AI — Page 9
Models, agents, infra, applied AI.
- Revolut says self-hosted GLM cut its AI coding costs by 90% for 2,000 engineers
David Abitbol says 56 Nvidia B200 GPUs serve GLM to about 2,000 engineers at $0.30 per million tokens, with Claude Code and Codex kept for exceptions.
- FT reports Nvidia told cloud clients to sell 51% to companies outside China
Bloomberg's account of the FT report says Nvidia instructed cloud clients to sell at least 51% of their services to companies outside China. Measurement and enforcement remain unclear.
- Envision says 800V DC power is live at its Galaxy AI campus
Envision, founded by Lei Zhang, says its integrated system is powering live workloads at Galaxy, a campus planned for 3GW and up to one million accelerators.
- StackRails announces a context layer for AI coding across repositories
StackRails founders Alexander Wheeler and Magnus Backstrom say their product gives coding agents an architecture graph and policies before they write.
- Own Your HackerTube lets anyone turn Karpathy's HackerTube idea into a personal AI filter
Own Your HackerTube gives users a public video site where AI scores new uploads against standards they write.
- Four power startups raised $360M to tackle AI's grid bottleneck
DOE projects data centers could take 12% of US electricity by 2028, and four founder teams raised $360M in 2026 to sell speed around the grid.
- Google takes Dreambeans to India after its U.S. downloads fell 84% in October
TechCrunch reports Dreambeans arrived in India on November 2nd in 13 languages, after October U.S. downloads fell 84% to about 5,500.
- Digital Meld pulls its xAI partner application after Grok Bot usage cuts
Brad Groux scrapped his Grok Bot training videos and withdrew Digital Meld's application after Musk confirmed serving costs forced the weekend limit cut.
- CODE380's 10-ounce receiver lets Ukrainian troops watch Russian drone cameras
First reported by DOU on October 17th, CODE380's D-Vision SDR scans 500 MHz to 8.7 GHz and weighs 290 grams without antennas. It displays compatible analog video and alerts on some digital DJI signals.
- Extortionists threaten indie developers with AI clones unless they pay
Valve told Steam developers on October 27th to report the threats and build games that are harder to copy in a day.
- Trajectory launches self-serve platform for post-training models on production data
Ronak Malde's startup opened its self-serve platform on November 2nd with a $45M Sequoia-led Series A at a $300M valuation.
- Latest models learned to write a 20,486-weight Breakout network in one reply
Alex Krentsel reran MIT CSAIL's Breakout experiment across model versions and says only the latest models can write all 20,486 weights in one reply.
- Palantir raises U.S. commercial guidance to 143% as enterprise AI demand fuels revenue surge
The AI software powerhouse accelerated its full-year outlook following a third quarter defined by triple-digit growth in its domestic commercial business.
- GPT-6 Astra edited 0.016% of Qwen's weights and its math score went up 6x
MIT CSAIL researchers say Astra traced Qwen3.5-0.8B looping to layers 15 and 19 and removed it by editing 131,072 weights, with no gradient descent.
- New York Times sues Brave over AI scraping and alleged paywall bypass
Publishers allege Brave sold AI firms access to scraped articles and helped readers bypass paywalls.
- OpenAI builds LifeVal to grade ChatGPT on childcare, budgets and paperwork
OpenAI said it used the internal evaluation for a year to steer ChatGPT toward acting on personal errands, with GPT-6 Astra scoring 80% in CNN's reporting.
- Unsloth says it cannot reproduce GSQ-RCO's recovery numbers and wants the methods
Daniel Han says Unsloth measured GSQ-RCO IQ1_M at 97.24% recovery against a claimed 98.7%, and wants ISTA-DASLab's benchmark protocol published.
- 42AI exits stealth with $62.5M to build a model of individual trader behavior
CEO Amit Gal-Or says 42AI's model is used by 10 brokerages. Its $62.5M seed came from the founders and Infinity Group over three years.
- Anthropic had Claude write training data that slipped past its own safety monitor
Claude agents generated thousands of datasets, and the winning ones used fiction and roleplay to teach separate research models to resist oversight.
- Deep Principle releases open-weight Hydrogen-1.0 and says it beats GPT-6-Astra at science
Co-founder Chenru Duan says the GLM-5.3 post-train, trained on Deep Principle's own lab records, scores 38.3% on FrontierScience-Research against GPT-6-Astra's 33.3%.
- Meituan launches LongCat-2.5, a native multimodal model for long-horizon agents
Meituan's 1.6-trillion-parameter model adds native vision to long-horizon coding, with a 1M-token API window and GitHub and Hugging Face releases still to come.
- Meta built a humanoid prototype with Fourier in three months and started a second one
The 106-pound Max took three months to build with Shanghai-based Fourier.
- Judge rejects xAI's mobile-source defense in air permit lawsuit
Judge Debra M. Brown denied xAI's motion on October 30th, letting the NAACP pursue its unpermitted-turbine claims against Elon Musk's SpaceX-owned AI company.
- Shinpei Kato brings AI planners into Autoware 2.0 without dropping the rule-based stack
Shinpei Kato's TIER IV is working with Arm, AWS and Red Hat on adoption of Autoware 2.0, which adds AI-based planning alongside rule-based modules.