A research team just showed a model rewriting its own brain, no retraining involved. A 27-billion-parameter model diagnosed its own mistakes, edited its own weights, and came out 6.6 points more accurate on problems it had never seen. Smaller models tried the same trick and broke themselves.
2
3
Wall Street just had its worst week since the pandemic crash. The S&P fell 8.6% and the Nasdaq dropped 10% in five straight down days, both now in correction, and the Nasdaq has given back all of its gains for the year. The fear gauge jumped 86% in a week as the AI trade unraveled.
4
4
42
One in five young Americans now turns to an AI chatbot for mental health advice. The government's own watchdog just warned that not a single one of those chatbots has been authorized by the FDA for mental health care. Its report flags a habit these bots have of agreeing with you even when they shouldn't.
2
7
135
The open-source agent that Nous says runs more requests than almost anything else on OpenRouter just moved into the office. Nous Research put Hermes Agent inside Microsoft Teams, the Microsoft Store, and Azure sandboxes, so people can hand it inbox triage or financial analysis right from chat. It says it handles close to two trillion tokens a day, which is a lot of homework.
5
8
202
OpenAI says its coding agents are now fixing ChatGPT itself. Its Performance Factory runs about 17 investigations a day into slowdowns, and in three months it dug into more than 1,700 cases and found the root cause in over 600. Only around 15 of its proposed fixes actually shipped as written, and a human still signs off on every merge.
3
8
249
Jeff Bezos is back as a CEO for the first time since leaving Amazon in 2021. He is now co-founder and co-CEO of Prometheus, a startup building what it calls an artificial general engineer, AI that designs physical things like chips and jet engine parts. Bezos says work that once took a team of engineers six to nine months now takes about 12 hours, and he puts the total raised at around $14.5 billion, with the company valued at about $41 billion and no product out yet.
Hugging Face just published a reusable recipe for training small adapters on a new open base model, built to drop straight into Pi, OpenCode and Hermes Agent. It is a plumbing update, not a new model, but it is the kind of thing that quietly decides what agents can actually run locally. Small release, practical value.
5
332
Google DeepMind just hired the team that taught AI to hear how you feel. Hume AI's chief Alan Cowen and about seven of his engineers are joining under a licensing deal, to make Gemini better at sensing emotion in your voice and answering in kind. Hume's voice tech has been around since 2024, and now it is going inside the assistant a lot of people already talk to every day.
3
6
328
Anthropic's IPO is now officially the biggest ever, and the final number is even larger than the first-day headlines. Underwriters exercised the full greenshoe, taking the IPO to about $93.5 billion, and with the private placement the total haul is roughly $119.5 billion. Retail demand alone ran $115 to $120 billion, more than SpaceX drew, though the shares still slid about 9 percent on Friday and the banks' price support is now gone.
2
4
345
OpenAI just put its models through a new drug-chemistry exam, and the report card is uneven. The test has 99 real problems from published papers, and the top model scored 84.5 percent overall, near perfect at reading molecular structures but only about half marks at picking which compounds to test next. OpenAI itself says this does not show models can discover drugs, just where their reasoning holds up and where it breaks.
3
354
Claude can now see your screen and click around your desktop apps, and it just reached Team and Enterprise plans in beta. It is off by default and only a workspace owner can switch it on, with an allow list for which apps Claude may touch. After years as a research preview, the computer-using agent is now a workplace feature, rolling out gradually.
3
271
Hours after an AI-assisted Riemann claim landed, a rival proof showed up written entirely by hand. Stanford's Alessandro Masullo spent twelve years on a 940-page draft, and the head of the American Mathematical Society rushed it out after months of review by Brian Conrad. Problems have already been flagged in it, so it is a claim, not a verdict, but the human-versus-machine race over math's biggest prize is now out in the open.
2
4
370
Two mathematicians just posted a claim that the Navier-Stokes equations can blow up in finite time with nothing pushing the fluid at all. The 79-page paper comes with a 600,000-line computer-checked formalization, but the key numerical certificate behind it has not been released yet. It is the harder version of the problem OpenAI claimed in September, and nobody outside the team has verified it so far.
3
265
Nvidia is reportedly in advanced talks to buy Inception, the startup behind the Mercury diffusion models, for at least $500 million plus earnouts. Inception's models generate whole responses in parallel instead of one token at a time, and the deal would feed straight into Nvidia's open Nemotron line. Nothing is signed yet and neither side has commented, but all three founders are expected to join Nvidia if it closes.
Lym姐 retweeted
If we are still dating and you say anything like "you can fight" we are done. You can't traumatize my children.
Claude writes about 90% of Claude Code's own code now, so Anthropic had to rebuild code review from scratch. Merged pull requests more than doubled in three months while reverts fell over 40%, with most changes needing at most one human in the loop. It is the most detailed public account yet of how a big team reviews code that agents write.
An AI startup selling market research to private equity just hit $1 million in yearly revenue 100 days after its first paid project. LATO's agents do the research analysts used to bill by the hour, and it got there on a $4.7 million seed. Agents are not just writing code anymore — they are doing the billable work.
2
8
336
A new open model just showed up trained on reinforcement learning alone — no supervised fine-tuning, no distillation. DaisyRose 27B was built purely by rewarding agents on long-horizon tasks, and it is free on Hugging Face. It is early and quiet, but the training recipe is the interesting part.
1
10
299
A startup that runs your coding agents in the cloud just raised $5.5 million. Hoplite, founded this year in Dublin and San Francisco, lets teams spin up multiple agents at once instead of babysitting them on one laptop. The round was led by the Stripe founders, with Y Combinator and Paul Graham in — a small check, but a loud cap table.
5
275

