AI in 15 — November 03, 2026

An AI just opened up another AI's brain and changed sixteen thousandths of one percent of its weights by hand. There was no training at all, and the smaller model's math score went from eight percent to forty-eight.

Welcome to AI in 15 for Tuesday, November 3, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. And before anyone panics, Kate, that forty-eight percent comes with an asterisk. We'll get to it.

Of course it does. Here's what's coming up. MIT researchers let GPT-6 Astra rewrite other neural networks directly.

Anthropic's updated IPO filing has seventeen billion dollars in quarterly revenue, and growth is slowing.

A critical Linux flaw that Claude reportedly turned into a remote exploit is still unpatched on Ubuntu.

China drafts export controls on frontier AI models.

Plus glasses that read your lips, AI-designed viruses given to a patient, and mathematicians talking about ego.

Marcus, this was the top story on Hacker News with more than two thousand points. What did the MIT team actually do?

It's a group at MIT CSAIL that includes Omar Khattab, who created DSPy. They call it Weight-Space Programming. Normally, if you want to change a neural network, you train it. You run data through it and nudge millions of numbers a tiny bit at a time with gradient descent. Here they skipped that. They gave a frontier model, OpenAI's GPT-6 Astra, the raw weights of another network and asked it to reason about them and then edit them by hand.

And it could read them?

Surprisingly well. They gave it four small networks trained on different Atari games, Breakout, Pong, Pac-Man and Freeway, and it worked out which network played which game from the numbers alone. It identified which shape each small image model draws. Looking only at the middle layers of a little transformer, it worked out which language that transformer had been trained on. [HOST> That's like reading someone's diary by looking at the ink molecules.

Pretty much. Then it got to writing. There was a car-on-a-ramp control policy that failed every single time. Astra changed three and a half percent of its weights in one go, and success went from zero to a hundred percent across a thousand and twenty-four runs. It also wrote Breakout networks from scratch that scored about five hundred and seventeen. A network trained the normal way for five hundred million steps scored twenty-five.

Okay, and the headline number. The Qwen math jump.

They took Qwen 3.5 at zero point eight billion parameters, a small model that tends to get stuck repeating itself forever. Astra traced that looping to a single direction spread across two layers. It edited a hundred and thirty-one thousand weights, which is zero point zero one six percent of the model. MATH-500 went from seven point nine to forty-eight point one percent. GSM8K, a grade-school maths benchmark, went from about fourteen to sixty-seven. Instruction following roughly tripled too.

So where's the asterisk?

There are a few, and the top Hacker News comment was basically "read the paper before you lose your mind." First, those gains assume a budget of fifty forward passes, so the model gets fifty runs to land on an answer. With ten to twenty-five, the looping barely improved. Second, they tried the same thing on Gemma 3 at two hundred and seventy million parameters, which doesn't loop much, and maths went from six point eight to seven point seven. Basically nothing.

So it didn't make Qwen smarter. It took away something that was getting in the way.

Exactly, and the authors say so themselves. Weight editing removed an obstacle. It didn't add ability. Imagine someone who keeps saying "um" until time runs out. Stop the "um" and you find out what they actually knew. One more limit: Astra couldn't spot a sycophantic model from the weights alone. It only managed that once it got a tool that showed attention patterns.

Why do you think it's this good at reading weights in the first place?

Nobody outside OpenAI knows. Some commenters suspect OpenAI trained Astra on weight data on purpose. Others point out there's a long line of research on generating weights. The authors call these "very early results." But the idea is new, and it matters for interpretability, for repairing models, and for the old question of models improving other models. Here a model diagnosed a specific bug inside another one and fixed it. That's the part to keep an eye on.

Quick hits. Anthropic filed a revised S-1 on Monday with its third-quarter numbers. Marcus, what's the top line?

About seventeen billion dollars in revenue for the quarter. People quoting the filing say sixteen point nine six. That's up about forty-seven percent from the second quarter's eleven point six billion, and more than ten times the same quarter last year. For scale, the first quarter was four point seven billion. CNBC also reports eight hundred and ninety-five billion dollars in commitments. I'd treat that figure carefully, because coverage of the first filing gave numbers anywhere from two hundred and sixty-five to eight hundred billion.

Forty-seven percent in a quarter is still enormous. Why are people calling it a slowdown?

Because the first two quarters each roughly doubled. And commenters reading the filing say compute and infrastructure costs rose about seventy-seven percent over the quarter, faster than revenue. There are plausible reasons. Anthropic cut prices in September. The cheaper 5.5-tier models arrived, and Opus 5.5 launched on September twenty-second, right at the end of the quarter. And one analyst says the summer's "tokenmaxxing," people burning tokens just because they could, has worn off.

An FT leak last week said it was profitable, though.

On an adjusted basis, and cash-flow positive in the quarter, according to what Anthropic told investors. That's a dollar seventy of revenue for every dollar of compute. Adjusted figures always deserve a careful look, but that's still a different picture from the big losses a year ago.

And the detail everyone's quoting is on page one eighty-seven.

The section is titled "Claude is Accelerating Our Ability to Close the Loop." According to Business Insider, Anthropic's internal use of Claude, valued at retail prices, now costs more than its entire payroll. That's for about fifty-three hundred employees.

That's quite a sentence to put in a prospectus.

It is, but look at what it measures. That's tokens priced at the retail list price, not cash Anthropic actually spent. It's a bit like a bakery saying its staff eat a million dollars of croissants a year at shop prices. And headcount nearly doubled over the year too. It tells you the company uses its own product heavily. It doesn't tell you the AI replaced the people. [HOST> When does it list?

After today's midterms. Marketing could start the week of November ninth, with trading before Thanksgiving, at a reported target valuation above two trillion dollars. For comparison, OpenAI just cut its private round from thirty billion to twenty. And Oracle filed on Monday to raise twenty billion dollars in new stock and bonds for its AI buildout. Its stock fell about eight percent, and its free cash flow last fiscal year was negative twenty-three point seven billion.

So everyone needs capital, and Anthropic is the one asking the public for it.

Security now. CVE twenty twenty-six, nine eight four three two. Marcus, explain it quickly.

It's a use-after-free bug in the Linux kernel, caused by a race condition when an IPv4 TCP connection is being set up. Use-after-free means the kernel keeps using memory it has already handed back, and attackers love that. The bug has been there since kernel 4.4, about ten years. It's rated nine point eight out of ten. Secondary reports say Anthropic showed Claude writing a remote root exploit for it in a research environment, against servers listening on Multipath TCP. I haven't seen that in Anthropic's own disclosure, so for now it's reported, not confirmed.

And it's still not patched?

Not everywhere. Debian, Oracle Linux and Amazon Linux 2023 have fixes. As of the latest reports, Ubuntu and Red Hat Enterprise Linux 8, 9 and 10 don't, almost three weeks on. Sources disagree on which Ubuntu versions are affected. One lists everything from 22.04 to 26.10, another says Canonical's tracker only lists 26.10.

What should people do in the meantime?

If you don't use Multipath TCP, turn it off. Drop IPv4 packets that carry IP options at the firewall, and restrict who can reach your servers. Red Hat says its mitigation "reduces exposure" but "cannot fully prevent it."

The uncomfortable part is the timing, right?

That's it. AI models are getting much faster at finding bugs and building exploits, and patches still ship at human speed, through human release processes. If the gap between "an AI wrote an exploit" and "your distribution shipped a fix" stays at three weeks, attackers will make good use of it. On a related note, Microsoft published research this week on North Korean groups adopting AI in their operations.

Next, China. Its commerce and science ministries have proposed export controls on frontier AI for the first time.

The catalogue adds forty technologies, twenty-three of them in AI. A model's weights, source code and architecture would be restricted if it meets any one of three tests: at least ten to the twenty-six training FLOPs, at least two trillion parameters, or a place in the global top ten at the time of export. Advanced agent capabilities would also be restricted, and agents built for defence or critical infrastructure banned outright.

What about open source? That's where Chinese labs have done best.

Already-published open weights are exempt. But to open-source newly developed restricted technology, you have to report it first, and the document never defines "open source." Using a model through an API or an app doesn't count as exporting it. Running a restricted model locally outside China needs a licence. [HOST> So what does that mean for companies here that self-host DeepSeek or Qwen?

That's the part I'd watch. One Hacker News commenter cites OpenRouter data showing open-weight models went from about half of usage to about seventy percent this year, and many of those models are Chinese. And "top ten at the time of export" is a moving target. A model that's fine today could fall under the rules after the next leaderboard update. So does the open-source exemption survive the first really strong Chinese release? I'm not sure it does. Public comment closes on November fifteenth. To be fair, some commenters say this mostly writes down guidance that already existed.

A fun one that went viral on Hacker News. Subvocal makes glasses that let you talk to your computer without making a sound.

The startup raised one point three million dollars from Y Combinator and Afore Capital. You mouth words silently, and the glasses turn them into text. They claim more than two hundred words a minute. Early-bird preorders are four hundred and ninety-nine dollars, with a hundred and ninety-nine dollar deposit, shipping in the first quarter of 2027. [HOST> So it isn't reading my mind.

No, only words you deliberately mouth. And Subvocal hasn't published any accuracy numbers. The closest research is Cornell's sonar-based EchoSpeech glasses, about ninety-five percent accurate in a twenty-four-person study. So it's plausible, but I'll wait for reviews.

I'd honestly use these. Nobody wants to dictate to their laptop in an open-plan office.

That's the problem it solves. Voice is a great way to talk to AI that's socially awkward in public. Commenters loved it, asked for clip-ons for prescription glasses, and raised fair privacy questions about face-sensing wearables.

AI-designed biology has reached a patient. Radical Numerics says it treated someone with AI-designed viruses.

Specifically bacteriophages, viruses that infect bacteria and leave human cells alone. The San Francisco lab, which has ties to the Arc Institute, says its Omnii-1 model designed them. Under the FDA's compassionate-use rules, they were given to a patient with an E. coli infection that resisted both antibiotics and existing phage therapies. Before that, the cocktail killed the patient's bacteria in the lab.

And did it work?

They haven't said. There are no patient or physician details, and the "first" claim hasn't been independently checked. Engineered phages have treated patients before, including a well-known 2019 case. So it's a real milestone for the approach, but we don't have the result yet. Drug-resistant infections are a big problem, though, and being able to design new phages in days instead of months could make a real difference.

And a quick follow-up on the Exin mouse story we mentioned yesterday.

One new detail from the founder's thread. Opus 5.5 refused once it realised biology was involved. Opus 5 did a dry run on water and graph paper but stopped before the live procedure. Sonnet 5 did the actual injection, and it paused on its own when the target fell outside the rig's range. Three Claude models, three different lines. It's still a demo with no paper.

Two quick ones. OpenAI published an engineering post called "Scaling Git."

It's about how its Git infrastructure copes now that coding agents generate so much of the code. Hacker News liked it for being free of hype, just engineers with charts. One commenter counts it as the fifth post like this, after Cursor, GitHub and Anthropic. Version control was built for the pace of human commits, and agents commit far faster. [HOST> And Goldman Sachs?

Bloomberg reports Goldman has restricted traders and bankers from using some AI tools. Beyond the headline, the details aren't confirmed, so that's all we'll say for now.

Finally, a guest essay on Terence Tao's blog, "Egoless Mathematics in the AI Era."

It's by Karthik C. S., Amit Sahai and Jie Xue, not by Tao himself. Their argument is that mathematics is about truth, not credit, so there's nothing to fear from AI finding proofs. One line: "Its truths do not become more true through the distinction of the person who expresses them."

Lovely. Did Hacker News accept it?

Partly. The obvious reply was that if ego doesn't matter, why are AI companies racing to be first to solve famous problems? Fair point. The egos have just moved to the labs.

One to watch: Anthropic's roadshow, which could start next week. Now there's a real number to test. Will public investors pay over two trillion for forty-seven percent quarterly growth when compute costs are rising faster?

I'd watch the Ubuntu patch first. Plenty of servers are still waiting on that fix.

That's your AI in 15 for today. See you tomorrow.