The Day in AI
In short
- Simplex AI open-sourced Linger, a local personal agent that it says runs a 125B model on a 12GB GPU at 45.5 tokens per second.
- Levent Alpöge, a mathematician at Anthropic, says he proved the Riemann Hypothesis with Claude and posted a Lean formalization that has not been independently verified.
- METR updated its time horizon plot with longer, more open-ended tasks and found models do worse when tasks need judgment.
- Meta plans to sell its pocket-sized Muse Charm device for $300 with up to 12 months of free Verizon 5G, according to The Information.
- Palo Alto Networks agreed to buy AI startup Applied Compute for $4.5 billion, Quartz reports.
The day in brief
Simplex AI, founded by former DeepSeek-VL lead Zhang Bo, open-sourced Linger on 19 November, a local agent it says runs a 125B model on a 12GB GPU at 45.5 tokens per second.
Linger reads local email, calendars and files and operates the user's logged-in browser. It asks for confirmation on a card before any write action. The engine underneath, PolyStrata, splits a mixture-of-experts model's weights across GPU memory, system RAM and SSD. Only a few experts run for each token, so the rest can sit in slower memory. That lets the 125B Qwen3.8-Flash-Next model run on a 12GB card with 64GB of RAM. Simplex reports the 45.5 tokens per second as a steady decode rate that excludes time to first token, and nobody has reproduced it yet. Pandaily reports a vendor figure of 179 tokens per second on a 32GB RTX 5090. Local use is free, the license is Apache 2.0 with added commercial restrictions, and a hosted version is waitlist only.
On 20 November, Levent Alpöge posted a document on his own website saying he proved the Riemann Hypothesis with Claude, and that the method extends to the Generalized Riemann Hypothesis. He says the arguments are formalized in Lean, the proof assistant that checks each step mechanically, without assumptions. Mathematicians on Zulip are examining it. TechCrunch and The Verge report that most seem confident the Lean result will hold. A human-readable explanation of the proof has not been published.
On the same day, Ravi Vakil posted an all-human proof of the Generalized Riemann Hypothesis by Alessandro Masullo, a Stanford postdoc. Both claims are now waiting on outside review.
METR's time horizon plot scores models by how long human testers take on the tasks the models can complete. The update adds tasks that take humans far longer, and tasks that are more open-ended. MIT Technology Review reports that models are weaker on judgment-heavy tasks than on coding. The Information reports that the doubling time is 3.6 months and that recent models, including OpenAI's GPT-6 Astra, can finish tasks that take a person about a work week. MIT Technology Review describes the original plot as doubling roughly every four months.
Also in the news
- OpenAI released GPT-Live-1.1 in the API on 19 November, with better pronunciation, accent consistency and transcription in multilingual voice conversations.
- Meta will reportedly sell Muse Charm for $300 and has delayed its smartwatch to next year to focus on it, according to The Information and Techmeme.
- Palo Alto Networks agreed to buy Applied Compute for $4.5 billion, according to Quartz.
- DeepSeek is reportedly raising $16 billion and running into a shortage of computing capacity, according to Digitimes.
- Alibaba is training its next-generation Qwen model on its own T-Head chips, TechNode reports.
- Nvidia is close to buying the diffusion model developer Inception for more than $500 million, according to The Information.
- MIT CSAIL posted a paper showing language models rewriting their own weights.