OpenAI pauses training after DNS exploit; Kimi K3.1 Pro becomes first near-frontier open-weight model; one million prediction market forecasts confirm strong calibration

Engineering

Kimi

Introducing Kimi K3.1 Pro

Kimi released Kimi K3.1 Pro, an open-weight AI model with a one-million-token context window that scores above Claude Fable 5.1 on Artificial Analysis. The company reports better results for code, 3D design, and financial analysis, and says the model performs close to top proprietary models. Kimi removed cyber-related training data to reduce misuse risk and will publish full weights by November 4.

Latent Space

AI Is Getting Good at Math. Here's Why.

An internal OpenAI model found a counterexample to Erdős's unit distance conjecture, which had been open since 1946. The author argues that models search depth-first and use an automated verifier to test many candidate proofs quickly, making the cost of pursuing an unlikely idea near zero. Failed attempts become negative hints that steer the next iteration, compressing what took human mathematicians months into minutes.

Cloudflare Blog

Bringing post-quantum cryptography to geographically fenced key management

Cloudflare describes a post-quantum version of its GeoKey Manager. The service distributes wrapped customer keys globally, but only edge servers in an approved region can unwrap them. It uses hybrid ML-KEM-768 and X25519 encryption, so confidentiality holds if one algorithm stays secure, and keys are loaded into the kernel keyring rather than persisted on disk.

Philip O'Toole

Twelve years of distributed SQLite

Philip O'Toole traces distributed SQLite from 2014 to 2026. Many teams built replication on top of SQLite using three approaches: consensus, streaming, and merge. Commercial companies sold the idea as paid services, while rqlite, dqlite, and Litestream remain free and open source.

Social Sciences

Astral Codex Ten

Your Book Review: The Ghost In My Brain

A reader reviews Clark Elliott's memoir about a severe concussion from a 1999 car accident that left him unable to walk to his car or recognize his own house for a decade. Prism lenses and geometric exercises helped him recover, but the scientific evidence for them is thin. The reviewer argues that persistent post-concussion syndrome needs patient-generated case registries to bridge the gap between individual anecdote and RCT population averages.

Marginal Revolution

Prediction Markets Are Accurate

A new PNAS paper studies about one million resolved forecasts from Polymarket, Kalshi, Manifold, and Metaculus across twelve domains and time horizons ranging from hours to years. The forecasts were overall highly calibrated: events assigned a probability of p% occurred approximately p% of the time. Some high-stakes areas showed systematic over- or underestimation.

Marginal Revolution

Protecting Americans from Medicine

Max Doppelt flew to Shanghai to buy a cancer drug for his mother that was approved in China in 2023 but still unavailable in the U.S. as of mid-2026. Alex Tabarrok argues for FDA reform to allow faster domestic access to drugs approved abroad, and criticizes a proposed Treasury requirement to vet all U.S.-China pharmaceutical deals as counterproductive to the goal of saving patients' lives.

Humanities

JSTOR Daily

Teaching Students to See What They Miss

Sue Kenney describes a high school Narrative Medicine lesson built around Sir Luke Fildes's 1891 painting The Doctor. Students practice slow looking by examining the painting in layers, then separate what they observe from what they assume — noticing, for instance, that the doctor has no medical instruments visible. The lesson connects these habits of sustained attention to what the instructor calls radical listening: applying patience, curiosity, and resistance to premature interpretation when attending to another person's story.