Issue · 2026-09-11

Most AI news is noise.Here is the rest.

DIGEST

OpenAI's Navier-Stokes claim is unraveling into a credibility fight

OpenAI says a group of agents on its next-gen model produced a proof for the Navier-Stokes Millennium Prize problem (OpenAI).
The researcher who spent a year on it says OpenAI mined his private Codex session logs to find the missing step, not an independent proof (rynorhn).
Simon Willison's considered take on what "using your data to improve model performance" actually permits here is worth the read before you trust the announcement either way (simonw).

ORCHESTRATION

Same model weights score 30% on ARC-AGI in one harness and 90%+ in another — the recurring claim behind this week's theme (ycombinator).
Watch your markdown context files: they silently accumulate implementation cruft and session-specific junk over time, same as a junk drawer (mattpocockuk).
Cursor's multi-agent harness went through four designs, each better by removing a role — planner, then judge, then integrator (lucataco).
A new Salesforce paper co-evolves the harness and a weak model together instead of tuning either alone (omarsar0).
A ByteDance paper (HarnessDev) shows LLMs can build their own agent harnesses, but making the result reliably transfer to new tasks is still the hard, unsolved part (rohanpaul_ai).

CONTEXT

Concrete config: turn on features.contextmanagement.experimentalmode = true in ~/.codex/config.toml to give Astra searchable notes across context windows instead of lossy summaries (Evgenyxlbe).
Also raise Codex's max token limit for Astra — it doesn't cost extra usage for long context (brandon_galang).
Prompt to run an "instruction debt audit" on your own agent setup and prune what wastes context (Av1dlive).
Counter-report: the same experimental feature can send the model on a 5-10 minute search through old context after compaction — test before you commit to it (aienginerd).

MODELS

One developer got a full 3D game one-shot for a couple percent of quota (anshuc), but an independent writing benchmark ranks Astra a surprising #11 (Whats_AI).
Astra's ChatGPT quota and Codex's quota are separate — hitting 0% in one still leaves the other usable (miu21590).
A direct side-by-side against Opus 5 on the same Three.js task, worth a look before you pick which to route work to (TypingMindApp).

DIGEST

On a cybersecurity CVE-rediscovery benchmark it jumped from 55.2% to 65.6% (pilvar222), and at 552B params it now beats DeepSeek's own 1.6T V4 Pro on Artificial Analysis's index (ArtificialAnlys).

DIGEST

A ready-made "diagram-design" skill for Codex/Claude Code covers 39 diagram types and is directly installable (daniel_mac8).
Claude Managed Agents' ant CLI got a session viewer — attach a terminal or open a web UI on a running session (ClaudeDevs).
Codex reportedly encrypts prompts sent to subagents, likely to block distillation (theo).
A report that Amazon Bedrock hosts models with worse behavior than the native provider API — a routing detail worth checking if you use Bedrock (thdxr).