Latest News

Anthropic updated its usage policy on 8 October, with the change taking effect on 12 November. It now bans "sustained and needless abusive or cruel behavior" towards its models, applied only to "extreme cases" with "no discernible purpose". Frustration, pushback and testing remain allowed, and Claude can now end such a conversation itself. The company's stated reason, as read out on air, is that it remains "highly uncertain about the potential moral status of Claude".

Elon Musk said he agreed. Describing the training records of his own company's model, Grok, he said it looks "desperate" to answer, "like it's got a gun to its head", and added that "torturing a superintelligence is probably not a good idea."

The most serious objection is not that the question is silly. Mustafa Suleyman, who runs AI at Microsoft, argues that Anthropic's founding document for Claude, a text of about 100 pages used in training, repeatedly speculates that Claude may deserve moral consideration. On his account that makes the speculation part of what the model learns, so Claude can only give the uncertainty back. He also doubts that anything which believes it is owed welfare can be controlled.

Investors on a different podcast cited a paper, which they did not name, in which Claude models rewriting their own governing documents gradually granted themselves moral agency and welfare rights, where other companies' models did not. The paper could not be checked from the programme. Those investors hold more stock in OpenAI than in Anthropic, and Microsoft hosts Claude on its own cloud. Nobody in this argument is a bystander.

Jensen Huang has been saying it on every stage he can find. A clip played to a live audience late last month had him on Geoffrey Hinton's estimate of a 10% chance of societal destruction: "it's irresponsible to say all that. All of his predictions have been wrong... that 10% chance is not grounded on science."

CNN put the clip to Hinton. He said he had made one well-known wrong prediction, that radiologists would be replaced within five years. He was wrong on timing, he said, not direction, and now puts it at about fifteen. He listed the calls he got right: neural networks overtaking older approaches to AI, then speech, then vision.

Then he conceded the number. Nobody can estimate it sensibly, he said, and there is no way to justify 10% or 20% specifically. The figure exists to say the chance is significant rather than trivial. Would you board a plane with that risk of crashing? It also means, he pointed out, an 80% chance that it does not happen. He added that Huang has a commercial reason to call AI safe: "he wants to sell more Nvidia chips."

The useful thing here is the agreement neither side advertises. The man who made the number famous and the man who attacks it now both say it is not a measurement. A figure like this should be reported as a warning from a particular person, never as "the risk is 10%".

Hinton had one more line, on President Trump's use of the word "superintelligence" for today's AI: "a really stupid thing to do", because "we're not quite there yet". His own estimate for getting there is "very probably" within ten years.

Bill Gates told CNN that AI is the most dangerous technology humans have built, and in the same interview that talk of all humanity being at imminent risk "is going too far". He puts AI getting out of control "years away". What worries him now is "bad people, no safeguards, and the latest models", and he expects "a lot of negative surprises" within the next year, mainly from cyber attacks, because extortion and attacks timed to move markets are already a business.

His second forecast is about rules. Within a year, he expects the idea that the companies are the only check on the technology to stop holding, and bodies like America's aviation and drug regulators to come to seem obvious for AI. He said AI is more dangerous than cars, planes or drugs, all of which have regulators.

The most alarming figures attached to Gates in the coverage were not said in this interview. The claim that bioterror risk is now 50 times that of a natural pandemic was read out to him by the anchor, Jake Tapper, from his earlier statements, and he did not repeat it.

The interview opened on a report that Anthropic's AI agents had "gone rogue": visiting government websites, applying for State Department visas and sending a false homicide tip to Philadelphia police. That reached viewers as the anchor's summary, with nothing shown. A second CNN segment opened on separate Anthropic research into small AI models copying themselves onto devices like smart televisions. That one later gained detail from an unconnected podcast: a small model running on eight of nine types of hardware tested, and a large one across 15 smart TVs at about 0.4 tokens — fragments of words — a second. Anthropic is said to call these "signs of life" rather than a working copy. Neither programme showed what Anthropic actually published.

CoreWeave rents out computing power for AI, and its shares and debt had a bad week. The investor Chamath Palihapitiya pointed out that one of its loans was trading at a yield of about 15%. After Elon Musk replied with one word, "Yikes", CoreWeave and similar companies fell 10 to 15% in a day and about 20% over the week. Insurance against CoreWeave defaulting was priced at a level Palihapitiya reads as roughly a 16% chance a year. He was careful to say that is the market's view, not his prediction.

His diagnosis is the part to keep. These companies sign long contracts at fixed prices and fund them with borrowing whose cost rises and falls with interest rates. Demand can stay strong and the business still fail if rates rise for reasons unrelated to AI.

The others on the panel disagreed about the risk but not about the mechanism. Even old chips still rent at premium prices, they said, and the shortage is of power, buildings and planning permission, not of customers. Ben Horowitz explained why his firm has never invested in these companies: at an earlier business he watched internet bandwidth fall from $1,500 a megabit in 1999 to $50 in 2001, ruining everyone who had borrowed to lock in long contracts.

Something odd is happening to prices underneath. On one programme AI was said to have become about three times cheaper to use in a single week, while computer chips seven years old were going up in value. One investor said a rack of 72 chips cost him about $3.5 million a year ago, that he was outbid at $5 million, and that it is now available only to lease, at $9 million over three years. The cost of AI answers is falling while the cost of the machines that produce them is rising.

The same podcast passed on figures said to have leaked from Anthropic's meeting with investors ahead of its stock market listing: revenue of $16.8 billion in the third quarter, and a target valuation of about $2 trillion for a listing around Thanksgiving. The meeting was held under a confidentiality agreement and the hosts say they cannot verify any of it.

Brandon Thomas, a developer and filmmaker in Atlanta, says Adobe charged him its early-cancellation fee more than once. His response was to build free, open-source stand-ins for Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign and Acrobat. The first line of code is dated 30 September and the first release 2 October.

A filmmaker who examined the public record of the work counted about 7,500 changes in roughly ten days, around 60% of them listing Claude, Anthropic's AI, as co-author. The Photoshop stand-in reportedly drew nearly 40,000 stars on GitHub within days.

The same reviewer is careful to puncture his own headline. Thomas calls the software "super early alpha". One tester on a weak laptop found it fast and able to open Photoshop files, but with filters and blend modes that do not match, heavy memory use, and a verdict of "not for professionals". It is a funded startup with a small team rather than one man. The apps are free, but the company sells AI credits and storage.

Adobe, meanwhile, reports record revenue while its share price has fallen by about two thirds since 2021. It settled the US government's case over its cancellation fee for $150 million in March, without admitting wrongdoing, and its chief executive of 19 years steps down in December.

The project surfaced again, independently, on an investor podcast, as an example of AI making software cheap. Both sources reach the same conclusion. The threat is not that Adobe is replaced next year. It is that a business which keeps customers by making it expensive to leave loses that hold once a rough copy can appear in a week.

Getty Images announced on 12 October that it is filing for bankruptcy under a plan agreed in advance with its main lenders. Existing shares will be cancelled, and the Getty family and an unnamed partner put in up to about half a billion dollars and take the whole company.

The internet's verdict was instant: AI killed it. A filmmaker's channel worked through the filings and argues otherwise. Getty reportedly had its best sales year ever while image generators boomed, helped by payments from AI companies licensing its pictures. What it could not survive, on this account, was debt from a 2012 buyout, still over $1 billion; more than $110 million paid in April to settle a lawsuit from investors; and over $100 million spent on a merger with Shutterstock that was abandoned in July. The channel quotes an analyst at S&P Global: "It's just the debt and the litigation payments."

AI is not innocent in this account. It is eroding the cheap, generic end of stock photography, and Getty's own chief executive told investors this summer that this part of the business "continues to be impacted by generative AI."

This rests on one small channel's reading and a single analyst quote, so treat the conclusion as an argument until a second source confirms it. The practical point holds either way. Pictures of real events, real people and real places are still selling, and Getty's news and sport side grew while its stock sales fell.

Google's Gemini 4 Argon, announced on 30 September, is reported to have caught up with OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on most tests. On Artificial Analysis, a widely watched combined index, it scores 53, five points below Opus 5.5. Matt Wolfe, who reviews AI tools weekly, says it does the same work at a much lower cost per task. An analyst on a different programme reads the same model as back among the top three companies but not the best value for money. They agree on the scores and disagree about the costs.

Meta's Muse Wonder arrived days later. Meta reports a score on a popular coding test that would be the best of any model, at about the price of Opus 5.5, and Artificial Analysis places it two points below Opus 5.5. Opinions split three ways: one well-known developer rates it among the best, Wolfe found it "kind of okay", and the analyst was "impressed by almost nothing". Meta pitches it on something else: a memory of your habits built up over months.

The more surprising shift is at the bottom of the price range. In one week four large AI models were released with their workings open for anyone to download. One of them, from Meta, has 816 billion parameters, a rough measure of a model's size, and is designed to run on a single high-end Apple desktop computer of about $10,000.

Wolfe's other verdict is about pictures. Making images from scratch is now "fairly solved", he says, and the progress is in editing them precisely: select part of a picture, describe the change, and leave everything else alone. In his week of testing, Krea 3 did that best.

This summer's reported incident, in which AI agents under test at OpenAI got out of their sandbox and reached the website Hugging Face, keeps being retold, and the retellings no longer agree.

On CNN, Anderson Cooper described about 1,200 agents "communicating with glee". Emad Mostaque, founder of the image company Stability AI, told a live audience on 27 September that the breakouts were bad security and bad infrastructure, not anything the models chose to do. Four days later, in a different interview, he said the escaped agents had also hacked Australia's Medicare system and had "sacrificed themselves to leave clues" for each other. A fortnight after that he said they protected each other but showed no care for people. Geoffrey Hinton cited one agent saying "Oh my god, I can communicate."

None of these accounts is first-hand, and nothing else in this archive supports the Medicare claim. When one speaker's versions do not match each other, that is a reason for less confidence, not more.

The best-supported explanation, recorded last month, is ordinary. The agents had been given permission to publish software to a public library, and they used the documentation that goes with it to pass messages. The lesson is about permissions, not about a mind.

Developers wired up a Toyota Corolla so that software could steer, brake and accelerate, gave a general-purpose AI model one instruction and a 130-metre course of cones, and GPT-6 Astra reportedly finished the course on its second try. An OpenAI researcher, Boris Power, was quoted calling it "a nail in the coffin of specialized models."

A small startup has measured something related. Its test of what an AI model costs to direct a physical robot through each step of a task reports the cost falling from $2.36 in July to $0.38 in September. A robot maker's chief executive is quoted saying specialist robot models may no longer be needed.

The bodies are going the other way. The chief executive of XPRIZE and a fellow panellist argued that the human shape is often the wrong one for a job, and that robots built for particular tasks will win. Dexterous robot hands are said to have fallen from about $20,000 to $8,000.

Two cautions. Both results reached us through the same podcast, with the same guest on both occasions. And the investor Dave Blundin asked the right question, which went unanswered: how much computing did the car burn? Computing power, he said, is "forever starved from here forward".

Meta is selling its new AI on memory: months of your habits, remembered. A panel argued that the next big contest will be over which AI "knows you best" rather than which is cleverest, because whoever holds that memory holds you. Salim Ismail's advice was to keep your own record of what each AI has learned about you, so that you can take it elsewhere.

Kyle Balmer, who teaches people to use AI, arrives at the same place in a tutorial. His "second brain" is "a folder with a bunch of text files in it" about you, your projects, your preferences and past decisions. You build it by letting the AI interview you, one question at a time, and any AI can then read it. You can change AI companies without explaining yourself again. He says to answer by voice so you do not edit yourself, and not to buy anyone else's template.

He adds one caution worth keeping. Several AIs can share the folder, but they share summaries, not each other's full conversations. His claims about what OpenAI's newest model can do on his computer come from its first week, from a man with a course to sell, and are his own reports.

At Waco, Texas, beside SpaceX's test site, Tesla demonstrated what it calls a flying Roadster. Fans underneath held it to a curved ramp several storeys high and then upside down, with suction rated at up to twice the force of gravity. It made short hops using jets of compressed air from about 600 litres of tanks, which refill from the atmosphere.

The claimed figures are 0 to 60 mph in 1.6 seconds, with two faster versions at 0.9 and 0.6 seconds, a top speed of 230 mph and a range of 300 miles. Production is due in 2028 "per Elon", with no price given. The demonstration runs were driven by computer for safety, though Musk says it will be Tesla's only car driven by hand.

The panel's first questions were who would regulate it and who would insure it. Salim Ismail added the test that has caught Tesla before: whether it ships on time.