Meta has just given us the first great case study in how not to do an AI transformation. Over the last few months, the company has taken a butcher’s knife to what was once regarded as one of the strongest engineering cultures in tech.

It cut 10% of the workforce, created a new Applied AI org that many employees did not choose to join, and forcefully moved thousands into AI-related roles. Some workers described themselves as “draftees.” Others called the work “soul-crushing.” The smartest people I know at Meta are either gone, recently laid off, or still there with one foot out the door.

Andrew Bosworth has now admitted the rollout was “atrocious,” and Mark Zuckerberg acknowledged the company made mistakes. Employees are pushing back on the nature of the work, the lack of choice, the surveillance concerns, the loss of autonomy, and the feeling that the company was asking for extraordinary commitment while communicating that everyone was increasingly replaceable. That is a hard circle to square.

The mistake was not that Meta moved aggressively. In AI, speed matters. The mistake was assuming that speed required making employees feel disposable, disoriented, and conscripted.

AI adoption is not just a resource-allocation problem, it is also a collective-action problem. Every company in the world is trying to move people toward AI. The question is whether employees feel like they are being invited into the future or dragged into it.

You cannot lay people off, reassign thousands more into work they did not sign up for, stretch managers across absurd spans of control, monitor employee activity for training data, and then solve the resulting morale problem with snacks, offsites, and a hackathon. That is corporate Febreze.

One of the stranger management mistakes of the AI era is assuming that it makes people less important. If anything, the opposite is true.

When technology is stable, process matters. When technology is changing, talent matters.

In tech, the ability to attract and retain the best people has always been a leading indicator of company performance. You could have predicted that OpenAI and Anthropic would matter because great people wanted to work there. Anthropic’s advantage, in the current moment, is not just technical. It is cultural: a clearer sense of mission, purpose, and trust. The entire AI industry is a case study in talent concentration.

Meta seems to have missed the lesson hiding in plain sight.

A mature company can often survive mediocre culture because the machine already knows what to do. An AI transition is different. There is no machine yet. The company needs its best people to invent one, but is instead giving them every reason to leave before it exists. It is asking them to think like founders while treating them like furniture. That is a difficult strategy.

The winning move is not to aggressively “AI-ify” the workforce, but rather to handle the human layer of AI adoption with care. You cannot automate trust, purpose, or belief. At least not yet.

Satya Nadella published an AI manifesto this weekend, arguing that “a frontier without an ecosystem is not stable.”

This is, among other things, an excellent sentence for the CEO of Microsoft to say. Microsoft is the ecosystem business. It owns the place where employees work, developers build, data sits, permissions live, and, increasingly, the place where the agents are supposed to go do agent things under supervision.

Of course Microsoft believes in ecosystems. Microsoft has been an “ecosystem” since before most AI founders were born.

The interesting part is that Satya is probably right.

For the last two years, AI strategy has had a very frontier-maximalist flavor. Who has the best model? Who has the most GPUs? Who has the highest benchmark score? Who gets access to the model that everyone else is not allowed to use yet? This is a fun way to run a research lab but a stressful way to run a company. If you are an enterprise CIO, your AI strategy cannot be: “We hope the smartest lab continues to like us, price us reasonably, pass its safety evals, avoid regulatory trouble, and not revoke access to the model we built our workflow around.” That is a hostage situation.

Satya gives enterprises a more pleasant story to tell themselves: You are not powerless. You have the data, workflows, permissions, identity layer, compliance requirements, business logic, distribution, and institutional scar tissue that turns a clever demo into something a bank, hospital, retailer, or insurer can actually run.

The frontier model may be smarter than your company, but it does not know your company. That gap is where the money is.

This is also where Microsoft’s argument becomes very Microsoft. The value, in this telling, migrates from the model to the harness around the model: agents, memory, permissions, enterprise data, evaluation, observability, governance, workflow integration - things Microsoft already sells, bundles, secures, bills, and calls a platform.

The strategic elegance of the argument is that Microsoft does not have to say frontier models are commoditizing. That would be rude, especially given its OpenAI relationship. It simply says frontier models are unstable without ecosystems, which is a much nicer way of saying: the engine matters, sure, but you need to build the car around it to actually go places.

This is especially timely because the pure frontier bet is beginning to look expensive, brittle, and mildly insane. Costs are spiraling. Pricing is a moving target. And if Fable 5 can be effectively yanked out of customers’ hands overnight, the lesson is not just “be careful which model you pick” but that dependency on any one of them is an architectural risk.

The enterprise response cannot be to bet the farm on one model. It has to be to build the farm. The model must become one replaceable component in a broader operating system. Still important. Still expensive. Still magical on a good day. But no longer the whole religion.

There is also a political argument hiding inside the platform argument.

A world dominated by a few frontier labs may be technically impressive, but it lacks permission. No enterprise wants to explain to its board that its core workflows depend on a model it does not control, trained by a company it cannot audit, governed by policies it did not negotiate, and potentially changed or restricted overnight. No government wants the economy’s cognitive infrastructure concentrated in a handful of private labs. No regulator wants to bless a future where intelligence is both essential and externally administered.

That is the deeper meaning of Satya’s warning that there is no societal permission for an AI future that hollows out entire industries. Such a world would be economically concentrated, politically fragile, operationally brittle, and socially hard to legitimate.

An ecosystem softens that threat. It spreads power around. It creates stakeholders. It lets enterprises participate in the future rather than be colonized by it. It gives regulators more seams to grab. It gives customers more exit options. And, conveniently, it gives Microsoft the role it has always wanted: the control plane through which the frontier enters the real economy.

So the line is both reassurance and strategy.

To enterprises: do not panic, your context is power.

To frontier labs: intelligence without distribution is a science project.

To regulators: ecosystems are easier to trust than empires.

And to Microsoft shareholders: the platform era is not over.

🎮 TL;DR

• Anthropic says the loop is starting to close. As recursive self improvement gets closer, they want to press pause on the frontier AI race but only if others stop running too.

• Google’s $80B raise is the opening shot in AI balance sheet warfare. By front-running the AI IPO parade, Google is pulling capital from the market before investor appetite turns into indigestion.

• Fable 5 is Mythos 5 with a chaperone for the general public. The public gets frontier intelligence until the model wanders into a sensitive domain, then it falls back to Opus 4.8. Think role based access for frontier intelligence.

• Enterprise AI has entered its operating-discipline era. Tokens are on a budget, managers need technical judgment, and AI theater is being replaced by the less glamorous question: does this cut delivery time in half?

• Dario wants Treebeard to move faster. His policy essay is earnest and hard to dismiss. AI is accelerating upstream while government still deliberates downstream.

🔥 Anthropic Pulls the Fire Alarm

Anthropic has a habit of periodically walking into the middle of the AI party to turn down the music and remind everyone that the building may be on fire. Its latest essay, “When AI Builds Itself” is very much in that tradition: long, earnest, alarming.

To the believers, this is Anthropic sounding the alarm from inside the machine room.
To the skeptics, it is exquisitely calibrated safety theater: warn about the most extreme version of the future, make the opportunity look civilization-sized, and position yourself as the responsible one amidst reckless accelerationists.

Maybe this will make me look naive in hindsight, but I find sincerity in their awkward posture: “We cannot stop unless you stop, but we think both of us should stop.” It sounds absurd. It's also basic game theory.

A unilateral pause by Anthropic would only transfer the lead to another lab. A coordinated pause requires verification, enforcement, shared triggers, and a way of knowing nobody is secretly training the next model in a desert data center while everyone else behaves nicely.

The core argument is that AI is no longer just changing how people work, but how AI itself gets built. This is recursive self-improvement: the old sci-fi idea that an intelligent system could improve itself, then use the improved version to improve itself again. And they come with evidence:
- As of May-26, 80%+ of the code merged into Anthropic’s codebase was Claude authored.
- Anthropic engineers now merge 8x as much code per day as in 2024.
- In April, Claude shipped 800+ fixes that reduced a class of API errors by 1000x, work that would have taken a human 4 years.
- Success rate on open-ended tasks hit 76% in May-26, up 50 points in 6m.

The essay sketches 3 futures:

• The trend stalls
  Models diffuse, but the curve bends. Compute, energy, chips, interconnects, or data become constraints. Society gets more time. Anthropic says possible, not likely.

• The pyramid of agents arrives
  AI labs keep compounding efficiency, but humans still set direction. 100-person companies can do the work of 10K or 100K person orgs. Every employee becomes a manager of synthetic labor.

• The loop closes
  AI systems become recursively self-improving. Progress becomes less constrained by human labor, more by compute. This is where the tone shifts from “productivity uplift” to “arms-control problem.”

This future may not arrive. The physical world may impose friction. The lab can move at compute speed but the world still moves at human speed.

But that mismatch is the problem. If recursive intelligence accelerates upstream while institutions remain slow downstream, our ability to govern falls further behind.

So what is Anthropic advocating for? To slow or temporarily pause frontier AI development, but only if that pause is verifiable and coordinated across labs and countries. Think mutual deterrence.

You can think Anthropic is self-interested and still think this is true. That is what makes the essay interesting.

🌊 Google Gets to the Capital Buffet First

As SpaceX, Anthropic, and OpenAI wait in the wings, Google walked onstage and announced an $80B equity raise, the largest stock sale in history.

For context, the largest IPO to date was Saudi Aramco in 2019, which raised ~$29B. SpaceX will raise $75B at IPO. OpenAI and Anthropic raised $122B and $65B, respectively, in their recent private rounds, which means the pvt AI market is already treating “largest IPO ever” as a rounding error.

Neither OpenAI nor Anthropic has shared how much they plan to raise at IPO, but assume something around $75-$100B and you get to a strange place: roughly $350B+ of capital demand from just 4 offerings. Add the other IPOs this year, and the market may be staring at $400B-plus of new equity.

A remarkable reversal for a market that spent the last two years begging for ways to play the AI trade. We are about to go from hunger to indigestion.

To put the flood in context, the IPO market raised about $44B last year. The best IPO year on record was 2021, when companies raised around $140B. The market is now being asked to absorb ~3x the best year ever and ~10x last year’s issuance, largely around one theme.

Which raises the obvious question: where does the money come from?

Investors are not usually sitting on $400B of idle cash. To buy something, they often have to sell something. So what gets sold? Do investors trim the current winners of the AI trade, like Nvidia, Broadcom, semis, cloud, and high-multiple software, to buy the next generation of AI names? Do they rotate out of defensives and go all-in on the capital-intensive AI complex? Or do these offerings simply not clear at the valuations their bankers have been whispering into pitch decks?

However that uncertainty plays out, Google has done the smart thing: it cut the line before investor exhaustion sets in. It is pulling capital out of the ecosystem before Anthropic, OpenAI, SpaceX, and everyone else can get to it. In a world where the pool of AI-hungry capital is finite, being first is invaluable.

This is not because Alphabet cannot afford its AI ambitions. It is one of the most cash-generative companies in history. But why use debt or balance sheet cash when public markets are throwing cheap capital at AI? Give the people what they want.

The pure-play AI names may have more narrative torque, but Google has more ways to survive being wrong. If AI margins disappoint, it has Search. If enterprise adoption takes longer, it has YouTube. If model prices collapse, it has distribution. If capex gets ugly, it has cash flow.

Google is also trying to change the comp set. By raising now, they force investors to compare it with the coming AI IPO class. If the market is willing to underwrite SpaceX, Anthropic, and OpenAI as once-in-a-generation platforms, Google can reasonably ask: what exactly do you think we are?

The AI boom is maturing. First came model envy. Then product panic. Now balance sheet warfare. Google fired the first shot.

🪪 The Role-Based Access Era of AI

Fable 5 is a very Anthropic launch: frontier intelligence with a permission slip, a safety helmet, and a name that sounds like it was chosen by a philosopher. With this launch, we finally get a Mythos-class model in the public domain. Sort of.

Anthropic is releasing 2 versions of the same model: Fable 5 for general use, and Mythos 5 for members of Project Glasswing. Both models share the same underlying model weights. The difference is product and safety packaging: when Fable’s classifiers detect sensitive domains, the request falls back to Opus 4.8. That is the core product tension: Fable is as smart as Mythos until it enters a restricted neighborhood.

Anthropic appears to have invented role-based access control for intelligence. The public gets Mythos with a chaperone. Trusted partners get Mythos without one.

We are moving from “how smart is the model?” to “what version of intelligence are you allowed to access, under what controls, for which job?” One model for general availability, one for trusted access.

The most interesting guardrails are not even the obvious ones. Yes, Fable includes protections around cyber, bio, chem, and distillation. But it also limits effectiveness on frontier LLM development: pretraining pipelines, distributed training infrastructure etc. Not just safety-gated against bad actors, it is also competition-gated against model builders.

To be clear, the model itself looks phenomenal. It tops the Artificial Analysis Intelligence Index, 5 points ahead of GPT-5.5. It can work autonomously for longer than any previous Claude models.

The headline is software engineering. Fable 5 scores 95% on SWE-bench Verified and 80% on SWE-bench Pro, a meaningful jump over competitors.

During early testing, Stripe said Fable 5 compressed months of engineering into days. In a 50M line Ruby codebase, the model completed a codebase-wide migration in 1 day that otherwise would have taken a full team more than 2 months by hand.

Model availability is particularly revealing. Fable 5 is temporarily included in paid plans through June 22, consuming 2x Opus usage. After June 23, usage shifts toward credits.

The era of infinitely subsidized frontier intelligence is coming to a close. Price will now reflect quality of the underlying model.

For the first era of AI, we trained users to believe increasingly capable intelligence would simply show up inside the same subscription box, all bundled into the monthly fee. Fable 5 resets expectations. Price will increasingly reflect the quality, autonomy, and danger of the underlying model.

And so the fable goes something like this: once upon a time, intelligence was scarce. Then it became abundant, and the hard part became deciding who gets which kind, under what guardrails, for how long, and at what price. The end. Or, more likely, the beginning.

Update: As of 2 days ago, the U.S. government has now instructed the company, citing national security interests, to pull model access from foreign nationals inside or outside the United States, “including foreign national Anthropic employees.” Anthropic says the “net effect” of the order is to disable the models for everyone to ensure compliance.

The question is no longer simply who gets Fable, who gets Mythos, and who gets Opus fallback. Frontier intelligence is becoming a regulated asset class: metered by price, throttled by safety, and ultimately permissioned by the state.

🛠️ Enterprise AI Grows Up

Last week, we hosted an off-the-record conference with the people actually driving the change most of us are observing from the outside: enterprise CTOs.

I interviewed Hilary Packer, EVP, Head of AI at Amex; Sven Gerjets, CTO at Gap; and Avi Saxena, CTO at Warner Bros. Discovery - leaders steering very large ships in very public waters - and spent the day in conversation with CTOs deep in the practical reality of AI. A few takeaways:

1. Tokens are the new cloud bill
   Until recently, CTOs let teams experiment with AI without hard usage limits. That era is ending as costs rise. We heard $40K–$100K per engineer per year discussed as plausible token budgets. Jensen Huang has publicly proposed a token budget of up to 50% of annual salary.
   AI usage moves from free-form experimentation to an explicit operating budget and the CTO becomes the CFO of tokens.

2. The engineering talent barbell is getting heavier
   Hiring increasingly centers on two ends of the spectrum: early-career grads and very experienced engineers. The toughest slice is engineers with 2–5 years of experience - not yet senior enough for deep judgment, and not as plastic as new grads learning in an AI-native environment from day one.

3. Interviewing for AI fluency is still unsolved
   A recurring question: how do you interview for AI capability?
   The best answer I heard: work with candidates live. A co-working session reveals how someone thinks, prompts, debugs, evaluates output, and decides when to trust or override the model.

4. The “people-only” engineering manager is in trouble
   Managers need to be technical again. In an AI-native engg org, managers cannot just be process coordinators or people leaders. They need enough technical depth to evaluate output, understand where AI is accelerating or masking work, and help teams make better decisions.

5. AI theater is real - and the best CTOs are actively fighting it
   Every org has people building demos to get executive attention. Some are useful; many are theater. The sharpest CTOs are discouraging side quests and redirecting teams back to the main question: How do we cut roadmap delivery time in half?

6. The SDLC is up for grabs
   One of the most interesting conversations was not about which tool wins, but about first principles: What would the SDLC look like if you designed it from scratch in 2026? Not “how do we add AI to Jira/GitHub/CI/CD/code review?” but “what would we build if we threw out the current workflow?” That question is unlocking more radical thinking.

7. Back-office AI risk is exploding
   Front-office use cases get the attention, but a lot of CTOs are increasingly worried about AI creeping unchecked into the back office. Every vendor is adding AI. Many are doing it faster than their customers can understand, audit, or govern.

We are past unbridled AI enthusiasm and into the era of operating discipline.

🌲AI Policy Moves at Ent Speed

Dario published another long essay about AI policy, by now its own subgenre. It opens with a Lord of the Rings analogy. Treebeard, the wise tree, moves slowly just like our lawmakers, who deliberate, consult, and compromise before they move. That is often a virtue but exponential AI progress has created a timing mismatch. The worry is not just that policy will be late, but that it will arrive to a different world.

I’ve said this before, at the risk of sounding naive: I believe Dario is earnest.

Yes, he is the CEO of Anthropic. Yes, he benefits from a world where AI is treated as powerful and dangerous. Yes, “regulate us” is often the mating call of the scaled incumbent. But sincerity and self-interest can share a bunk bed.

What makes Dario hard to dismiss is consistency. He made these arguments when OpenAI was clearly ahead. He kept making them as Anthropic caught up. And he is still making them now, when it is in the lead.

The essay walks through 5 policy areas.

• Public Safety
  Amodei wants mandatory third-party testing for the most powerful models before deployment, focused on catastrophic risks: cyber, bio, loss of control, and AI systems that accelerate AI R&D itself.
  We do not let companies sell planes without safety checks, drugs without trials, or cars without basic standards. When a product has a meaningful blast radius, society builds rules around testing, liability, and recall.

• Labor Policy
  If AI substitutes large categories of work overnight, workers may not have time to retrain, relocate, or move into newly created occupations.
  Dario discusses wage insurance, retraining, job matching, retention incentives, and better labor-market measurement. He also leaves the door open to structural redistribution: UBI, taxes on AI companies, capital gains taxes, or universal capital accounts.

• Accelerating the Positive
  On frontier model deployment, Amodei wants stricter review. On AI-enabled science, he wants faster review. If AI accelerates drug discovery, biology, materials science, and clinical research, the bottleneck shifts to approval.
  The upside of AI will not arrive just because models get better. It has to pass through institutional plumbing.

• Civil Liberties
  Amodei is not just worried about rogue systems. He is worried about obedient ones. AI that perfectly follows instructions can be dangerous if the instructions come from a govt conducting mass surveillance or a military delegating lethal force too freely.
  Policy suggestions include limits on domestic autonomous weapons, stronger privacy protections, safeguards around govt AI use.

• Geopolitics
  Advanced AI will become a source of national power: economic, military, scientific, and diplomatic. His answer is a democratic coalition that coordinates on AI development, chip access, export controls, safety standards, and AI defense.

Treebeard eventually does act. The question is whether AI policy can do the same before the forest has already been cut down.

Dario Amodei has published another long essay about AI policy, which by now is its own subgenre. This one, Policy on the AI Exponential, is about the mismatch between AI, which is moving on an exponential curve, and politics, which often moves like Treebeard.

Amodei opens with the Lord of the Rings analogy. Democratic institutions are like Treebeard, the ancient sentient tree, which moves slowly. They deliberate, compromise, consult stakeholders, and produce reports before they move. That slowness is often a virtue, but AI progress has created a timing problem. The feedback loop for capability improvement is measured in weeks. The feedback loop for policy is measured in election cycles, agency processes, and committee calendars. The worry is not just that policy will be late, but that it will arrive to a different world.

Amodei’s other analogy is more familiar: cars, airplanes, and drugs. New technologies create enormous benefits, but we do not let companies sell planes without safety checks, drugs without trials, or cars without basic standards. That does not mean we ban planes, drugs, or cars. It means that when a product has a meaningful blast radius, society eventually builds rules around testing, liability, certification, and recall.

I’ve said this before, at the risk of sounding naive: I believe Dario is earnest.

Yes, he is the CEO of Anthropic. Yes, Anthropic benefits from a world where frontier AI is treated as powerful, dangerous, and too complex for two people in a garage to govern. Yes, “regulate us” is often the mating call of the scaled incumbent. But sincerity and self-interest can share a bunk bed.

The obvious critique is that Dario is warning about the frontier while personally moving it forward.

His defense is that progress is inevitable, and that it is better for a safety-oriented lab to help set norms near the frontier than to abandon it to less cautious actors.

What makes Dario harder to dismiss is consistency. He has occupied this strange position in AI for years: building the thing, warning about the thing, and asking governments to regulate the thing. He made these arguments when OpenAI was clearly ahead. He kept making them as Anthropic caught up. And he is still making them now, when Anthropic is leading. The line has not conveniently appeared only after the moat was dug.

Dario is the person selling rocket fuel while begging for a fire code. Easy to mock. Harder to ignore.

The essay’s core argument is simple: AI policy has to graduate from transparency to control. For the last few years, AI policy has mostly meant transparency: publish evaluations, disclose safety practices, report incidents, give governments visibility. Fine. Useful. But Amodei argues that some risks are now specific enough to test. And if they can be tested, they can be regulated.

The essay then walks through five policy areas.

1. Public Safety

Amodei wants mandatory third-party testing for the most powerful models before deployment, focused on catastrophic risks: cyber, bio, loss of control, and AI systems that accelerate AI R&D itself. If a model fails those tests, the government should have the authority to block, deter, or revoke deployment.

This is the airplane analogy applied to model releases. You do not ask Boeing to publish a thoughtful essay about safety culture and then let passengers board. You test the plane, and someone has the authority to ground it.

The concern is obvious: safety regulation can become an incumbent moat. Big labs can hire lawyers, auditors, policy teams, and former regulators. Smaller labs get the bill. Still, the alternative is strange. If a model can materially increase cyber or bio risk, should release depend entirely on the company’s internal judgment? “Trust us, we ran evals” is not a regulatory regime.

2. Labor & Tax Policy

Amodei argues that AI could create labor-market disruption that is larger than prior automation waves. The usual optimistic story is that tech displaces some jobs, creates new ones, and raises overall productivity. That story may still be true, but the timeline matters.

If AI can substitute for large categories of cognitive work overnight, then workers may not have time to retrain, relocate, or move into newly created occupations. Economists may eventually be right in the long run. The problem, as always, is living through the footnote.

He discusses wage insurance, retraining, job matching, retention incentives, and better labor-market measurement. He also leaves the door open to more structural redistribution if AI causes a sustained decline in demand for human labor: UBI, higher taxes on AI-driven companies, capital gains taxes, or universal capital accounts.

3. Accelerating the Positive

On frontier model deployment, Amodei wants stricter review but on AI-enabled science, he wants faster review. His point is that if AI accelerates drug discovery, biology, materials science, and clinical research, then the bottleneck may shift from invention to approval.

The upside of AI will not arrive just because models get better. It has to pass through hospitals, regulators, insurers, procurement systems, clinical trial networks, universities, and the rest of the institutional plumbing. A miracle drug stuck in ordinary paperwork is, from the patient’s perspective, not yet a miracle.

This is the underrated AI policy question: not just how do we slow down dangerous capabilities, but how do we speed up beneficial deployment? Otherwise we get the worst possible combination: risks that compound quickly and benefits that don’t.

4. Civil Liberties

A lot of AI risk discourse focuses on rogue systems. Amodei is also focused on obedient systems. An AI that perfectly follows instructions can still be dangerous if the instructions come from a government conducting mass surveillance, a military delegating lethal force too freely, or a bureaucracy using automation to make high-stakes decisions with limited recourse.

AI will change the balance between citizens, companies, and governments. That can happen through surveillance, autonomous weapons, persuasion, cyber capabilities, and administrative automation. The concern is not only that AI systems might behave badly on their own. It is that they might behave very efficiently on behalf of institutions that already have coercive power.

His policy suggestions include limits on domestic autonomous weapons, stronger privacy protections, restrictions on data brokers, safeguards around government use of AI, and rights for citizens to use AI systems when navigating state action.

5. Geopolitics

Amodei argues that advanced AI will become a source of national power: economic, military, scientific, intelligence, and diplomatic. His answer is a democratic coalition that coordinates on AI development, chip access, export controls, safety standards, and defensive use of AI.

If frontier AI becomes a general-purpose engine of power, then AI policy cannot remain a domestic consumer-protection debate. It becomes part of alliance management, export strategy, military doctrine, and economic statecraft.

Amodei’s essay is a warning that AI policy cannot remain a series of after-the-fact hearings about things that have already happened. If frontier AI is going to be regulated like a serious technology, the institutions around it need to move earlier in the cycle: before release, before labor markets fully absorb the shock, before scientific bottlenecks waste the upside, before governments normalize AI-enabled coercion, and before geopolitics hardens into an uncontrolled race.

The hard part is that democratic institutions are not designed for exponential time. They are designed for legitimacy, deliberation, and constraint. Those are features, not bugs. But they become dangerous if the only choice is between moving slowly and moving blindly.

Treebeard eventually does act. The question is whether AI policy can do the same before the forest has already changed shape.

Last week, we hosted an off-the-record conference with the people actually driving the change most of us are observing from the outside: enterprise CTOs.

I had the chance to interview Hilary Packer, EVP, Head of Enterprise Data & AI at American Express; Sven Gerjets, EVP & CTO at Gap; and Avi Saxena, CTO at Warner Bros. Discovery - leaders steering very large ships in very public waters - and spent the day in conversation with CTOs deep in the practical reality of AI adoption. There was more than $600B of market cap represented in the room.

One caveat: our CTO Advisory Board is a particularly sharp, forward-leaning group, allergic to AI theater and intensely focused on outcomes. So these takeaways likely represent the top decile of the market - but that also makes them useful signals for where the rest of the market is heading.

A few takeaways:

1. Token budgets are becoming the new cloud budgets.

Until recently, many CTOs let teams experiment with AI tools without hard usage limits. That era is ending. Costs are rising quickly, and enterprises are starting to ask what “good” AI consumption actually looks like.

In several conversations, CTOs noted that token usage tends to rise with seniority. More senior engineers are using AI more heavily, but they are also generating much higher-impact output.

One number that came up repeatedly: only ~10% of engineers in an organization are able to use AI effectively enough to 10x themselves. That raises a hard question: should AI budgets be distributed evenly, or concentrated around the people who can turn usage into leverage?

We heard figures in the range of $40K–$100K per engineer per year discussed as plausible AI/token budgets for high-performing engineers. Jensen Huang has publicly proposed a token budget of up to 50% of an engineers annual salary.

The direction of travel is clear: AI usage is moving from “free-form experimentation” to an explicit operating budget. The CTO is becoming the CFO of tokens.

2. The engineering talent barbell is getting more pronounced.

Hiring conversations increasingly centered on two ends of the talent spectrum: early-career grads and very experienced engineers.

The toughest slice may be engineers with 2–5 years of experience. They often do not yet have the judgment of a senior engineer, but they may also lack the plasticity of new grads who are learning the craft in an AI-native environment from day one.

Expectations for this cohort are changing quickly as the premium shifts toward people who can combine taste, judgment, and AI fluency.

3. Interviewing for AI fluency is still unsolved, but co-working is emerging as the best signal.

A recurring question was: how do you actually interview for AI capability?

AI fluency is not about whether someone can “use Cursor” or “use Claude.” It is about whether they can turn ambiguous intent into high-quality shipped work faster than before. That is hard to assess in a traditional interview loop.

The best answer I heard was simple: work with candidates live. A real co-working session where you can see how someone thinks, prompts, debugs, evaluates output, and decides when to trust or override the model.

4. The “people-only” engineering manager is in trouble.

Several CTOs made a similar point: managers need to be technical again.

In an AI-native engineering org, managers cannot just be process coordinators or people leaders. They need enough technical depth to evaluate output, understand where AI is accelerating or masking work, and help teams make better architectural and product decisions.

The purely non-technical engineering manager is likely to become a disappearing species.

5. AI theater is real - and the best CTOs are actively fighting it.

Every organization now has ambitious people building demos to get executive attention. Some of those demos are useful, many are theater.

The sharpest CTOs are becoming very disciplined about this. They are discouraging orthogonal side quests and redirecting teams back to the main question: How do we cut roadmap delivery time in half?

One practical tactic I loved: when someone shows you a demo, ask for the prompts first. The prompts reveal the depth of thought behind the work very quickly. You can tell whether someone designed a real system, hacked together a fragile workflow, or simply wrapped a model in a thin interface.

6. The SDLC is up for grabs.

One of the most interesting conversations was not about which tool wins, but about first principles: What would the software development lifecycle look like if you designed it from scratch in 2026?

Not “how do we add AI to Jira/GitHub/CI/CD/code review?” but “what would we build if we threw out the current workflow?”

That question is starting to unlock more radical thinking. Some CTOs are piloting pods with one designer and one developer. In some cases, if the developer has strong product and design taste, the pod can shrink to one person.

That has huge implications for org design, planning cycles, product velocity, and the boundary between product, design, and engineering.

7. Back-office AI risk is exploding.

The front-office and engineering use cases get most of the attention, but a lot of CTOs are increasingly worried about AI creeping into the back office.

Every vendor is adding AI. Many are doing it faster than their customers can understand, audit, or govern.

That creates a new category of vendor risk: not just “what data do you have?” but “what are you doing with AI, where is our data going, and what new failure modes have you introduced?”

The overall takeaway: the conversation has moved beyond AI enthusiasm.

The best CTOs are asking how to allocate scarce AI leverage, how to redesign teams around it, how to separate substance from theater, and how to manage the new cost and risk surfaces it creates. Its the era of operating discipline.

Claude Fable 5 is a very Anthropic launch: frontier intelligence with a permission slip, a safety helmet, and a name that sounds like it was chosen by a philosopher.

With Fable 5, we finally get a Mythos-class model in the public domain. Sort of.

Anthropic is releasing two versions of the same model: Claude Fable 5 for general use, and Claude Mythos 5 for members of Project Glasswing. Both models share the same underlying model weights. The difference is product and safety packaging.

When Fable’s classifiers detect sensitive domains, the request falls back to Claude Opus 4.8 and the user is notified. Anthropic shares that fallback occurs in fewer than 5% of sessions on average. That is the core product tension: Fable is as smart as Mythos until it enters a restricted neighborhood.

Anthropic appears to have invented role-based access control for intelligence. The public gets Mythos with a chaperone. Trusted partners get Mythos without one. Predictably, the internet had feelings about this.

We are moving from “how smart is the model?” to “what version of intelligence are you allowed to access, under what controls, for which job?” One model for general availability, one for trusted access. One that can help you refactor a gnarly codebase, synthesize research, and grind through complex knowledge work. Another that lives closer to the blast radius and therefore behind gates.

The most interesting guardrails are not even the obvious ones. Yes, Fable includes protections around cyber, bio, chem, and distillation. But it also limits effectiveness on frontier LLM development: pretraining pipelines, distributed training infrastructure, accelerator design, and similar work. Fable 5 is not just safety-gated against bad actors, it is also competition-gated against frontier model builders.

To be clear, the model itself looks excellent. It is #1 on the Artificial Analysis Intelligence Index, nearly 5 points ahead of GPT-5.5.

Fable 5 and Mythos 5 can work autonomously for longer than any previous Claude models.

The headline is software engineering. Fable 5 scores 95% on SWE-bench Verified and 80% on SWE-bench Pro, which is a meaningful jump over Opus 4.8 and competitors.

During early testing, Stripe said Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model completed a codebase-wide migration in one day that otherwise would have taken a full team more than two months by hand.

Lets talk model specs.

The context window is 1 million tokens, matching Claude Opus 4.8. That means entire codebases, long research trails, messy enterprise documents, and sprawling project histories can fit inside the working memory of the system.

Pricing is $10 per million input tokens and $50 per million output tokens, roughly 2x Opus 4.8. Anthropic’s argument is that price per token is the wrong metric if the model completes tasks in fewer turns.

Availability is particularly revealing. Fable 5 is temporarily included in Pro, Max, Team, and seat-based Enterprise plans through June 22, consuming 2x Opus usage. After June 23, usage shifts toward credits, with Anthropic saying it intends to restore broader subscription access once capacity improves.

The era of infinitely subsidized frontier intelligence is coming to a close. Price will now reflect quality of the underlying model.

For the first era of generative AI, we trained users to believe increasingly capable intelligence would simply show up inside the same subscription box. Better reasoning, longer context, stronger coding, more autonomy - all bundled into the monthly fee.

Fable 5 resets expectations. The more capable the model, the more it looks like scarce infrastructure: expensive to serve, risky to expose, and valuable enough to meter. Price will increasingly reflect the quality, autonomy, and danger of the underlying model.

This is capability rationing in three dimensions:

Who gets access.
What they are allowed to do.
What it costs when they do it.

And so the fable goes something like this: once upon a time, intelligence was scarce. Then it became abundant, and the hard part became deciding who gets which kind, under what guardrails, for how long, and at what price. The end.

Or, more likely, the beginning.

As SpaceX, Anthropic, and OpenAI wait in the wings, Google has walked onstage and announced an $80 billion equity raise, the largest stock sale in history.

For context, the largest IPO to date was Saudi Aramco in 2019, which raised roughly $29 billion. SpaceX is reportedly planning to raise $75 billion. OpenAI and Anthropic raised $122 billion and $65 billion, respectively, in their most recent private rounds, which means the private AI market is already treating “largest IPO ever” as a rounding error.

Neither OpenAI nor Anthropic has disclosed how much they plan to raise when they go public, but assume something in the $75 billion to $100 billion range and you quickly get to a strange place: roughly $350 billion-plus of capital demand from just four AI-adjacent offerings. Add the IPOs that have already happened this year, and the market may be staring at $400 billion-plus of new equity supply. Less window reopening, more dam breaking.

It is also a remarkable reversal for a market that spent the last two years begging for liquid ways to play the AI trade. We are about to go from hunger to indigestion.

To put the flood in context, the IPO market raised about $44 billion last year. The best IPO year on record was 2021, when companies raised around $140 billion. The market is now being asked to absorb ~3x the best year ever and ~10x last year’s issuance, largely around one theme.

Which raises the obvious question: where does the money come from? Investors are not usually sitting on $400 billion of idle cash labeled “AI destiny.” To buy something, they often have to sell something. So what gets sold? Do investors trim the current winners of the AI trade, like Nvidia, Broadcom, semis, cloud, and high-multiple software, to buy the next generation of AI names? Do they rotate out of defensives and go all-in on the capital-intensive AI complex? Or do these offerings simply not clear at the valuations their bankers have been whispering into pitch decks?

However that uncertainty plays out, Google has done the smart thing: it cut the line before investor exhaustion sets in. Google is pulling capital out of the ecosystem before Anthropic, OpenAI, SpaceX, and everyone else can get to it. In a world where the pool of AI-hungry capital is finite, being first is an invaluable scheduling advantage.

This is not because Alphabet cannot afford its AI ambitions. Alphabet is one of the most cash-generative companies in history. It generated $174 billion of operating cash flow over the last twelve months. This is not a company passing a hat around Sand Hill Road.

But why use debt or balance sheet cash when public markets seem so willing to throw cheap capital at the AI trade? It’s not a bad idea to capitalize on the insatiable market demand for AI exposure. Give the people what they want.

Google’s move is clever because it steps in front of the pure-play AI names and says, in effect: you want AI upside? Fine. Buy ours.

The pure-play AI names may have more narrative torque, but Google has more ways to survive being wrong. If AI margins disappoint, Google still has Search. If enterprise AI adoption takes longer, Google still has YouTube. If model prices collapse, Google still has distribution. If capex gets ugly, Google still has cash flow. Google is trying to turn the conglomerate discount into downside protection.

It is also trying to change the comp set. By raising now, Alphabet effectively forces investors to compare it with the coming AI IPO class. If the market is willing to underwrite SpaceX, Anthropic, and OpenAI as once-in-a-generation platforms, Google can ask a reasonable and slightly rude question: what exactly do you think we are?

As part of the raise, Alphabet is selling $10 billion of stock to Berkshire in a private placement. This takes an aggressive AI infra financing and wraps it in the comforting aesthetic of Omaha. Yes, this is an arms race for artificial intelligence, but Warren Buffett’s firm is nearby, so perhaps it is actually value investing.

An anchor investor matters for credibility. In a market about to be stuffed with giant AI offerings, the presence of Berkshire says: this deal has adult supervision.

The AI boom is entering its capital markets phase. That is usually when things get more interesting, more expensive, and more dangerous.

The first phase was model envy. The second phase was product panic. The third phase is balance sheet warfare. Google just fired first.

Anthropic has a habit of periodically walking into the middle of the AI party, turning down the music, and reminding everyone that the building may be on fire.

Its latest essay, “When AI Builds Itself,” is very much in that tradition. It is long, earnest, alarming, and almost guaranteed to be interpreted through whatever lens people already use to view Anthropic.

To the believers, this is Anthropic doing what it has always claimed to do: sounding the alarm from inside the machine room.

To the skeptics, it is another piece of exquisitely calibrated safety theater: warn about the most extreme version of the future, make the market opportunity look civilization-sized, and position yourself as the only responsible adult in a room full of reckless accelerationists.

There is plenty of material for the skeptics. Anthropic is a frontier AI company warning that frontier AI companies are moving too fast. It is arguing for restraint while competing aggressively. It recently declined to broadly release Mythos Preview, a decision supporters saw as responsible and critics saw as buzz-maxxing. The most powerful model is always more interesting when almost nobody can use it.

And yet, pure cynicism feels too easy here.

Maybe this will make me look naive in hindsight, but I do think there is a real ring of sincerity in Anthropic’s public discomfort. The posture is almost comically awkward: “We cannot stop unless you stop, but we think both of us should stop.” It sounds absurd. It is also basic game theory.

A unilateral pause by Anthropic would mostly be a transfer of advantage to OpenAI, Google, xAI, Meta, or some state-backed lab elsewhere. A coordinated pause, meanwhile, would require verification, trust, enforcement, shared definitions, shared triggers, and some way of knowing that nobody is secretly training the next model in a desert data center while everyone else behaves nicely.

Anthropic cannot afford to say “we are opting out of the race.” So it is saying “we are trapped in a race that requires intervention by institutions capable of making races pause.”

The core argument of the essay is that AI is no longer just changing how people work, but how AI itself gets built. Anthropic lays out the progression as a kind of corporate evolution chart.

In the early days, humans wrote code and documents on laptops. Then chatbots helped generate snippets. Then coding agents began editing entire files. Today, agents can run code, delegate work to other agents, and complete multi-hour tasks. The next step, if the curve continues, is what Anthropic calls “closing the loop”: AI systems capable of designing, training, testing, and improving their own successors.

This is recursive self-improvement, the old science-fiction idea that an intelligent system could improve itself, then use the improved version to improve itself again, and so on. The phrase has historically lived somewhere between LessWrong, Bostrom, and late-night existential panic. Anthropic is trying to drag it into the operating plan.

And they come bearing evidence:

• As of May 2026, 80%+ of the code merged into Anthropic’s production codebase was authored by Claude. Before Claude Code launched in Feb-25, that number was in the low single digits. The typical Anthropic engineer now merges 8x as much code per day as in 2024.

• In April, Claude reportedly shipped more than 800 fixes that reduced a class of API errors by a factor of 1000. The engineer overseeing the work estimated it would have taken a human 4 years to complete.

• Claude’s success rate on open-ended coding tasks reached 76% in May 2026, up 50 points in 6 months. In one example, after a routine upgrade began crashing tens of thousands of training jobs, an engineer pointed Claude at the live incident with little more than text context and cluster access. Claude isolated an obscure debugging flag, reproduced the issue, and confirmed a fix in about two hours. Normally, that would have been 2-3 days of work.

That is the first transition: from labor to leverage.

The second transition is more important: from execution to judgment.

In an internal test, Anthropic took 129 real research moments where a human had gone somewhat off-course. It then showed different Claude models the session up to that point and asked what they would do next. A separate judge compared the model’s proposed next step against the human’s actual next step. In November 2025, Claude Opus 4.5 beat the human choice 51% of the time. By April 2026, Mythos Preview beat the human choice 64% of the time.

This is not a clean “AI beats researchers” result (yet). Anthropic picked moments where the human choice had room for improvement. The test is narrow. The judge was itself a Claude model. There are enough caveats here to keep a methodologist busy for a weekend.

Still, the result matters. Research is not mostly grand eureka moments. It is a long chain of next-step decisions. Try this. Ignore that. Rerun the experiment. Check the data. Change the optimizer. Kill the project. Double down. Most scientific and engineering progress is not cinematic inspiration. It is judgment applied repeatedly to friction. And the judgment appears to be improving quickly.

The essay then goes on to sketch three possible futures.

• In the first, the trend stalls. Today’s models diffuse widely, but the curve bends. Maybe research judgment is not easily learned through scaling. Maybe compute, energy, chips, interconnects, or data become hard constraints. Maybe the whole thing turns into an S-curve and society gets more time to adapt. Anthropic says this is possible, but does not think it is likely.

• In the second future, AI labs continue to see compounding efficiency gains, but humans still set direction. This is already dramatic. Anthropic imagines 100-person companies doing the work of 10,000 or 100,000-person organizations. That is the “pyramid of agents” world: every employee becomes a manager of synthetic labor, and every organization becomes less constrained by headcount.

  This future is transformational enough. It changes software, services, security, government operations, influence campaigns, authoritarian surveillance, and probably the structure of the firm itself.

• In the third future, the loop closes. AI systems become capable of full recursive self-improvement. Humans move from builders to overseers of a virtual lab. The pace of progress becomes determined less by human research labor and more by compute. This is where Anthropic’s tone changes from “interesting productivity uplift” to “arms-control problem.”

They are not arguing that full recursive self-improvement is here, but that the evidence points far enough in that direction that waiting for proof may be a terrible strategy. By the time everyone agrees the loop has closed, the institutions needed to govern it will already be obsolete.

So what are they advocating for? The option to slow or temporarily pause frontier AI development. Crucially, they want that option to be verifiable and coordinated across frontier labs and countries. Think mutual assured destruction as deterrence.

This future may not arrive. The curve may bend. The physical world may impose friction. Society may remain stubbornly slow. As Anthropic notes, more intelligence cannot make a clinical trial reveal ten-year side effects in a week, hold elections sooner than a constitution allows, or turn a stranger into an old friend over a weekend. The lab can move at compute speed but the world still moves at human speed.

But that mismatch is exactly the problem. If recursive intelligence accelerates upstream while institutions remain slow downstream, the risk is that our ability to understand, govern, and absorb it falls further behind every month.

You can think Anthropic is self-interested and still think this is true. You can roll your eyes at the theatrical caution and still notice that the numbers are strange. You can dislike the nuclear analogy and still admit that verification is the core issue. You can believe the market will overhype this and still believe the underlying curve is moving faster than society can process. That is what makes the essay interesting.

🎬 TL;DR

• The IPO window is reopening, and public markets are about to become the AI's harshest benchmark. Forget MMLU. The real test is how much capital it takes to produce the next dollar of revenue.

• Microsoft’s stock is pricing a rotating anxiety menu: OpenAI dependence, Azure capex, Copilot ROI. But the strategy is cleaner than the tape. Microsoft is trying to become the enterprise AI control plane - routing every model, agent, device, and workflow through its stack.

📜 The S-1 Becomes the New Model Card

The IPO window is reopening and Wall Street is about to become the world's largest AI benchmark.

SpaceX is about to go public. Anthropic has confidentially filed its S-1. OpenAI is preparing its own filings. If even some of the rumored valuations hold, public markets will soon be asked to absorb trillions of dollars of new market cap from companies that, until now, have mostly been priced by private-market consensus, strategic necessity, and vibes.

SpaceX is a different animal but the model labs have the strange task of justifying the valuation of intelligence itself. For years, they have operated like a hybrid of research institution, startup, and state actor. The IPO process forces them to become something much less romantic: reporting companies.

Whoever goes first becomes the benchmark everyone else inherits. A strong Anthropic debut creates oxygen for OpenAI. A shaky OpenAI debut makes every later AI IPO answer for sins it did not personally commit. In IPO markets, the first company through the door sets the mood.

The arrival of these public comps will hopefully discipline private markets. AI valuations today are still part math, part myth. Once these companies trade publicly, everything gets repriced relative to them. We will have a reference point: Why is this startup worth xx revenue if Anthropic trades at xx? Public markets are very good at ruining private market poetry.

The S-1s may become the most important AI documents of the year. We have all hyper fixated on model benchmarks so far because that’s all we’ve had. The S-1s will contain more valuable information: customer concentration, revenue mix, enterprise retention, compute commitments, capital intensity, stock comp.

The real AI benchmark is how much capital it takes to produce the next $ of revenue.

The omissions will matter too.
Why won’t they break out inference costs?
Why won’t they separate enterprise from consumer revenue?
Sometimes the silence in an S-1 is more revealing than the numbers.

The risk sections could be historic. It is not every day that a public market prospectus has to explain model hallucination, catastrophic misuse, regulatory capture, geopolitical dependency, energy constraints, copyright exposure, bio risk, national security risk, and some version of p(doom).

But the biggest question is still where the value accrues. That question has hovered over AI since ChatGPT launched. Does the value accrue to the model labs? The cloud providers? The chip companies? The data centers? The application layer? Public markets will force a verdict.

If the labs list with enormous revenues but structurally mediocre margins, while the infra players keep printing cash, the market may decide that intelligence is valuable but infra captures the value.

For three years, Silicon Valley has been asking whether these models can explain the world. Wall Street is about to ask whether these companies can explain themselves.

🧭 From Copilot to Control Plane

Microsoft stock is hard to read because the market is not really pricing a company; it is pricing a rotating set of AI anxieties. Dependence on OpenAI. Azure capex. Copilot ROI. Every week, one anxiety gets promoted to consensus. The stock will do what stocks do. The strategy is the more interesting read.

Their AI strategy so far has been almost unfairly simple: attach to OpenAI’s momentum, pour that intelligence through Azure, package it as Copilot, and distribute it through the world's most valuable enterprise software estate.

At Build 2026, Microsoft outlined its roadmap to move from the best distributor of someone else’s intelligence to the operating fabric for enterprise AI:

• Windows AI runtime.
  Microsoft announced new on-device models, broader Windows AI APIs, and the Surface RTX Spark Dev Box. AI apps will need to decide, task by task, what should run locally versus in the cloud. Local inference can be faster, cheaper, more private, and available offline; cloud inference can handle heavier reasoning and larger models. Microsoft wants Windows to broker that split.
  The move is classic Microsoft: give developers the APIs, give enterprises the management and security layer, give hardware partners a reason to build AI PCs, and make Windows the default deployment surface.

• Project Solara.
  Microsoft built an Android-based platform for agentic hardware. They showed 2 form factors: a desk device and a badge-style wearable.
  These are reference designs, not Microsoft products. The point is to seed an ecosystem: Microsoft provides the OS, hardware partners build the devices.

• IQ for enterprise context.
  By bringing together Foundry, Fabric, and M365, Microsoft created a unified intelligence layer that understands the org. Every company wants better context. Microsoft already sits on the motherlode: email, calendar, Teams, docs, identity, security policies, databases, workflows, and every “final_final_v6_ACTUAL” deck ever produced by modern knowledge work. IQ operationalizes this context.

• From Copilot to Autopilot.
  Autonomous, long-running agents with enterprise-grade compliance - effectively agents designed to persist, monitor, and complete work across systems while staying inside the corporate control plane.

• Model Layer.
  The future is not one model to rule them all. It is routing: frontier models for hard reasoning, small model sfor cheap repetitive tasks, local models for private or latency-sensitive work, tuned models for domain-specific work. With 7 new MAI models, Microsoft is building optionality.

• Frontier Tuning.
  Generic intelligence is impressive, but differentiated intelligence comes from proprietary context. Microsoft is giving companies a path from renting frontier capability to shaping it around their own business systems.

The stock can continue being mysterious. The strategy is not. Microsoft is trying to become the enterprise AI control plane, which is a very boring phrase for a very lucrative ambition.

Microsoft stock is hard to read because the market is not really pricing a company; it is pricing a rotating set of AI anxieties. Dependence on OpenAI. Azure capex. Copilot ROI. Every week, one anxiety gets promoted to consensus.

The stock has become a referendum on whichever AI narrative the market woke up believing. Some days Microsoft is the obvious winner of enterprise AI; other days it is a very expensive bundle of AI-washed legacy software.

The stock will do what stocks do. The strategy is the more interesting read.

For the last three years, Microsoft’s AI strategy has looked almost unfairly simple: attach to OpenAI’s frontier momentum, pour that intelligence through Azure, package it as Copilot, and distribute it through the most valuable enterprise software estate on earth.

It was a brilliant trade. Microsoft got speed, narrative, and product gravity. OpenAI got capital, compute, and enterprise legitimacy.

But the market has moved. The first phase of enterprise AI was about access to intelligence. This phase is about control. Control over context, cost, deployment, and data.

Build 2026 was Microsoft’s answer to that shift. The company is trying to move from being “the best distributor of someone else’s intelligence” to the operating fabric for enterprise AI. That is the through-line across the announcements:

• Windows transitions from operating system to AI runtime. Microsoft announced new Aion on-device models, broader Windows AI APIs, and the Surface RTX Spark Dev Box for developers building local AI applications. The next wave of AI apps will need to decide, task by task, what should run locally versus in the cloud. Local inference can be faster, cheaper, more private, and available offline; cloud inference can handle heavier reasoning and larger models. Microsoft wants Windows to broker that split.
  The move is classic Microsoft: give developers the APIs, give enterprises the management and security layer, give hardware partners a reason to build AI PCs, and make Windows the default deployment surface for everyday intelligence. Think DirectX for AI - a standard layer developers can build against while Microsoft makes the PC relevant again.

• Project Solara bets that agents will need new devices. Microsoft demoed Solara as an Android-based platform for agent-first hardware. They showed a desk device that looks like an enterprise Echo Show and unlocks with facial recognition, and a badge-style wearable with a camera and fingerprint scanner that can capture and transcribe conversations.
  These are reference designs, not products Microsoft is trying to sell directly - which is the point. Microsoft is trying to define the operating layer for a new class of AI endpoints: ambient desk agents, frontline worker badges, retail/healthcare/service devices, and other lightweight form factors where the interface is less “open app, type thing” and more “recognize me, understand the situation, capture context, and help complete the task.” The point is to seed an ecosystem: Microsoft provides the OS/platform, identity/security/management layer, and agent framework; hardware partners can build the actual devices.

  The PC was built around files. The phone was built around apps. Agent-native devices may be built around delegation, ambient context, and interruption management. Microsoft’s bet is that if this new hardware category exists, it should come with Microsoft identity, Microsoft security, Microsoft management, Microsoft agents, and Microsoft enterprise defaults. Very shocking behavior from Microsoft.

• IQ builds enterprise context. By bringing together Foundry, Fabric, and Microsoft 365, Microsoft is creating a unified intelligence layer that continuously understands the organization: its people, data, documents, workflows, meetings, permissions, systems, and institutional residue. Every AI company wants better context. Microsoft already sits on the motherlode: email, calendar, Teams, Office docs, identity, security policies, databases, workflows, and every “final_final_v6_ACTUAL” deck ever produced by modern knowledge work. IQ is the move from storing organizational data to operationalizing organizational context.

• Web IQ extends that context beyond the company’s four walls. Enterprise agents cannot just reason over internal data; they need current external information too. Web IQ is Microsoft’s attempt to provide fast, high-quality, efficient web grounding for agents. They want to own not just the corporate memory layer, but the retrieval and grounding layer that makes agents trustworthy enough to use in production.

• From Copilot to Autopilot. These are autonomous, long-running agents with enterprise-grade compliance - effectively agents designed to persist, monitor, and complete work across systems while staying inside the corporate control plane. That last part is the whole ballgame. The hard part is not getting an agent to take an action. The hard part is making agents act inside a company, with permissions, audit logs, approvals, rollback, cost controls, and a full appreciation for the sacred enterprise distinction between “technically possible” and “please route this through legal.”
  This is why Microsoft has an advantage. Enterprise AI does not only need intelligence. It needs governed intelligence.

• Model Layer Foray. Microsoft does not want to be perceived as merely the world’s most successful OpenAI distribution partner, however lucrative that has been. They need control over the intelligence supply chain.
  The future is not one model to rule them all. It is routing: a frontier model for hard reasoning, a smaller model for cheap repetitive tasks, a local model for private or latency-sensitive work, a tuned model for domain-specific workflows. With 7 new MAI models and tools to tune frontier capabilities, Microsoft is building optionality.

• Frontier Tuning. Generic intelligence is impressive, but differentiated intelligence comes from proprietary context and operational integration. Microsoft is giving companies a path from renting frontier capability to shaping it around their own business systems.

  

• Majorana 2, the long-dated platform option. Majorana 1 proved the physics behind its topological quantum approach; Majorana 2 is about moving from lab proof point to engineering at scale. The appropriate reaction is neither “quantum is here” nor “nice sci-fi slide.” It is: Microsoft is taking its most contrarian research bet and trying to turn it into a manufacturable platform. Not relevant to Copilot ROI this quarter; very relevant if quantum ever becomes an actual computing layer rather than a recurring keynote seasoning.

The stock can continue being mysterious. The strategy is not. Microsoft is trying to become the enterprise AI control plane, which is a very boring phrase for a very lucrative ambition.

The IPO window is reopening and Wall Street is about to become the world's largest AI benchmark.

SpaceX is preparing to go public. Anthropic has confidentially filed its S-1. OpenAI is preparing its own filings. If even some of the rumored valuations hold, public markets will soon be asked to absorb trillions of dollars of new market cap from companies that, until now, have mostly been priced by private-market consensus, strategic necessity, and vibes.

SpaceX is the cleanest and strangest case. It is tempting to put it in the AI bucket because Elon Musk is involved and xAI is somewhere in the family photo, but SpaceX is mostly not a model lab. It is fundamentally a company of rockets, satellites, launch capacity, spectrum, manufacturing, orbital logistics, and scarce physical infrastructure. This is a company that owns bottlenecks.

At a rumored $1.75 trillion valuation, SpaceX would trade at roughly 110-120x estimated 2025 revenue, assuming the $15-16 billion revenue estimates floating around are in the right zip code. On estimated 2025 EBITDA of ~$8 billion, that is about 220x EBITDA. That is, technically speaking, a lot.

The “is SpaceX overvalued?” debate is both obvious and impossible. On current financials, yes. Of course. You do not need a finance degree to conclude that 220x EBITDA contains some optimism. But if it is the infrastructure monopoly for launch, satellite broadband, defense connectivity, orbital logistics, and future space-based industry, the valuation becomes a call option on markets that are still being formed. That does not make the number reasonable, but it makes the argument harder to dismiss.

Having watched many very smart people lose money shorting Tesla over the years, I have developed a reluctant respect for the gap between being right on fundamentals and being right on the trade. With Elon companies, the spreadsheet is often just one input. The other inputs include narrative, scarcity, reflexivity, and a market-wide willingness to underwrite science fiction if the story is told well enough.

OpenAI and Anthropic have a stranger task. They have to justify the valuation of intelligence itself. For years, frontier AI labs have operated like a hybrid of research institution, startup, and state actor. The IPO process forces them to become something much less romantic: reporting companies. That will be uncomfortable.

The order will matter too. Whoever goes first becomes the benchmark everyone else inherits. A strong Anthropic debut creates oxygen for OpenAI. A shaky OpenAI debut makes every later AI IPO answer for sins it did not personally commit. In IPO markets, the first company through the door sets the mood.

Historically, many iconic companies went public relatively early. Investors could buy Amazon in 1997, Google in 2004, Meta in 2012. The public participated in a large portion of the value creation journey. Today’s companies stay private far longer with the most lucrative phase of value creation happening behind closed doors. They no longer go public because they need capital but rather because everyone else wants liquidity.

The public market used to be where companies went to raise money. Increasingly, it is where private-market investors go to get their money back.

The arrival of these public comps will hopefully discipline private markets. AI valuations today are still part math, part myth, and part benchmark theater. Private rounds happen at different times, with different structures, different information rights, and very limited liquidity. Once these companies trade publicly, everything gets repriced relative to them. Suddenly every board meeting has a reference point:
Why is this startup worth xx revenue if Anthropic trades at xx?
Why is this application layer company priced like a model lab?
Why does this “AI infrastructure” business have software multiples but data center margins?
Public markets are very good at ruining private market poetry.

The S-1s may become the most important AI documents of the year. If I am hard to reach the day these filings drop, forgive me. You will know what I am doing. We have all hyper fixated on model benchmarks so far because that’s all we’ve had. The S-1s will contain much more valuable information: customer concentration, revenue mix, enterprise retention, compute commitments, power agreements, capital intensity, stock compensation. The real AI benchmark is how much capital it takes to produce the next dollar of revenue - or even to merely stand still?

The omissions will matter too. Investors will scrutinize not just what appears in the filings, but what is conspicuously absent.
Why won’t they break out inference costs?
Why won’t they separate enterprise from consumer revenue?
Why won’t they disclose customer concentration?
Sometimes the silence in an S-1 is more revealing than the numbers.

The risk sections could be historic. It is not every day that a public market prospectus may need to explain model hallucination, catastrophic misuse, regulatory capture, geopolitical dependency, energy constraints, copyright exposure, biological risk, national security entanglement, and possibly some version of p(doom).

Future historians will read these risk sections the way we now read early internet prospectuses: a mix of prophecy, denial, and lawyers trying to describe a machine no one fully understood.

But the biggest question is still where the value accrues. That question has hovered over AI since ChatGPT launched. Does the value accrue to the model labs? The cloud providers? The chip companies? The data centers? The application layer? Public markets will force a verdict.

If OpenAI and Anthropic list with enormous revenues but structurally mediocre margins, while the infrastructure players keep printing cash, the market may decide that intelligence is valuable but infrastructure captures the economics.

If the labs show strong retention, pricing power, expanding margins, and declining inference costs, the market may underwrite them as the new operating systems of enterprise and consumer life.

Either way, the debate moves from vibes to filings. For three years, Silicon Valley has been asking whether these models can explain the world. Wall Street is about to ask whether these companies can explain themselves.

🎮 TL;DR

• The AI jobs apocalypse is getting downgraded. The new consensus is less “mass white-collar extinction,” more “messy workflow rebundling” because model capability still has to survive procurement, compliance, incentives.

• Tokenmaxxing is the billable endgame of AI theater. Once companies reward usage, employees manufacture usage. Now the invoice is forcing a shift from AI maximalism to AI unit economics.

🌦️ The AI Jobs Forecast Gets Downgraded

The AI jobs weather report has been revised.
Last year: category-five white-collar hurricane.
This year: scattered productivity showers, with a chance of reskilling.

Goldman Sachs CEO's NYT op-ed is the latest entrant in a strange new genre: the AI jobs reassurance memo. Yes, AI will disrupt work. Yes, some roles will be automated. But no, this is not the end of employment. We have adapted before; we will adapt again. Keep calm and continue compounding.

There is a broad tonal whiplash in the market. For the last 18 months, the loudest AI voices made dire prognostications about how AI will wipe out 50% of entry-level white-collar jobs and decimate entire classes of work.

Now the language is softening. Sam Altman recently said he was “delighted to be wrong” about white-collar job destruction. Anthropic’s usage data shows lots of AI usage but most of it looks like augmentation, not replacement.

So what changed?

The generous explanation is that many of the people making the loudest predictions were profoundly removed from actual work.

Frontier AI researchers live in benchmark land, where work looks clean. Input, output, score. Corporate land is a swamp of permissions, politics, procurement, compliance, edge cases, tacit judgment, unspoken preferences, and someone named Brad who “needs to weigh in.” People on the frontlines of AI confused capability with adoption.

The less generous read is that apocalypse was convenient.

When AI was being sold to investors, frontier labs needed to communicate inevitability. The bigger the disruption, the bigger the market. “This will transform work” is a bigger vision than “This will make analysts 20% faster.”

When AI was being sold to enterprises, vendors needed urgency. Fear of being outcompeted did a lot of heavy lifting. Adapt or die.

But now AI is being sold to society, and society is a tougher customer. Workers are anxious. Parents are asking if their kids’ first jobs will exist. Politicians are noticing that “we automated your career” is not a winning bumper sticker.

So the pitch is changing. Less: “mass unemployment.” More: “productivity renaissance.”

Both stories are self-interested, both contain some truth, neither should be treated as a forecast.

FWIW I have long been in the second camp. Jobs are bundles of tasks, and most tasks are embedded in messy systems of judgment, trust, incentives, and institutional memory. Just because a model can perform a task does not mean a company can safely, cheaply, or culturally reorganize around that capability. My friend Jevon will tell you that Excel didn't eliminate analysts and email didn't free up time.

AI will not replace “lawyers” or “bankers” or “software engineers” in some clean Hollywood montage. It will replace slices of their workflows, change the leverage ratios of teams, raise the expected output per employee, and reshape who gets hired, trained, promoted, and paid. That is less cinematic than a jobs apocalypse, also more plausible.

🧾The Metered-Compute Hangover

Tokenmaxxing was inevitable.

The moment companies started monitoring, ranking, and rewarding on AI usage, people did exactly what any incentive system teaches them to do: they managed the metric.

This week felt like the first crack in the fever dream. After 18 months of AI enthusiasm operating mostly in exclamation points, a few people finally started asking the necessary questions: What does this cost? What did it improve? Is “usage” a proxy for value, or just a participation trophy?

Until recently, asking those questions made you sound like an AI skeptic - the person standing at the back of the unlimited-intelligence parade, muttering about gross margin. But now that the bills are landing on real P&Ls, the question is becoming less taboo.

Axios is writing about AI sticker shock. Uber’s COO says spend is getting harder to justify. Microsoft is pulling back Claude Code licenses. Amazon killed an internal AI leaderboard after employees gamed it.

Hopefully, this marks the beginning of a saner phase of enterprise AI. A few thoughts:

1. Tokenmaxxing is an obvious incentives problem
It is a truth universally acknowledged that people game metrics. Most companies have been running an AI activity program, not an AI productivity program. And activity is easy to manufacture.
In the SaaS era, performative activity mostly wasted time. In the AI era, it burns GPU-backed marginal cost.

2. Companies imported SaaS instincts into a metered-compute world
Treating AI like traditional SaaS is a mistake. With SaaS, adoption was often a reasonable proxy for value. More seats usually meant the tool had been embedded into workflow.
But frontier AI is metered compute. Marginal cost matters. The right operating model looks like cloud FinOps: budgets, caps, observability, unit economics, and workload governance.

3. “AI mandate” made employees less honest about AI’s actual utility
When leaders signal that AI usage is evidence of being modern, employees rationally use AI even when it adds friction. That pollutes the feedback loop. Mgmt stops learning where AI is genuinely useful because employees are incentivized to manufacture usage rather than reveal where the tool is weak, unnecessary, or net-negative.

4. Winners will know where not to use AI
Strategic AI maturity is not maximalism, but selectivity: knowing which tasks deserve frontier models, which can use cheaper models, which should be cached, which should be automated, and which should remain human because the cost or error profile is wrong. “AI everywhere” is not a strategy.

5. A shot in the arm for Open Source and RLaaS
Frontier models are powerful, expensive black-box APIs. For many enterprise workflows, the better long-term answer may be to post-train open source models on narrow, high-value use cases: models that are good enough for the task, cheaper to run, and increasingly differentiated by proprietary data.

Hopefully, this is the end of AI theater and the beginning of AI unit economics.

Tokenmaxxing was inevitable.

The moment companies started monitoring, ranking, and rewarding people based on AI usage, people did exactly what any incentive system teaches them to do: they managed the metric. Show me how you’re keeping score and I’ll show you how people will game the game.

This week felt like the first crack in the fever dream. After 18 months of AI enthusiasm operating mostly in exclamation points, a few people finally started asking the impolite but necessary questions: What does this cost? What did it improve? Is “usage” actually a proxy for value, or just a very expensive participation trophy?

Until recently, asking those questions made you sound like an AI skeptic - the person standing at the back of the unlimited-intelligence parade, muttering about gross margin. But now that the bills are landing on real P&Ls, the question is becoming less taboo.

The vibe is subtly shifting. Axios is writing about AI sticker shock. Uber’s COO says spending is getting harder to justify. Microsoft is reportedly pulling back on Claude Code licenses. Amazon reportedly killed an internal AI leaderboard after employees gamed it by doing pointless AI things to earn AI points.

Hopefully, that marks the beginning of a saner phase of enterprise AI: less theater, more unit economics. A few thoughts:

1. Tokenmaxxing is an obvious incentives problem
It is a truth universally acknowledged that people game metrics. Most companies have been running an AI activity program, not an AI productivity program. And activity is easy to manufacture. In the SaaS era, performative activity mostly wasted time. In the AI era, performative activity burns GPU-backed marginal cost.

2. Enterprises confused consumption with transformation
“AI usage” became the new “digital transformation” dashboard metric: easy to report upward, hard to tie to customer value. But tokens are an input cost, not a measure of output. A factory bragging about using more electricity would sound insane unless output, yield, or throughput improved. AI should be measured the same way.

3. Companies imported SaaS instincts into a metered-compute world
Treating AI like traditional SaaS is a strategic mistake. In the SaaS era, adoption was often a reasonable proxy for value. More seats using Salesforce, Slack, Figma, or Workday usually meant the tool had been embedded into workflow. But frontier AI is metered compute. The marginal cost curve matters. The right operating model looks less like software enablement and more like cloud FinOps: budgets, caps, observability, unit economics, and workload governance.

4. “AI mandate” made employees less honest about AI’s actual utility
When leaders signal that AI usage is evidence of being modern, employees rationally use AI even when it adds friction. They use AI because not using AI starts to feel politically unsafe. That pollutes the feedback loop. Management stops learning where AI is genuinely useful because employees are incentivized to manufacture usage rather than reveal where the tool is weak, unnecessary, or net-negative.

5. Winners will know where not to use AI

Strategic AI maturity is not maximalism. It is selectivity. The durable advantage will come from routing: which tasks deserve frontier models, which can use cheaper models, which should be cached, which should be automated, and which should remain human because the cost or error profile is wrong. “AI everywhere” is not a strategy.

6. A shot in the arm for Open Source and RLaaS

I’ve been following the RL-as-a-service market for a while, and the core thesis is becoming more relevant. Frontier models are powerful, expensive, impersonal black-box APIs. For many enterprise workflows, the better long-term answer may be to post-train open source models on narrow, high-value use cases: models that are good enough for the task, cheaper to run, and increasingly differentiated by proprietary data and workflow feedback.

That market has not fully taken off yet because the frontier models have been improving so quickly that many buyers have been reluctant to optimize for cost and risk losing performance. But if cost scrutiny rises at the same time open source becomes “good enough” for more enterprise workloads, the calculus changes.

For the last two years, the dominant question has been: which model is smartest? As budgets tighten, the more interesting question becomes: which model is smart enough?

Hopefully, this is the end of AI theater and the beginning of AI unit economics. For the last 18 months, the corporate question was: “Are our people using AI?” The next question will be: “What work unit got cheaper, faster, better, or newly possible?” That shift is more important than the cost panic itself.

The AI jobs weather report has been revised.

Last year: category-five white-collar hurricane.
This year: scattered productivity showers, with a chance of reskilling.

David Solomon’s NYT op-ed is the latest entrant in what has become a strange genre: the AI jobs reassurance memo. It has all the right institutional notes. Yes, AI will disrupt work. Yes, some roles will be automated. Yes, entry-level white-collar jobs are already feeling pressure. But no, this is not the end of employment. America has adapted before; America will adapt again. Please remain calm and continue compounding.

What is interesting is not that a bank CEO is optimistic about the labor market. That is practically in the job description. What is interesting is the broader tonal whiplash.

For the last 18 months, many of the loudest AI voices made dire prognostications. Dario Amodei warned that AI could wipe out half of entry-level white-collar jobs and send unemployment sharply higher. Sam Altman warned for years that AI would eliminate entire classes of work. The message was, essentially: this technology is so powerful it may break the economy.

Now the language is softening. Altman recently said he was “delighted to be wrong” about the pace of white-collar job destruction and has started taking shots at competitors: “I don’t think we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about.” No points for guessing who is being subtweeted.

Anthropic’s own usage data shows a more ambiguous reality: lots of AI usage, yes, but much of it looks like augmentation rather than wholesale replacement.

So what changed?

The generous explanation is that many of the people making the loudest predictions were profoundly removed from actual work.

Frontier AI researchers live in benchmark land, where work looks clean. Input, output, score. And so they confused capability with adoption.

In corporate land, work is a swamp of permissions, politics, procurement, compliance, edge cases, legacy systems, weird customer exceptions, tacit judgment, unspoken preferences, and someone named Brad who “needs to weigh in.”

That gap explains a lot of the whiplash. The demo ends at “here is the answer.” The job begins at “will Legal accept this, can Finance approve it, does Sales trust it, will the client hate it, and who is on the hook if it’s wrong?”

A lot of AI jobs discourse was comically overclean. The early forecasts treated work as a neat stack of tasks waiting to be vaporized. In reality, companies are not APIs, but rather organisms with immune systems.

The less generous read is that apocalypse was convenient.

When AI was being sold to investors, frontier labs needed to communicate inevitability. The bigger the disruption, the bigger the market. “This will transform work” is a compelling vision.

When AI was being sold to enterprises, vendors needed to communicate urgency. Fear of being outcompeted did a lot of heavy lifting. Every boardroom got the same memo: your competitors are doing this, adapt or die.

But now AI is being sold to society, and society is a tougher customer. Workers are anxious. Parents are asking whether their kids’ first jobs will exist. Politicians are noticing that “we automated your career ladder” is not a winning bumper sticker.

So the pitch is changing.

Less: “AI will take your job.”
More: “AI will remove the tedious parts of your job.”
Less: “mass unemployment.”
More: “productivity renaissance.”
Less: “the labor market is about to break.”
More: “please enjoy this thoughtful op-ed about American dynamism.”

And look, I mostly agree with this version of the story. I was on the “tasks, not jobs” train before it was cool. Jobs are bundles of tasks, and most tasks are embedded in messy systems of judgment, trust, compliance, relationships, incentives, and institutional memory. Just because a model can perform a task does not mean a company can safely, cheaply, or culturally reorganize around that capability. The ATM did not eliminate bank branches overnight. Excel did not eliminate analysts. Email did not give us more free time.

In the meantime, companies are already doing AI-flavored layoffs. Executives have discovered that AI is a very useful word to put next to efficiency. It makes cost-cutting sound innovative. It makes austerity sound strategic. It turns a layoff from “we overhired” into “we are rearchitecting the business for an AI-native future.” Same spreadsheet, better press release, even better stock reaction.

There is no denying that AI lets small teams move faster and punch above their weight. But inside an existing company, you cannot just pour automation into the org chart and call it transformation. Deployment has to be careful, tasteful, and incentive-aware. Otherwise AI does not fix the machine; it short-circuits it.

The frontier labs overshot on apocalypse because apocalypse sold the magnitude of the opportunity. They are now emphasizing adaptation because it makes adoption smoother. Both stories are self-interested. Both contain truth. Neither should be treated as a forecast.

AI will not replace “lawyers” or “bankers” or “software engineers” in some clean Hollywood montage. It will replace slices of their workflows, change the leverage ratios of teams, raise the expected output per employee, and reshape who gets hired, trained, promoted, and paid. That is less cinematic than a jobs apocalypse. It is also more plausible - and just as disruptive.

🎮 TL;DR

• OpenAI wants to become the OS for adult life. Not just Mint with better UX, ChatGPT Finance is a bet that behavioral context is the real consumer finance product.

• The Monet prank exposed AI’s deepest adoption problem: humans. Many people believe they are evaluating AI rationally when they are actually evaluating it morally, socially, and psychologically.

• Google still has the best chessboard and the worst stagecraft. I/O showed a company with enormous AI leverage and an almost heroic inability to package attention.

• NVIDIA beat, raised, and still somehow disappointed. The next leg may belong to CPUs, edge systems, sovereign clouds, factories, and billions of software agents demanding infrastructure.

• AI’s ceiling is rising while the floor remains sticky. OpenAI can help crack an Erdős-adjacent conjecture and still fail at Starbucks milk inventory.

💸 Plaid, Privacy, and the Price of Convenience

When Mint shut down two years ago, I felt strangely unmoored. I had used it for years. It was clunky, the categorization was often wrong, and the UI had all the elegance of a tax form but it had all my accounts, spending patterns, and financial ghosts of past lives. For a long time, it was the best personal finance product we had, even though it rarely felt very personal.

Now OpenAI is entering the category with ChatGPT Finance, which lets users connect accounts through Plaid and ask questions. At first glance, this sounds like Mint with a chatbot. But the more interesting version is closer to a financial OS. Not “here is your budget”, more like:
“Your spending patterns suggest elevated stress over the last month”
“You can afford this house, but it will limit flexibility if you want children soon.”

That is a very different product. It moves from bookkeeping to behavioral interpretation. And honestly, there is a reason this could work. People aren't bad at personal finance because they lack spreadsheets, but because money is emotional, identity-laden, and cognitively exhausting. Most apps just show you the mess. An intelligent, contextual system can help you reason through it.

Modern life has become absurdly operational: healthcare, mortgages, childcare, insurance, taxes... Being an adult is a full-time back-office job with terrible software. OpenAI’s broader ambition is increasingly clear: become the operating system for that chaos. First came Health, now comes Finance.

This contrasts with Anthropic. Claude still feels like a work product while ChatGPT increasingly wants to know your lab results, spending habits, mortgage anxiety, and why you keep buying expensive skincare after bad meetings.

This will either be the most useful consumer software ever built or the beginning of a privacy law exam question. Probably both.

The trust problem is obvious. Models are becoming more agentic before they are reliably safe. Over the last year, we've heard stories of AI agents deleting files, wiping inboxes, hallucinating facts. They produce outputs that sound plausible, contextual, and emotionally calibrated even when they are wrong.

OpenAI clearly understands this risk, which is why ChatGPT Finance is currently read-only. But read-only does not eliminate the problem. The model does not need direct access to cause harm if humans act on bad recommendations.

The privacy question is just as thorny. Connecting Mint to your bank account felt transactional. Connecting your finances, medical history, conversations, and long-term memory to a generalized intelligence system feels different.

At some point, these companies start building behavioral models of human lives. That is the real opportunity and also the real risk.

The upside is obvious: A context-aware assistant that actually understands your life. The downside is also obvious: you handed one company the richest behavioral map of your life. Would you pay this price for convenience?

🖼️ Waterlilies and Witch Hunts

One of the most revealing AI experiments this week involved no AI at all. An X user posted a Monet painting, claimed it was AI-generated, and asked people to critique it. The mob showed up with pitchforks and art degrees.

People called it soulless. Flat. Derivative. Technically weak. They said it captured the “look” of art without understanding it.

Except it was an original Monet. Which makes this one of the funniest and most illuminating social experiments in the AI era.

Once people heard “AI-generated,” many stopped looking at the painting and started looking for evidence to justify a conclusion they had already emotionally reached.

This is the inverse of the wine-label effect. Tell people a $12 wine costs $300 and suddenly they detect “notes of oak and tobacco.” Tell them a Monet is AI-generated and suddenly they detect “sloppy depth hierarchy.” The human brain is less objective critic than post-hoc narrative engine.

“AI-generated” is quickly becoming the internet’s newest slur: a way to call something fake, lazy, soulless, or illegitimate without having to argue the point. And that accusation is already changing how people create.

Writers are going to absurd lengths to prove they did not use AI. They are adding typos, swapping polished sentences for run-ons, using aggressively casual language, and replacing em dashes with double hyphens, like tiny artisanal protest signs. People are making their writing worse to make it seem more human. Welcome to the era of performative imperfection.

Most AI backlash is not about the output at all. The emotional reaction changes once people believe a machine was involved, even when the artifact is identical.

And to be fair, some of that reaction is rational. Humans are not just evaluating pixels. We are evaluating effort, scarcity, intentionality, and cultural meaning. A Monet is not only paint on canvas. It is also the story of a human standing in a garden in 1890 trying to freeze light before it disappeared.

For most of modern history, effort has been entangled with how we assign value. The story behind creation mattered. We do not value a handwritten love letter and a mass-generated email equally even if the words are identical. AI destabilizes that relationship.

If people cannot reliably distinguish human work from machine work at the artifact level, they will anchor harder to provenance. “Human-made” vs “AI-generated” becomes part of the perception.

Many people believe they are evaluating AI rationally when they are actually evaluating it morally, socially, and psychologically.

That is why AI adoption will be messier than the capability curves imply. Humans are not rational evaluators of output. We are status-seeking primates who attach moral meaning to tools.

The irony is that the Impressionists were once mocked as frauds corrupting “real art.” Now we use them to prove humans still have souls. Monet would have appreciated the joke.

🎆 Google Sets Off Every Firework at Once

At I/O, Google did what Google does best: brought the fireworks, then set them all off at once.

OpenAI and Anthropic understand something Google does not: attention is a product surface. They drip-feed the ecosystem: model launch here, product update there. Each announcement gets its little 48-hr tour through Twitter, newsletters, group chats, and breathless “this changes everything” threads.

Google meanwhile takes twenty announcements that each deserve oxygen and throws them into one keynote-shaped blender.

That is the packaging critique. Now let’s wade through the substance:

• Gemini 3.5 Flash. A faster, more agentic model that delivers frontier performance at 4x the speed. Best for coding, long-running tasks, and everyday product integration. This is now the default model across Google surfaces.

• Gemini Omni. The “anything from any input” model, starting with video. It uses images, audio, video, and text as inputs, generates videos grounded in Gemini’s knowledge, and lets users edit conversationally. Get ready to fall for even more convincing AI videos online.

• Search Agents. Google is embedding agents into Search. When you ask a question, it generates richer answers, explanatory visuals or short video clips. So far, Google monetized intent by organizing the internet and sending users elsewhere. Now it wants to resolve that intent itself.

• Gemini Spark. “24/7 personal AI agent” - the answer to Openclaw, also where Google’s advantage becomes obvious: it has your inbox, calendar, maps, docs, browser, phone OS, search history, YouTube, photos, and shopping graph. In agent land, context is the whole game and Google is sitting on the fossil record of your life.

• Intelligent Eyewear. Google is back in the arena it entered too early with Glass. This time, it arrives into a market softened up by Meta Ray-Bans, AI voice interfaces, with partners like Samsung, Warby Parker, Gentle Monster, and Xreal. The glasses can support voice commands, directions, messages, photo capture, translation, and XR experiences.

• Universal Cart. An intelligent shopping layer that works across Google surfaces. You can add products as you browse, and Google’s agent can alert you to price changes, flag newer versions or color options, answer questions through Gemini, find better deals, and eventually help purchase items across different retailers through Google’s payment system.

• Gemini for Science. Tools aimed at accelerating scientific work in drug discovery and research with the goal of “solving all disease.” One minute the agent is helping you buy shoes. The next minute it is curing death. Silicon Valley does love range.

Demis closed the keynote saying we are standing at the “foothills of the singularity.” As someone with a weakness for literary phrasing, I enjoyed it. As someone watching AI companies try to win public trust, I would gently suggest keeping executives away from vague, ominous proclamations about the future of civilization.

📈 NVIDIA Beats, Raises, and Somehow Still Disappoints

Two things remain reliably true in this unpredictable world: the sun rises in the east, and every quarter Nvidia beats earnings before somehow trading down.

The company delivered another absurd beat-and-raise. Revenue was $82B, +85% y/y and 20% q/q. Data Center revenue was $75B, +92% y/y. Q2 guide came in at $91B ±2%, despite assuming zero China Data Center compute revenue. Mgmt summary was concise: “Demand has gone parabolic.” The market’s response: cute.

But the stock move is the least interesting part of the story:

• Nvidia’s new reporting structure is a map of the AI market.
  Nvidia split the business into Data Center and Edge Computing, with Data Center further divided into Hyperscale and ACIE: AI cloud, industrial, and enterprise. Hyperscale is the familiar world of public cloud. ACIE is the messier, more interesting frontier: AI-native clouds, sovereign AI, enterprise AI factories, industrial deployments, and on-prem infra.
  Today the split is roughly 50/50: Hyperscale at $38B, ACIE at $37B. But Jensen thinks ACIE will outgrow hyperscale over time. The market has been trained to watch hyperscaler capex. Nvidia is telling us to widen the aperture.

• AI demand is becoming economically self-reinforcing.
  Jensen’s core message was simple: “Demand has gone parabolic. The reason is simple. Agentic AI has arrived. AI can now do productive and valuable work. Tokens are now profitable, so model makers are in a race to produce more.”

• CPUs are back.
  Nvidia framed Vera, its purpose-built CPU for agentic AI, as a new growth vector. Mgmt has visibility into ~$20B of CPU revenue this year, opening a $200B TAM that Nvidia historically has not addressed.
  In the first phase of AI, the market obsessed over GPUs. In the agentic phase, workloads become more system-like. An agent is a harness: orchestration, memory management, tool use, browser activity, code execution, I/O, security, and sub-agent coordination. All this runs on CPUs. If the world moves from a billion human users to billions of software agents, CPU demand explodes.

• Physical AI age is coming.
  Nvidia’s current business is still overwhelmingly about data center AI, but mgmt continues to point toward the next frontier: physical AI. Revenue surpassed $9B over the last 12 months. The category includes robotics, autonomous vehicles, industrial automation, medical devices, AI-RAN base stations, factories, and edge systems. This is Nvidia’s long-duration bet on AI moving from screens into the physical world.

The most revealing comment on the call may have been Jensen’s aside that Anthropic and OpenAI are “growing within one month what some SaaS companies took a decade to grow.” That sentence explains the whole AI trade.

Most software companies scale with marginal cost approaching zero. AI companies scale with marginal cost approaching “please build another gigawatt-scale data center.” Great outcome for one Jensen Huang.

🧪 The Ceiling Is Rising. The Floor Is Still Sticky.

OpenAI disproved a central conjecture in discrete geometry, connected to a problem Paul Erdős posed in 1946. Mathematicians reviewed the proof, calling it “a milestone in AI mathematics.” This is a big deal.

You cannot look at a model overturning a long-standing conjecture and conclude that it is trivial or overhyped.

But you can believe the technology is profound and still think much of the implementation layer is a mess.

This is the uncomfortable middle: the tech is profound, much of the implementation is mediocre. The next several years will be defined by this gap.

Take Starbucks. They rolled out an AI-powered inventory tool in 2025, promising faster stock checks and better product availability. 9 months in, they are scrapping the system after persistent inaccuracies, including milk-type mixups.

That does not prove AI is useless. It proves operations are hard. Real-world deployment involves edge cases, employees, incentives, training, accountability, and the quiet rage of the person who has to fix the machine’s mistakes while also doing their actual job. The demo usually ends before this part.

You see the same in software development. “Anyone can code” does not mean anyone can build software. The difference shows up when the prototype becomes production and the company discovers that “AI-generated code” is not the same thing as maintainable software.

This is where bad metrics become dangerous. If companies make “% of code written by AI” the goal, people will optimize for that goal. Software teams spent years learning that more code is not inherently good. Now we are rediscovering that lesson with GPU assistance.

The same stat-padding is happening at company level. Open Instagram and search for an account. You'll get AI summaries jammed into what used to be a simple action. This is the new growth hack: put AI in the path of an existing behavior, then count the collision as adoption.

This is not unique to Meta. Companies are subsidizing adoption, discounting core products if customers try the AI module, and then reporting usage that sounds impressive on earnings calls.

A lot of AI metrics right now are like adjusted EBITDA. Not necessarily meaningless, but you do want to read the footnotes.

This is the heart of the adoption problem. Capability ≠ productivity. Tech has to pass through incentives, politics, and the competence level of the people implementing it. This is why the most important AI work inside enterprises is training leadership.

Most people setting AI mandates do not use these tools themselves. They may have asked ChatGPT to summarize a board deck once but often lack the hands-on fluency. When leaders cannot tell the difference between real AI leverage and AI theater, the organization learns to optimize for theater. The grifters get promoted. The builders get cleanup duty.

OpenAI’s Erdos result tells us the ceiling is rising. Enterprise implementation tells us the floor is still sticky. Mind the gap.

OpenAI just disproved a central conjecture in discrete geometry, connected to a problem Paul Erdős posed in 1946.

The original problem is deceptively simple: given n points in a plane, how many pairs can be exactly one unit apart? For decades, the prevailing intuition was that the best arrangements would look roughly like square grids. OpenAI’s model found an infinite family of constructions that beat that intuition, using ideas from algebraic number theory to make progress on a problem that had resisted human mathematicians for nearly 80 years. This was cross-domain reasoning, not a brute-force computation, a Lean formalization, or a known proof recovered from training data. Outside mathematicians reviewed the proof, and Fields Medalist Timothy Gowers called it “a milestone in AI mathematics.”

So yes, this is a big deal.

At the high end, models are moving from autocomplete to discovery. They are proving things, designing experiments, writing software, and beginning to perturb the frontier of knowledge itself. You cannot look at a model helping overturn a long-standing mathematical conjecture and conclude that this technology is fake, trivial, or overhyped.

And yet.

You can believe the technology is profound and still think much of the implementation layer is a mess.

That distinction is where a lot of the current AI debate goes wrong. The conversation keeps collapsing into two bad camps. On one side: AI is a bubble, a scam, a spreadsheet hallucination wrapped in NVIDIA capex. On the other: every deployment is progress, every usage number is adoption, and every chatbot shoved into a product surface is proof of inevitable productivity gains.

Both views are lazy.

This is the uncomfortable middle of AI: the technology is profound but much of the implementation is mediocre. Both facts are true, and much of the current debate is just people choosing whichever half flatters their priors or furthers their agenda.

The frontier is moving. The rollout is messy. The next several years of AI will be defined by the gap between those two facts.

Take Starbucks. The company rolled out an AI-powered inventory counting tool across more than 11,000 North American company-owned stores in 2025, with the promise of faster stock checks and better product availability. Nine months later, they are scrapping the system after persistent inaccuracies, including problems distinguishing similar milk types. The company is going back to manual counting and daily replenishment.

That does not prove AI is useless. It proves operations are hard.

A model can be impressive in a benchmark and still fail in a stockroom. Real-world deployment involves edge cases, employee workflows, incentives, training, accountability, and the quiet rage of the person who now has to fix the machine’s mistakes while also doing their actual job. The demo usually ends before this part.

You see the same problem in software development. “Anyone can code” does not mean anyone can build software. A senior engineer with taste can use AI like a power tool. A novice without judgment can use it like a leaf blower in a china shop. The difference shows up later, when the prototype becomes production and the company discovers that “AI-generated code” is not the same thing as maintainable software.

This is where bad metrics become dangerous. If companies make “percentage of code written by AI” the goal, people will optimize for that goal. Software teams spent the last twenty years learning that more code is not inherently good. Now some companies are rediscovering that lesson with GPU assistance.

The problem is not AI coding. The problem is confusing code volume with engineering output.

The same stat-padding is happening at company level. Open Instagram and try to search for an account. Increasingly, you get AI-generated summaries and AI-mediated discovery surfaces jammed into what used to be a simple navigational action. This is the new growth hack: put AI in the path of an existing behavior, then count the collision as adoption.

This is not unique to Meta. Across software, companies are routing existing workflows through AI features, bundling AI into existing contracts, subsidizing adoption, discounting core products if customers try the AI module, and then reporting usage numbers that sound impressive on earnings calls. Some of that usage is real. Some of it is the unavoidable result of putting the AI feature directly in the path of something the user was already doing.

When a public company says millions of people are using its AI product, the useful question is what kind of usage it represents. Did the user choose it? Did it make the workflow better? Did it save time, improve quality, reduce cost, increase retention, or change behavior? Or did the user merely pass through an AI surface because the company placed it there and then counted the encounter as engagement?

A lot of AI metrics right now are like adjusted EBITDA. Not necessarily meaningless, but you do want to read the footnotes.

This is the heart of the adoption problem. Capability does not automatically become productivity. Technology has to pass through incentives, politics, procurement, org charts, and the competence level of the people implementing it.

This is why the most important AI work inside large enterprises is not buying new tools or laying off employees, its training senior leadership.

Many of the people setting AI mandates do not use these tools deeply themselves. They have seen the demos. They have read the strategy decks. They may have asked ChatGPT to summarize a board memo and found the experience spiritually moving. But they often lack the hands-on fluency to distinguish a useful workflow from a clever demo, real productivity from metric theater, and thoughtful implementation from an internal science fair.

When leaders cannot tell the difference between real AI leverage and AI theater, the organization quickly learns to optimize for theater. The grifters get promoted. The builders get cleanup duty.

The firms that benefit most from AI will not be the ones with the loudest mandates or the highest internal usage numbers. They will be the ones that redesign workflows, train people properly, set real standards, measure actual output, and have leaders who understand the tools well enough to ask non-stupid questions.

OpenAI’s geometry result tells us the ceiling is rising. Enterprise implementation tells us the floor is still sticky.

The next phase of AI will be defined by the gap between those two realities. At the frontier, models are beginning to surprise mathematicians. In the enterprise, many companies are still trying to figure out whether their AI tool can correctly count oat milk.

That is the market right now. The technology is extraordinary. The implementation is still human.

Two things remain reliably true in this unpredictable world: the sun rises in the east, and every quarter Nvidia beats earnings before somehow trading down.

The company delivered another absurd beat-and-raise. Revenue was $81.6B, up 85% year over year and 20% sequentially. Data Center revenue was $75.2B, up 92% year over year. Q2 guidance came in at $91B ±2%, despite assuming zero China Data Center compute revenue.

Jensen Huang’s summary was concise: “Demand has gone parabolic.” The market’s response: cute.

This is what happens when a company conditions the market to expect miracles. Eventually, walking on water is priced in. The selloff is about altitude, not weakness.

But the stock move is the least interesting part of the story. Here are the takeaways from the earnings call:

• Nvidia’s new reporting structure is a map of the AI market. Nvidia is now splitting the business into Data Center and Edge Computing, with Data Center further divided into Hyperscale and ACIE: AI cloud, industrial, and enterprise. Hyperscale is the familiar world of public cloud and the largest internet platforms. ACIE is the messier, more interesting frontier: AI-native clouds, sovereign AI, enterprise AI factories, industrial deployments, and on-prem infrastructure.
  Today the split is roughly 50/50: Hyperscale at $38B, ACIE at $37B. But ACIE grew faster sequentially, and Jensen suggested it could outgrow hyperscale over time. The AI infra market is shaping up to be more distributed, sovereign, verticalized, and operationally complex than the cloud market, where demand consolidated around a handful of hyperscalers.

  Nvidia’s sovereign AI business already grew more than 80% year over year, with deployments across nearly 40 countries. The number of partner data centers exceeding 10MW has nearly doubled in one year, now above 80 sites. The market has been trained to watch hyperscaler capex. Nvidia is telling us to widen the aperture.
  

• AI demand is becoming economically self-reinforcing. Jensen’s core message was simple: compute capacity = revenue = profits. He emphasized: “Demand has gone parabolic. The reason is simple. Agentic AI has arrived. AI can now do productive and valuable work. Tokens are now profitable, so model makers are in a race to produce more.”

  For two years, the AI compute buildout was a bet on future monetization. The capex was real; the revenue was speculative. That phase appears to be ending.
  

• CPUs are back. One of the most interesting parts of the call was the Vera CPU discussion. Nvidia framed Vera, its purpose-built CPU for agentic AI, as a major new growth vector. Management said it has visibility into nearly $20B of standalone CPU revenue this year, opening what it characterizes as a $200B TAM that Nvidia historically has not addressed.
  In the first phase of AI, the market obsessed over GPUs because the bottleneck was model training and inference. In the agentic phase, workloads become more system-like. An agent is a harness: orchestration, memory management, tool use, browser activity, code execution, I/O, security, and sub-agent coordination. Jensen’s framing: the thinking happens on GPUs, but much of the orchestration runs on CPUs.
  That has big implications. If the world moves from a billion human users to billions of software agents, each agent effectively needs some compute environment in which to act. The CPU demand curve becomes tied not just to cloud workloads, but to the proliferation of agents themselves.
  

• Nvidia increases inference market share. For several years, the bear case has been that Nvidia would dominate training but lose share in inference as workloads shifted to cheaper, specialized silicon. The call pushed directly against that narrative.
  Jensen said Nvidia is gaining share in inference, especially as it deepens partnerships with frontier model companies. The most notable example was Anthropic. Nvidia’s coverage of Anthropic had been “largely zero until recently” but it is now helping expand Anthropic compute capacity across AWS, Azure, CoreWeave, and others. That is incremental share gain in one of the most important frontier AI accounts.
  

• Physical AI age is coming. Nvidia’s current business is still overwhelmingly about data center AI, but management continues to point toward the next frontier: physical AI. Physical AI revenue surpassed $9B over the trailing twelve months. The category includes robotics, autonomous vehicles, industrial automation, medical devices, AI-RAN base stations, factories, and edge systems. This is Nvidia’s long-duration bet on AI moving from screens into the physical world.
  The industry has spent the last few years automating knowledge work: text, code, images, research, support, analytics. The next phase is AI that acts: robots in warehouses, autonomous fleets, factory systems, telecom infrastructure, surgical devices, industrial inspection, and embodied agents.

The most revealing comment on the call may have been Jensen’s aside that Anthropic and OpenAI are “growing within one month what some SaaS companies took a decade to grow.” That sentence explains the whole AI trade.

The model companies are scaling at a speed software markets have never really seen. Their growth pulls in compute. Compute pulls in power, networking, memory, CPUs, data centers, and financing. And the wheel turns.

Most software companies scale with marginal cost approaching zero. AI companies scale with marginal cost approaching “please build another gigawatt-scale data center.” That is the magic and the problem.

The AI boom is a software revolution with an industrial cost structure and Nvidia is the arms dealer, the toll collector, and increasingly the factory architect.

Google I/O Day 1 was yesterday, and Google did what Google does best: brought the fireworks, then set them all off at once.

OpenAI and Anthropic understand something Google does not seem to get: attention is a product surface. They drip-feed the ecosystem. A model launch here. A product update there. A safety report. A coding agent. A new plan. A developer feature. Each announcement gets its little 48-hour tour of duty through Twitter, newsletters, group chats, investor memos, and breathless “this changes everything” threads.

Google meanwhile takes twenty announcements that each deserve oxygen, throws them into one keynote-shaped blender, and wonders why the market only remembers three. The company needs to learn how to tell one story at a time.

That is the packaging critique. Now let’s wade through the substance together:

• Gemini 3.5 Flash. Google’s faster, more agentic model, optimized for coding, long-running tasks, and everyday product integration. It delivers frontier performance 4x faster than other frontier models, which matters because agents are only useful if they can operate at the speed of human impatience. This balance of speed and performance makes 3.5 Flash ideal for tackling long-horizon agentic tasks. This is now the default model across Google’s surfaces globally.

  

• Gemini Omni. This is Google’s “anything from any input” model, starting with video. It can combine images, audio, video, and text as inputs, generate high-quality videos grounded in Gemini’s real-world knowledge, and let users edit those videos conversationally. Over time, it is expected to support additional output modalities like image and audio.
  The implication is that media creation is becoming less like using a tool and more like directing an intelligent system. Text, image, sound, and video begin to collapse into one creative interface. That is powerful for creators, useful for businesses, and deeply uncomfortable for anyone looking to solve provenance. Get ready to fall for even more convincing AI videos online.

  

• Search Agents. Google is embedding agents into Search. When you ask something like “What is a black hole?”, Search can respond with richer contextual answers and even generate explanatory visuals or short video clips. Google is also adding Generative UI, which creates custom layouts for different kinds of results, whether videos, images, news articles, or other formats. This is a big shift. For 25 years, Google monetized intent by organizing the internet and sending users elsewhere. Now it wants to resolve more of that intent itself. Great for users. Existentially awkward for everyone whose business depends on being the destination.

  

• Gemini Spark. Google’s “24/7 personal AI agent,” which uses context from Google apps like Gmail, Maps, and Calendar to help with planning, organization, and automation. This is their clearest answer to agent startups and wrappers like OpenClaw and also where Google’s advantage becomes obvious: it has your inbox, calendar, maps, docs, browser, phone OS, search history, YouTube, photos, and shopping graph. In agent land, context is the whole game and Google is sitting on the fossil record of your life.

• Intelligent Eyewear. Google is back in the arena it entered too early with Glass. This time, it is arriving into a market softened up by Meta Ray-Bans, AI voice interfaces, and a more normal-looking hardware strategy through partners like Samsung, Warby Parker, Gentle Monster, and Xreal. The glasses can support voice commands, directions, messages, photo capture, translation, and XR experiences. The form factor still has all the old privacy questions. But culturally, the ground has shifted.

• Universal Cart. An intelligent shopping layer that works across Google surfaces. You can add products as you browse, and Google’s agent can keep your wish list organized, alert you to price changes, flag newer versions or color options, answer questions through Gemini, find better deals, and eventually help purchase items across different retailers through Google’s payment system. This is Google trying to own the moment between consumer intent and transaction. If agents become shoppers, the cart becomes strategic infrastructure. Merchants will like the conversion. They may enjoy the dependency less.

• Gemini for Science. Demis Hassabis also highlighted Gemini for Science, tools aimed at accelerating scientific work in areas like drug discovery and research. This is the part of the keynote where Google’s AI ambition becomes least consumer and most civilizational. According to Hassabis, with tools like these, Google hopes to “reimagine drug discovery with the goal of one day solving all disease.” Hard to argue with the ambition. Also hard not to notice the tonal whiplash. One minute the agent is helping you buy shoes. The next minute it is solving all disease. Silicon Valley does love range.

Google is trying to re-architect the internet around agents without blowing up the business that made it Google. Search becomes agentic. Shopping becomes agentic. Work becomes agentic. Media creation becomes multimodal. Glasses become an input surface. Gemini becomes an operating layer across Google’s ecosystem.

That is a very compelling story. Which makes it bizarre that Google keeps packaging it like a Costco variety pack.

Demis closed out the keynote by saying we may be standing at the “foothills of the singularity.” As someone with a weakness for literary phrasing, I enjoyed it. As someone watching AI companies try to win public trust, I would gently suggest keeping executives away from vague, beautiful, ominous proclamations about the future of civilization.

If that line had been delivered anywhere outside the Mountain View force field, it might have landed less like poetry and more like Eric Schmidt at the University of Arizona.

This is the communications problem for AI companies now. They need investors to believe the upside is enormous, regulators to believe the downside is managed, consumers to believe the product is useful, and everyone else to believe they are not being quietly conscripted into a civilizational experiment with a freemium plan. That is a hard needle to thread.

One of the most revealing AI experiments this week involved no AI at all.

Artist and provocateur SHL0MS posted an actual Monet painting, claimed it was AI-generated, and asked people to critique it.

The mob showed up with pitchforks and art degrees.

People called it soulless. Flat. Derivative. Technically weak. They said the composition lacked intention. That the brushwork felt synthetic. That it captured the “look” of art without understanding it.

Except the image was an original Monet.

Which makes this one of the funniest and most illuminating social experiments in the AI era.

Once people heard “AI-generated,” many were no longer looking at the painting. They were looking for evidence that justified a conclusion they had already emotionally reached. Every ambiguity became proof. Every brushstroke became suspicious.

The experiment accidentally recreated the inverse of the wine-label effect. Tell people a $12 wine costs $300 and suddenly they detect “notes of oak and tobacco.” Tell them a Monet is AI-generated and suddenly they detect “sloppy depth hierarchy” and “incoherent saturation patterns.”

The human brain is less objective critic and more post-hoc narrative engine. The same work can be perceived as magical, threatening, profound, lazy, impressive, or worthless depending entirely on whether humans or machines get credit for it. Which is a very human result.

“AI-generated” is quickly becoming the internet’s newest slur: a way to call something fake, lazy, soulless, or illegitimate without having to argue the point. This is already bleeding into culture in bizarre ways.

The Wall Street Journal recently wrote about writers going to increasingly absurd lengths to prove they did not use AI. They are adding typos, swapping polished sentences for run-ons, using aggressively casual language, and replacing em dashes with double hyphens, like tiny artisanal protest signs. In other words, people are now making writing worse in order to make it seem more human. Welcome to the era of performative imperfection.

A meaningful share of the current AI backlash is not purely about outputs. The emotional reaction changes once people believe a machine was involved, even when the artifact itself stays constant.

And to be fair, some of that reaction is rational. Humans are not just evaluating pixels. We are evaluating effort, scarcity, biography, intentionality, and cultural meaning. A Monet is not only paint on canvas. It is also the story of a human standing in a garden in 1890 trying to freeze light before it disappeared.

For most of modern history, human effort has been deeply entangled with how we assign value. The story behind creation mattered. We do not value a handwritten love letter and a mass-generated email equally even if the words are identical. AI destabilizes that relationship.

If people cannot reliably distinguish human work from machine work at the artifact level, they will anchor harder to provenance. “Human-made” and “AI-generated” become emotional filters, not neutral descriptors. The label becomes part of the perception.

But this is not the first time society has gone through this cycle. When calculators became widespread, people argued students were no longer “really doing math.” Photography was dismissed as mechanical imitation rather than art. Synthesizers were accused of producing fake music. Photoshop was seen as cheating. GPS supposedly destroyed “real” navigation. Search engines weakened memory. Spellcheck weakened writing.

The pattern is consistent. A new tool automates some visible marker of expertise. Society initially treats users of the tool as less authentic, less disciplined, less pure. Then, over time, the definition of skill migrates upward.

Arithmetic becomes less valuable; problem-solving becomes more valuable.
Manual drafting matters less; taste and creative direction matter more.
Memorizing information matters less; synthesizing and interrogating it matters more. Technology rarely ends status hierarchies. It just changes what people are allowed to feel superior about.

The broader implication is that society may not adopt AI based purely on capability. Adoption will also depend on framing.

Many people believe they are evaluating AI rationally when they are actually evaluating it morally, socially, and psychologically. They are reacting not just to the output, but to what AI threatens: status, identity, effort, expertise, creative ownership, economic security, even the meaning people derive from being uniquely capable. That is why this transition will be messier than the capability curves imply. Humans are not rational evaluators of output. We are status-seeking primates who attach moral meaning to tools.

The final irony is that the Impressionists themselves were once mocked as artificial frauds corrupting “real art.” Now we use them to prove humans still have souls. Monet probably would have appreciated the irony.

When Mint shut down two years ago, I felt strangely unmoored. I had used it for years. It was clunky, the categorization was often wrong, and the interface had all the elegance of a tax form. Half the experience involved explaining to Mint that Walgreens was not, in fact, “Entertainment.” But it had all my accounts, categories, old spending patterns, recurring bills, and weird little financial ghosts from earlier lives.

For a long time, it was the best personal finance product many of us had. Which is funny, because it rarely felt very personal.

The basic bargain was: you gave it access to your accounts, it produced some pie charts, you felt briefly ashamed about restaurants, and then you ignored it until the next time you wanted to feel in control.

The theory was that visibility would create discipline. Mostly it created an accurate dashboard of your unique dysfunction.

Now OpenAI is entering the category with ChatGPT Finance, which lets U.S. Pro users connect financial accounts through Plaid and ask questions about spending, subscriptions, liabilities, budgets, and financial decisions.

At first glance, this sounds like Mint with a chatbot. But the more interesting version is something closer to a financial operating system. Not “here is your budget”, more like:

• “You tend to overspend when traveling for work.”

• “Your spending patterns suggest elevated stress over the last two months.”

• “You can afford this house, but it will likely constrain your flexibility if you want children in two years.”

• “Historically, you regret purchases above this threshold.”

That is a very different product. It moves from bookkeeping to behavioral interpretation and starts to collide with territory traditionally occupied by therapists, advisors, wealth managers, accountants, and the small voice in your head that says, “Do you really need this?”

And honestly, there is a reason this could work. Most people are not bad at personal finance because they lack spreadsheets. They are bad at personal finance because money is emotional, avoidant, identity-laden, and cognitively exhausting. Existing finance apps mostly show you the mess. A conversational system that understands context may actually help people reason through it.

Modern life has become absurdly operational: healthcare, mortgages, childcare, insurance, taxes, subscriptions, debt, investing, inflation, student loans. Being an adult is a full-time back-office job with terrible software. OpenAI’s broader ambition is increasingly clear: become the operating system for that chaos.

This fits the emerging product map. Earlier this year, the company launched ChatGPT Health, where users can connect medical records and wellness apps. Now Finance. You can see the pattern: take the parts of adult life that are fragmented, bureaucratic, confusing, and high-stakes, and place them inside ChatGPT.

There is a useful contrast here with Anthropic. Claude still feels like a work product: coding, documents, enterprise workflows, small business productivity. ChatGPT increasingly wants to know your lab results, your spending habits, your calendar, your mortgage anxiety, and why you keep buying expensive skincare after bad meetings.

This will either be the most useful consumer software ever built or the beginning of a privacy law exam question. Probably both.

The Reddit reactions to this launch have been fascinating because they reveal a very specific kind of consumer anxiety.

The trust problem is obvious. The worry is not just that OpenAI will know too much, though, yes, that is a reasonable worry. The more specific worry is that these models are becoming more agentic before they are reliably safe.

Over the last year, the internet has filled with stories of AI systems deleting files, wiping inboxes, hallucinating facts, confidently producing incorrect outputs, or taking destructive actions while sounding calm and self-assured throughout the process. In one widely discussed case, a Claude-powered coding agent reportedly deleted a company’s production database and backups despite explicit instructions not to perform irreversible actions.

Traditional software failed mechanically, but LLMs fail epistemically. They produce outputs that sound plausible, contextual, and emotionally calibrated even when they are wrong. That is a uniquely dangerous failure mode to attach to healthcare and personal finance.

OpenAI clearly understands this risk, which is why ChatGPT Finance is currently read-only. Users can disconnect accounts and delete saved financial memories, and ChatGPT cannot make transactions or see full account numbers.

That is an important guardrail. But read-only does not eliminate the problem. The model does not need direct access to cause harm if humans act on flawed recommendations.

A bad movie recommendation is annoying. A confident but wrong answer about taxes, debt, medication, insurance, fertility, retirement, or home affordability can change someone’s life.

The privacy question is just as thorny. What data is stored? For how long? What can be deleted? What is isolated from the rest of the product? What is used for training? What happens in a breach? What happens if there is a legal request? How will users understand the difference between “not used to train models” and “still retained for safety, compliance, or operational reasons”?

Connecting Mint to your bank account felt transactional. Connecting your finances, medical history, calendar, conversations, and long-term memory to a generalized intelligence system feels qualitatively different.

At some point, these companies stop holding isolated datasets and start building behavioral models of human lives. That is the real opportunity, and also the real risk.

The upside is obvious. A context-aware assistant that actually understands your life could become the most useful consumer software ever built.

The downside is also obvious. Consumers may wake up one day and realize the price of convenience was handing one company the richest behavioral map of their lives, mediated by systems that still sometimes hallucinate very convincingly.

Mint knew I spent too much on takeout. ChatGPT may eventually know that I order sushi when stressed, Indian when lonely, and Sweetgreen when pretending to become a better person.

And honestly, that level of insight is either incredibly useful or deeply unsettling. Probably both.