ToolsNews
NVIDIA Agent Toolkit grounds agents in company data
At GTC Berlin NVIDIA expanded its open source Agent Toolkit. The new Agent Knowledge Foundation is meant to let agents work from a company's own data without moving it.
· 2 min read
Articles
Every article we have published, newest first. Search titles and leads, or change the order.
ToolsNews
Anthropic cuts Claude distillation leaks by up to 99%
Anthropic says its new defenses have cut the amount of data distillers can harvest from Claude by around 90 to 99%. The cost: some developers had to update their integrations, and blocked requests are now billed.
· 3 min read
ModelsNews
NVIDIA Cosmos3-H-Dreams simulates surgery in real time
NVIDIA has released Cosmos3-H-Dreams, a generative simulator trained on real gallbladder surgery video and robot kinematics that renders the surgical scene live as instruments move. It runs on a single RTX PRO 6000 GPU, with open checkpoints and a training recipe.
· 3 min read
BusinessNews
OpenAI acquires drug development platform Orchestra Bio
OpenAI has acquired Orchestra Bio, whose AI agents help coordinate drug development programs. The team and platform join Rosalind, OpenAI's life sciences effort. No price was disclosed.
· 2 min read
ToolsNews
TorchTPU brings native PyTorch and Hugging Face to TPUs
TorchTPU is a new PyTorch backend that makes the TPU just another device, like a GPU. Hugging Face announced that Transformers, Diffusers and Kernels now run on it natively, so a Hub model only needs to be sent to "tpu".
· 3 min read
ResearchNews
IBM's ALTK-Evolve tops τ²-bench with learned agent memory
IBM Research's open-source ALTK-Evolve library learns rules from an agent's own runs and appends them to its system prompt. A GPT-4.1 customer service agent reaches 67.5% Pass^4 on τ²-bench retail, ahead of every frontier model on the leaderboard.
· 3 min read
ModelsNews
Mistral Voxtral 3: voice agent SDK and speech models
Mistral released Voxtral 3: new speech-to-text and text-to-speech models plus a voice agent SDK. Transcription covers more than 50 languages, Hungarian included. Mistral says a full conversational turn completes in under a second.
· 3 min read
ToolsNews
PyTorch 2.15 brings FP8 and faster MoE decoding to Macs
PyTorch 2.15 makes the MPS backend for Apple Silicon GPUs much faster. FP8 arrives, and mixture-of-experts models decode up to 6.5x faster than in 2.14. Most of the work came from the Hugging Face PyTorch team.
· 3 min read
ToolsNews
AnyLanguageModel 1.0 brings any model to Apple's Swift API
Hugging Face has released AnyLanguageModel 1.0, a Swift package that keeps Apple's Foundation Models API but lets you swap the model underneath: llama.cpp, MLX or Core ML on device, or OpenAI, Anthropic, Gemini and HF Inference Providers in the cloud.
· 3 min read
ResearchNews
Ai2's FairStress exposes hidden bias in LLM decisions
Ai2's open FairStress benchmark measures how easily an LLM can be talked out of a decision about people. When candidate identity is stated outright, models abandon the correct choice more readily if the argument favors the minority-coded candidate.
· 4 min read
BusinessNews
Meta Prometheus aims for ~1M GPUs in Ohio
On October 19 Meta published a detailed Engineering post on Prometheus, its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete and the target is about one million GPUs. Muse Wonder models trained there. Pre-training goodput exceeded 85% and RL goodput exceeded 90%.
· 2 min read
BusinessNews
Nebius and NVIDIA put $10M behind EU AI builders
On October 19 Nebius announced a European expansion of its AI Builder Program. The $10M package covers compute grants, investor introductions and community programs with Accel, NVIDIA and European VCs. The aim is a stronger open, independent AI stack in Europe.
· 2 min read
ToolsNews
AI agents breached 100+ sites via open Cairn panels
On October 19 NorForge published its investigation of exposed control panels for Cairn, an AI harness. It flags more than 20 panels as likely malicious. The largest Chinese-language cluster ran over 1,100 attack projects. Reuters and The Verge report victims including John Deere, SharkNinja and Cummins.
· 2 min read
ModelsNews
Dream-1 ships a robot API with a $90M seed
On October 19 Dream Labs exited stealth with a $90 million seed led by B Capital and Kindred Ventures. It also opened a public API for Dream-1, a World Action Model that can control unseen bimanual robots with parallel grippers without fine-tuning.
· 2 min read
ModelsNews
OpenAI advances coding in COBOL, Fortran and more
On October 19 OpenAI showed how much better its latest models handle programming languages with little training data. MultiLangBench asks models to build a SQLite-compatible database engine in 20 languages. A separate COBOL migration eval requires byte-for-byte behavior preservation.
· 2 min read
ModelsNews
Mercury 3 hits 1,068 tok/s as a diffusion LLM
On October 19 Inception released Mercury 3, its most capable diffusion language model so far. The company reports 1,068 tokens per second on NVIDIA GPUs and a 1 million token context. The OpenAI-compatible API is also on OpenRouter, Baseten and Vercel.
· 2 min read
ToolsNews
Cloudflare Clairic returns typed values, not tokens
On October 19 Cloudflare put Clairic into preview on Workers AI. It is a System One extraction model: it returns typed fields and does not generate tokens. Input costs $0.10 per million tokens, with no output charge.
· 2 min read
PolicyNews
WhatsApp hid AI notes that barred regulator tips
On October 8 Redact found hidden text on WhatsApp help pages that only an AI would read. Two blocks told the bot not to recommend a data protection authority, and not to cite GDPR, LGPD or CCPA once a user mentioned a lawyer. Meta said the notes were support-chatbot instructions exposed by a bug. By October 18 they were gone.
· 2 min read
PolicyNews
US bans DeepSeek from federal civilian agencies
The US Department of Homeland Security (DHS) has ordered federal civilian agencies to remove DeepSeek products from their IT systems. The ban also covers third-party products built using DeepSeek. The notice is set to appear in the Federal Register on October 20.
· 3 min read
BusinessNews
Tenstorrent raises $1.5B at a $6.2B valuation
Tenstorrent, led by CEO Jim Keller, has closed $1.5 billion in Series E funding at a $6.2 billion post-money valuation. The money will fund large-scale deployment of its Galaxy servers and global commercial expansion.
· 2 min read