AI — Page 6
Models, agents, infra, applied AI.
- Swarms v17 InterLink gives AI agents a shared conversation instead of passed strings
Released on November 2nd, 2026, Swarms v17 InterLink adds six communication channels and a validated fan-out workflow, plus breaking changes that can silently ignore old arguments.
- Julius Research releases DeepFinance and the best model scores 54.8% on 27-hour deal work
Julius Research tested 16 model configurations on 21 anonymized PE and IB tasks averaging 27.1 human hours each, and none scored above 54.8%.
- Stanford researchers compress LLM memory into 16 tokens and beat full context on five benchmarks
Fallahpour and colleagues at Stanford report 44.6% on PersonaMem-v2 against 30.9% for the next best method, using 16 tokens instead of the full collection.
- Thinking Machines launches Inkling 2 Small on Tinker with closed weights
Mira Murati's lab plans to release the 594B-parameter model's weights under Apache 2.0 next week, following external safety testing and red-teaming.
- Google plans $17B Wyoming data center as voters back limits elsewhere
Google's planned Laramie County campus pairs a 702 MW fuel-cell development with carbon capture, air cooling and a pledge to pay its full power costs.
- Pompliano accuses Anthropic of fishing for Silvia secrets before Claude finance launch
Pompliano says Anthropic courted Silvia engineers in April, then warned the company days before Claude added Plaid-linked finance tools.
- A mystery seller puts $500M of data center loans up for auction days after the midterms
The dozen-loan sale includes a $25M slice tied to the Stargate site in New Mexico where Oracle has already cited force majeure over power delays.
- OpenAI pledges to pay power costs at campuses it owns and operates
OpenAI's November 5th compact promises early resident input, annual reports and limits on NDAs with public officials amid ballot-box opposition to AI data centers.
- World Labs puts Atlas in real time with Frameboy demo at 15 fps
Fei-Fei Li's World Labs streamed Atlas from cloud GPUs at up to 15 fps on November 5th and added the live mode to its closed-beta Viewfinder app.
- Google shows Gemini turning a paper sketch into a preview app in three steps
Google's November 5th walkthrough uses Canvas to read a photographed drawing, generate code and revise the running preview by chat.
- METR says Anthropic's AI capability growth rate has roughly doubled
The nonprofit estimates growth on well-specified tasks was about twice its March 2024 to October 2025 pace, and says AI likely drove most of the acceleration.
- NVIDIA details four open-weight Cosmos 3 checkpoints for driving and robot simulation
NVIDIA's 16B post-trained models cover multi-view driving data, synchronized camera-LiDAR streams, simulator-to-video rendering and two-arm robot prediction.
- NVIDIA open-sources cuRoboV2-Fast, a robot planner that works in 4.4 ms
Sundaralingam, who built the first 100 ms cuRobo prototype in 2022, is releasing NVIDIA's 4.4 ms planner with Adithyavairavan Murali and Stan Birchfield.
- Meta Muse voice and video chat reaches everyone in the US
Wang announced the US-wide rollout on November 5th, two months after Muse launched as Meta's task-running personal agent.
- PulseAI says its wearable ECG analysis model received FDA clearance
Founder Alan Kennedy's Belfast-based PulseAI announced U.S. regulatory clearance on November 5th. A 296-person study reported 92% fewer inconclusive Apple Watch ECGs using its algorithm.
- Reported GLM-5.5 pipeline claims 1 trillion crawled web pages for Z.ai
A November 4th post attributes six internal data engines to Jie Tang's GLM-5.5 work, with no official Z.ai announcement confirming the figures.
- Google ships Gemini Omni 1.2 Flash with 15-second clips at $0.07 a second
The November 5th release adds native 15-second generation, multi-turn edits across full clips, stronger reference consistency, and a $0.07-a-second 720p rate.
- Governments from New York to Amsterdam are slowing AI data center permits
Reuters' November 5th roundup collects moratoria, grid audits and water rules affecting AI data center development.
- Mistral expands Studio to put Le Chonk to work inside enterprise walls
Mistral linked its open-weight flagship to managed agents, workflows and apps that enterprises can run in their own cloud, VPC or on-premises systems.
- Aina opens preorders for a ring to approve AI agent actions
Founder Apoorv Shankar, who sold his first ring company to Ultrahuman, is taking preorders for an agent-control ring. WIRED reports a $399 introductory price and January 2027 shipping.
- Moonshot sought Blackwell clusters in Europe after US curbs blocked them in China
CNBC says the Beijing lab discussed renting Blackwell clusters from European neoclouds in 2026, but no signed lease or actual use was reported.
- Benzinga brings licensed market data to AI agents through Firecrawl Alexandria
Jason Raznick's Benzinga will feed six real-time market products into Eric Ciarla's Firecrawl Alexandria, with paid, credited access for developers building financial agents.
- Analysts say SpaceX's 10 GW build in 2027 is plausible if the power arrives
SemiAnalysis estimates 3.6 GW is already under construction as SpaceX targets an additional 10 GW of compute in 2027.
- Anthropic tells Claude Code to build the genre, not the game
Anthropic's Claude Code version 2.1.317 lets developers name a game to describe a genre or mechanics, while offering an original alternative to a direct copy request.