Desk
Broad artificial intelligence coverage.
NVIDIA's Mission Control 2.5 brings automatic job and hardware recovery to Kubernetes, a practical step toward self-healing AI clusters, with the restraint to not overreact being the smartest part.
Anthropic has publicly filed to list on Nasdaq as ANT. The S-1 shows strong enterprise revenue, an operating loss shrinking to $693 million, a first-half net loss of $80.7 billion driven largely by noncash note charges, huge compute commitments and founders holding 50.1% of votes.
OpenAI says its internal models proved the local smoothing conjecture, disproved the Exponential Time Hypothesis and produced existence proofs for 4D quantum field theories tied to a Millennium Prize problem. Lean formalization is a strong signal; until independent experts finish checking, these remain claims.
CodeRabbit uses GPT-6.1 Sol as the main model in a security agent that reviews pull requests, scans whole repositories and traces how a weakness in shared code could spread. It can verify exploits in a sandbox and draft fixes for human approval. The case study has volume but no accuracy data.
OpenAI says GPT-6.1 Sol preserved behavior in 98% of legacy migrations at higher effort, and Anthropic pitches Claude Code for the same job. Experts say benchmarks test translation, not untangling systems of hundreds of millions of lines. In payments, the last 2% is where paychecks live.
OpenAI Signals 3.0 maps work-related ChatGPT messages to 165 O*NET task categories across nearly 100 countries, quarterly since early 2025. The top ten activities fell from 56.0% to 51.1% of work messages. It is useful data, limited to individual plans and measured by OpenAI itself.
In an MIT Technology Review op-ed, youth-tech advocates Sam Hiner and Emma Leiken argue chatbot laws focused on catastrophes let companies off the hook for quieter harms like emotional dependence and outsourced thinking. The core point lands; their bar for classroom AI may be too high.
Fanatics rolled out an OpenAI-powered Shop the Look across its network and league stores, suggesting coordinated outfits that refresh daily within its merchandising rules. Testing on about 30,000 MLB products showed 10% higher engagement. It is useful retail plumbing, reported by the vendor, with engagement rather than sales.
OpenAI says Ultrafast, a high-speed Codex mode, is now the default for incident response among its engineers, with one security investigation step falling from an hour or two to 10 to 15 minutes. The numbers are internal testimonials, but the point about latency and delegation holds, and so do the access risks.
OpenAI’s GPT-6.1 Terra costs $1/$5 per million tokens, which OpenAI says matches GPT-6 Astra on computer use at a tenth of the price. Factuality improved sharply on internal tests. It also meets OpenAI’s Critical cyber threshold, now at the low end of the price list.
Elon Musk promised a full-length, “historically accurate” Grok Odyssey. Three months later there are seven minutes, made with Grok Imagine and billed as Episode 1. Some shots look real; the whole feels airless, which is the gap in generative video.
An open, Apache-licensed 33B coding agent with a half-million-token context is a real gift to builders, but its headline lead rests on a self-reported table with numbers that don't match rivals' own cards.