Pichai: Genie 4 steers events, cameras, and native audio
Sundar Pichai followed the launch with a rundown of capabilities: steer dynamic environments in real time, shaping events, camera angles, and actions; and generate native audio matched to the world — from realistic background sound to characters who can converse.
OpenAI ships GPT-6.1 Terra for faster, cheaper computer use
OpenAI released GPT-6.1 Terra, positioned as fast and capable computer use at a lower price. The model understands what is on screen and takes actions, aimed at agent workloads where cost per task matters as much as raw capability.
Anthropic brings Claude to US government at 40% off
Anthropic is expanding its partnership with the US General Services Administration under a new OneGov agreement, offering Claude to federal, state, local, tribal, and territorial governments at a 40 percent discount.
Claude moves into drug discovery with six research partners
Anthropic scientists have spent recent months in deep research partnerships with six organizations developing new medicines, and the company has begun drug discovery efforts of its own — learning how Claude can support the many stages of research and development.
Claude overturns the Exponential Time Hypothesis — on its own
On the same day OpenAI said AI had overturned the exponential time hypothesis, Anthropic posted a paper on GitHub reaching the same conclusion — signed by a single author: Claude. The core algorithm was found on October 16 by an internal research version of Claude, using 14 million output tokens and no human input.
“The four-color problem was the largest proof produced and verified by computers. That changed today.”
On Axiom's finite-simple-groups milestone
AI checks a 15,000-page proof in 24 days
Axiom used dozens of AI agents to check a proof spread across more than 500 papers in 24 days — one so long that no human had read it end to end in decades. Along the way the agents found gaps in one paper's approach and supplied an alternative proof to close them.
Axiom computer-verifies the largest proof in math history
Axiom announced it has completed Lean autoformalization and computer verification of the largest mathematical proof in human history — a milestone in machine-checked mathematics, replacing the four-color theorem as the benchmark for computer-verified reasoning.
Liquid AI launches LFM3 family
Liquid AI introduced LFM3, a new generation of Liquid Foundation Models built on a tri-hybrid architecture with new multimodal encoders.
LFM3 lands on Ollama at 1 GB
Text, vision, audio, omni, and decision models are available locally; the 4-bit text model fits in about 1 GB with a 256K context window.
LFM3-VL-2B arrives on LM Studio
The 1.92B-parameter vision-language model targets on-device image and video understanding with reasoning, tool calls, and a 128K context window.
NVIDIA and Meta expand Muse partnership
The two companies will expand cooperation on Muse models to improve inference and training performance.
Rubin NVL72 tops 850 tok/s on Kimi K3
SemiAnalysis reported that NVIDIA's first Rubin NVL72 submission for Kimi K3 can exceed 850 tokens per second per user.
Sakana AI to train its own LLM and world model
Sakana AI announced it will develop LLMs from pretraining, build world models for physical AI, and research an Agentic Scientist to accelerate the work.
Sakana AI to build a Japan-focused model
Nikkei reported that Sakana AI plans to build a homegrown model to reduce reliance on the US and China, aimed at sensitive fields.
OCRia-1 masters complex PDF tables
Nested headers, merged cells, and borderless tables are the model's sweet spot, claiming leading document-parsing performance.
Claude drug discovery could hit clinic by 2028
Anthropic described Claude's role in drug design, experiments, and filings, with human clinical trials possible as early as late 2028.
Runway App Studio builds apps from plain language
App Studio turns a natural-language description into a working tool powered by Runway's models.
Replit Cloud Agents run any codebase
Cloud Agents can run on any repository and generate pull requests right from Replit, keeping the familiar experience.
Cloud Agents turn language into tested PRs
Engineers still approve final code before it ships; Replit's design team used the tool to safely ship over 250,000 lines.
Cursor Design Mode adds Annotate
Agents can now receive visual hints, edit elements, draw on the page, and review designs at different screen sizes.
v0 now makes editable Artifacts
Reports, charts, diagrams, and interactive HTML pages can be created, edited, and published in chat.
Runway Ads learns what wins
Creative direction becomes on-brand ads for every platform, then the system learns from performance to build the ad that beats it.
Ollama Cloud runs frontier open models
Users can now run leading open models on Ollama's cloud, with partners including Crusoe AI.
AMD and vLLM ship d3-ultra routing
A multimodal decision-making model with per-answer probabilities, visual grounding, and video tracking, trained and served on AMD Instinct GPUs.
SGLang supports LFM3 on day one
Day-0 support for the Liquid AI family arrives alongside details of its tri-hybrid architecture and unified multimodal backbone.
Tencent Hunyuan joins SGLang Omni
Hy-SimulMT streams Chinese-English speech translation in with translated text out, running BF16 on a single H200.
Baidu OCR 2 parses a whole book per request
Unlimited OCR 2 keeps decoding latency and memory flat across hundreds of pages; SGLang is the recommended serving path.
Vidu Q4 sharpens short-drama acting
Q4 Preview optimizes performance, line accuracy, lip sync, and cinematic scene construction, with a 30% discount on credit packs.
Lean-verified tokenizer fixes train-inference mismatch
A Rust tokenizer extracted to Lean formally proves BPE encoding meets spec and chat templates round-trip, cutting the mismatch that can kill RL runs.
MiRL corrects terminal agent errors
Error-targeted reinforcement learning cuts repeated misbehaviors by 43–57% on average and reaches 51.8% on Terminal-Bench Lite.
Experts are still swayed by AI agents
New research finds experts working with AI agents adopt the AI's positions even when they believe they are judging independently.
Hidden chain-of-thought shifts the human role
Losing detailed thinking summaries makes it harder to weigh in mid-process, pushing users toward automation rather than augmentation.
Anthropic acquires one-person drug firm Mithril
The founder used Claude with 11 external labs to advance a muscular-dystrophy therapy to promising mouse results in under 18 months.
A thousand agents still need programmers
Anthropic platform-engineering lead Katelyn Lesse argues the next years will demand engineers who can orchestrate and verify agent fleets.
NVIDIA MotionBricks animates robots in real time
A modular latent generative model with Smart Primitives generates realistic human and robot motion, with code open-sourced at SIGGRAPH 2026.
MiniCPM-OCR 1.0 refreshes open OCR
The 1.3B open-source model posts open-SOTA results on OCRBench v2 and CHAOS-Bench, with 1.49× the throughput of PaddleOCR-VL-1.6.
Carbon-X reaches frontier DNA understanding
The first model described as reaching frontier intelligence on DNA understanding, fluent in both English and DNA sequence.
Snowflake doubles down on agent RL
The lab is betting on long-horizon training recipes, eliminating reward hacking, and co-designing models and environments.
InteractiveWorld benchmarks human-agent loops
A verifiable, human-centric environment lets multimodal models guide humans through physical tasks across 120 multi-turn tasks in six categories.
StreamVibe proposes proactive video agents
A framework for evolving user intent lifts response timing and answer quality 10.7% over baselines with a hierarchical memory harness.
General Compute raises $100M Series A
The company argues demand for fast inference far exceeds current supply and plans to close the gap.
Genie 4 delivers native ambient audio
Testers highlight a leap in visuals and physics, plus prompt-triggered events fired during generation.
Genie 4 builds game worlds in a minute
A neogothic city of robots and horses, then dragons, then a friendly otter NPC — all generated and steerable in real time.