AI engineering, October 23, 2026

Top News · 230 HN points

Theorem's research pipeline uses Claude to turn existing fuzz tests into specifications and generate formal proofs for compiled production code. The Rocq kernel checks those proofs, and a subset is being published for independent verification and reuse.

Here is what makes this credible:

• The work covered 239 entry points across 123 OSS-Fuzz libraries, generating about 180 million lines of Rocq proofs.
• Opus 5.5 proved half of 36 verification tasks using about 40% fewer tokens than Opus 5.
• Sharing proof work cut token use by 17% compared with agents working alone.
• Theorem offers to run the pipeline on other open-source projects and is publishing a Claude skill.

One catch: Proofs can require hundreds of millions of tokens. Coverage is limited to specified crash properties for inputs up to 64 MB, under a single-threaded, static RISC-V Linux setup with a fixed memory budget. This does not establish that entire libraries are bug-free.

Sources: anthropic.com

Top News · Reported · ctee.com.tw · translated from Chinese

NVIDIA's new DSX POD combines Vera Rubin compute, networking, power, and liquid cooling in a modular containerized system. NVIDIA says moving integration into supplier factories can cut deployment from six months to as little as 14 days once site facilities are ready.

Here's what changed:

• A design unit connects seven liquid-cooled compute racks with about 17 km of Spectrum-X network cabling.
• Wiring, validation, and burn-in happen before shipment, reducing on-site connection and troubleshooting work.
• The system uses 800 VDC power, and Vera Rubin supports coolant temperatures up to 45°C.
• Ingrasys, a Hon Hai subsidiary, is a global partner for Vera Rubin-based modular AI factories.

One catch: The deployment timeline assumes site facilities are ready. Global delivery is expected in early 2027.

Sources: ctee.com.tw

Top News

If you need typed answers from text and images, OpenAI's Decisions API is now generally available with gpt-6-luna. OpenAI claimed it was 10× faster than the Responses API when it announced the public beta. The GA release doubles the question limit per request and adds shared-input caching, image instructions, and structured-data input.

Here's what changed:

• Requests support up to 400 questions, up from the public-beta limit of 200.
• Shared input can be cached across requests to get prompt-cache discounts on repeated input.
• Images can appear in individual question instructions, so each question can carry its own image.
• Structured data can be sent through input_json.

Why care? You can batch more questions into each call and reuse shared context across requests at discounted prompt-cache rates.

Try it: Send a request to v1/decisions using gpt-6-luna.

Sources: developers.openai.com

1. Modular releases MAX source code for royalty-free commercial use across CPUs and GPUs
2. Modal releases Spindle, an open-source Tinker-compatible training server charging for GPU time instead of tokens
3. Ollama adds Recursive Flint coding model with a 262K-token context window on macOS · 8 Stars today
4. CoreWeave ships NVIDIA GB300 NVL72 instances, reporting over 6x H100 throughput per GPU on DeepSeek R1 · Official · coreweave.com
5. Amazon SageMaker adds training session APIs for custom LLM training loops without managing GPUs
6. Claude Opus 5.5 passes 58% of runs on FrontierCS 2 computer science research benchmark · 2 Upvotes