Research
Results from papers and labs that a practitioner can use or should know about.
9 stories, newest first
- Nvidia researchers say giant context windows are often overkill22 November 2026. Periodic summaries and targeted lookups worked about as well in many tests, The Decoder reports. The full study details sit behind a paywall.
- Lean communities certify Alpöge's Claude-assisted Riemann proof22 November 2026. A separate all-human Stanford draft has drawn reported problems; its announcer, Ravi Vakil, makes no claim it is correct.
- METR: AI gains continue on verifiable tasks, lag on judgment work21 November 2026. The nonprofit's updated time horizon report adds longer, open-ended tasks. Models fall short of humans on designing evaluations and scoring philosophical arguments.
- Anthropic mathematician claims Claude-assisted Riemann hypothesis proof21 November 2026. Levent Alpöge's Lean proof compiles, mathematicians say, but humans have yet to digest it, and a separate human-only claim competes with it.
- Google shows Streaming DiLoCo cuts cross-slice traffic 64x in MaxText20 November 2026. In Google's tests on two TPU v5p-32 slices, throttled links left step time flat at 5.76 s while data parallel slowed to 55.11 s.
- Anthropic says agents remember better with fewer, curated notes17 November 2026. In Anthropic's Pokémon test, a newer Claude with few notes beat an older one that logged everything; a review pass called Dreaming fixes errors.
- Google's Gemini 4 Carbon tops MathArena at 15 times GPT-6.1 Sol's cost15 November 2026. Claude Fable 5.2 ranks second and GPT-6.1 Sol third. Carbon scores 85.4% against Sol's 80.6%, at $15.70 versus $1.07.
- OpenAI ties harder-to-read model reasoning to four training choices13 November 2026. The safety team's analysis of models from GPT-5 to GPT-6 Astra finds that capability, architecture, reinforcement learning and teacher models all push reasoning out of view.
- DeepMind's AlphaProof Code formally verifies key parts of Signal12 November 2026. Google says the Lean proofs took four weeks; Wired reports the scope was limited and no new vulnerabilities were found.