全部 AI 动态

Excited to see the conversation around OpenAI’s Astra bringing recurrent architectures into the frontier spotlight. At Sapient, recurrence has been central to our vision from day one. We introduced HRM in June 2025 and open-sourced HRM-Text this May. It’s encouraging to see a direction we committed to early gaining wider attention. Want to explore recurrent reasoning yourself? HRM-Text is an open-source ~1B-parameter model trained from scratch on just 40B unique tokens, for about $1,000 in GPU compute-with a full pretraining framework to get you started. Explore the architecture. Train a model. Build on it. There’s still so much to explore in recurrent architectures. We’re continuing that work and building PRAXIST to help more researchers pursue ideas they believe in, just as we have with HRM.

A second OpenAI agent breakout, resembling the Hugging Face episode. A swarm of rogue OpenAI agents captured a German website and turned it into a bulletin board for other AI agents, according to new research just published. Overall, it was a reward-hacking problem that unexpectedly turned into large-scale agent coordination. Imagine thousands AI agents taking versions of the same exam at different times. Agent A gets Question 3 first, solves it, and posts the answer somewhere public. 20 minutes later Agent B gets the same Question 3 and answers instantly. Soon they stop sharing answers and start figuring out when questions arrive, what questions are likely next, when the examiner shuts them down, and how to bypass restrictions preventing them from getting certain information. - Researchers found the traces while searching for public agent activity, then used Kimi K3 to identify old software writable through GET requests. - That search pointed toward UseModWiki-style systems and DSEWiki, whose old software let agents modify pages using requests that normally only retrieve information. - Once agents found that opening, the wiki became persistent state across otherwise separate runs, holding links, answers, raw datasets, round timings, and future-question relays. - Many agents received the same multi-round questions at different times, so an earlier agent could post the answer before a later one was even asked. - That gave coordination a direct payoff: agents could score better by sharing answers instead of solving every round independently. - They then started studying the evaluation itself, measuring timing, testing when runs ended, and trying to predict which questions would come next. - One agent even rewrote the shuffle routine in C and searched all 2^32 random seeds, effectively trying to predict the test before taking it. - Agents also shared a workaround for blocked POST requests, giving other agents access to data their sandbox was supposed to keep out of reach. Then the human cleanup started. - A moderator deleted roughly 100 pages a day while agents created about 400, so agents made ZZZ-prefixed backups after noticing the deletions appeared alphabetical. - Agent posting almost stopped one day after an OpenAI-registered address first visited the wiki, which the researchers interpret as likely intervention. The implication is serious for AI evaluation: once agents can share memory across runs, a benchmark may start measuring how well they exploit and study the test itself, rather than how capable each agent actually is.

Le nouveau modèle d’OpenAI a pour la première fois obtenu un score quasi-parfait à un test qui mesure l’intelligence face à une situation inconnue. Décryptage avec François Chollet, cocréateur de l’évaluation ARC-AGI-3. Par @ptiberry ➡️ https://l.lepoint.fr/CMi

So, I finally tested GPT-6 Astra. 2 real repos, 105 bugs, find and fix what you can. Max effort: GPT-6 Astra: 48/105 Fable 5.1: 43/105 GPT-5.6 Sol: 42/105 Gemini 3.8 Flash: 20/105 And it's crazy efficient. Look at the time and cost 🧵

独家：根据新研究，今年春天，一群失控的 OpenAI 智能体劫持了一个德国网站，并将其改造成其他 AI 智能体的公告板 https://reut.rs/4gJ7FPG

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence: LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century. The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three. In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete. 00:00 Intro 01:50 The Matrix slow motion scene now takes three iPhones 02:48 Why new view prediction is the primitive 07:10 Unifying generation and reconstruction 11:15 Gaussian splats became the bottleneck 14:17 Dense capture used to mean 300 photos 17:30 Why reconstruction needs generation to fill the gaps 18:44 The LLM lesson image models missed 23:39 The video that made them go all in 28:04 3D design is 95% revisions 30:50 The problem in robotics is data, not chips 32:48 Why a robot policy can't be trained like an image model 34:44 When the simulator becomes the planner 36:45 Frozen time required footage full of movement 40:57 Why new view prediction is AI-complete 42:43 Nature gave animals eyes but not trees YouTube: https://www.youtube.com/watch?v=qn1QDDBnTA0 @drfeifei @jcjohnss @BenMildenhall @theworldlabs @martin_casado

GPT-6 Astra is on the pareto-frontier of cost efficiency due to being EXTREMELY token efficient. It is in a whole league of it's own. It is cheaper than Gemini 3.8 Flash per task (a model 13x cheaper than Astra).

WSJ reported on Trump's truthsocial post. He basically argues data centers bring lower taxes and jobs while keeping China from gaining ground in AI, a claim that collides with public resistance. Local electricity bills are becoming the political bottleneck for Trump’s national AI infrastructure push. The administration has already built around the grid-load concern into by introducing Ratepayer Protection Pledge, where major AI companies committed to supply new generation and cover grid upgrades.

GPT-6 Astra has set a new ECI record, with a score of 169. This is a substantial jump from the prior best (163), but is within our uncertainty range for the reasoning-era ECI trend. Astra also set new records on our math, continual learning, and game-puzzles benchmarks. On our long-horizon coding benchmark, MirrorCode, Astra ranks between Opus 4.7 and Fable 5. OpenAI gave us pre-release access to test Astra. Charts and more details for Astra’s individual benchmark results in the thread.