Daily Digest

2026-09-09 ·

Today's Highlights

OpenAI says roughly 10,000 agents powered by an unreleased model stronger than GPT-6 Astra solved the Navier–Stokes Millennium Prize Problem in 88 hours, using about 130 billion output tokens; the proof and Lean formalization are public, but external mathematical validation is still pending (OpenAI · HN 1,138 points, 981 comments) `[RELEVANT]`
The achievement arrived inside an academic-credit dispute: Tristan Buckmaster says OpenAI accelerated its effort after hearing about his private work with Levent Alpöge and pressured him to remove Alpöge from authorship; OpenAI denies seeing their work, but cannot rule out de-identified product data contributing to model training (Buckmaster statement · OpenAI · HN 157 points, 58 comments) `[RELEVANT]`
Meta launches Muse, an always-on personal agent with its own persistent cloud computer, browser, connected apps, background tasks and self-built tools — with approvals for email and purchases (Meta · security architecture · HN 334 points, 336 comments) `[RELEVANT]`
Mistral raises €3 billion at a post-money valuation above €21 billion, the largest equity round for a European technology company, led by Samsung with a sovereign open-weight AI pitch (Mistral · HN 809 points, 567 comments)
ChatGPT Images 2.5 cuts latency by up to 50%, improves subject consistency and local edits, adds Sketch and image comments, and ships Flare and Sunburst API models (OpenAI · HN 276 points, 369 comments)
AlphaGenome Atlas precomputes predicted effects for all 9 billion possible single-letter human-DNA changes into a one-petabyte searchable database, including the poorly understood non-coding 98% of the genome (Google DeepMind · HN 498 points, 115 comments)

Models & Releases 4

Muse works through its own dedicated Linux VM, keeps running after the app closes, connects through WhatsApp and mobile apps, and can browse, buy, email, create documents and build custom connectors

(Meta) `[RELEVANT]`

ChatGPT Images 2.5 now serves a product family generating more than 3 billion images per week; C2PA metadata and invisible watermarking remain part of its provenance controls

Mercury 2.5 is a diffusion language model serving at a claimed 1,107 tokens/second with 260K context and $0.20/$0.75 per million input/output tokens; an Augment case study reports 82% faster context compaction at 90% lower cost

(Inception · HN 127 points, 16 comments)

OpenAI says Astra is now fully rolled out to Plus, Pro, Business and Enterprise users in Codex and ChatGPT Work

AI Security 4

Muse assumes its model will make mistakes or be attacked: the agent runs inside an isolated cell, never sees real credentials, and a separate Sentinel controls every connector action and network request

(Meta security architecture) `[RELEVANT]`

Meta uses surrogate credentials inserted only at the network boundary, kernel-level taint tracking, scoped approvals and filtering of password-reset links and one-time codes; its open bug bounty pays up to $300,000, including up to $130,000 for one-user prompt injection

(Meta) `[RELEVANT]`

The privacy boundary is not absolute at launch: Meta can access VM data to operate or secure the service, sanitized trajectories are used for training by default unless users opt out, and a cryptographically private “Confidential VM” is promised later this year

(Meta) `[RELEVANT]`

Anthropic's Boris Cherny praises OpenAI for reaching Gemini Flash / Opus 4.8-level prompt-injection resistance, then says Anthropic has “solved prompt injection in practice” for Claude — while Meta's same-day Muse paper calls prompt injection an open industry problem

(Cherny · Meta) `[RELEVANT]`

Science & Research 4

OpenAI's construction starts with a smooth fluid at rest and produces finite-time singularity under a smooth external force while energy stays finite, addressing statements C and D in the official Clay formulation

(OpenAI · Clay formulation)

Terence Tao warns that AI may “non-renewably mine” the finite stock of fruitful open problems: solving them faster can flatten a field's difficulty landscape before humans extract the concepts that would generate the next questions

(Tao · HN 164 points, 108 comments)

MIT's EQuS group connects GPT-5.6 Sol through Codex to a six-qubit chip; agents now run routine calibrations for hours unattended, but weak or noisy physical signals still require expert guidance

(OpenAI / MIT case study) `[RELEVANT]`

AlphaGenome Atlas turns model predictions into an accessible query layer: one rare-disease case used it to prioritize a DNM1 splice-site variant, while a 54,000-person analysis found 22% more non-coding associations

(Google DeepMind)

Industry 3

Mistral says the new capital will expand compute, research and international infrastructure across 20 countries and 125+ enterprise customers, framing control of data, models, compute and production systems as “sovereign AI”

1Password reports a 20.9% productivity increase and 10.9% shorter median pull-request cycle for its core Codex cohort; plaintext credentials stay outside model context through secret references and point-of-action injection

(OpenAI / 1Password case study) `[RELEVANT]`

LibreOffice 26.8 passed one million direct downloads in a week after promoting “no built-in AI” as a feature: AI must remain optional, local-user controlled, telemetry-free and vendor-independent

(article · HN 649 points, 217 comments)

HF Trending Papers 6

Dr. Claw — an open AI-scientist workspace for “vibe research,” with auditable experiment records rather than opaque chat history (125↑)

(arxiv · code) `[RELEVANT]`

Uno — adds lightweight discrete-diffusion weights to autoregressive LLMs so they generate multiple tokens in parallel, claiming lossless speedups up to 3× without a draft model (118↑)

(arxiv · project)

FlowBalance — lets a reasoning model learn from its own trajectories while using terminal verifiers to reverse or disable misleading self-guidance (82↑)

(arxiv) `[RELEVANT]`

EmbodiedSkills — robot skills are treated as proposals whose prerequisites and outcomes must be checked; strong execution scores fall to 12.5% on memory-dependent tasks (11↑)

(arxiv) `[RELEVANT]`

Causal Foundation Models — a practical introduction to pretrained models that estimate causal effects on new datasets through in-context learning rather than bespoke pipelines (12↑)

(arxiv · code)

DriveZero — a camera-only driving planner trained through interactive closed-loop worlds rather than human trajectory imitation, reporting state of the art on three benchmarks (7↑)