The Day in AI
In short
- MIT CSAIL researchers report that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights directly and raised its MATH-500 score from 7.9% to 48.1%.
- Anthropic released FlashProve, an open-source prototype that proves a response came from committed weights, and reports proving at 110 tokens per second on Llama 4 Maverick.
- Alibaba Cloud launched Memory on Model Studio, a long-term memory service for agents that it reports scores 85.15% on HaluMem-medium question answering.
- Baidu launched Apollo World, a driving simulator for its Apollo Go robotaxis that Baidu says generates up to 100,000 km of simulated driving a day.
- OpenAI hired Jill Hazelbaker, president and chief corporate affairs officer at Uber, to lead its global communications as Chief Corporate Affairs Officer.
The day in brief
MIT CSAIL researchers reported on 2 November that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights by reasoning about them in natural language, lifting the model's MATH-500 score from 7.9% to 48.1%.
Earlier methods set weights with an optimizer such as Adam or Muon, a network trained to write weights, or a compiler. In this work Astra reads the raw numbers, reasons about them in plain language and writes changes directly, with no gradients and no training loop. It found the reasoning loop problem that Qwen's own model card warns about and fixed it. The authors report that the same edit also roughly tripled the score on IFEval, a test of instruction following, though the edit targeted math.
In a second test on the Kinetix physics environment, Astra changed 3.5% of the weights and took a car-on-ramp task from 0% to 100% success over 1,024 episodes, on a single try. All of these figures are the authors' own.
Anthropic released FlashProve on 2 November, an open-source research prototype for proving which model produced an answer. A provider publishes a short commitment to an exact set of weights. Each prompt and response then gets a zkSNARK proof, a compact cryptographic proof that anyone can check on a CPU without seeing the weights. Anthropic reports proving at 110 tokens per second on Llama 4 Maverick, with proofs of 18.78 MB that verify in about 21 seconds on one thread.
The cost is high. Anthropic puts proving at about 72 times the per-GPU time of FP8 vLLM on eight H100s, rising to about 94 times at the highest load it tested. The design is software only and doesn't depend on trusted hardware.
Alibaba Cloud's Memory service lets agents keep a user's preferences, plans and earlier commitments across sessions and build them into a structured profile, with rules and fields the developer defines. It connects to Codex, Claude Code, DeepSeek Harness and OpenClaw through an API, the Model Context Protocol (MCP) or a command-line tool. Alibaba reports 85.15% on HaluMem-medium question answering against 81.71% for Zep. It also reports first place among eight systems on PersonaMem-v2, at 56.82. SenseTime's SenseNova 6.8 Flash, released the same day for office and design agents, also puts memory across conversations at the centre of its pitch.
Also in the news
- Baidu launched Apollo World, which simulates road scenes with camera and lidar data, and says its 100,000 km a day equals about three years of real-world testing.
- Jill Hazelbaker joins OpenAI as Chief Corporate Affairs Officer, a role The Information says has been empty since the start of the year.
- Dan Driscoll, the former Secretary of the Army, reportedly joined OpenAI in a new role as Senior Counselor for National Security Modernization.
- Apple opened Siri AI integration through App Intents to all developers in the iOS 27.3 betas, according to Techmeme's summary, and the API details are still unknown.
- Apple began beta testing Audio Intelligence for Apple Watch Series 12 in iOS 27.3, with four opt-in features for sounds, songs, conversation catch-up and notes.
- Google took its Dreambeans AI stories app to India on 2 November with 13 Indian languages, after Sensor Tower counted about 5,500 US downloads in October.
- DeepSeek added a creator mode to DeepSeek Harness that builds working plugins from a user's idea, with demos including a 3D LLM visualizer.
- OpenAI's ChatGPT consumer product lead Adam Fry told Big Technology that ChatGPT has 1.2 billion weekly users and search volume about 20% of Google's.
- Anthropic warned that it is very close to recursive self-improvement capability, meaning AI that builds AI, TechCrunch reported.
- CrowdStrike agreed to buy Fiddler AI for about $400 million and Dreadnode for about $95 million, according to the Wall Street Journal.