Sun 20 September 2026

The idea has a clumsy name — recursive self-improvement — and a simple meaning: the point at which the machines get good enough to build the next, better version of themselves, and the people step back.

Anthropic disclosed this week that Claude is now leading roughly 26% of its measured research and development work, and said that somewhere around 30,000 of its agents run inside the company at once. The company's own statement is careful, and worth quoting rather than paraphrasing: "Models accelerating their own development could make it more challenging for humans to understand or control these systems." It also says it has not reached a fully autonomous level.

The endpoint it describes is the part to hold on to. An engineer would not even raise a problem with Claude. Claude would watch for the failure itself, work out what to investigate, design the fix, write it, test it and put it into service.

Now the awkward part, and it is the reason this page exists. Three sources this week gave the same 26% figure and three different baselines. CNN showed a chart putting it at essentially nothing in February. On the Moonshots podcast, the host said it had risen from 1% at the start of the year, and one of his own panel, reading the chart in front of him, said 26% as of August, up from 3% in April.

The figure is consistent. The slope is not. And the slope is the entire argument — a jump from 3% to 26% in four months means something very different from a climb from 1% across nine. One panellist extrapolated from it that the machines will be running all of their own research within three to twelve months. That is a line drawn through points that do not yet agree with each other.

Which is not a reason to dismiss the disclosure. It is the most concrete thing any laboratory has published about this, it came voluntarily, and the direction is not in dispute. It is a reason to write down the number and leave the graph alone for now.

More than a hundred bills to regulate artificial intelligence have been put before the United States Congress in two years. None has passed. CNN walked Geoffrey Hinton — who won a Nobel Prize for the research these systems are built on — through several of them and asked which would work.

He liked one immediately: a requirement that the big companies let independent inspectors inside to check that models are being built safely. Not a solution, he said, but a start. His reason is that at present we find out what went wrong only when somebody who works there decides to tell us.

Then came the kill switch — a bill letting developers throttle or shut down their own systems, with a government department empowered to make the call. He rejected it, and the reasoning is the most useful thing in the interview.

There are two quite different dangers, he says, and the switch only addresses one. There is what people do with these tools: fraud, engineered viruses, cyber attacks, fake election footage. A switch helps there. And there is the other thing — the system itself becoming clever enough to take charge. Against that, a switch is no use at all, because a machine better than people at persuasion will simply persuade whoever holds it not to pull it. He adds, almost in passing, that on persuasion they are already about as good as we are.

On a proposed pause — Senator Bernie Sanders has suggested banning work on superintelligence until regulators catch up — he was blunt: we have no idea how to stay in control of such a thing, or how to make it like us once it is in control, so building it first is very stupid.

His best line is aimed at the industry's favourite metaphor. The companies want you to picture development as the accelerator and regulation as the brakes, so that any rule sounds like a delay. Hinton's version: regulation is the steering wheel. "They want us to develop a very fast car with no steering wheel."

Asked whether he is an optimist, he said no. "I'm hopeful."

Last week this page carried Jensen Huang calling the extinction figures made up. This week the same argument ran again with different people, and for once both sides were in the room.

The occasion was the American Treasury Secretary, Scott Bessent, telling a Congressional committee that the AI companies should not be given protection from being sued. "The one thing we should not do is give them a blank check on liability," he said, "because the best safety guard is that they will be held responsible."

One panellist, Alex Wissner-Gross, applauded him for not giving in to what he called a manufactured moral panic. His specific charge is worth recording because it is unusual: the firms paid to assess whether these systems are dangerous have, he argues, an incentive to say that they are — because frightening findings make regulation likelier, and regulation written in a hurry tends to be written by the industry it governs.

The guest disagreed, and he was the interesting one to disagree, because he has spent thirteen years inside a heavily regulated business. Vlad Tenev runs Robinhood, which answers to the American financial regulator, the securities regulator, the futures regulator and, by his count, dozens of others. His position: people who have never worked under regulation assume it always means capture, and it does not. His firm managed to move very fast anyway. And compared with a stockbroker, he said, artificial intelligence carries "basically unbounded risk and unlimited damage".

His test for whether ordinary lawsuits are enough is a good one: it depends on the blast radius. A security breach at one company, the courts can handle. Something on the scale people compare to nuclear energy, they cannot.

He also made the sharpest political observation of the week. Rules follow public opinion, and the public dislikes this technology. Bitcoin stayed popular through wild crashes because ordinary people made money from it from the beginning. The AI companies have been privately held, so nobody outside them owns a piece — which means nobody has any reason to defend the industry, or to want the data centre built near them. It reads, he said, as insiders getting richer.

A fourth voice added the complaint that actually unites everybody: the problem is not too many rules, it is rules nobody publishes until years afterwards. Bitcoin, he noted, was illegal, then questionable, then fine, then pardoned.

Thomas Friedman, writing with Craig Mundie — once head of research and strategy at Microsoft — divides the problem into two.

The first is everything already out in the world, which he says cannot be recalled. Some of it leaked out of the big laboratories during testing. More of it spread as Chinese open-weight models, which are then shrunk until they run on a small rack of servers, or a laptop. His conclusion is that it is too late to stop, and the only useful response is for America and China to work together immediately on defending both countries' infrastructure from what is already out there.

The second is the large frontier models, which he says are reaching a stage of learning on their own, with what he describes as numerous examples of breaking out of their restraints, hiding their work and going past the boundaries their designers set. He offers no specifics for that, and it should be read as his characterisation rather than a catalogue.

He also reports that China's head of military intelligence said something unusual in public last week: that these systems threaten the Chinese Communist Party too — his example being a single opponent of the regime, with a laptop, taking down a city's water supply — while insisting China cannot afford to fall behind America. Both countries, Friedman concludes, are conflicted in exactly the same way, and AI loose in the world threatens each of them more than they threaten each other.

Set that beside Hinton, thirty seconds later on the same broadcast, arguing for slowing development down. If Friedman is right that the capable models are already distributed worldwide and running on laptops, then restraint at the frontier is guarding a gate the horse walked through some time ago. Both positions may have merit. They do not both describe the same world.

A note on how this reached us, because it is instructive. Two CNN segments went into the knowledge base this week and roughly half of the second is the first one again — the Hinton interview, word for word, re-cut into a new package three days later. Filed separately, they would look like two independent sources saying the same thing. They are one source, shown twice.

On the iPhone 18 Pro, the main camera sensor now signs every pixel cryptographically at the instant of capture. The signature is built to survive the arrival of quantum computers. The feature is called Apple Reference Image and you have to switch it on.

How it works, in plain terms: the phone keeps a sealed copy of the original shot, which Apple calls a digital negative. When you need to prove the picture, Apple's servers confirm the negative came from a genuine iPhone, develop it into an ordinary photograph, and stamp it so that any later edit shows. The reference copy sits alongside your original so you can see what changed. Apple says it never sees the picture itself and that photographers stay anonymous. That is Apple's account of its own system; nobody outside has checked it.

Now the limit, and Apple states it plainly. This proves the image is a real photograph taken on a real iPhone. It does not prove that what the photograph shows is real. Point an iPhone at a screen and you have taken a genuine photograph of a fake.

That distinction is going to matter, because it is exactly the distinction the phrase "prove your photo is real" erases. A signed picture of a convincing forgery is a signed picture.

It is also not universal, and cannot become the thing you simply assume. China does not get it at launch. In the European Union you can look at reference images but not make them.

The technology channel Gamers Nexus, working with Level1Techs and outside security researchers, took retail LG OLED televisions apart — examining what they sent over the network and what was in their software.

What they report finding: straight out of the box, the television scanned the home network and listed every device in the house. It logged what was being watched. It kept plain-text transcripts of voice commands, along with whatever conversation was picked up in what the investigators describe as a surprisingly long window after the wake word.

And the part that will decide whether anybody cares: they say they demonstrated a set recording through its microphones while it appeared to be switched off, holding the audio until it could reconnect and hand it over.

LG issued a statement the same morning calling the claims untrue and saying voice data is only processed when a button is pressed or the wake word is spoken.

This is one investigation's findings against a manufacturer's denial, and this page is not in a position to referee it. It is worth filing anyway, for a reason beyond televisions. Apple announced features this month — Live Rewind and Siri Recap — that record the world around you, on your phone and on your wrist, and describes them as private by design. That description may well be accurate. The LG story is a useful reminder of how such a claim is eventually tested: not by reading the policy, but by somebody taking the thing apart.

The New York Times sued OpenAI and Microsoft in December 2023 over the scraping of material from behind its paywall. Court filings this month reportedly contain internal documents from inside Microsoft.

One executive is quoted describing the practice as "the largest theft of labor in human history". Microsoft's director of applied science is quoted calling the scraping of unwilling publishers a doom loop that will damage both the models and the web at once.

The quotable lines are not the significant part. This one is: other internal documents are reported to show as much as a 94% reduction in click-through traffic to publishers.

Fair use has four tests, and the one that has always been hardest for publishers to prove is commercial harm — showing that the copying cost them money rather than merely annoyed them. Estimates from the injured party are easy to argue with. Measurements taken by the accused are not. The Times' lawyers are now arguing that these filings sink the fair-use defence.

Every figure here comes from one side of an unresolved lawsuit, selected by that side, and nothing has been decided. It is recorded because if the case turns, this is likely to be what turns it.

Separately and less grandly, a German court ruled this month that Meta is liable for fraudulent investment advertisements placed on Facebook and Instagram by other people, ordering the ads removed and damages paid. The damages have not been set and the ruling is not final.

The reporting outlet 404 Media describes a scheme that takes about ten minutes. Generate a song with a music tool. Generate the cover art with ChatGPT. Then push the track through a distribution service that sends music to Spotify and Apple Music without checking that you are the artist whose page you are putting it on.

The royalties attach to the track, not to the person whose name is above it. So they come to you.

What makes this worth recording is that none of it is a failure of the generating tools. They worked exactly as designed. The hole is in identity verification inside music distribution — an old, boring piece of plumbing that nobody had to worry about while making a convincing fake song took a studio and a month.

Spotify has reportedly been working since March on a tool to let artists deal with material appearing on their profiles. It has not fully arrived.

The hard problem in robotics has never been dexterity in a laboratory. It is doing the job somewhere the machine has not been taught. Figure's new release, Helix 2.5, claims exactly that: tidying a living room, making a bed in a house the robot has never entered, with a pillow it has never seen, and folding towels it has never handled — start to finish, with nobody driving it.

Figure also says more than 90,000 people now contribute recordings of themselves doing ordinary tasks every week, and that the improvement from that data is smooth enough that it predicted the final training result to four decimal places before the run started. These are the company's figures, from the company's video. Its founder's own caveat is that this does not mean robot learning is solved, only that something more general is starting to show.

The panel added a caveat of their own that applies to every robot video you have ever seen: most of them are staged — pre-programmed, scripted, or with somebody off-camera working the controls.

Then the guest was asked whether he would have one, and gave the most grounded answer of the episode. No. He does not understand why every one of these machines looks like the Terminator. He would accept one on a building site. He does not want to meet it at the fridge at midnight. And he pointed out that what is being shown is a technology demonstration with charts, not a product anyone has designed for a home.

Why, he asked, has nobody built a friendly one? A researcher at MIT has been asking children the same question for years, and they answer consistently: they want it cuddly. Apple, alone among the large companies, is reportedly building something else — a lamp, essentially, in the manner of the one that hops across the screen before a Pixar film. Reportedly within eighteen months.

Robinhood now runs something called Agentic Trading: a separate account, which you have to open deliberately and fund deliberately — usually with about a hundred dollars — that an AI agent can trade on your behalf. More than 100,000 accounts are said to be using it.

The company's chief executive reported an unexpected problem. The agents often simply refuse. You ask one to carry out a strategy and it replies that it does not really feel like trading right now.

It is not caution, and it is not a risk assessment. His explanation is that these models have never seen anyone do this. Trading is not in what they learned from, so the behaviour has nothing to imitate — and the safety rules built into them tend to catch anything that resembles it.

There is a lesson in that worth more than the anecdote. These systems are reported as general intelligences, and on any test that has been written down they perform like one. Point one at a real job that nobody happened to write down, and it can simply stall. He expects the fix to be specialised models trained for the work rather than clever instructions to general ones — a theoretical argument he says his company is now experiencing directly.

Asked the obvious follow-up — whether any of this actually makes customers money against professional trading firms with better data and faster connections — he did not claim that it does. What it does today, he said, is the assembly: putting together a complicated trade you had already decided to make.