AI — Page 16

Google says it co-developed TorchTPU with Meta's PyTorch team; [RadixArk worked with Google and Meta on SGLang-TorchTPU](https://developers.googleblog.com/announcing-new-vllm-and-sglang-inference-engines-on-tpu/), which Meta says is in production.

Quantinuum's October 19th result uses color codes on its trapped-ion Helios system; it measures quantum memory, while Apollo remains a 2029 target.

The 64GB models are scheduled to go on sale October 23rd; NVIDIA says eight-system configurations require manual setup, while Sync's guided assistant currently handles up to four.

The giant posted a 26% revenue jump and triple EPS growth while scaling infrastructure spend for its frontier AI models.

The October 21st Berlin keynote demo linked simulation software, AI models and NVIDIA hardware to train a policy for Hugging Face's Microduck robot.

Carmen Li's GPU-rental benchmarks would underpin cash-settled H100 and B200 contracts, but regulators had not cleared the market as of October 21st.

The waitlisted agent monitors connected work apps and drafts or completes tasks, while its early version can still misread threads and urgency.

The waitlisted research preview checks connected accounts about every three hours, prepares morning briefs and drafts documents across iPhone, Android and the web.

The tutorial lays out how one runtime can coordinate voice models, interruptions and persistent sessions; MiniMax's post does not claim the company uses it.

Co-founder Daniel Lord-Doyle says Mary Technology's integration carries a matter's evidence, team findings and review history into Harvey as Mary Technology expands in the U.S.

The report says current interpretability methods are not dependable enough to catch every dangerous behavior. A separate assessment published September 9th revisits cyber incidents Anthropic first disclosed July 30th; those tests ran without safeguards included in released models.

The October 21st announcement puts GPT-6 models and Codex workflows in Google Cloud's marketplace, with private preview underway and general availability planned in the coming weeks.

Founder Sujal Dhakal says more than 5,000 families use Vene, which handles care coordination by phone and text and is pursuing provider-network deployments.

The two messaging-based AI products have overlapping audiences and features; Fae users are scheduled to begin moving to Tomo on November 2nd.

Led by Max Kirkby, the lab reports top benchmark scores for two open-weight models built to return calibrated probabilities instead of generated answers.

Richard Socher's first model at Recursive is open-weight and Apache 2.0 licensed, with performance and speed claims the company has not independently verified.

The public beta adds recurring scans across repositories and Git providers, fix tracking, an API and an MCP connector for Claude Enterprise customers.

The October proof removes dimension dependence from a 1995 bound, after a human-directed AI search helped identify the key mathematical criterion.

On a Portuguese essay grader running Qwen3-4B, Dharma-AI reports lower measured cost; its VocabX paper remains under review.

The model generates camera frames from simulator data; its published tests compare against Unreal Engine renders, not real-world footage.

The research system pairs candidates with evolving explanations; its tests rely on LLM judges, fictional customer profiles and company data last updated in 2024.

A model-switch prompt surfaced in Kimi Code, but it confirms neither a release date nor K3.1's availability, price or capabilities.

George Saon and IBM's speech team added a 2B multilingual model and revised two others, with performance figures largely reported by IBM.

NVIDIA reports a 27.1% throughput gain for its 1B NVFP4 embedding model over BF16 on GB200. Separate agent simulations report 13% to 14% fewer inference tokens with reranking.