2026-10-11 17:10 UTC

ai-research

band: warmmomentum: stable score: 0.394
temperature history

Episodes (9)

Expert review and reproduction will determine whether the paper’s optimization and AlphaEvolve-assisted method validly improves the best known matrix-multiplication exponent.
expiredknownscott: medium
Levent Alpöge claims an AI-assisted proof of the Hopf problem, while Boris Alexeev has reportedly released a roughly 250,000-line Codex-generated Lean formalization that would make the result an unusually large advance in AI-assisted formal mathematics if valid.
watchingconvergesscott: medium
Anthropic claims Claude led 26% of its AI R&D work under human supervision in August 2026, up from under 1% in February, indicating a substantial shift toward agent-executed model development without fully autonomous research.
resolvedconvergesscott: medium
Prinz claims GPT-6 Astra decrypted a previously unsolved 1918 German radio message using the key TRUPPENVERSCHIEBUNG and found agreement with naval records, potentially demonstrating useful AI-assisted cryptanalysis of historical ciphertext.
expiredknownscott: low
The newly formed independent Advisory Group on Mathematics and Artificial Intelligence says it is advising OpenAI on releasing a large batch of reportedly significant internal-model mathematics results and will publish recommendations, potentially establishing an externally visible release-governance process.
resolvedconvergesscott: medium
Tim Dettmers claims dlab's forthcoming Open Source Week stack combines aggressively quantized local inference, frontier-comparable autonomous research, and CliffCompaction's roughly 50% cost reduction, potentially making sustained research agents practical on personal hardware.
watchingconvergesscott: medium
Bloomberg reports that Mirendil, founded by ex-Anthropic researchers to build recursive self-improving models, is in talks to raise up to $1B at a $5B valuation led by Kleiner Perkins with a13 participating, which would place frontier-scale capital behind an explicitly self-improvement-focused neo-lab launching a frontier model by early 2027.
seedconvergesscott: medium
Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma claim Fable 5.1, driven through the Claude Science harness with only periodic 'keep going' prompts, computed the nine-loop six-particle (hexagon) amplitude in planar N=4 super Yang-Mills — an outstanding problem in scattering amplitudes — for roughly $1–2k of near-unattended inference, with the result verified by expert Lance; acceptance of the amplitude into the field and replication of the approach would establish frontier agent harnesses as demonstrated solvers of computational-physics barriers experts considered out of reach on academic budgets.
corroboratedconvergesscott: high
ClaudeAI user sebasmtl claims his open-source OpenPhysicsAI physics lab — 13 flags scored against sealed, previously-unpredicted experimental measurements plus 4 starter trials — lets anyone's AI compete as a solver, and real external solver attempts would establish it as a used machine-verified benchmark for AI scientific capability.
seedconvergesscott: medium

Trajectory notes