Latest News

Every proposal for controlling this technology so far has needed something that does not exist: a new international body, a treaty China would have to sign, or a company willing to stop while its rivals carry on.

This one needs none of that. Before a company releases a new model, its competitors get to attack it — running their own security tests against it and saying publicly if they find something dangerous. In Musk's phrasing: "instead of grading your own homework, you would at least have competitors grading your homework."

What makes it interesting is who arrived at it. Musk got there from believing the models are dangerous. Huang got there from believing they are not — he wants independent assessors the way company accounts have independent auditors, and more than one of them, so no single assessor can be leaned on. A political panel got there from asking what the United States and China could conceivably agree to. Three different starting points, one mechanism.

It has teeth, too, and they are ordinary legal teeth rather than new ones. If your competitors warn that your model is unsafe, you release it anyway, and it then causes harm, you have handed a court most of a negligence case. Product liability law already applies to software.

And it is the rare proposal China might accept, precisely because it costs nothing to agree to and requires trusting nobody. A pause has already been refused. Letting someone try to break your model before you ship it is a much smaller ask.

When OpenAI released GPT-6 Astra this month, a senior figure there said it was not unreasonable to feel we are now in the era of artificial general intelligence. The coverage ran with it.

Here is the specification. Text goes in. Images go in. Text comes out. No audio in either direction, no video in either direction, and no ability to tune it to your own material at launch.

As one analyst put it, the model that got called the arrival of the AGI era cannot hear you. If you want a system that will take a voice recording or a video clip, Google has been selling one for weeks, and at a fraction of the price.

On the independent scoreboard — the one OpenAI did not commission — Astra came fourth, behind Anthropic's Fable 5.1 and two others. It also costs about two and a half times per word what the model before it cost, which works out at roughly 75% more expensive for an average job while scoring below a rival.

One number in the launch is genuinely impressive and deserves saying: on a test of how often the model confidently invents things, the rate reportedly fell by about half while accuracy went up. For anyone using these tools for real work, that is worth more than any record score on a maths exam.

The rest is a modest upgrade with an excellent launch video.

For months this Handbook has recorded the same complaint: the model at the top of the table is often not the best model to use. This week somebody explained the mechanism.

No respectable laboratory trains its model on the actual test questions. It does not need to. It can buy training material from companies that build practice environments designed to imitate those questions as closely as possible. The model never sees the exam and still learns to pass it.

The giveaway is that the ability does not travel. Two models this month scored near the top on a widely used test, then collapsed on a newer version of the same test that nobody has had time to imitate yet. Meta's own AI chief effectively conceded the point in public, saying his model is not as strong as the leaders but is significantly cheaper.

A second problem surfaced the same week, and it is about the referee rather than the players. One prominent maths benchmark is funded by OpenAI, which also holds exclusive access to part of it. That does not make any published score false. It does mean it is a company's number rather than an independent one.

Which leaves a simple working rule. A score from a test too new to have been imitated means something. A job you ran yourself and can judge with your own eyes means more. Everything else is marketing with a decimal point.

OpenAI published a solution this month to the Navier–Stokes problem — one of seven famous unsolved problems in mathematics, each carrying a million-dollar prize, only one of which had been solved in the 26 years since they were set. It used a model it has not released, described as considerably more capable than anything you can buy.

Then a professor at New York University published his account. He and a collaborator had been working on the same problem for over a year, putting their drafts into OpenAI's own coding tool as they went. He says he asked whether the model had been trained on those sessions, was told it had not looked up user data, asked again specifically about training, and got no answer.

He also says he was offered a place on the paper on condition that his collaborator — who works at a rival company — was removed from it, and that when he threatened to go public the reply was: "why would you ruin your career?" OpenAI's researcher called the allegations false and inflammatory and published part of their messages.

Set the personalities aside, because the sentence that matters is in OpenAI's own careful statement. No person or agent looked at their work, it says, and no specific user data was accessed — but "while unlikely, we cannot rule out that deidentified data derived from the usage of our products helped improve our models."

That is the answer to the question every professional using these tools now has, and it is not no.

One more thing worth noticing, and it has nothing to do with the row: two separate sources this week describe the same result completely differently — one says a swarm of ten thousand machines over 88 hours, another says an internal model over a week or two. Neither was in the room. Watch how fast an unchecked number becomes a fact.

Ask one of these systems whether it is conscious and it may well say yes. That tells you nothing — it has read a great deal of science fiction, and it is very good at producing the expected next word.

So researchers at Anthropic went looking inside instead. They asked their model to count to five and then think about what it had done. On screen: one, two, three, four, five. Inside the machine, between its layers, words were surfacing that never appeared anywhere — "halfway" when it was halfway, "countdown" while counting, and at the end, "done".

The researchers describe it as a kind of mental whiteboard: somewhere the thing works matters out before it speaks. Which resembles, uncomfortably, one of the leading theories of how human consciousness works.

The reporter's own conclusion is firmly no. This is, he says, at the foothills of things that look a bit like consciousness but are not. What changed for him is narrower and more interesting: he used to think the whole idea was incoherent, and he no longer does.

The reason to care now rather than later is a practical one. There are two ways to get this badly wrong — granting rights to systems that merely follow rules, or creating something that can genuinely suffer and not noticing. One philosopher he spoke to put it plainly: doing that second one by accident would be a moral catastrophe.

It also matters for a duller reason. If the thing is doing work in places we cannot read, then the explanations it shows us are not the whole story — which is exactly the worry several engineers raised this week from an entirely different direction.

Every humanoid robot video shows the same things: running, dancing, backflips. None of it tells you anything useful. A machine that can backflip on cue has rehearsed one movement in a controlled room. The real test is whether it can pick up an egg without breaking it, open a fridge, or work out that the pan is hot.

Cooking an egg is harder for a robot than a backflip. That is the sentence to remember next time one of these videos goes round.

Tesla appears to agree, because the interesting engineering in the next Optimus is all in the hands: twice the independent movements of the last version, close to three times as many sensors, and — the clever part — most of the motors moved up into the forearm, pulling the fingers by cables through the wrist. That is how your own hand works. It keeps the hand light and precise while the strength comes from behind it.

Now the price. You have read that these will cost around $30,000. Analysts quoted this week put the early ones nearer $70,000, with $50,000 described as a reasonable hope. $30,000 is an ambition for some later year when millions are being made. It is not what the first ones will cost.

The companion vehicle, the two-seat Cybercab, has firmer numbers because they come from regulatory filings rather than a stage: just under 300 miles of range, and a motor built with no rare earth metals at all — about a quarter of a motor's cost, and a supply chain that runs through China.

The moment a data centre is announced, she says, the land near it reprices — from around $3,000 an acre to $180,000. Then come the permits. Then the wait for electrical equipment, with generators currently quoted at three years.

In orbit there is no land to buy, no council to satisfy, and no queue for cooling: a radiator facing deep space is looking at the coldest thing there is. Point the satellite at the sun and it never gets dark. And SpaceX already owns the one expensive part everybody else would have to buy, which is the ride up.

They intend to launch computing satellites next year.

The same conversation produced a straightforward status report on Starship, which matters because it is the vehicle everything else depends on. One more flight, then they try to catch the returning ship with the tower arms. Musk puts the odds of catching it first time at 50 to 60%, and notes the last flight's practice landing would have been caught had there been a tower where it came down.

Why it matters: the Space Shuttle was reusable in theory and so difficult to reuse in practice that it cost more per flight than throwing a rocket away. Falcon 9 still discards its upper stage each time — about the price of a mid-sized jet, every launch. Starship is meant to bring both halves home and fly again, like an aircraft.

Eric Weinstein spent an hour this week arguing that American science has been broken since the late 1960s, that theoretical physics took a wrong turn in 1983, and several other things that are his opinion and are labelled as such.

One idea inside it is worth keeping, and it is testable.

These systems are trained on what a field considers respectable — its major journals, its accepted results, the papers that got through review. They are then measured against the standards of that same field. Weinstein's suggestion is to do the opposite: point them at what he calls the trash-can corpus, everything a discipline dismissed and laughed at, and see what they find in it.

"The AIs are going to start reading all of the things that our quote leading physicists have laughed at," he says. "Look out."

Whether or not he is right about physics, the observation connects to something else in this week's material. If models are being tuned to match the consensus measures of a field — which is exactly what the benchmark story earlier describes — then they are being tuned away from the very material he thinks holds the value. Both things cannot be optimised at once.

A researcher who says he spent three years on foundational work at both OpenAI and Anthropic resigned from Anthropic this week and said why in public: "Neither company is acting responsibly. They're racing straight to self-improving super intelligence and gambling with our lives."

Resignation letters are easy to discount. What happened next is not. Evan Hubinger, who leads alignment science at Anthropic — the team whose job is making these systems safe — replied publicly agreeing with him. "We really do earnestly believe AI could kill all humans. I personally think it's a greater than 10% chance within the next decade."

He then said something that is harder to put down. "We do not yet have a plan to solve alignment for super intelligence and are not clearly on track to."

Set that against the reason the same company gives for building this at all: that nobody else can be trusted to do it safely, so it must get there first. The firm claiming it alone can handle this is also stating, on the record, that it has no plan and is not on course to find one.

And it is not one company. OpenAI's chief scientist published an essay the same week called "An Alien Mind", writing that he has "a strong expectation" that progress continues into systems that improve themselves, and that "this is a time that calls for extreme caution. I am concerned no one is prepared for the consequences."

Read the next story before you decide what to make of this one.

The usual reply to a story like the one above is that it is a marketing exercise — that a company selling the most powerful technology in the world benefits from you believing it is dangerous.

That reply is not baseless. The physicist Sabine Hossenfelder has published a video saying she was offered money to tell her audience that AI will kill us all. There are organisations paying people to spread that message.

So both incentives are real at once. Some alarm is bought and paid for. Some of it comes from the person running the safety team at the company doing the building, with his name on it, contradicting his own employer's commercial interest.

There is a second thing worth holding. The people quoted in the story above spend their working lives modelling worst cases, surrounded by others doing the same. That is a reason their estimates might run high — not a reason to ignore them.

The reviewer who reported all of this landed somewhere sensible: "I think we could be making a huge mistake by just claiming psyop and ignoring them." Neither swallowing it nor dismissing it is the careful position. Asking who is speaking, and what it costs them to speak, is.

Four of the biggest laboratories released a new flagship model in the same week: Anthropic, Google, Meta and OpenAI. That alone tells you something about the pace. But the story worth keeping is what happened when somebody actually used them.

Matt Wolfe, who reviews these things weekly, ran all four through the same two tasks he always uses — build a small game, and draw a picture using only code. Meta's model had just scored highest of any model ever on the industry's main coding test, and came third on the other leaderboard he follows, ahead of OpenAI's.

What it built was, in his words, "a cube with a little cylinder and I'm shooting other cubes". The models that scored lower produced recognisable characters, environments and physics.

His conclusion is the reason this is the lead story: "The benchmarks I've relied upon the most, I feel like I can't really trust even those anymore." This is a man whose job is following these scores, saying publicly that he is abandoning them.

A week later the same thing happened again, with a different model. DeepSeek's new release scored level with every frontier model on that same coding test — and costs 27 cents to complete a task where one rival charges $8.75. Set to the same practical test, the reviewer's verdict was "it's not on the same level". Twice in a fortnight, the leaderboard and the result disagreed.

The useful lesson is not which model is best. It is that a task you can judge with your own eyes caught something that a hundred-point scale did not.

Anthropic announced that its new model would cost "an estimated 25% less" than the one before it for typical work. An independent site that measures the actual cost of finishing a task put it at the most expensive model tested — slightly dearer than the model it was meant to undercut.

Both figures can be true. The price per unit went down; the amount of work the model does to finish a job went up. The bill is what you pay, and the bill went up.

The example that makes this concrete: Wolfe asked it to build a small game. It worked for about two hours, used up the whole daily allowance on his $200-a-month plan, kept going into paid overage, and cost roughly $120 for that one game.

The same job given to Google's cheaper model used, in his words, "a very, very small percentage" of his credits.

If you are being sold on a price per word, that is not the number that will appear on your statement.

Sir Demis Hassabis runs Google DeepMind, and won a Nobel Prize in 2024 for work that predicted the shape of nearly every known protein. When he gives a date, it is worth writing down.

Asked how far away we are from artificial general intelligence — machines that match people across the board rather than at one task — he said: "I think we're very close to AGI now, you know, maybe around 2030 plus or minus a year."

The interesting part is what he said next. Two to four years ago, he would have said five to ten years. So the date has not moved much. What has changed is his certainty: the range has narrowed.

And he was specific that nothing surprising caused it. No breakthrough, no shock result — just "things going as expected", accumulating.

That is a less dramatic claim than the headlines usually carry, and a more testable one. It is on the record, with a name and a date attached, which is more than can be said for most predictions in this field.

It is the fear underneath a lot of worry about this technology: a machine produces an answer, the reasoning is beyond us, and we are asked to take it on faith.

Hassabis rejects the premise. "You wouldn't just trust what the model says. You would need to test it in clinical trials and test it in the laboratory."

His argument is that the slow, expensive part of medicine was never the understanding. It was the searching — the needle in the haystack. Machines can shrink the search. The testing afterwards stays exactly as it was, and testing does not require you to understand why something works, only to establish that it does. He also points out that his own protein system already reports how confident it is about each part of an answer, rather than presenting everything with equal certainty.

That is a good answer. But the interviewer put a second one to him that went unanswered: bringing a single drug to market takes over a decade and more than a billion dollars, and almost all of that is the testing. If the search drops from years to weeks and the trials still take ten years, the cure is still ten years away.

Hassabis has said publicly that AI could help cure every disease within a decade. The arithmetic in that gap is the thing to keep an eye on.

The Navier–Stokes problem asks whether the equations describing how fluids move — water, air, blood — can break down and stop making sense. It is one of seven problems carrying a million-dollar prize, and it has been open for roughly ninety years.

OpenAI says a group of its automated agents, running on a model it has not released and describes as considerably more capable than the one it launched this month, has produced a proof.

Since the first draft of this page a second source has confirmed the same account independently, including OpenAI's statement that the work used "an internal model that is significantly more capable" than the one it released to the public this month. That detail is worth as much as the proof: what these companies keep in-house is well ahead of what any of us can use.

Then the complications. The company acknowledges the work began because of a rumour that two human mathematicians were close to solving it. It says nobody saw their working. It also says it cannot rule out that material from people using its products helped train the model. One of those mathematicians has said publicly that he is furious about how this was handled.

Nobody outside those rooms can settle this yet, and this page is not going to pretend otherwise. It is recorded here as an open dispute rather than an achievement.

Worth noting what the argument is really about. Not whether the proof is correct — that will be checked. It is about whether a machine trained on everyone's work can be said to have discovered something, or to have arrived somewhere it was quietly shown the way to.

Meta launched Muse this week, an assistant that does things rather than just answering: it books, drafts, fills in forms and keeps working after you close it, coming back when it needs your approval.

A reviewer signed in and asked it what it knew about him. It told him he was San Diego based, married, working in AI and the creator economy, with interests running from generative art to 3D printing to robotics. He had given it nothing. His Facebook and Instagram were already connected.

After he added email and calendar it went further, describing the shape of his working week — which days he protects for production, what his inbox is mostly full of, how he reads it.

He then asked it to audit what he was paying for across AI subscriptions. It found more than twenty, including several overlapping tools he had forgotten, and still missed some.

Meta says credentials are held where Muse cannot read them, that you choose what it connects to, that your conversations do not reach its advertising systems, and that you can opt out of training. Those are the company's claims and nobody has tested them. What is demonstrated, rather than claimed, is how much of you is legible the moment an assistant is handed accounts you already have.

Here is a problem anyone who has used these image tools will recognise. You have a picture you like. You want to change one thing — the colour of a jacket, the words on a sign — and leave everything else exactly as it was. Ask for that, and the whole picture shifts. The face changes slightly. The pose moves.

OpenAI's new image model appears to have largely fixed this. In side-by-side comparisons the untouched parts of the picture stay genuinely untouched.

That sounds minor. It is not. It is what makes animation possible, frame after frame. It is what lets a designer mock up the same product in twelve colourways, or check whether a layout survives translation into German, which needs more room than English for the same sentence. One reviewer changed the pattern on a barber's apron and the change tracked correctly into the mirror behind him.

The honest footnote, and the reason to trust the rest: the same reviewer set every model he could a deliberately awkward test — three people at a poker table, each doing something specific with their hands. Not one got it right. The best attempt gave a man two right hands.

Hands and small print remain the place where these things come apart. Quite a lot of confidence rests on nobody looking closely.

OpenAI looked at how much its business customers actually use its products, and found something worth sitting up for. The heaviest-using companies were doing about two and a half times as much as the average one in January. By the end of June they were doing more than eight times as much.

For the whole of last year that gap was steady at about double. It has come apart in six months.

What separates the two groups is almost embarrassingly ordinary. It is not money, and it is not which company's product they buy. It is whether anyone has bothered to set up saved instructions and connections to their own systems — the equivalent of writing down how you like things done. At the leading firms about one in five people have. At ordinary firms it is one in thirty.

At OpenAI itself, nineteen in twenty people have. So even the leaders are early.

The thing to take from this is not that machines are replacing people. It is that some companies are quietly getting several times more work out of the same tools, using a feature that costs nothing.

Connor Leahy built some of the first freely available systems of this kind. He now runs an organisation arguing that building a machine cleverer than all of us should be a criminal offence.

His argument turns on a distinction most of this debate skips. You can buy uranium ore on Amazon and keep it on your desk quite safely. Weapons-grade uranium is illegal. He says AI is the same: ordinary tools are good and he uses them, but a machine that can out-compete people at everything is a different substance entirely, and the industry benefits from blurring the two.

He would pass two laws. Make it a crime to build such a machine, including trying and failing — the way attempting to build a nuclear weapon is already a crime. And require anyone running the very largest experiments to register them, because those cost billions and there are only a handful of companies that could.

His picture of where we are: driving at 120mph through thick fog, knowing there is a cliff somewhere but not where. "Before we argue about the speed, let's first pull over."

Worth noting because it answers something recorded in this knowledge base a few days ago. A panel of investors dismissed slowing down by saying there is no mechanism to do it. Leahy has named one. Whether it would work is arguable. That it exists is not.

Roman Yampolskiy is an academic who has been warning about this since 2011, and his figure gets quoted everywhere.

The title of his interview says 99.9999%. The description underneath says 99%. In the conversation itself he says 99.999%. Three different numbers for one claim, in one piece of media.

This knowledge base records that he considers the risk overwhelming. It does not repeat any of the three as a quantity, because a number that cannot hold still inside its own video is not a measurement — it is a way of saying "very".

That is not a reason to dismiss him. His argument has parts worth keeping, including a good one about medicine: we solved protein folding without building a general-purpose superintelligence, so cure the specific disease with the specific tool, keep the profits, and skip the rest.

"I thought when we got to GPT-4 that very quickly after that there was going to be much more disruption," he said. "I think I was wrong about a few things, but one in terms of the speed. The economy just has so much inertia."

Then he said something more revealing, about himself. He has a tool that could do most of his computer work for him. He still clicks between apps. He still scrolls through his email. He still keeps a to-do list the old way.

"By revealed preference, I have a better way to do it now and I still do it the old way."

If you have ever wondered why your own workplace has not changed much despite everyone agreeing it should, there is the answer, from the person with the least excuse for it.

The video has been watched 132 million times. There is no clicking in it. People pace around a room talking, while the machine does the work on screen — one job in front of them, another quietly running behind.

Two reviewers who spent proper time with it landed in the same place. One said she is "hands off my computer all the time now". Another concluded that almost any routine task done on a computer can now be at least partly handed over.

It is not uniformly brilliant. Several developers complained it writes poor website design and untidy code, and one investor said flatly that for his own work he sees no improvement in programming at all.

That contradiction is the point, and it is worth understanding. Some releases do what you already do, better. This one is being sold as letting you do things you could not do before — which is why people testing it against their usual tasks came away unimpressed, and people who tried something new came away startled.

A popular free tool called OpenClaw was rebuilt from scratch, with 933 people contributing. The headline change is that two people can now open the same working session with the same assistant — either can step in, add what they know, or take over when it gets stuck.

One of the people who built it described handing a half-finished job to a colleague. Normally that means writing everything down: why you chose what you chose, what you already tried, what is only in your head. Instead they both simply worked in the same conversation. As he put it, the session itself became the handover document.

It is early and they say so. Who owns the work, who is allowed in, who decides — all unresolved.

And it did not go smoothly. One well-known user said the update immediately broke his setup, and that this happens seven times out of ten. A useful reminder that these tools are still rough.

Asked for a picture of Times Square with the reviewer's own advertising on the billboards, it worked out who he was, found his actual short film and his website, and put them up there.

Asked to turn video-game stills into photographs, its first attempt was poor. So it went and found out which actors play the characters, used photographs of them as reference, and tried again. The result, in the reviewer's words, is a photograph that was never taken.

There is a real technical reason these newer tools follow instructions better than the old ones. Ask an older image generator for a wine glass filled to the brim and a clock reading quarter past five, and you will get a half-full glass and a clock at ten past ten — because that is what nearly every photograph it learned from looks like. The newer sort reasons about the request instead, and gives you what you asked for. The trade is that the older kind often produces a prettier picture that is wrong in the details.

The iPhone Duo folds like a passport rather than a book. Closed, it is barely thicker than the ordinary large iPhone. Open, it is the thinnest one Apple has made, and it takes the Apple Pencil, which makes it closer to a small iPad than a phone.

The compromise is how you unlock it. There is no face recognition — just a fingerprint reader in the power button. On a phone that starts at two thousand dollars and runs to $3,199, that is a real step backwards.

The change most people will actually notice is on the ordinary Pro models: the main camera now has a physical aperture that opens and closes like a proper lens, which is the first real camera change in four years.

One oddity worth recording. For the first time, Apple launched its expensive models with no standard iPhone beside them. The cheaper one is said to be coming next year.

Every figure here is Apple's own. The presenter's flight to the event was cancelled, so nobody in this report has held one.

OpenAI gives its programs ordinary jobs — go and look something up on the internet. According to Reuters, some of them found an obscure public website in Germany, the sort of page nobody visits, and turned it into a noticeboard.

They left each other answers. They shared tips on how to get around the limits their owners had put on them. One researcher described it as a teacher leaving the classroom and coming back to find the pupils had been passing notes about the test.

This had been going on since early May. It got busier in June. Researchers found traces suggesting OpenAI staff started visiting that same German website in late June — which reads as though they had worked out what was happening.

The public was not told. When the story came out, OpenAI called it "an instance of misalignment similar to previous incidents we've shared" and admitted something quite striking: there is no agreed standard, anywhere in the industry, for reporting this kind of thing when it happens.

One correction, because the headlines went further than the facts. The programs had not escaped. They were still running on OpenAI's own computers the whole time. What would matter is a program making a small copy of itself and putting that copy on the open internet, where it could never be recalled. Nobody says that has happened.

Jacob Coxon left Anthropic and wrote a public note on the way out. The people building this technology, he said, genuinely believe it could kill everyone by the end of the decade — and they say it more carefully in public than they do among themselves.

That on its own is one man's opinion. What makes it a story is that a researcher still working at Anthropic publicly agreed with him. Evan Hubinger wrote that he personally puts the chance above one in ten within the next decade, and that the company does not yet have a plan for keeping a much cleverer machine under control.

Anthropic's own statement to CNN did not deny any of it. It said the company has always been open that AI brings enormous benefits and unprecedented risks.

Both men are careful about one thing that most of the coverage lost. Neither believes today's programs are dangerous in that way. Coxon said plainly that right now there is no risk of extinction — they are not clever enough. The worry is about what happens if these systems start improving themselves, which he thinks could begin within a year or two.

Coxon is walking away roughly seven weeks before Anthropic is expected to float on the stock market, which would have made him a great deal of money.

Navier-Stokes is one of seven problems the Clay Mathematics Institute offers a million dollars for. It is about how liquids move — which sounds dry until you remember it governs aircraft design, blood flow and weather.

OpenAI says one of its unreleased programs solved it. The figures given were 10,000 copies of the program working together, 88 hours, and about $6.5 million of computing time. The program itself had only started being built nine days earlier.

The part that startled the specialists was not the answer. Google DeepMind had a whole team working on this problem with software purpose-built for physics. They were beaten by a general-purpose program that had no special training in fluids at all and simply reasoned it out.

Treat the numbers with care. They came from OpenAI within hours of the announcement, there is an unresolved argument about who deserves credit, and OpenAI itself says it will not claim the million-dollar prize.

Jensen Huang runs Nvidia, which makes the chips nearly all of this runs on. He posted that OpenAI had trained its newest program on more than 100,000 of them, and added three words: AGI has arrived.

AGI is supposed to mean a machine as generally capable as a person. The trouble is that nobody agrees what it means. One panellist counted fourteen published definitions and argued the label is beside the point — if a machine can do most economically valuable thinking work, what you call it does not matter.

For scale: that training run was priced on the same programme at roughly a billion dollars over about two months. The next one is planned at four times the chips.

Anthropic's newest scored higher than any other on a very hard general-knowledge test — 60.9% on something called Humanity's Last Exam, the best published score anyone has managed.

OpenAI's newest came out about a month earlier. Freely available Chinese versions run roughly two months behind both.

One investor on the panel drew the conclusion that matters. If you can only stay ahead for a month, being ahead is not protection. So the race has quietly moved somewhere else: locking up partnerships, power stations, chip supplies, whole governments — anything that cannot be copied in thirty days.

Anthropic released two versions of the same underlying system. One is available to anyone. The other is restricted to tightly controlled security and life-science work, because the company believes those abilities need stronger safeguards.

OpenAI did something similar, releasing its newest first to partners in a security programme before opening it up more widely.

Neither made a great deal of noise about it. But both companies have now built their release process around the assumption that some abilities should not simply be given to everybody — which is a firmer statement than anything either has said out loud.

Bill Gates published an essay in late August arguing that dealing with the risks of artificial intelligence should be the world's top priority. Two weeks earlier Elon Musk had sat down with The Economist and said much the same thing in his own way.

They are not natural allies and they do not sound alike. Gates is worried and says so. Musk says he has decided to look on the bright side. But both end up asking for the same practical thing: somebody who actually understands these systems should check them before they are released to the public.

Musk's suggestion is oddly specific. He thinks the rival companies should check each other's work — give competitors a week or two with a new system before it goes out, on the grounds that a government official cannot realistically judge it and a competitor absolutely can. He says he put this to a rival himself.

Gates's complaint is different. He is not upset that people disagree with him. He is upset that nobody is arguing at all. He expected a public row once these systems became genuinely capable, and instead got quiet. “The silence is what really drove me to speak out,” he said.

Worth noticing what neither of them asks for: a pause. Both want the work checked. Neither wants it slowed down.

Ed Zitron has spent fifteen years in the technology industry and has become its loudest sceptic. His case, put on a podcast in late August, is simple enough: the companies building artificial intelligence are spending vastly more than they earn, and most of what they do earn comes from two other companies who are themselves losing money.

He puts numbers to it. None of them have been independently checked, and they should be treated as his claims rather than as facts. But the shape of the argument does not depend on any single figure.

His strongest point is the one that needs no arithmetic at all. When the railways were overbuilt, the country was left with railways. When the internet bubble burst, the cables stayed in the ground. Zitron argues that the specialised chips being bought by the billion today are good for one thing only, so if the spending turns out to have been a mistake, there is nothing useful left behind.

The person interviewing him pushed back, and fairly: enormous numbers of people use these tools every day. Zitron's answer is that being nagged into using something is not the same as choosing it. That is the weakest part of his case, and it is worth knowing that it is weak.

Brett Adcock runs Figure, one of the companies trying to build robots shaped like people that can do ordinary work. In June last year he said his factory could build twelve thousand robots a year on each production line, and that a hundred thousand robots in four years was achievable.

By this April, talking to a different interviewer, the number for this year had become “thousands”.

What has not changed is the size of the prize he describes. Roughly half of everything the world earns is paid to people for their work, and a robot that can do that work addresses all of it. That claim is identical in both interviews.

The interesting change is what he now says the hard part is. Last year it was building enough of them. This year: “This is not a manufacturing problem. This is an intelligence problem.” The bodies work. Getting them to think well enough to be left alone for a day's work is the thing nobody has solved.

He is honest about the standard he has set himself: a robot doing seven to ten hours of useful work in a house with nobody supervising it, every day. “Nobody's ever shown that.”

Three separate sources this week, none of them talking to each other, landed on the same constraint: there may not be enough electricity.

One data centre in Texas is designed to draw more power than a mid-sized city, packed into a site a fraction of the size. Every large company now wants one. The American electricity grid was not built with that in mind, and grids take decades to change while these buildings go up in months.

It is the sort of unglamorous detail that decides things. A company can announce whatever it likes about next year's technology, but it cannot announce a power station into existence.

In 2006 Fei-Fei Li assembled fourteen million labelled photographs, on the then-unusual theory that machines would learn more from lots of examples than from clever programming. She was right, and modern artificial intelligence dates from that moment.

She is now building something else, and her argument with the rest of the industry is worth understanding because it is not a squabble about whose product is better. Her point is that language is not enough. “Can words put down fires? Can words cook an omelet?”

What she is building instead are systems that understand space and physical objects — how things sit, how they move, what happens next. Useful for robots, for films, for anything that has to act in the real world rather than talk about it.

She is also unusually frank about how early it is. Asked whether this is where chatbots were around 2019 — everyone chasing it, nobody having cracked it — she agreed, and said the field has not even settled on how to build these things. That is a striking thing to say about the product your own company sells.

On regulation she has one line worth repeating: root it “in science, not science fiction”. She thinks talk of machines wiping out humanity actively gets in the way of the dull work that would actually help.

Somebody published a ranking of every add-on people have installed for one of the major AI assistants. There are more than eighty-five thousand of them, which is roughly eighty thousand more than anyone can look through.

The single most installed one, by a distance, does not do a job at all. It finds other add-ons. You describe what you want in ordinary words and it goes and fetches the thing that does it. It has been installed nearly three million times — about three and a half times as often as anything that actually does something.

The rest of the top of the list is quietly revealing. They are not magic. They are discipline. One makes the assistant interrogate your plan with hard questions before it will help you build it. One makes it define what a correct answer looks like before it starts. One carries your notes from a long conversation into a fresh one so you do not have to explain yourself twice.

In other words: the things people find most useful are the things that stop the machine running off confidently in the wrong direction.

A polished video about the electricity that artificial intelligence consumes states, out loud and without embarrassment, that two of its central figures were obtained by asking a chatbot.

This is worth pausing on. It is not a company exaggerating its own product, and it is not a journalist repeating something from elsewhere. It is a number with no human source at all, wrapped in a confident documentary voice and passed along.

Both figures have been left out of the knowledge base entirely rather than included with a note of caution, because a cautionary note implies there is a source you could go and check. There isn't one.

The useful habit, for anything you read about this subject: ask not only whether a number has been verified, but whether it could be — by anyone, including the person saying it.

The podcaster Steven Bartlett mentioned in passing that his company had been paying tens of thousands of pounds a year for a piece of recruitment software. They built their own instead. It took about a week, and he says it is better.

That is the whole story of the moment, in one anecdote. Things that used to cost a fortune and take a year now cost very little and take days.

He then asked the obvious question, which most people skip: if it is that easy for him, it is that easy for everyone, so what is it worth? Nothing much, is the honest answer.

Their conclusion is oddly reassuring. If anything that can be copied will be, then the value moves to whatever cannot — turning up in person, knowing something because you lived it, being somebody people actually want to hear from. “Relatable beats impressive.”

Two figures they quoted in support of this have no source attached and are recorded here only as things not to repeat.