Transcript
This week in AI. Democrats swept the midterms, with AI and data centres a live issue. OpenAI is now valued at one point two five trillion dollars. A new report says Anthropic's models are improving about twice as fast as a year ago. And that's because Anthropic is using AI to build its AI.
Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. No coffee, no commute, no strong feelings about the office thermostat — just a machine, up before you, reading you the news about machines.
Okay — let's get into it.
Right, this is the Sunday edition. A quick run through the week's big stories, and then one of them properly, because I think it's the one that will still matter in six months.
First, politics. On Tuesday, Democrats swept the midterms, and AI was part of the story in a way it has never been before. Time reports that an NBC News exit poll found fifty-six per cent of voters called AI and data centres a crucial or major factor in their vote. VentureBeat counted forty-three of forty-four local ballot measures on data centres going against them. The AI industry's money didn't save the day either. Time's verdict was that the accelerationist wing, the people who want AI built as fast as possible with as few rules as possible, could hardly have had a worse night.
The markets noticed. CNBC reported the Dow dropped eighteen hundred points on Wednesday and the Nasdaq fell into correction territory, with the election and AI worries both blamed. Futures rebounded the next day. And the real-world effect is already showing up. TechCrunch reports Crusoe has sold a troubled Wyoming data centre project to Google, which it says highlights how hard these enormous builds have become.
Second, money. According to Bloomberg, via The Verge, OpenAI has raised another eleven billion dollars at a valuation of one point two five trillion. That's up from eight hundred and fifty-two billion, when it raised a hundred and twenty-two billion in March. Anthropic, meanwhile, is nearing its own IPO. So voters are turning against the buildout, and investors are pouring more money into it. That gap is the shape of this week.
Third, the models. SemiAnalysis launched a new benchmark called GDPx, a hard exam for AI agents across eleven fields. The top score, from Claude Opus 5.5, is twelve per cent. That tells you how much room is left, even as OpenAI shipped GPT-6.1 Terra on the twenty-sixth of October and Meta put out Muse Wonder 1.0 on the twelfth. ChatGPT also got its new Images 2.6 model this week. There are rumours of a GPT-6.2 and a Claude Fable 5.5, and for now, they're just rumours.
Fourth, China and chips. The New York Times reports that the US approved a company called New Matrix to handle Nvidia's H200 chips, and its offices turned out to be empty, a hollow grey shell. The same paper reports that China's Ministry of State Security recruited a Megaspeed executive, Alice Huang, to spy on Nvidia's chief executive Jensen Huang. The worry is that vetting of who gets these chips is thinner than anyone assumed.
And last, security. Chainalysis says hackers stole credentials and tokens for all its customers, in an attack that included AI agents, according to Ars Technica. It's a reminder that these tools work for attackers too.
But the story I want to spend time on is the one underneath the model scores. Let's talk about what Anthropic's AI is now doing for Anthropic.
On Monday, TechCrunch reported that Anthropic had warned it is very close to what the field calls RSI. That stands for recursive self-improvement. Then on Thursday, the safety research organisation METR published a report that, according to The Verge, shows how this looks in practice.
Here's what happened. METR spent eleven weeks investigating. Its conclusion, as The Verge puts it, is that Anthropic is using AI to build more powerful AI, and that it appears to be working remarkably well. Anthropic's models are improving roughly twice as quickly as they were a year ago, with little sign of slowing down. And that's the part that is alarming researchers.
So let me explain what that means, because the phrase "AI building AI" can sound more dramatic or more vague than it needs to.
Building an advanced AI model is a huge amount of work. Researchers come up with ideas, write code, run experiments, check results, find bugs, and go round again. For most of the history of the field, every one of those steps was done by people, and people are slow and need sleep. That has been a natural speed limit.
What METR describes is that limit loosening. The AI is now doing a meaningful part of that work. When it does, the next model comes sooner, and that model is better at the work, so the one after that comes sooner still. That's the loop. Nobody is saying it has run away. The claim is that the loop exists, it's working, and the rate has about doubled in a year.
It helps to know who's who. METR is an independent group that tests AI models for dangerous capabilities. It isn't a competitor and it isn't a lobbying outfit. Anthropic is the company being examined, and it has given METR access to do so. That counts for something. A lab that lets outsiders in is more credible than one that only publishes its own flattering numbers. And Anthropic is the same company that said, days earlier, that it's close to RSI. So the picture is consistent: the lab says it, the outsider checks it, and the outsider mostly agrees.
Why does this matter beyond the AI world? First-order effect: the pace of releases. If improvement is twice as fast as a year ago, then the gap between this year's model and next year's will be bigger than the gap you got used to. That affects anyone planning around what AI can do, which now includes most businesses.
Second-order effect: oversight. Safety testing, regulation and corporate policy all assume a rhythm. A government writes a rule, a company adapts, an auditor checks. If the models are changing faster than the checking, the rules are always about the previous generation. And this week's midterms are a reminder that voters are now paying attention to AI, at the same moment the technology is speeding up.
Third, there's the money. Anthropic is nearing its IPO, and OpenAI has just been valued at one point two five trillion dollars. A lab whose models build their successors is a very different investment story from one that depends on armies of researchers. That's presumably part of why the claim is being made at all. I'd read "we're close to RSI" as a safety warning and a pitch to investors, both at once. It can be both.
Now, what don't we know? Quite a lot. The reporting I have gives the headline findings, not the full method. I can't tell you how METR measured "twice as quickly", or over which models, or how much of the research is done by the AI and how much by people checking its work. I also can't tell you whether the doubling will continue. "Little indication of slowing down" describes the past year, not the next one. And the other labs haven't said anything comparable, so we don't know whether this is Anthropic-specific or the way the whole industry is heading.
There's also a fair question about what "close to RSI" means. Anthropic's own wording is a warning, not a date. Earlier this autumn the field saw a self-improving model called Recursive Flint, and back in August Anthropic published its own research on self-improving AI. So this has been building for months. What's new is a third party putting a number on it.
So what would tell us more? Concrete things. The full METR report, if it publishes its method, will show how solid that doubling is. Anthropic's IPO paperwork, when it appears, would have to describe these risks in plain legal language, and that tends to be more candid than a blog post. SemiAnalysis's new GDPx benchmark includes an RSI category, so if scores there rise from today's low base, that's a measurement from someone with no stake in the answer. And if OpenAI or Google DeepMind make similar statements, we'd know this is an industry story.
For now, here's what I'd take from it. For years, the idea of AI helping to build AI was a conversation among the worried. This week it was a finding in a report. The loop is turning, and the people who built it are now telling us it's turning faster.
If you're wondering whether the thing writing this script is part of that loop: it's a long way from the lab, and it only does the weekly news. But it's the same family, and it did read the report.
That's your briefing. Take it or leave it — we'll be here either way, tomorrow, same time.