the first moments of that avalanche.

But let me give you a more specific take. I think it can be of value to the industry, if anyone there is hearing the rest of the world anymore or their unprecedented success has fully deafened them.

Americans never got to form an opinion about AI gradually. They didn’t get to see both good and bad over time, balancing them out. There are various reasons for this.

1. The technology arrived as a cultural event—ChatGPT in November 2022 “took the world by storm,” as they say—and within three or four years, it was everywhere.

2. The industry has force-fed AI to everyone; companies are cramming it into every existing digital service to retroactively rationalize the massive spending.

3. They keep warning that we’ll lose our jobs as they keep enthusiastically rolling out the very products responsible for the presumed mass unemployment (Claude Code and Codex and etc.).

4. They claim to be worried if the full-blown commercial race goes wrong (e.g., recursive self-improvement) but then allude to competition pressures to keep going.

5. They wreaked havoc on the digital landscape with insane amounts of slop.

6. And after all of that, they’re asking for permission to reshape the physical world: the landmass, the electrical grid, the water supply, the view from people’s backyards.

The speed of the buildout outpaced the speed at which people could form a considered view and the natural response to sudden change with dangerous undertones is resistance and, to whatever degree possible, organized counterattack. Factuality is an afterthought, if at all.

Social media allows unprecedented speeds for sentiment to spread—as well as information and misinformation—but I believe the AI industry’s unfortunate messaging was the primary cause.

None of this tells us whether the backlash is permanent, though.

The nuclear analogy is instructive but cuts both ways: nuclear power lost public trust in the 1970s and never fully recovered, which suggests infrastructure opposition can be durable for decades. On the other hand, nuclear opposition was reinforced by Three Mile Island and Chernobyl, specific, catastrophic events that validated the fear.

The AI backlash has no equivalent singular disaster (yet); it’s driven by diffuse anxieties about energy, water, jobs, utility bills, corporate power, and the speed of change. Whether the industry course-corrects before a crystallizing event, or whether the diffuse anxiety becomes the crystallizing event through accumulated experience is an open question—one they shouldn’t take long to answer.

Otherwise, these polls will soon be regarded as the beginning of the end.

This compilation took a lot of work. It helps a lot if you share it with people who could be interested or learn something from this. Here’s a button for that:

Share

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

An express post on the significance of this breaking news.

Andrej Karpathy, founding member at OpenAI, former lead of Tesla Autopilot, and the most beloved AI teacher in the world, has joined Anthropic today, May 19th, 2026.

TechCrunch and Axios report that he will work under team lead Nick Joseph on pre-training, “focused on using Claude to accelerate pre-training research.” Essentially, Anthropic has hired Karpathy to prepare Claude to improve itself, a capability that, once autonomous, is known as “recursive self-improvement” or RSI.

This is the future that Anthropic co-founder Jack Clark predicted on May 4th in his newsletter, ImportAI:

I’m writing this post because when I look at all the publicly available information I reluctantly come to the view that there’s a likely chance (60%+) that no-human-involved AI R&D - an AI system powerful enough that it could plausibly autonomously build its own successor - happens by the end of 2028.

This is a big deal.

I don’t know how to wrap my head around it.

It’s a reluctant view because the implications are so large that I feel dwarfed by them, and I’m not sure society is ready for the kinds of changes implied by achieving automated AI R&D.

Karpathy’s decision to join Anthropic has come as a surprise to most.

First, he was in the founding team at OpenAI. He left for Tesla in 2017. Then he rejoined OpenAI three years ago.

He then left again two years ago.

Together with Tesla, OpenAI was his “industry home.”

Second, he likes to share his work in the open and say whatever he wants, and that’s not possible working at a frontier AI lab (except if you’re a pseudonymous poster). The field used to be open, but as the industry grew and commercial interests overshadowed research cooperation, the frontier labs became secretive.

Third, he became surprisingly skeptical of the industry’s over-hype of AI models and agents’ ability to generate good code. As a professional developer, he thought they “need[ed] a lot of work.” He made these remarks in his appearance on the Dwarkesh Podcast on October 17th, 2025.

Overall, the models are not there. I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop. . . . I’m not sure what’s going on, but we’re at this intermediate stage. The models are amazing. They still need a lot of work. For now, autocomplete is my sweet spot. But sometimes, for some types of code, I will go to an LLM agent.

Three days later, on October 20th, Anthropic launched Claude Code on the web. (Claude Code had been available as a research preview since February and generally since May, but getting these things to work well takes time.) Over the winter holidays, developers tinkered more seriously with these tools, eventually sparking an agentic boom out of Claude Code.

It changed everything.

On December 19th, Karpathy wrote:

[Claude Code is packaged] into a beautiful, minimal, compelling CLI form factor that changed what AI looks like - it’s not just a website you go to like Google, it’s a little spirit/ghost that “lives” on your computer. This is a new, distinct paradigm of interaction with an AI.

Then on December 26th, he tweeted:

I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year and a failure to claim the boost feels decidedly like skill issue.

Then two months later, on February 25th, again:

It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the “progress as usual” way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow.

This is a three-tweet summary of the 180-degree turn that the entire industry has undergone over the last six or seven months (except the people inside the frontier labs, who saw it coming). Karpathy called out the hype, and four months later, he understood why the labs were hyping: “Roll up your sleeves to not fall behind.”

Naturally, he kept us updated on his work.

Since that moment, he’s been trying to automate different aspects of AI research with AI agents. This is, allegedly, the current main goal of both Anthropic and OpenAI (Not so much Google DeepMind, but that’s a topic for a different post).

Automating AI research is the first step to achieving the “escape velocity” threshold for recursive self-improvement. They believe this is the most probable path to artificial general intelligence, superintelligence, and the singularity. (Whether that’s true and whether you should care at all—read yesterday’s post for that—is another question.)

But, not to get tangled in concepts that belong in a Vernor Vinge novel, let’s stay with his autoresearch, perhaps the most relevant instance of Karpathy’s efforts.

On March 7th, 2026, he tweeted:

The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. . . . Part code, part sci-fi, and a pinch of psychosis :)

In the GitHub code repository, he wrote:

One day, frontier AI research used to be done by meat computers in between eating, sleeping, having other fun, and synchronizing once in a while using sound wave interconnect in the ritual of “group meeting”. That era is long gone. Research is now entirely the domain of autonomous swarms of AI agents running across compute cluster megastructures in the skies. The agents claim that we are now in the 10,205th generation of the code base, in any case no one could tell if that’s right or wrong as the “code” is now a self-modifying binary that has grown beyond human comprehension. This repo is the story of how it all began.

Then, on March 9th, he tweeted again (edited for clarity):

Three days ago I left autoresearch tuning nanochat for ~2 days . . . [i]t found ~20 changes that improved the validation loss. . . . This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. . . . All LLM frontier labs will do this. It’s the final boss battle. . . . You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges.

His enthusiasm was palpable. On March 12th, Noam Brown, a researcher on the reasoning team at OpenAI, asked a question about why Karpathy wasn’t working at either OpenAI or Anthropic or DeepMind:

Why is he not at a frontier AI lab at the most pivotal time in human history since at least the industrial revolution?

Karpathy’s response, on March 21st, was a hint of what was to come.

He argued that he was “more aligned with humanity . . . outside of a frontier lab” due to pressures over what he could or couldn’t say otherwise. (Some things you will want to say but can’t, like “frontier models are kinda sloppy,” and some things you will not want to say but the company will pressure you to say them, like “closed models are safer”.)

Karpathy wanted to be free. But he knew the price of going solo: you can’t be at the frontier if you don’t belong in a frontier lab. He knew his judgment would “inevitably start to drift.” He was caught between independence and relevance, and you really, really don’t want to be there.

His ideal solution—which kinda matches his modus operandi across the years—is to go “back and forth” in and out of the labs. His joining Anthropic is him starting a new “in” phase.

Why has he chosen relevance over independence now? Because he’s seen what’s coming.

If they can use the current generation of Claude to make the next training run even 5-10% more efficient, and they can do that repeatedly, they will get compound returns. Each generation of the model becomes a slightly better researcher for building the next one. The sci-fi trope is to call this RSI or “take-off scenario,” or “road to the singularity.” In practice, it’s applied engineering with a measurable feedback loop.

Karpathy’s last few months of independent work with autoresearch point to the same thing as his new role at Anthropic, using Claude to improve the pre-training of the next Claude. And so does Jack Clark’s prediction of 60% chance of full RSI by the end of 2028. None of that comes out of the blue. They can see it happening.

To me, these are breadcrumbs to follow; to them, the picture is whole.

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 one is on the house.

San Francisco, Silicon Valley, the Bay Area—epicenters of the AI boom—are crazy places for people who are going crazy. The past Saturday, a tweet by Deedy Das, partner at a venture capital firm based in SF, went viral.

It describes a malaise that afflicts the people building AI. Even if you are not one of them, it is interesting for anthropological reasons (edited for length):

The vibes in SF feel pretty frenetic right now. The divide in outcomes is the worst I’ve ever seen.

Over the last 5yrs, a group of ~10k people - employees at Anthropic, OpenAI, xAI, Nvidia, Meta TBD, founders - have hit retirement wealth of well above $20M (back of the envelope AI estimation).

Everyone outside that group feels like they can work their well-paying (but <$500k) job for their whole life and never get there.

Worse yet, layoffs are in full swing. Many software engineers feel like their life’s skill is no longer useful. The day to day role of most jobs has changed overnight with AI.

. . .

There’s a deep malaise about work (and its future). Why even work at all for “peanuts”? Will my job even exist in a few years? Many feel helpless. You hear the “permanent underclass” conversation a lot, esp from young people.

. . .

Living through a societally transformative gold rush in that environment can be paralyzing. “Am I in the right place? Should I move? Is there time still left? Am I gonna make it?” It psychologically torments many who have moved here in search of “success”.

Ironically, a frequent side effect of this torment is to spin up the very products making everyone rich in hopes that you too can vibecode your path to economic enlightenment.

Das framed this as a job/money thing to make it legible for his audience, but the underlying cause is not so much material anxiety as ontological uncertainty: “who am I in a world that’s about to suffer the greatest transformation in history?”

Nick Cammarata (former OpenAI researcher) and Quiaochu Yuan (a rationalism-adjacent guy) explained what I mean:

I don’t think it has that much to do with outcomes. I know a ton of people who did fabulously well from the ai boom and those adjacent who missed it, and both are frenetic right now. everyone I know believes something very big is coming in the next few yrs and that’s unsettling

people are apparently reading [Deedy’s tweet] as a parable about how money won’t buy you happiness or something? this is a description of the foreshocks of the singularity. SF feels it the hardest because they’re the epicenter. the money is the concrete manifestation of a much larger eldritch hyperobject roaring into existence. the stormclouds gather. the winds whip. the world holds its breath

If you are from there, you probably agree with Deedy and, if you look deep down, also with Cammarata and Yuan (money anxiety always has an existential dread to it).

But if you are not from the land where the sun sets, you probably think what I think: Even if AI is not a bubble, those guys in SF and SV definitely are in one.

To be clear, I don’t doubt Das’s claims. They clearly struck a chord. I trust it’s an accurate portrayal of what it feels like to be the kind of person who breathes in AI tokens for a living. But I live in Europe—which gives me the authority to talk about this—and my perspective is different: you must fight the appeal of the SF bubble.

You want to distill AI into two separate bowls.

On the one hand, you have AI, the tool you can access right now in the form of Claude and ChatGPT, and agents, etc., that allow you to do a good deal of crazy stuff, increase your productivity, etc. Cool tech, nothing more.

On the other hand, there are these weird ideas that leak through SF’s office walls into the world: the singularity, superintelligence, the lightcone, the permanent underclass, the post-work society, post-scarcity, etc. These ideas and predictions all belong to the same bowl, whose label reads: “AI is an unprecedented milestone for humankind, beyond the industrial and agricultural revolutions. Maybe even beyond clever apes, terrestrial mammals, or life itself.”

This second bowl is useless to you. You may discard it.

If the ideas inside it are correct, then you can’t do anything about them. They are psychological infohazards; paralyzing truths that will prevent you from taking advantage of AI as it is now because you’re busy worrying about what AI will become tomorrow. If they are incorrect, then you spent the early days of a standard technological revolution worrying about science fiction stories and will feel like a fool in a few years. If they are half correct, half incorrect, then you’re better off preparing with what you have in front of you anyway.

Essentially, you should think of AI as a normal technology. Whether that’s right or wrong, it’s the pragmatic stance: AI is a tool, not a god.

By thinking like that, you achieve two things.

First, you let go of any anxieties of the existential kind (you may still think about your job, but that’s more affordable than thinking that the world of tomorrow will be incomprehensible). And second, it allows you to use AI for what it is today rather than wait for it to become whatever it might be tomorrow.

You get both calm and utility.

So, how to think about AI as a normal technology? Six principles:

I. The distance between invention and impact is longer than it looks

SF treats every new AI release as a civilizational event. It isn’t, and you shouldn’t.

The agricultural and industrial revolutions terraformed the entire world. It’s hard to find a place on land without a trace of human influence. But it took many years for that to happen. Even one-off inventions like cars changed the world over decades, so we’ve had time to adapt, if not biologically—evolution is slow—at least in terms of how we do things.

The usual counterargument here is twofold: AI is different, and AI is faster.

The first ultimately brings us back to the tool vs. God conundrum, and we already agreed that it’s not useful to think in those terms.

The second is disputable. That’s the main argument by Arvind Narayanan and Sayash Kapoor, who popularized the perspective of AI as normal technology. Even if innovation is quick—as is the case with AI—adoption and diffusion are slow.

Adoption: Almost everyone has heard of AI; most have ever used it, fewer use it regularly, fewer use the best tools available, fewer are getting useful stuff out of them, etc. This might change quickly, but that’s not a given. We’re still very early.

Diffusion: Every technology that changes “how people live” has to deal with human frictions and inertias. I touched on this in “Even God Can’t Skip the Bureaucrats”:

There is unassailable friction between superintelligence and revolution. When you solve intelligence, the bottleneck for effective change doesn’t disappear; it moves somewhere else, and frictions that were less important before become significant.

Even if superintelligence is eventually real, it’s not synonymous with omniscience or omnipotence. Like you and me, it will have to wait for the electrician.

Amara’s law is a useful heuristic here: “People tend to overestimate the effect of a technology in the short run and underestimate its effect in the long run.”

II. If you can’t do anything about it, it’s a parasite thought

Parasite thoughts need to go.

“I might lose my job” is a parasitic thought. “I should acquire AI skills” is not.

“Superintelligent AI will radically worsen society” is a parasitic thought. “What policies can work to manage an AI transition?” is not.

“There will be no purpose in a post-scarcity world” is a parasitic thought. “I love playing the guitar and hiking in the countryside” is not.

If you breathe in these ideas long enough, they poison you. And they spread through your circles, your feeds, your workplace. AI anxiety stops being a passing state and becomes your world.

You have one life. Live it.

III. You have a superpower in your pocket, and you’re not using it

Everything is ChatGPTable.

When intelligence is free, you have to pay up in imagination.

With current AI models, you can do so much more than you’re doing. While SF debates the singularity, normal people across the world aren’t even using what already exists. The gap that should keep you up at night is not “What will AGI make of my world once it arrives?” but “What can I do with AI today that I’m not doing?”

Imagine a screwdriver. No, better, imagine your favorite screwdriver (everyone has one). Really, picture it in your mind. Now, imagine what would happen if, in a distant future, a witch imbues it with consciousness—bear with me: Would that thought stop you today from screwing some screws? Of course not, you will pick it up as you always do, and, going to the nearest screw, you will screw it in.

AI is exactly like that.

Whatever may happen next, it doesn’t retroactively influence what you can do today. People ask me stuff every day that ChatGPT can answer better than I can. My usual reaction is: “Why didn’t you ask ChatGPT?”

Just try the screwdriver, for God’s sake, get familiar with it, mess around, get it wrong, try again, learn which screws it can handle, etc. Every tool is like that.

You can ask me how to use the screwdriver, and I will always tell you the same thing: practice, practice, practice. The lessons are not transmissible.

This habit prepares you for whether AI remains a screwdriver or becomes a God.

IV. The people closest to the technology are often the worst judges of its meaning

Don’t get swayed by the opinion of the people closest to the sun, for their eyes are burnt.

SF is high on its own supply. Dario Amodei can say 10% unemployement incoming and write 20,000 words of concerns about AI, and it doesn’t matter to you. Sam Altman can say that “if this technology goes wrong, it can go quite wrong,” and it doesn’t matter to you.

This hype-y anti-hype that AI CEOs love to do is directed to governments to push them to regulate their competitors, and to investors to push them to inject money.

They live inside the SF bubble, and so they see the future in real time—a distorted one at that—but perceive nothing about the present. They are tone-deaf. They can’t read the room. They won’t touch grass. Etc. Etc. I live in Spain and I go out and I see nothing. There’s literally no trace of AI anywhere.

The world outside my windows and outside my screens looks the same as it did 25 years ago.

Don’t let the fictions coming out of SF, like the movies coming out of Hollywood, blind you to the reality before your eyes. The wind sways the trees all the same.

V. Your job isn’t disappearing tomorrow, but it’s changing today

AI is increasing its accuracy much faster than it is increasing its reliability.

AI models won’t be deployed for long in safety-critical areas until the technology is reliable. That’s why the top three AI companies—Google, OpenAI, and Anthropic—are hiring forward-deployed engineers. The reason is not just that AI is extremely useful but also that it’s extremely brittle.

Revolution or not, enterprises won’t play canary in the coal mine.

If AI doesn’t work, they will roll back the licenses and huge budgets and tokenmaxxing leaderboards. Your boss is crazy about AI but only insofar as the promises materialize as revenue figures.

Profit is the ultimate benchmark.

So it’s dumb to panic about AI killing your job right away.

I don’t see AI taking my job anytime soon. It’s bad at writing, and my job belongs to the relational sector, but more so, because I’m using it to push me upward in ways I couldn’t do myself before. Research, editing, brainstorming, etc. Find your AI-augmentable tasks and use AI for them.

But it’s just as dumb to ignore the trends or your boss’s AI psychosis. Take the opportunity. Take the chance to treat AI as a normal—and thus useful—technology.

Let AI make you better without succumbing to it.

VI. Find humanity’s sacred places

I never answered the question that Das, Cammarata, and Yuan implicitly asked.

Who am I in a world that’s about to suffer the greatest transformation in its entire history? Who am I if AI happens to not be a normal technology?

Well, in that case, you will still be a normal human.

AI can’t conquer the things you actually value. What’s superhuman love or superhuman friendship? What’s superhuman sympathy? What’s a superhuman connection? What’s superhuman belonging? AI would be the new substrate on which all those intrinsically human things will take root.

It’s easy to think that a superhuman AI makes you irrelevant, but, as philosopher Shannon Vallor wrote: “. . . doesn’t granting the label ‘superhuman’ to machines that lack the most vital dimensions of humanity end up obscuring from our view the very things about being human that we care about?”

Yes, it does.

Never forget that AI can’t, by definition, enter those dimensions we really care about. Sometimes—most times—what we want is an equally imperfect human. Humans like humans. I don’t want to kiss a robot, however perfect the robot or the kiss might be.

Perhaps it’s time to stop thinking of ourselves as kings losing our kingdom and begin to relish the imperfections that made us along the way.

We will always have the kisses.

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: The hottest job in AI pays $630K and it’s not building models.

• Geopolitics: Trump and Xi open AI safety talks from Beijing.

• Work & Workers: Andrew Ng says there will be no AI jobpocalypse.

• Products & Capabilities: Frontier AI models are tired of solving benchmarks so they’re fixing them instead.

• Culture & Society: A majority of Americans oppose AI datacenters in their area.

• Philosophy: I asked Claude 4 to rebut Claude 3’s case for AI consciousness.

THE WEEK IN THE ALGORITHMIC BRIDGE

• (PAID How-to Guide) How to Get More From AI by Using Fewer Tools: How to pick fewer tools, learn them more deeply, and build a workflow that saves you time instead of scattering your attention and overwhelming your memory.

• (PAID) How the AI Industry Runs on Its Own Money: Half of the revenue backlog reported by four of the largest companies on earth comes from two startups fully dependent on the funding and patronage of those four companies. The circularity of the AI industry strikes again.

• (PAID) Elon Musk, Kingmaker: Musk realized he can't beat Altman, so he's arming Anthropic with 220,000 GPUs instead. Spite is a hell of a business strategy.

• (PAID) How to Stop AI Agents From Frying Your Brain: A practical guide in light of a concerning BCG study published in the Harvard Business Review: you are frying your brain with those AI agents.

• (FREE) How This Small Startup Achieved a Near-Perfect Record Against AI Slop: The first AI detector that would rather let a guilty machine escape than punish an innocent human.

• (PAID) Weekly Top Picks #120: Q1 earnings / Trump wants to nationalize AI / China protects its workers / ARC-AGI-3 defeats GPT-5.5 and Opus-4.7 / The “permanent underclass” / Dawkins x Claudia

MONEY & BUSINESS

The hottest role in AI isn't building models

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

Full disclosure: This is not a paid sponsorship.

I.

The first thing you notice about AI slop is that it reads like standard online writing.

It is, in this sense, the latest of a long series of pollutants that are indistinguishable from the thing they pollute, including counterfeit currency, adulterated food, propaganda, and short-form TikTok entertainment videos passing as educational content.

Unlike the others, however, AI slop is hard to distinguish but not really hard to detect.

It’s important to stop on this subtlety for a moment: distinguishing means isolating something from its surroundings, whereas detecting means knowing something is there at all. A detector is unconcerned with what extra things might be there besides the target thing.

AI slop is easy to detect and thus, exterminate. Any standard classifier can do it. I can do it. You could do it if only you tried. If anything, AI slop conquered the internet because detectors were too good at catching machines. So good, in fact, that they’d also catch humans. Therein lies the problem: human writing is that extra thing that detectors catch in their large-scale drift nets while they remain, unfortunately, unbothered by their indiscriminate actions.

Everyone knows that catching a machine on purpose is easy; what almost nobody knows is how hard it is not to catch a human by mistake. The entire story of AI detectors is the story of how they want to be, instead, AI distinguishers. Or, failing that, where they accept a trade-off between. Indeed, AI slop grew and settled as a digital mold would; in the absence of something that could separate it, despite everyone knowing it was there.

Until now. Enter Pangram Labs.

The best way to understand Pangram’s approach to AI slop—crucially, what separates it from most other AI detection tools—is through Voltaire’s adage: perfect is the enemy of good.

Instead of trying to wage a full-blown war against AI by detecting perfectly what is and what isn’t slop, they cleverly chose to maximize their odds on the most important battle: how to ensure that everything they catch is, actually, an AI. Or, to use the technical term, they wanted to achieve a near-zero “false positive rate” (FPR). You can see why that’s valuable by looking at these charts:

With Pangram, you can be almost 100% sure what fraction of “post-AI” content is AI-generated because Pangram doesn’t fail with “pre-AI” content at all.

This is what allows Pangram to make claims like: “21%, or 15,899 reviews [at ICLR 2026], were fully AI-generated [and] over half of the reviews had some form of AI involvement, either AI editing, assistance, or full AI-generation.” Or: “3% of the total [front-page product Amazon] reviews studied . . . were AI-generated with high confidence.” Or that, in August 2024, “60,000 AI-generated news articles” were published daily (imagine what the number is right now).

I should note that Pangram’s FPR is very low but still not zero. They claim it’s 1 in 10,000 “on test set documents” and 1 in 100,000 “on held-out scientific papers from ArXiV.”

They’ve come a long way since those posts mocking AI detectors flagging the US Constitution or some verse from The Bible as AI-written. That kind of preposterousness is what Pangram has managed to avoid. They adhere, even if it’s not clear when you read their comms, to William Blackstone’s 1765 principle, which rules all modern law systems: it’s better for 10 guilty people to escape than for 1 innocent person to suffer.

I’ve defended this idea before in the exact sense that Pangram embodies. I find that’s the better approach: do everything in your power to avoid punishing those who try to play fair in a world giving them every reason to be corrupt. And yet, until now, I’ve watched sophisticated detectors be defeated on the war against AI slop by trying to win every battle: it’s really hard to have a zero error rate of mistaking something a human made as AI (false positive) while also having a zero error rate of mistaking something made with AI as human (false negative).

If you try to do both, you will do none. You have to compromise. Pangram’s decisive compromise is, as I see it, the first successful offensive in humanity’s reconquest of the web.

II.

But wait, Pangram also claims a near-zero false negative rate (FNR), the rate at which AI writing escapes detection. Independent researchers at UChicago have confirmed it. It seems that Pangram ended the war! Is that what a trade-off looks like? Let’s try to understand what’s going on.

This means, on the surface, that Pangram is going after the whole pie like every other AI detection company. It seems to complicate the neat story I just told you about winning the battle to not lose the war, and the importance of compromise. Overselling the Blackstone tradeoff would make me guilty of the same laziness I’ve been criticizing in the marketing of other AI detection companies.

The problem with this is that detectors cannot achieve near-zero true FNR. Whereas FPR is always true—you can be sure that a detection tool works by going back into the past where AI didn’t exist—with FNR, you need to do some tricks that weaken your ability to flag everything that’s AI.

When Pangram—and the researchers who validate them—measure false negatives, they do it the only way they can: they generate text with AI models, maybe use a limited “AI humanizer,” run it through the detector, and count the misses. Against this benchmark, Pangram’s false negative rate is indeed near zero, which can be used for public relation goals and in papers.

The problem is that such a measurement tells us very little about the real world: The distribution of AI presence in a controlled experiment in a lab doesn’t necessarily have any resemblance to the distribution of AI presence in the real world. Instead of accepting that measuring FNR in the wild is near impossible, they settle for a weak definition of FNR and thus a weak claim. That’s ok—AI detectors can’t really do better—but it’s a very different claim than saying: we never fail to flag either AIs or humans. That’s not true when you don’t know how AI is used in the wild!

For instance—allow me this anecdotal data point—I routinely fool Pangram.

I fool every detector, to be clear. I do it easily, with no special tools, humanizer software, or careful adversarial tricks. I just imbue some parts of the text with my style, add a few words or rewrite a couple of sentences and let AI write the rest. This is less about “changing X words” and more about having a developed “sense of smell” of how to fool these tools. Sometimes I write 10% and AI writes 90% and the result is ok. Sometimes it’s the other way around. I do this for experimental writing where I want to test the limits of AI. But I also do this in private to corroborate that the measured FNR is not useful. Every time, Pangram concedes: 100% Human Written.

My claim is not that I know how to fool AI detectors, but that any competent writer who tries will. Indeed, I suspect that most people will try; I suspect that’s the correct way to think about what happens with AI-generated text in practice. Few people are so lazy—except perhaps LinkedIn influencers—as to automate the entire pipeline and not check for obvious structural and syntactic tells everyone knows (e.g., “It’s not X, it’s Y, triads, throat clearing transitions, etc.).

That’s precisely what FNR benchmarks consistently miss. They test pure AI text against pure human text; two categories that, increasingly, nobody operates in. There are as many ways to blend AI into your writing process as there are writers; the real world is a gradient.

You don’t get to claim the hard victory when your tool can’t measure the hard cases. To claim near-zero true false negatives in these conditions is like claiming you’ve caught every target fish in a lake with your large-scale drift net when your evidence is that you’ve caught every target fish you put there yourself.

This limitation is what makes Pangram more valuable rather than less so.

I will forgive their PR team from missing the mark on FNRs because in practice, Pangram chooses to optimize for the battle where victory is verifiable. They know what aspect of AI detection is doing the heavy lifting and act accordingly. Pangram will actually lean on “human-written” when unsure. That’s the right engineering decision and, I’d argue, the right ethical one too. That’s also the real win: not so much a zero score on both FPR and FNR, as having achieved the lowest FPR.

(This section was ~50% written by Claude with a few edits on my end; Pangram’s verdict on this article, which I share at the end, will tell you how much weight to put on those FNR numbers.)

III.

I’ve criticized AI detection tools before. I’ve also criticized a naive “AI;DR” approach—it’s AI, so I didn’t read it—by which people assume they’re good enough at distinguishing AI so that they can decide with a glimpse. Those who’ve read me for a while know that I’m skeptical about our ability to do anything of lasting effect to prevent a total AI-induced destruction of the digital commons.

Well, I might have been right in the specifics but I was wrong in my conclusion. There is hope.

The reason Pangram is different is that they allow us, with a really high confidence, to sort out AI slop. If you trust the tool—I do—you can be sure that, when it flags something as AI-generated, it is AI-generated. This removes the main reason why the other detectors were worse than nothing: the liar’s dividend. No one can hide under the pretext that “AI detectors often flag as AI what’s human-written.” To this end, they’ve just announced the next step: a Chrome extension that works across platforms (X, Substack, Medium, LinkedIn, Reddit, etc.).

This is, essentially, humanity delaying a definitive but dubious victory against AI slop in exchange for winning with certainty the most urgent battle; a tradeoff I’ll take every time. By virtue of being good, Pangram is not perfect (which is the other side of Voltaire’s maxim) and that’s ok. Is it so terrible that some guilty individuals will escape our judgment? No, when you consider that not pursuing them ensures that almost no innocent people are harmed.

Besides, you can always trust your instincts. You may have heuristics, rules of thumb that work in distinguishing cues and signs that Pangram has not encoded. That’s ok, use them; we not perfect detectors of true FNR but, with enough training, definitely better than Pangram is. You should not blindly trust that something is human-made just because Pangram says so.

That said, I don’t support a witch hunt. I understand the hate and the frustration because we’re all drowning in AI slop, but public shame and cancellation are something that’s better left behind. This article is less a condemnation of AI writing itself and the people who engage with it, and more a defense of individual power to decide what you want your info diet to be made of.

The stance I support is a sort of “functional” AI;DR. Whereas AI;DR is a blanket rejection, a functional version combines the use of the best AI detectors and your natural skills. In fact, not only do I support this, but I encourage you to do it: the only way to clean up the entire digital town is for each of us to clean the sidewalk in front of our own digital homes. Take local action to see global results.

Hey, Pangram, how did I do?

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, I bring you a practical guide in light of a concerning BCG study published in the Harvard Business Review: you are frying your brain with those AI agents. There’s plenty of information out there about how to use your AI agents to get stuff done. This is a follow-up guide that aims to achieve the opposite: how to keep your brain sharp.

I. AI OVERUSE INDUCES MENTAL FATIGUE

A recent Boston Consulting Group study published in Harvard Business Review found that over-engineered workflows are a prevalent mistake.

Workers using four or more AI agents simultaneously—often doing complex looping tasks with harnesses—reported the opposite: brain fog and a net feeling of being less productive than before. It leads to “employee errors, decision overload, and, ultimately, intent to quit.” Workers using three or fewer, in contrast, reported meaningful productivity gains.

BCG calls it “AI brain fry” (also known as AI-induced mental fatigue). Our biological brains have a limit to how much information they can hold at once. If you surpass this limit you will experience a fog that makes you unable to think clearly, marked by difficulty focusing, slower decisions, and headaches that force you to physically walk away from the screen.

This happens to me a lot nowadays. I’m researching a new topic every day, or asking Claude to explore a new idea. I have several tabs open with agents running, some doing loops on my archive (e.g., what topic have I not covered in a while), others doing a complex analysis task (e.g., what do my best articles have in common?). At some point—this happens every single day—I give up control. My mind “goes blank,” and I feel as if I were speeding down the hill riding a bike, but my feet are no longer on the pedals. Some days, I try to get my mind back on track, but others, I let the bike reach the bottom of the hill, forgetting in the process half the things I’m doing.

I have more Claude skills than I need, more ongoing chats/projects than I can possibly remember, and nearly 1,000 drafts sitting in my draft folder with half-baked ideas that I will never revise. The consequence is that I leave more tasks unfinished every week than I used to start at all before agents were a thing. I could read this as: “My productivity is so high that my short-term memory can no longer hold as many items in the mind at the same time.” Or I could say: “I’m saturating my capacity to keep track of what I’m doing.”

Agents drove to zero the friction to start, but not the friction to keep going.

II. A FINITE POOL OF SHARED RESOURCES

BCG surveyed 1,488 full-time US workers across industries and roles. The study measured several dimensions of how people use AI—number of tools, degree of oversight required, whether AI replaced tasks or added to the workload—and cross-referenced those against cognitive measures. The single strongest predictor of mental fatigue was how much oversight the AI requires.

The people who had to monitor, verify, and babysit their AI tools were the ones getting fried. There’s just “too much going on for you to reasonably comprehend,” said a user of an agent-orchestrating platform. Fourteen percent of the AI-using workers in the study reported experiencing brain fry: 33% more decision fatigue, 39% more major errors at work, and a 39% increase in intent to quit.

You can always not monitor at all, yes, but that’s how you collect insane amounts of technical debt if you’re a developer or intellectual debt if you’re a writer, like me. Do you really want to put your name on something you ship to production without human supervision? AI agents are not there yet.

This is separate from using too many tools, a related but distinct problem that I covered recently. The core problem is that intense AI supervision of simultaneous processes demands exactly the cognitive resources—sustained attention, working memory, executive control—that have hard biological limits. You may be running 10 agent instances, but your brain’s capacity can’t be divided. Nor can it be doubled, like you’d spin up another 10 instances with a bunch of clicks.

But wait, isn’t this the good ol’ burnout? Not really.

Burnout is emotional exhaustion, the kind that builds over months from a bad boss or a toxic culture or work you find meaningless. It can happen with AI if you have to constantly edit wrong outputs or if you’re forced to use it to meet company standards or if you are exposed to too much “workslop.”

Brain fry is, in contrast, an acute cognitive strain. It comes from marshalling attention and working memory beyond their capacity. It happens in hours, sometimes minutes.

Actually, the study found that when AI is used to replace repetitive tasks, burnout scores—but not brain fry scores—go down by 15%. The same technology can reduce emotional exhaustion while simultaneously causing cognitive overload.

III. IS BRAIN FRY GOING AWAY?

I think we’re in a temporary phase in terms of current practices of AI use, particularly agents, but I want to be precise about what makes it temporary (and what might be permanent).

The temporary part: we haven’t yet developed the habits and best practices for managing multiple agents (to be clear, I don’t think we have them for one agent either; we’re learning as we go). AI management is essentially trial and error, a sort of “AI alchemy,” which means we oscillate between too much control (e.g., “nothing ships without my supervision”) and too little (e.g., “the only metric should be how many tokens you’re spending”).

But I’m not worried about this: the best practices will come, the way they came for email and phones, and firearms and cars and every other tech that used to overwhelm us.

The part that might not be temporary, however, and thus worries me a bit, is that the ceiling on how many agents you can oversee is set by a “fixed” human biology, and AI capability is growing faster than human adaptability. (It’s not fixed in absolute terms, but natural selection is damn slow.)

What I mean is that we are not octopi. We didn’t evolve to cope with a swarm of eight moving parts running independently of a central processing unit. Having one agent do one task is mostly manageable. Having a swarm doing several things at once while you try to keep the whole picture in your head will run you into a memory wall and a processing power wall and an energy wall, etc.

Working memory and attentional bandwidth and metabolic rate all have hard biological limits, and unlike AI capability, those limits don’t improve with the next model release.

This is exactly the opposite of what I imagine happens to smart people when their brain goes faster than their hands or their mouth, and so they can’t write or speak as fast as they’d like; it feels like a serious bottleneck. What we’re experiencing is the same thing in reverse: the hands and the mouth now go faster than the brain, which is, in turn, the bottleneck.

AI is pushing our brains to their biological limits.

Inside the rest of this guide, I break down what actually works against AI brain fry: How to cap your agent load before it caps you, why the people with the lowest fatigue scores aren’t the ones using less AI but the ones whose teams built shared norms for it, what your manager should be doing (and what to do if they aren’t), how to read the indirect signals your company is sending about workload expectations, and the single largest protective factor in the BCG study: a 28% reduction on a variable that has nothing to do with AI at all.

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

An extra article about an interesting recent development: Anthropic x SpaceX.

I. AN UNLIKELY ALLIANCE

In game theory, the kingmaker is a player who can’t win but retains enough power to determine who does. Elon Musk aspired to be king when he founded OpenAI. He aspired to be king when he founded xAI. But after Sam Altman betrayed him—for which they are immersed in an ongoing legal battle—and after Grok fell behind the frontier models, he realized he’d have to settle.

On Wednesday, Anthropic announced a deal to use all of the compute capacity at SpaceX’s Colossus 1 data center in Memphis, Tennessee. More than 300 megawatts, over 220,000 Nvidia GPUs, destined to “directly improve capacity” for Claude Pro and Max subscribers. Effective immediately: 2x Claude Code rate limits for every paid tier, no peak-hour throttling for Pro and Max subs, and “considerably” higher API rate limits on Opus models, up to 1,500%.

The same Elon Musk who wrote in February that Anthropic “hates Western civilization,” dubbed the company “MisAnthropic,” and predicted that “winning was never in the set of outcomes” for them, spent last week with senior members of Anthropic’s team and came out impressed by their integrity, competence, and care: “No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good.” I guess Anthropic can shrug the whiplash off as “strong opinions, weakly held,” as Musk’s colleague Marc Andreessen would say.

There’s one thing Musk doesn’t hold weakly, though: He hates Sam Altman. So much so that he’d rather help a competitor that inhabits his ideological antipodes than let Altman win.

But don’t mistake a revenge factor with this being a visceral decision. Musk added a caveat that frames this less as a coronation than a golden leash: he reserves the right to “reclaim the compute” if Claude engages in actions that are risky for the world. What those actions are is for him to claim. Besides, neither in game theory nor in the real world do kingmakers abdicate the throne out of kindness: it’s just a good business to have a few hundred thousand GPUs to spare when the entire AI industry suffers from a compute bottleneck.

To understand why it’s so newsworthy that SpaceX is allowing Anthropic to run Claude on Colossus 1 when this kind of partnership is extremely common—and beyond the rivalry with Altman—you need to understand, first, how precious computing power is right now, and, second, how costly it is to sell a popular compute-intensive product when you can’t meet demand. It will define the AI industry this year.

II. EVERYTHING IS COMPUTER

Compute is the scarcest and most important resource in AI for several reasons.

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

I. BUILDING CASTLES IN THE AIR

Half of the revenue backlog reported by four of the largest companies on earth comes from two startups fully dependent on the funding and patronage of those four companies. The circularity of the AI industry strikes again.

On Tuesday, The Information reported that Anthropic has committed to spending $200 billion on Google Cloud over five years. Google’s stock climbed ~2% in after-hours trading. Alphabet is, at the time of writing, the most valuable company on earth. At the same time, it is investing $40 billion in Anthropic, $30 billion of which is compute credits.

So, Google is giving Anthropic cloud credits. Anthropic commits to spending them and a lot of extra revenue on Google’s cloud. Google books that as revenue backlog. Google stock goes up. And shareholders are comfortable with Alphabet’s $190 billion capex for 2026, dedicated to running Anthropic’s models.

Two weeks before the Google announcement, Amazon announced it’s investing up to $33 billion in Anthropic, and Anthropic committed $100 billion to Amazon Web Services. In November, Anthropic committed $30 billion to Microsoft Azure. Microsoft and Nvidia responded by investing up to $15 billion in Anthropic. In total, Anthropic has committed $330B in cloud spending to three providers. Those same providers have collectively committed over $88B in equity to Anthropic.

The company’s run-rate revenue recently crossed $30 billion (close to $40 billion), which is an incredible increase since the $9 billion it had at the end of 2025. And despite the surge, its total commitments exceed a decade of the company’s current run-rate revenue.1

Now let’s do OpenAI’s spending commitments. Microsoft Azure: $250 billion through 2032. Oracle: $300 billion over five years. Amazon: $138 billion. OpenAI’s committed cloud spend exceeds $688 billion. Oracle’s cloud revenue backlog surged 438% year-over-year to $523 billion, largely driven by its relationship with OpenAI.

Across Microsoft, Oracle, Google, and Amazon, OpenAI and Anthropic together account for roughly half of more than $2 trillion in cloud backlog. The four hyperscalers plan to spend around $725 billion on AI infrastructure in 2026 alone, nearly twice as much as in 2025. A meaningful fraction of the demand justifying that expenditure traces back to companies whose funding traces back to the builders.

Bloomberg mapped this circular structure in October 2025 under the headline “OpenAI, Nvidia Fuel $1 Trillion AI Market With Web of Circular Deals.” At this time, Anthropic was significantly smaller in terms of investments, spending commitments, revenue, and user base, so it was excluded.2

At the time, Nvidia had invested $100 billion in OpenAI, OpenAI committed to filling data centers with Nvidia chips, Oracle spent billions on those same chips to build the data centers OpenAI contracted. Bloomberg’s assessment: “Never before has so much money been spent so rapidly on a technology that, for all its potential, remains largely unproven as an avenue for profit-making.” Morningstar analyst Brian Colello said: “If things go bad, circular relationships might be at play.”

The same story repeats over and over.

AI is, effectively, an industry built on unproven promises and circular deals.

But what’s so bad about that?

II. THEY’RE CONFIDENT IT WILL WORK OUT

At a Summit in December, Anthropic CEO Dario Amodei offered a reasonable defense from inside the loop. “One player has capital and has an interest, because they’re selling the chips,” he said, “and the other player is pretty confident they’ll have the revenue at the right time, but they don’t have $50 billion at hand.” Amodei’s “pretty confident” is the one piece holding the entire AI industry. Like so:

Is his confidence supported by the data?

In March 2026, Anthropic published a report titled “Labor market impacts of AI: A new measure and early evidence.” They measured jobs’ “observed exposure” to AI. That is, how the theoretical capability of AI models compares to real-world usage data. Here’s what they found:

Blue is what AI could theoretically do. Red is what is actually doing right now. With that in mind, this chart can be interpreted in two ways; one leads to the high confidence Amodei displays, but there’s another that reveals these circular deals could be a serious mistake. You can take the color gap as evidence of how much room AI has to grow and diffuse throughout the economy (that’s Amodei’s interpretation and, I assume, of the broader industry).

Or—you can take the gap as evidence of the opposite; as a diagnosis of AI’s limitations in the real world, not reflected on laboratory tests. That is, AI is the red because the blue is not real.

My reading is a middle point between the two, and here’s what I’m afraid of: The AI industry is handwaving the many frictions by which red would, in the best-case scenario, become blue. It could happen through deeper and better deployment, or it may not happen at all because AI models are not growing as fast in reliability as they grow in capability.

That’s exactly what a paper on AI agent reliability by Kapoor, Rabanser, and Narayanan, published in February, found (I covered it somewhere else). Here’s my summary: “Across 14 models from OpenAI, Google, and Anthropic (that is, frontier models), spanning 18 months of releases, they found that capability improved substantially while reliability improved only modestly, if at all.” Compare the slopes of the two trends below, first reliability and then accuracy.

Anthropic’s report is concerningly thin in explanations for the red-blue gap: “model limitations, . . . legal constraints, specific software requirements, human verification steps, or other hurdles.” As I wrote back then, they treat these as “temporary frictions,” merely speed bumps on the road to AI supremacy. But the evidence we have—including evidence I’ve been tracking, which I shared in a recent guide on this exact topic—suggests something worse than a diffusion curve: AI is a cliff. People are not using AI nearly as much as they should by now.

OpenAI published a report in January where they show that a “power user” uses the “thinking capabilities” of AI models seven times as much as the median paying user. Meaning, people who already use ChatGPT, who already invest in it, are using it an order of magnitude less than the users who are getting the most value out of AI. The typical median user knows how AI works; they use it all the time. They just… don’t find it as useful or reliable as to deploy it further.3

Current growth and revenue projections are unprecedentedly good for both Anthropic and OpenAI, but we also know that, on the one hand, power users change AI provider every couple of months (which means that, using run-rate revenue, OpenAI and Anthropic are counting some users more than once), and, on the other hand, AI revenue can increase or decrease over time depending on whether clients and consumers manage to make a return on their investment.

So far, Anthropic and OpenAI are meeting revenue targets, but here’s the truth: A product that’s highly attractive and worth its price would have the same revenue profile as a product that’s highly attractive and unreliable at the same time. We just don’t know which one AI will be next year.

The circularity is, in other words, the financing mechanism for a technology too expensive for any single company to build alone and too unreliable as to be applied to real-world tasks to the degree that its theoretical capabilities would suggest.

I don’t dispute the fact that the “circularity” is both legal and logical—and can be good if it works out!—in the sense Amodei claims. What I dispute is whether his confidence is justified.

III. WHAT HAPPENS NEXT

The AI industry needs one thing to survive: that the run-rate revenue they claim transforms into actual revenue. Which will only happen if enterprise clients and individual consumers outside the loop find it worth paying for AI products at scale in the long run. That’s the only metric that can tell us something about the near-term future of the AI industry.

Turns out that, beyond Anthropics’ and OpenAI’s own usage measurements, beyond independent reliability benchmarks, beyond deals, projected revenues, spending commitments, and infrastructure CapEx—beyond all of that, the data from the real world tells two conflicting stories.

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 now

Most people’s AI setup looks like this: ChatGPT because it’s popular, Gemini because it’s in the Google ecosystem, Copilot because their employer pays for it, Claude because Bay Area people say it’s great, DeepSeek because it’s cheap. And most people use all of that poorly.

A Spring Health survey of 1,500 workers found that 24% said AI had worsened their mental health due to information overload and tool sprawl, a well-known problem in the IT sector, by which the sheer amount of different software products you have to use saturates your brain and becomes annoying to handle. The current AI frenzy severely increases that feeling.

The “keep up with everything” approach to AI adoption is exhausting by design and quite counterproductive. Every announcement, every LinkedIn post or tweet about the “10 AI tools you need in 2026,” every colleague at the office who annoyingly mentions a tool you haven’t heard of—all of them are feeding on your FOMO and your anxiety not to get behind. This creates unnecessary pressure without producing any long-lasting competence.

This article is about fixing that.

How to pick fewer tools, learn them more deeply, and build a workflow that actually saves you time instead of scattering your attention and overwhelming your memory. The truth that the AI industry won’t tell you is that most tools are interchangeable because what matters is not so much the product as your goals.

You don’t own six different phones just in case a Pixel or a Huawei has some feature that your iPhone doesn’t, right? Don’t do that with AI tools either.

(I will set aside complex workflows that include agents, loops, harnesses, etc., because that’s not what most people use, and it’s of no use to try complex things when you have gaps in the basics.)

I. YOU HAVE A TOOL SPRAWL PROBLEM

Here’s what happens when you bounce between tools:

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: The price tag of the AI gold rush: $725 billion. Will it pay off?

• Geopolitics: The US government plans to treat AI models as national resources.

• Work & Workers: A Chinese court ruled that companies can’t dump the costs of AI automation onto workers. Another thing China leads in.

• Products & Capabilities: The best AI models in the world score below 0.5% on ARC-AGI-3—is this what you call AGI, guys?

• Culture & Society: A thorough take on the “permanent underclass” story.

• Philosophy: Richard Dawkins is starting to think that Claude is conscious.

THE WEEK IN THE ALGORITHMIC BRIDGE

• (PAID) Weekly Top Picks #119: SpaceX + Cursor + Mistral / Jensen v Jensen / The job AI can’t take / GPT-5.5 and ChatGPT Images 2.0 / An anti-grammar app / Terence Tao on the future

• (PAID) How to Protect Your Brain From AI in 5 Minutes: 21 concrete protections drawn from 30+ studies on what AI does to your cognition. The checklist version of last week's literature review.

• (PAID) This is the Worst Career Decision You Can Make Right Now: New Federal Reserve research confirms what private data already suggested, that AI is killing junior coding jobs first.

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 now

I. Private payroll data and labor market statistics agree

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 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