01
Bloomberg Technology ↗
The Department of Homeland Security has ordered all federal civilian agencies to stop using DeepSeek products and remove any existing deployments from their IT systems. The order also bars the Chinese AI lab from taking part in federal government contracts. DHS warned that DeepSeek could transfer sensitive data on US government employees and vendors to the Chinese government and that its models could contain malware or security vulnerabilities. The move came days after DeepSeek released its V4.1 Flash model.
02
TechCrunch AI ↗
AWS released TorchNeuron in public beta, letting developers move PyTorch apps onto its Trainium chips with a one-line configuration change instead of a code rewrite. Amazon estimates about 85% of AI apps are built on PyTorch, where Nvidia's CUDA support has long made its GPUs the default choice. Voice AI company Deepgram says it now processes 70% more audio per dollar after moving to Trainium. Amazon also plans to contribute code upstream to make PyTorch less tied to any one chipmaker.
03
VentureBeat ↗
Chinese lab MiniMax released M3.1 Flash, an agentic model with a one-million-token context window priced at $0.15 per million input tokens and $0.60 per million output tokens. It scores 42 on the Artificial Analysis Intelligence Index, up from 29 for M3, while cutting average cost per task from about $0.51 to $0.22, and MiniMax claims strong results on coding, computer-use and cybersecurity benchmarks. On the AA-Omniscience test, however, it guessed rather than admitting uncertainty on most questions it got wrong, giving a 93% hallucination rate. MiniMax plans to release the weights on October 23.
04
Engadget ↗
Google researchers say four Gemini 4 Argon agents defeated Minecraft's Ender Dragon in a randomized survival world without assigned roles, strategies or game-specific tools. The agents coordinated through in-game chat and shared memory, splitting up jobs like tool-making, food gathering and portal building over a 27-hour run. Along the way they wrote 2,783 programs, exchanged 229 messages and died 605 times, using 1.35 billion input tokens. Google calls it the first fully autonomous end-to-end completion of Minecraft by AI agents.
05
Unite. AI ↗
Meta published a technical account of Prometheus, the gigawatt-scale supercluster in New Albany, Ohio, that trained its Muse family of models including Muse Wonder. The site already runs hundreds of thousands of GPUs as a single machine and is set to scale to roughly 1 million, with Vera Rubin racks drawing over 170 kilowatts each on the way. Meta says it cut training interruptions by 99% since mid-2025 and now keeps more than 85% of compute productive during pre-training. Its next supercluster, Hyperion, is being built in Louisiana.
06
TechCrunch AI ↗
Jim Keller's chip startup Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double the $2.6 billion it was worth in late 2024. The round was led by Trace Capital and Innovation X, with Fidelity, T. Rowe Price, SanDisk, Eclipse and Cadence Ventures joining. The money will scale chips built for serving open-weight models and video generation. Tenstorrent stayed independent after reportedly fielding acquisition interest from Intel and Qualcomm earlier this year at valuations of up to $5 billion.
07
Inception ↗
Inception launched Mercury 3, a diffusion-based language model more than 40 times larger than Mercury 1 with a one-million-token context window. The company reports 1,068 tokens per second on Nvidia GPUs, four to eight times faster than speed-optimized frontier models, and says it matches or beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks. It costs $0.20 per million input tokens and $0.70 per million output tokens, half price during launch. The model is available through Inception's API, OpenRouter, Baseten and Vercel.
08
Forbes Tech ↗
Dream Labs, founded by former Nvidia robotics researchers who worked on GR00T, raised a $90 million seed round led by B Capital and Kindred Ventures. Its Dream-1 world action model, trained on human videos, predicts video and actions together so it can drive different robot bodies through each machine's own controller. The company launched a public API, a fine-tuning service and a browser playground that controls real robots in its San Francisco office. Most of the funding will go to compute.
09
Unite. AI ↗
Town raised a $108 million Series B led by Forerunner Ventures and Menlo Ventures at a $1 billion valuation, just four months after a $55 million Series A led by a16z. Its assistants, called Townies, connect to each user's email, documents, calendar and more than 60 work tools, and can now hand off tasks to one another and work with outside contacts. Town says paying customers have grown fourfold since July and it hit seven-figure annual recurring revenue within weeks of launch. Customer data is not used for model training by default.
10
TechCrunch AI ↗
Tolan raised $30 million from Union Square Ventures, Khosla Ventures and Homebrew, doubling its total funding, with PitchBook estimating its valuation at over $268 million. The app gives users a cartoon AI companion that helps them manage daily tasks, relationships, emotions and goals. It has more than 250,000 active users who held over 70 million conversations in the past year, about 70% of them women and 40% parents. The company says annual recurring revenue has reached the double-digit millions.