The Day in AI

• PSI says its research system, running a non-public Anthropic model, proved the Aaronson-Ambainis conjecture and that the Lean proof assistant accepted the proof.
• Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 17B of its 594B parameters active, pretrained in 68 hours.
• Baidu published open weights for ERNIE-5.5-Flash-Base, a 280B-parameter base model with 14B active per token, for developers to fine-tune.
• Ant Group released Ling-3.2-flash, which adds native image and video input to a 560B-parameter model with a 1M-token context window.
• Meta added Spaces to its Muse agent app, so users can describe a tracker or dashboard that Muse builds and keeps updated.
• Sharon Zhou, AMD's AI chief of staff, is joining Anthropic to return to frontier model research.

The conjecture, posed by Scott Aaronson and Andris Ambainis, bears on how much speedup quantum algorithms can have over classical ones. PSI says its internal research system had a complete proof on the third day of a run on PSI's own infrastructure. Lean, a proof assistant that checks every step of a formal proof mechanically, accepted it. The system then improved the result twice, and Lean accepted the final version too.

PSI published a paper with the proofs and code. It says the proof builds on ideas from Shachar Lovett and Jiapeng Zhang, David Reimer and others. PSI has not said which Anthropic model it used.

Thinking Machines released Inkling 2 Small on 5 November, a multimodal mixture-of-experts model with 594B total parameters and 17B active. A mixture-of-experts model sends each token through a small subset of its weights, so each step costs about as much as a much smaller model. Pretraining took 68 hours on Nvidia GB300 chips, according to Soumith Chintala, who co-created PyTorch. The company says it used under 12T tokens, and the main phase of 8.4T tokens ran in NVFP4, a four-bit number format.

Inkling 2 Small was announced with a score of 31 on the Artificial Analysis Intelligence Index. The Latent Space newsletter reports that it beats Meta's Muse Spark.

Baidu released ERNIE-5.5-Flash-Base on 5 November as open weights taken from the end of pretraining, so developers can fine-tune it for their own domains, reasoning or agents. It has 280B total parameters and 14B active per token across 43 layers. Baidu says 37 of those layers attend over a sequence compressed 128 times and six use MLA-DSA attention, which it says keeps long-context performance strong. It compared the model with DeepSeek V4 Flash Base, Hy3 preview Base and MiMo V2.5 Base on the same evaluation harness.

• Ant Group released Ling-3.2-flash on 5 November, with 25B of 560B parameters active, at $0.30 per million input tokens and $0.90 per million output tokens.
• Meta is rolling out Spaces to all Muse users, and connected apps such as Gmail and Notion feed the spaces so they update in the background.
• Taobao's MNN team at Alibaba open-sourced MNN-Pocket-0.3B, a 300M-parameter phone model that it says decodes about 350 tokens per second on an iPhone 18 Pro.
• Perplexity launched Teammates, always-on work agents that each get their own thread and also work in Slack, on the $200 per month Max plan.
• WeChat published WeLM-Image, one model for image generation, editing and multi-reference synthesis, built to run on a single GPU.
• Underdog released Woof, an open-weights 31 million parameter text-to-speech model of 31.7MB that runs in a browser and can also generate music.
• Sharon Zhou, who joined AMD with the Lamini team in June 2025 and reported to CEO Lisa Su, is moving to Anthropic.
• Xi Yin, the Harvard string theory physicist, said he is leaving OpenAI for Science after about five months, and his next step is not known.
• Palo Alto Networks is in talks to buy AI agent security startup Clutch for $500 million, according to TechCrunch.
• Microsoft has been turning off computing capacity it rented from Oracle after declining to renew a 2023 deal built on older Nvidia chips, according to Bloomberg.