№ 107
Wednesday, September 9, 2026
AI & Tech Brief — September 9, 2026
№ 107
AI & Tech Brief — September 9, 2026
OpenAI says an internal model solved Navier–Stokes — one of the seven Millennium Prize Problems The proof shows that smooth 3D fluid flow can develop a singularity in finite time (resolving variants C and D of the official problem). OpenAI used a system of ~10,000 coordinating agents powered by an internal model “significantly more capable than GPT-6 Astra,” running for ~88 hours and consuming ~130 billion output tokens. The proof was formalized in Lean. But the announcement is overshadowed by a priority dispute: mathematicians Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) had concurrent work on the related forced Euler problem, and Buckmaster alleges OpenAI learned of their progress, redirected resources, and proposed publishing arrangements that would have excluded Alpöge. OpenAI denies accessing their Codex logs but acknowledges it “cannot rule out that de-identified data derived from their usage of our products helped improve our models.” 1,268 points on HN. Source: https://openai.com/index/navier-stokes-solution/ — discussion: https://news.ycombinator.com/item?id=49613262 — controversy: https://www.wired.com/story/openai-navier-stokes-math-discovery-academics/
Meta launches Muse, a personal AI agent for everyone Muse runs on a persistent, dedicated virtual machine with its own browser. It books appointments, fills out forms, handles customer service calls, and chats via its own app or WhatsApp. Users approve critical actions (sending emails, making purchases) before they happen, with a full audit trail. 530 points on HN. Source: https://ai.meta.com/muse/ — discussion: https://news.ycombinator.com/item?id=49615537
Google DeepMind: AlphaGenome Atlas — a predictive map of every possible DNA letter change The Atlas pre-computed the regulatory impact of all 9 billion possible single-nucleotide variants in the human genome using the AlphaGenome model, producing a 1-petabyte dataset with a single “AVI score” per variant. Already used by the Broad Institute to solve a rare disease case (DNM1 gene splice site) and by UK Biobank researchers who found 22% more non-coding genetic associations. Free web portal, no coding required. 567 points on HN. Source: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/alphagenome-atlas/ — discussion: https://news.ycombinator.com/item?id=49611251
Inception Labs releases Mercury 2.5 — the largest diffusion LLM ever trained 40% intelligence increase over Mercury 2, comparable to GPT-5.6 Luna (Low) and Gemini 3.5 Flash-Lite, at 1,107 tokens/sec on standard NVIDIA GPUs. 260K context, $0.20/M input, $0.75/M output (80% launch discount). Also announced: Mercury Voice (sub-170ms TTFT for voice agents) and Mercury Router (dLLM-based model routing). 206 points on HN. Source: https://www.inceptionlabs.ai/blog/introducing-mercury-2-5 — discussion: https://news.ycombinator.com/item?id=49616354
Terence Tao: open math problems are being “non-renewably mined” by AI Tao draws an analogy to a country surrounded by ocean but short on drinking water: the set of possible problems is infinite, but the set of fruitful ones — problems whose attempted solution generates insight and connections — is finite and being depleted. AI systems that beeline to answers without the productive struggle may strip-mine this resource. 375 points on HN. Source: https://mathstodon.xyz/@tao/117237320796901560 — discussion: https://news.ycombinator.com/item?id=49616968
DaVinci Resolve 21.1 adds AI assistant integration Blackmagic’s video editor now integrates with Claude, Claude Code, and ChatGPT Codex for organizing media, batch rendering, and creating highlight edits via natural language. Also adds 25 new Krokodove graphics tools and expanded camera format support. Free download. 408 points on HN. Source: https://www.blackmagicdesign.com/media/release/20260908-03 — discussion: https://news.ycombinator.com/item?id=49610181
OpenAI ships GPT Image 2.5 Sunburst and Flare
Two new image generation models in the API: Sunburst for precision editing, Flare for fast everyday generation. Both support new xhigh and max quality settings at GPT Image 2 token rates. Also GA’d: Prompt Cache Diagnostics in the Responses API for GPT-5.6+ models.
Source: https://developers.openai.com/api/docs/changelog
Gemini CLI v0.59.0: MCP OAuth SSRF mitigation and fail-closed workspace trust Security-focused release preventing server-side request forgery during MCP OAuth metadata discovery and enforcing fail-closed workspace trust in restricted mode. Source: https://geminicli.com/docs/changelogs/
“I-have-ADHD” — a skill that stops coding agents from burying the answer — 10 rules for LLM output: lead with the action, number steps, no preamble, no “Hope this helps!” Works as a Claude Code plugin. 462 points on HN, and the discussion is a goldmine of collective frustration with LLM verbosity. https://github.com/ayghri/i-have-adhd — discussion: https://news.ycombinator.com/item?id=49610631
Qwen3.8 27B quantization benchmarks: 4-bit holds up, 1-bit collapses — The 17 GB Q4_K_M quant matches the full 55 GB BF16 model on Terminal-Bench 2.1 and GPQA Diamond. At 1-bit, scores drop to random chance. The author spent ~$3,000 on Modal GPUs to find out. https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/ — discussion: https://news.ycombinator.com/item?id=49611128
How to build a printer — A developer turns a $30 Xteink X3 e-ink reader into a network printer by implementing IPP over HTTP on an ESP32 with 400 KB of RAM, using the display itself as the page buffer. Delightful. https://nishantjosh.dev/blogs/how-to-build-a-fking-printer/ — discussion: https://news.ycombinator.com/item?id=49617255
“How GPT-5.6 Sol helps run quantum computing experiments” (OpenAI blog) — An MIT grad student used Codex to automate qubit calibration. HN commenters note this is standard lab automation that Python scripts handled in 2011; the “quantum computing” framing oversells what is essentially routine measurement scripting. https://openai.com/index/codex-quantum-computing-experiments/ — discussion: https://news.ycombinator.com/item?id=49622561
“Tweet” and the bird logo enter the public domain — A Delaware court ruled X Corp likely abandoned the TWEET trademark and bird logo (but not TWITTER, thanks to “formerly known as Twitter” in the App Store listing). Mildly interesting IP law, not tech news. https://blog.ericgoldman.org/archives/2026/09/tweet-and-the-bird-logo-apparently-enter-the-public-domain.htm — discussion: https://news.ycombinator.com/item?id=49621751