One sentence got 116 million views this week.

"Racing straight to self-improving superintelligence."

A 27-year-old researcher quit Anthropic, told the Wall Street Journal that "by the end of next year things could be out of control already," and posted a thread. Within a day politicians were quoting it. Bernie Sanders' bill to ban superintelligence, announced five days earlier, suddenly had its whistleblower.

So let's take the claim seriously. Not the mood. The claim.

Can a machine make itself smarter? And if it can, is it doing it right now?

Turns out a lot of serious people spent this summer trying to answer exactly that. Not on X. In labs, with measurements. I read what they found so you don't have to.

Here's the reality: a machine can get better at anything you can score. It cannot yet get better at the things you can't. Superintelligence lives on the far side of that line. Nobody has crossed it. Nobody is close.

There are two kinds of "self-improving," and only one exists

The tweet skipped a distinction that decides everything.

Picture a golfer with a coach. The coach sets the target, watches every swing, and says better or worse. The golfer adjusts. Thousands of swings later, the golfer is very good at hitting that target.

That's the first kind of self-improvement. Fixed goal, fixed scoreboard, endless reps. Call it improvement inside a box. AI is already superhuman at this. Anthropic runs a standing test where it hands Claude a piece of training code and says: make this run faster without breaking it. In spring 2025 the model got a 3x speedup. This spring it got roughly 52x. A skilled human engineer gets about 4x in a full day's work.

That's real. That's remarkable. And notice who set the target. A person did.

Now the second kind. No coach. No target. The golfer has to decide what sport to play, invent the scoring system, figure out which drills matter, and know when to quit and try something else entirely. That's not practice. That's judgment. That's taste.

The first kind makes the machine better inside the box. The second kind is the box growing on its own.

"Self-improving superintelligence" needs the second kind. Every jaw-dropping number you've seen comes from the first.

When you take away the scoreboard, the wheels come off

Four research teams published tests in the last five weeks, each aimed at a different piece of the "machine builds the machine" story. Different labs, different methods, same result.

The most direct one gave top AI agents a simple job: here's a working AI training recipe, you have four hours, make it learn better. Then the researchers reran the agent's recipe from scratch and scored it against the original, on a scale where the original recipe is a 0.1 and perfection is a 1.0.

Average score: 0.17. Best in the world: 0.25.

Given the explicit job of making AI learn better, the strongest system on Earth closed under a fifth of the gap. And most agents didn't even try to change how the model learns. They tinkered with the plumbing and left the engine alone. Cranking up the "thinking" setting mostly made them more willing to touch the engine. Not more able.

A second team tested whether an AI assistant that remembers your past sessions actually gets better from remembering. Answer: sometimes, unevenly, and often it's hard to tell whether the improvement came from real learning or luck. The researchers had to build a separate tool just to tell the two apart.

A third team built the most ambitious closed loop of the bunch: a system that watches what it's bad at and uses that to pick what to train on next. It worked. A small model got several points better. But look inside the loop and you find a bigger "teacher" model supervising it, a curriculum designed by humans, and tests chosen by humans. The small model didn't get smarter than its teacher. It got closer to it. That's a student catching up, not a student surpassing.

The fourth team asked a question nobody in the hype cycle asks: if a machine is going to build its successor, something has to inspect what the successor learned. Can an AI do that inspection? The agents could run the tools. They could not read the results. In the researchers' own framing, they frequently misinterpreted their own measurements.

They can hold the microscope. They can't read what's under it.

Scientists outside the labs see the same thing

This isn't four cherry-picked studies.

A Princeton-led team ran a harder test. They took two real research papers that hadn't been published yet, so the AI couldn't have memorized them. They gave Claude six days, $3,000 in compute, and the open internet, and said: answer the same research question these scientists answered. Then they had the original authors grade the AI's paper like a journal reviewer.

Both rejected.

The AI did all the grunt work. Read the literature, ran hundreds of experiments, wrote it up. Then it ran nonsense experiments on tiny fake datasets, locked onto dead ends, couldn't back up and rethink, and when it got critical feedback it responded by watering down its claims instead of fixing its method.

Sound like anyone you've managed? It's the junior who can execute anything you spec and can't yet tell which thing is worth specing.

Anthropic cofounder Jack Clark read that study and called it a bearish signal for fast self-improvement. His own company hit the same wall trying to automate parts of its safety research. And Anthropic's own June report on this exact topic, packed with internal data, says two things plainly: AI now writes most of their code, and the gap on choosing what to work on is still large. Their words: not there yet, and not inevitable.

The company the whistleblower walked out of says it hasn't happened. Its own numbers agree.

The case for the tweet is stronger than you'd like

Let's be honest about the other side, because it isn't nothing.

The length of tasks AI can handle on its own is doubling every four months. Claude's success rate on messy, open-ended engineering problems hit 76% in May, up 50 points in half a year. In one internal test, the newest model picked a better next research step than the human 64% of the time. Six months earlier it was a coin flip.

And there's a real argument that you don't need taste at all. Most AI progress isn't a eureka. It's scale it up, watch what breaks, fix it, repeat. If 99% of the frontier is sweat, and sweat is exactly what these systems are superhuman at, maybe you grind your way to takeoff without ever having an original thought. Nobody knows. One of the Princeton researchers called it the trillion-dollar question.

But.

Every one of those impressive numbers, every single one, lives inside a box where a human wrote the scoreboard. The 52x had a fixed correctness check. The 64% was graded by another AI on a rubric a human designed. And the moment you take the scoreboard away, in every test above, the failure shows up at the exact same joint. Which change matters. What the result means. When to quit.

The sweat is automated. The judgment is not. And judgment is the part that would have to be automated for "out of control by next year" to be anything more than a feeling.

This looks less like a whistle and more like a moat

Now look at how the tweet traveled, because we've seen this chart before. Usually in a pitch deck.

The Wall Street Journal ran an exclusive 18 minutes before the tweet went up. That's not a man of conscience hitting post. That's a launch. The account had no prior activity, wiped or brand new. The first quote-tweets landed within minutes, and they came from AI-policy nonprofits that share a major donor who is also an Anthropic investor. For scale: when Ilya Sutskever, the most famous researcher departure in the industry's history, left OpenAI, his post drew under 6 million views. A researcher nobody had heard of drew 116 million in a day.

Dozens of researchers leave frontier labs every month. The Journal doesn't write about them. Politicians don't quote them.

I can't prove the money map from a screenshot. Treat that part as an allegation. But the parts I can verify are enough. The WSJ story did run first. The account was empty. And the Sanders bill, announced five days earlier, already had its ask written: a new cabinet-level agency with the power to define what AI is allowed to be, advised by a board of "experts."

Ask the oldest question in venture: who benefits?

Not the public. A cabinet-level agency with an expert board is the textbook incumbent moat. Compliance cost is a tax the leader can pay and the next entrant can't. Every regulated industry in America learned this. Banks. Pharma. Telecom. The rules get written by the people who already won, and the door closes behind them.

And the leader in this story is on the record. Anthropic's own June report says it would slow down or pause, if every other lab did too, verifiably. Read that as a founder would: a company one step ahead is asking for the race to be frozen in place. A "unilateral pause," in their words, would just change who the front-runner is. So they'd rather everyone stop at once. Of course they would.

If you showed me a startup with a viral launch timed to the minute, amplified by accounts funded by its own investors, landing the week a bill was introduced that would hand its allies a regulatory seat, I wouldn't call it traction. I'd call it a campaign.

And here's the part that should bother you most: the science doesn't support the fear the campaign is selling. Every study this summer says the machine can't cross the line the tweet says it's racing toward. You don't need to manufacture panic about a thing that's happening. You manufacture it about a thing that isn't.

Notice who agrees with the tweet. The doomers. And the boosters. The people who want AI banned and the people selling you a $50 billion datacenter share exactly one premise: the machine is about to take off. That premise sells fear. That premise sells fundraises. And that premise gets a seat on the board of the agency that decides who's allowed to compete.

Nobody gets paid on the boring answer. The boring answer is the true one.

We called this in May

Readers of the 14-part series will recognize the argument. Back in May, in the self-evolution chapter, the line was: for most real-world tasks, success is hard to score cleanly, and poor evaluation creates bad evolution. A week later: a system that optimizes beautifully on a deterministic benchmark may falter badly when uncertainty becomes central.

That was the bet. Four months and seven studies later, it's the finding. Every gain lives where the score is clean. Every failure lives where it isn't. The axis that matters isn't online versus offline, or one model versus many. It's scorable versus unscorable.

The weights are frozen. The system around them isn't. Give it a scoreboard and it gets better, run after run, sometimes 52x better. What it can't do is learn what to want. The model doesn't learn. The system learns whatever a human taught it to measure.

Even wired into teams, the conductor is still us. Anthropic's best multi-agent research result recovered 97% of a gap that two humans got 23% of the way through, and it ran on a rubric a human wrote. The Princeton agents had teammates, six days, and a budget, and still couldn't tell a good experiment from a bad one. Orchestration routes around execution gaps. It does not route around taste.

What this means if you're deploying capital

For everyone: when someone tells you a system is improving itself, ask one question. Who wrote the scoreboard? If a person did, you're looking at a very good optimizer. If the system did, you're looking at something nobody on Earth has demonstrated.

And when someone tells you the answer is a new federal agency, ask the follow-up. Who's on the board?

The bottom line

The machine can get better at what it can measure. It cannot yet get better at what it can't. Everything called superintelligence sits on the far side of that line, and as of this month the best systems in the world stand well short of it, misreading their own instruments and declining to touch the engine.

A viral tweet doesn't move that line. A bill doesn't move it either. What a bill can move is who gets to build on this side of it, and that's the part worth watching.

When the fear and the evidence point in opposite directions, don't ask who's scared. Ask who's selling.

Sources: AI4AI-Bench (arXiv 2608.20318); PAST-Bench (arXiv 2608.04003); NeoHorse-1 (arXiv 2609.08183); SAEScientist-Bench (arXiv 2609.09113); Kirgis, Kapoor et al., shadow evaluation (arXiv 2607.27191), via MIT Technology Review, Aug 18 2026; Whitfill & Cunningham, "The Economics of Recursive Self-Improvement," METR, Jul 22 2026; Favaro & Clark, "When AI Builds Itself," The Anthropic Institute, Jun 2026; Wall Street Journal, Sep 8 2026; Sanders/Casar Ban Artificial Superintelligence Act, announced Sep 3 2026.