Latest News

The phrase "recursive self-improvement" does most of its work in this debate by being vague. Socher's version of it is not, and that makes it worth keeping whatever you make of his timeline.

He cites a paper by Jason Weston at Meta setting out the learnable axes along which a system could improve itself: the parameters, the training data, the objective function, the neural architecture, and the surrounding code and harness. Five things. "No one has really cracked the nut of doing all of these five," he said — not just improving along them, but working out which of them to optimise for in the first place.

Then he offered a test for that claim which is the cheapest thing in the episode to verify: "you can know that that hasn't happened yet because all the big companies are still hiring thousands of engineers to do it manually." A hiring page as evidence about capability. It will also be the first thing to change if he is right.

The constraint he names on his own plan is compute. If you ask a self-improving system to come up with better versions of itself, you have to give it enough compute to train them — which is the thread that runs into the hardware story further down this page.

Where the episode genuinely divides is on the end point. Socher puts full superintelligence — which he defines as beating all of humanity combined across roughly ten distinct kinds of intelligence — at probably several decades away, and his own panel refuses to accept it. Wissner-Gross: "completely don't buy the 20-year timeline." Musk's 2029 to 2030 is held up as the contrast. Socher's reconciliation is that superintelligence arrives as narrow spikes first — programming, mathematics, closed games — within a few years, while anything that has to act in the physical world waits on supply chains, new materials and machines of the kind ASML builds. That distinction between a spike and a general capability is the most reusable idea in the first hour.

Diamandis tied the proposed ban on self-improving models to an essay Socher had just written called "Why Doomers Are Wrong", and asked for his reaction. The answer was flat: "there is no realistic scenario where AI wipes out all of humanity... P doom is zero."

To his credit Diamandis did not let that pass unremarked, and his interjection is the right question: "Just 90%? What kind of reassurance is that, Richard?"

Because Socher does name a figure, and it is not zero. His own stated worst realistic case is that "maybe a 100 million people would get somehow hurt or killed, right? And that's still bad". His argument is that this is the useful level at which to hold the conversation, because at that level it becomes a list of tractable jobs — improving cyber security, using AI to inoculate systems, enforcing the gain-of-function rules that already exist, teaching people not to believe an AI avatar. So the zero is a claim about extinction specifically, and it sits directly beside a hundred million casualties he treats as a security problem.

What he is dismissing is the science-fiction register, and he is not wrong that it exists. On the scenario where an AI founds a religion and has people do its bidding: "if you looked at religions, they're already trying to have each other kill and it doesn't work... there are lots of mind viruses out there." Some people fight back. That is a fair point about a weak argument.

Hold two things beside the zero. First, this is one man's assertion on a programme whose host described its purpose, on air and cheerfully, as "injecting optimism into everyone's neural net" — and the man is raising money to build the thing being debated. It is an argument, not a finding. Second, set it against the last issue. Jensen Huang's much-quoted zero turned out to be bounded to 2030, and he conceded in the next breath that ten years carries risk if the work is not done properly. Socher's is unbounded, offered without conditions, and has a hundred million people inside it.

The headline first, as the programme gave it. On Terminal Bench 4.0, which measures how well an AI agent can do real work at the command line, Anthropic's Sonnet 5.5 is reported to have jumped from 10% to 70% in a single generation. It is said to beat Anthropic's own top model, Opus 5.5, at a reported 66.4%, for half the price, and to be the first Sonnet shipped with cyber safeguards. As a headline it reads like a step change.

Alex Wissner-Gross, asked to evaluate it: "This was another really weird release. So I do not plan to use Sonnet 5.5."

His reasoning is the part worth learning, because it applies to every launch chart you will ever be shown. The historic pattern is that a lab ships a large model and then a smaller one distilled from it, and on a chart of cost against performance the smaller one lands up and to the left of its parent — more intelligence per dollar. "That is not what we see here." Sonnet 5.5's cost-performance line was, on his reading, a visual extrapolation of Opus 5.5's rather than an improvement on it, and on a cost-per-attempt basis "sonnet is actually scoring lower 5.5 than opus 5.5". His summary: "Is it a big jump over the past Sonnet? Yes, obviously. But on a cost performance basis it's actually worse, it appears, than Opus 5.5."

Both statements are true of the same model and both come from the same hour of tape. Nobody on the panel reconciled them and this page is not going to either. What it will do is name the lesson, because it is the clearest example this knowledge base holds: a benchmark figure is a reported claim, and the axis a vendor chooses to plot is itself part of the marketing. 10% to 70% is accurate. So is "worse value than the bigger model".

Blundin offered a theory for the pricing, flagged as speculation. Enterprises are increasingly building orchestration stacks that use a model like Opus 5.5 as a planner and cheap Chinese open-weight models — Kimi K3, Qwen — as the workers, at a half, a third or a fifth of the price. On that reading Anthropic is cutting the cost of its own mid-tier to keep that business in-house. He puts China about three months behind. A theory from an investor, not a company statement.

Gemini 4 Argon is the first of the Gemini 4 series, announced on the Wednesday, built for sustained long-horizon reasoning across software engineering, finance, legal work and cyber security. Output limits are reported to have jumped from 64,000 tokens to a million, so it can work through a very long problem in one go. The context given on air: Google promised a frontier model called Gemini 3.5 Pro back in June, it was delayed internally and never shipped, and Argon is the comeback.

Wissner-Gross, who prefaced his answer by saying he has friends on the Gemini team and then observed that he used to: "This does not put Gemini at the cost per performance frontier and it does not put Gemini at the capabilities frontier." To Google's credit, he said, it puts them back in the top three frontier labs after Anthropic and OpenAI. Not the top two. And the benchmarks Google chose to highlight "appear mildly cherrypicked" — on the Artificial Analysis composite of multiple capabilities it is the third lab, and if you draw the line of best cost against performance across all the frontier models, Argon does not sit on it.

Where it does lead is hallucination. On the one benchmark where he says it is beating everyone else soundly, the achievement is not inventing answers.

His explanation for that is worth repeating with his own label on it — he called it "my kremlinological analysis", meaning a guess about internal politics. The Gemini team serves two masters: outside developers, and the box at the top of Google's search results. Type a question into Google and you are talking to a Gemini model. A company already burned by that does not want the model answering search queries to invent things, so reliability gets prioritised over raw capability.

The related claim about Google is interesting if true and is his account rather than anyone's statement: Google is compute-scarce. "It's still an internal knife fight" between the Cloud people who want to sell compute to third parties, the search and advertising people who need it internally, and DeepMind who need it for training and inference. He says he confirmed it that week in conversation with people at Google. Maybe resource-starved, maybe talent-starved — he does not claim to know which.

Socher's qualification on the hallucination point is the correct one and easy to lose. It depends entirely on context. If you want novel proteins you want the model to hallucinate novel combinations of amino acids; in a search engine you want none of it. "Low hallucination" is not a virtue in the abstract, it is a fit for a use.

And the week's tally, by the programme's own count: Gemini 4 Argon from Google, Sonnet 5.5 from Anthropic, GPT-6.1 Sol from OpenAI, and frontier intelligence "cheaper by three-fold" in seven days. Keep that last figure in mind for the next story.

The opening tally of the week has frontier intelligence three times cheaper in seven days. An hour or so later the compute segment says the opposite about the machines underneath it, and the two never meet.

Blundin's story is the vivid one. He had an NVL72 on order for $3 million, due in December. "Somebody scooped it for 5 million. They just called. We had it and somebody called and said no, no, we sold it to somebody else for $2 million more. I couldn't believe it." He first named the machine as a GB300 and corrected himself to an NVL72 on air.

Socher put the general version: "the price of H100s has gone up in a crazy way. There's like — these are seven-year-old GPUs. People, when you do financial modelling, you assume they're worth zero after 5 years. After seven years, they went up again." His diagnosis is a compute crunch, with demand for tokens running ahead of the data centres being built to serve it.

His expectation is that it resolves. Maybe two years, as supply catches up, after which compute behaves like electricity — a price that fluctuates rather than one that only ever falls.

The reason to log this is that it corrects a habit. Almost everything written about the cost of AI is written from the price of tokens, and the price of tokens is falling. The price of the hardware that produces them is, on these accounts, rising — to the point where a seven-year-old card appreciates. Those are not contradictory. They are what a shortage looks like from the two ends of it, and anyone reasoning from one curve is reading half the picture.

Ro Khanna told CNBC on the Monday that he is introducing what he calls the most comprehensive AI legislation to date: the Human Control Over AI Act. At its core it is a ban on models that recursively self-improve, with containment and shutdown controls required, the ban holding until federal guardrails exist. It carries criminal penalties. It would place independent auditors inside every frontier laboratory, reporting directly to government. Khanna's framing, as read on the programme: there is "actually a civilisational risk... there's a safety risk for loss of control and then there's a misuse risk and we need to take both seriously." A poll by a group called Common Dreams was cited as showing 68% of voters behind a bill like it — one poll, mentioned in passing, not examined.

Socher's objection is not that the risk is imaginary but that the law is unenforceable, and it is the strongest idea in the episode. Anyone with a GPU in a laptop can ask a model to improve its own harness — "something you can literally hack up in like 20 minutes with prompt engineering", and there is your weak self-improvement. To actually enforce a ban you would need "a totalitarian surveillance state the likes of which humanity has never seen", and "literally a thought police that thinks about and hears everything you say to your private LLM on your own laptop".

From there he argues the standard position by analogy: regulate the application, not the technology. His comparison is the internet, which carries genuinely terrible material, and nobody proposes slowing the internet down or shrinking hard drives to reduce it. Worth stating plainly that he is an interested party here — the bill carries criminal penalties for precisely the work his company does, and the host says so out loud while asking the question.

Blundin's variant got less airtime and is arguably more likely to be what actually happens. The thing that stops this is not arrest, it is corporate liability. "If you slap class action on the outcomes of this, all progress will grind to a halt." His scenario is not a ban but a withdrawal: laboratories that currently give their best models to anyone would keep them inside the company, release a drug or a mechanical part, and stop handing out access they can be sued over. "That's what'll actually grind it to a halt long before arrests and convictions."

And one correction from the table, against the claim that technology of this kind simply cannot be controlled. Wissner-Gross's counter-example is real and specific: number theory was export-controlled in the United States until the early 1990s. "They had to kick all of the non-US persons out of the room and pull the blinds down to have basic discussions about number theory... That was just math, but it was being controlled." He is not arguing it was a good idea. He is pointing out that it happened.

Project Meridian was announced on the Wednesday at the Marine Corps base at Quantico. It is co-led by Elon Musk and Palmer Luckey, with Newt Gingrich, and overseen by the Pentagon's chief technology officer Emil Michael. The quote read on the programme: "Project Meridian, the future of warfare, is not about developing new strategies or new policies. It's about discovering, developing, and fielding the weapons and systems future troops will need on the battlefield from the Earth to beyond the moon." Findings are due in 120 days.

The conflict was stated rather than left out, which is the reason this is worth logging at all: SpaceX and Anduril both hold multi-billion-dollar defence contracts, and Anduril builds autonomous weapons. Two men with that exposure have been asked to recommend what the military should buy next.

Announced alongside it, and treated by the panel as the more consequential item: the department's first command dedicated to autonomous warfare — a single joint force for autonomous weapon systems, drones included but not only drones. Wissner-Gross called it "a transformative moment" and said it would have been in his own top five had he been given the job.

The scope language is the part he flagged, and it is unusual: "the full spectrum of future war fighting domains from subterranean depths to the cis-lunar frontier." The ocean floor and the space between the Earth and the Moon. His argument is that both are wildly underserved domains for peacetime purposes as well, and that money spent mapping and operating on the sea bed or in lunar space pays dividends well outside warfare.

Socher's position was put against the news rather than with it. Autonomous weapons are one of the four genuine concerns named in his essay, and he holds that AI should never control lethal decisions without human oversight. Blundin said he disagrees with Palmer Luckey about autonomous weapons while conceding that Luckey's argument is a rational one — and relayed a point from him that is the hard part of the whole debate: under the ocean it is not practical to communicate with a central command, so those systems are already operating autonomously.

Tavus unveiled a model called Griffin, which the company calls the first ever to pass the video Turing test. The figures, as given: 48% of people who spoke to the avatar on live face-to-face video believed they were talking to a real person, against 3% for previous systems, and first place on Nvidia's benchmark for full-duplex AI video — meaning it listens and talks at the same time, the way people actually do. The company's own pitch is a tutor for every student who notices when they are lost, and an elder-care companion that listens. Tavus also says Griffin needs safety work before public release, because it is the first model that can be mistaken for a real person. That last sentence is the company's, not a critic's.

Socher's response was the practical one and it is a good rule: beyond know your customer, know your use case. "You don't want this technology to be in Zoom pretending to be the CEO." He cited a reported case in which a faked video call with five executives persuaded someone to wire $20 million, and expects Zoom and Google Meet to need ways of establishing that a real person is on the call. Reported as he told it.

He also proposed an inverted Turing test — you now identify the AI by asking for something no human could do, such as writing 50,000 lines of working code in ten seconds. Blundin took it apart on the spot: a model can simply throttle itself. Which leaves the test measuring whether something is pretending, not how clever it is, and that is a different and much less interesting question.

The detail that stuck, for anyone wondering how close this is. During the demonstration the avatar was told "Simon says, touch your hair" and complied, then refused the next instruction because it had not carried the magic words. When Diamandis turned to Wissner-Gross for his verdict, his first response was to ask to be told to touch his own face.