AI — Page 3
Models, agents, infra, applied AI.
- Report ties Alice Huang network to $7.5B in Nvidia chip shipments to China
The Information ties Megaspeed executive Alice Huang to about 240,000 Nvidia chips for Chinese firms. Reuters could not independently verify the intelligence allegation, which China has denied.
- Google DeepMind's new benchmark tests agents on blueprints, schematics and polished decks
Built with Artificial Analysis, MKA uses 200 expert-designed tasks to grade visual knowledge work. Google says Gemini 4 Carbon ranks second on quality at a lower cost.
- Tyler Nishida sketches a chalkboard tutor with Claude Fable 5.2
The design engineer used Anthropic's model to prototype a board that presents answers step by step; the mockup describes a learner-controlled pace.
- Linux Foundation survey finds trust leads open-source firms' competitive advantages
A survey of 225 founders and enterprise leaders by the Linux Foundation and Serena finds AI boosting internal productivity, while enterprise sales remains a challenge.
- Egen joins Google Cloud's Gemini for healthcare rollout as an implementation partner
The Naperville services firm led by Alim Somani will implement Google's new healthcare AI for providers, payors and life sciences companies.
- DeepMind's AlphaProof Code proved selected Signal guarantees in four weeks
Pushmeet Kohli and Swarat Chaudhuri led a DeepMind effort that proved selected libsignal guarantees down to machine code, with humans reviewing specifications and flagged proof code.
- InnovateEDU joins IBM program to build AI literacy dashboard
Ramona Technologies, Udhyam and Colorado join InnovateEDU in a two-year IBM pro bono cohort, with no per-group grant value disclosed.
- Quantinuum upgrades Reimei to H2 for hybrid supercomputing in Japan
Quantinuum and RIKEN upgraded Reimei to a 56-qubit H2 and plan work on a fault-tolerant system and exploratory work with SoftBank on connecting quantum computers with AI data centers in Japan.
- Google DeepMind lists nuclear expert role to test Gemini weapons risks
Google DeepMind listings observed on November 9th, 2026, describe nuclear, radiological, biology and threat-modelling work; the original posting dates are unverified.
- Gravity puts ads under three LLMs and cuts token prices by up to 25%
Gravity's Inference.net launch lists Kimi at 20% off and DeepSeek and GLM at 5% off, with ads appended to about 4.2% of answers.
- Microsoft patent application proposes AI training on idle gaming PCs
Microsoft's application, filed April 30th, 2025, and published November 5th, 2026, describes splitting training into micro-tasks for opted-in devices and offering owners credits.
- Skild AI open-sources Steel physics solver, says it beats PhysX by up to 13x
Skild says Steel hit 3.90 million environment steps per second on 20 mixed Isaac Lab tasks at 245,760 environments. At a matched batch size of 61,440, Steel reached 3.61 million steps per second and used about 11 GiB of GPU memory to PhysX's 29 GiB.
- Anthropic's $200 Ultra plan measures at about 9x Pro per week
Tokenusage.dev counted 655M weekly tokens for the old Max 20x against 71M for Pro, a 9.2x ratio that matches Anthropic's newly published Ultra figure.
- Sanmina revenue jumps 79% to $14.53B after ZT Systems deal
Sanmina co-founder Jure Sola reported $3.86B in fourth-quarter revenue as acquisition costs and rising debt accompanied its expansion into AI infrastructure.
- Anthropic says Claude drafted routes to four targets from unpublished syntheses
Anthropic says three Claude models each wrote routes to four targets in 26 to 84 minutes. Professors reviewed the plans blind; none has been tested in a lab.
- OpenAI's GPT-6.2 timing remains unclear amid separate safety reports
A reported November 6th window passed without GPT-6.2 Astra or Sol appearing in OpenAI's catalog. OpenAI has not confirmed that safety issues delayed either model.
- Google puts its 7,000-language Explorer in a Berlin museum exhibition
The interactive installation visualizes LinguaMeta data on 7,000-plus languages at the Museum of Communication Berlin through February 27th, 2028.
- Anthropic says Fable 5.2 designed three working devices from one-paragraph prompts
Anthropic worked with Diode Computers on a speakerphone, a voice-assistant hub and a virtual pet, and says human engineers built all three without circuitry fixes.
- America.gov's in-house model claim leaves Gemini, Grok and Bedrock unexplained
The October 17th note claims a 5x speedup from an in-house model, while provider disclosures point to Gemini, Grok and an Amazon Bedrock code path.
- OpenAI sells $200 Codex Chromatic handheld that turns prompts into Game Boy games
The $200 handheld includes a programmable cartridge, returns November 19th, and gives Palmer Luckey's ModRetro a direct line into Codex developers.
- Correction: Thinking Machines identifies Fledge Alpha as Inkling 2 Small tests
Thinking Machines says its free OpenCode stealth model ran early Inkling 2 Small checkpoints, offering a possible explanation for shifts in token counts.
- OpenAI dots can now spread tasks across multiple local computers
OpenAI's persistent agents can now route a task to a second Mac or PC, testing Xcode builds and older SDKs while the developer keeps working.
- Alibaba gives Qoder users a free daily allowance for mystery model Qoduck
Qoder's November 7th promotion offers individual users a limited daily allowance. A reported tokenizer test suggests Qwen lineage, but the model's identity remains undisclosed.
- Anthropic is hiring an APAC national security lead to push Claude into allied governments
The Australia-based role requires 10-plus years in national security, eligibility for an Australian clearance, and close work with public-sector sales.