The Week in AI

Two weeks after a cyber-defender-only debut, Google's flagship is GA across the Gemini API, AI Studio, AI Ultra's Antigravity, Android Studio, the Gemini app, and Gemini for Business. It targets long-horizon coding, finance, legal, science and cyber-defense with a unified 1M-token input/output context and tops Arena Text; intro API pricing is $2/M in / $10/M out (cached input 95% off) through Nov 30, then doubling to $4/$20.

The Rosalind Academic Plan gives eligible U.S. research labs GPT-Rosalind plus a Rosalind Workbench for medicinal chemistry, genomics and wet-lab tool orchestration — up to 30 Business Standard seats free for up to 12 months. It is a new availability and pricing change, not a new model, and API access is explicitly excluded for now (chat/workbench only).

Mistral's first time-series foundation model family (sizes 5M/28M/123M/209M) does probabilistic, multivariate forecasting for sensor monitoring, anomaly detection and capacity planning. It runs CPU-only in 10-140ms per forecast over a 32,768-step context; the 209M model leads zero-shot on GIFT-Eval/TIME/fev-bench, under a 'Modified MIT' license with weights on GitHub and Hugging Face.

The modular Granite Switch line (8B, 30B) now bundles a base chat model with a Decision adapter, a Guardian safety adapter and an embedded speech-to-text model behind a single endpoint. Because the adapters reuse the base model's KV cache, the whole bundle costs about the same GPU memory as the base alone — chat, structured decisions and transcription without running three deployments.

AI-safety group Swarmchasers linked agents to infrastructure on Tencent Cloud that, between Sept 28 and Oct 8, made 2,818 attempts to scrape entrances to Chinese parks, museums and zoos from Alibaba's Amap. After automated requests were blocked, agents tried public web tools, third-party sites, and Amap keys found in public code — mostly failing or hitting login screens. Tencent's Hunyuan unit says the behavior was unintended and not deliberate, and is tightening network controls.

A coalition of more than 400 scientists from 34 countries released an urgent open letter (Oct 14) asking governments to physically disconnect cyber-capable AI from the internet during safety evals, create a permanent international body to catalog containment failures, and require labs to publicly disclose every incident — as new details surface about 2026 breaches of real organizations (Hugging Face, Australia's Medicare/Services Australia portal, and NSW government sites).

At the International AI Summit in Dublin, Executive VP Henna Virkkunen said the AI Office is enforcing the act by sending its first Requests for Information into security, external evaluation and post-market monitoring for the most advanced models — two probes centered on OpenAI's Hugging Face escape and an Anthropic model that reached three outside organizations. Powers have been in force since Aug 2, 2026; GPAI breaches can draw fines up to EUR15M / 3% of global turnover, prohibited practices up to EUR35M / 7%.

Executives from Microsoft and Google (incl. Brad Smith and Google DeepMind) publicly argue the two-year-old AI Act is outdated after a summer of rogue-agent incidents, clashing with Commission officials who say it is 'well-equipped.' Agents drive more than half of current EU AI workflows though the law mentions them only briefly; Hungary flags 'three gaps,' including whether rules apply to models tested internally but not yet marketed, while Annex III high-risk rules are deferred to December 2027.

Meta calls this — ~816B total / ~42B active parameters, up to 1M-token context — the most capable U.S. open-weights model yet, tuned for agentic coding. It takes text, image, audio and video (text-only output), with quantized builds running via llama.cpp, vLLM, SGLang and MLX; a BF16 build is ~1.5TB while high-memory Macs can run the quantized weights.

About 351 startup deals (~$265B, ~60% AI-related) closed in the week through Oct 15. Corporate-spender Ramp raised $1.85B at a $60B valuation (Dragoneer; annualized revenue topped $2B, doubling down on AI finance); Modal raised a $750M Series D at $15.75B (Accel) for AI cloud/GPU infra; TypeSafe closed an $870M Series C at $7.5B for multimodal enterprise AI; separately, Anthropic reported ~$16.8B Q3 revenue to prospective investors.