AI Daily · November 11 Wednesday

2026-11-10 08:00 — 2026-11-11 08:00 (UTC+8)

1. 01

   OpenAI introduces Control Room to stop research model output within 35 seconds

   OpenAI described Control Room, an internal system that tracks model workloads across its research infrastructure and stops them when safety or security concerns arise. OpenAI says the maximum time from a stop instruction to no further progress is 35 seconds. It currently covers internal research workloads, not OpenAI's external model-serving stack.

   Read the full report

   According to OpenAI, Control Room connects every research inference workload, including agentic sampling for training and evaluation and data processing, and works with its misalignment monitoring systems. Each job requests a job_id from Control Room, and inference engines check authorization on every request; once a job is stopped, engines refuse further output even if the workload code keeps running.

   OpenAI says the up-to-35-second lag comes from an inference-authorization caching layer it aims to reduce to zero. The system fails closed: if Control Room goes down or is disabled, all inference requests are denied. Fail-closed is live for all frontier model research workloads and will finish rolling out to remaining research workloads over the next few days. These are OpenAI's own claims and have not been independently verified.

2. 02

   Australia plans AI safety regulator to enforce rules on frontier AI companies

   On November 10, Prime Minister Anthony Albanese and three ministers announced mandatory safety and security obligations for all frontier AI companies wishing to operate in Australia, to be legislated through the Australian AI Standards and enforced by a proposed new regulator able to impose significant penalties. Penalty amounts and legislative details have not been released.

   Read the full report

   The joint statement lists six obligations: maintain security controls across development and deployment; identify and manage serious risks from their models; allow independent evaluation of models; rapidly notify the government of incidents; maintain and publish safety and security frameworks; and respond to updated standards and rules as technology changes.

   The government says the new regulator would work with the Australian AI Safety Institute for technical expertise, and that the protections align with interim recommendations from the parliamentary Joint Select Committee on AI. The statement gives no penalty figures or start date, saying further legislative details will follow after detailed work and consultation.

3. 03

   Zyphra open-sources ZONOS3-Mini text-to-speech model that runs in real time on phones and browsers

   Zyphra released ZONOS3-Mini, an open-source text-to-speech model with voice cloning and word-level control of emotion, pace and delivery. It has about 800M total parameters with roughly 100M active per token, and weights are on Hugging Face under Apache 2.0. Real-time performance claims come from Zyphra.

   Read the full report

   According to the Hugging Face model card, ZONOS3-Mini uses a mixture-of-experts design (64 experts, top-2 routing) and can run directly in the browser via WebGPU, with native CUDA, Apple MLX and C++/ggml runtimes and serving through SGLang-Omni. The card lists text normalization for nine languages including English, Mandarin, Japanese and Korean, and up to 10 minutes of audio per generation.

   Zyphra says the model runs in real time on phones and laptops. The model repository was created on November 9, and the company then announced the release on X.

4. 04

   Google makes Gemini 4 Carbon broadly available to developers, enterprises and Ultra users

   Google said on November 10 that Gemini 4 Carbon is now available to paying developers and enterprises via the Gemini API, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. The model was first announced on October 28 and had previously been limited to a set of cybersecurity defenders.

   Read the full report

   Google calls Gemini 4 Carbon its strongest model, aimed at cybersecurity and long-horizon agentic work such as complex multi-step coding, debugging and open-ended scientific research; it still recommends Argon for most software engineering and knowledge work. Google is also previewing /swarm in Antigravity, which turns a single prompt into a coordinated team of agents.

   Introductory pricing through November 30 is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards. Google says Carbon ranks first on third-party benchmarks including CursorBench 4.0; those rankings are Google's own claims.

5. 05

   NVIDIA releases Alpamayo 2 Edge, a 2B-parameter open-weights driving reasoning model

   NVIDIA released Alpamayo 2 Edge, a roughly 2B-parameter vision-language-action model that generates driving trajectories and causal explanations from surround cameras and navigation inputs, built for real-time in-vehicle use on DRIVE AGX Thor. Weights are on Hugging Face under the OpenMDW-1.1 license, with code under Apache 2.0.

   Read the full report

   According to NVIDIA's technical blog, the model pairs the Nemotron-based reasoning backbone of Cosmos 3 Edge with a diffusion action expert and reasons adaptively, generating reasoning text only when needed to save compute in routine driving. Closed-loop RL post-training in AlpaGym improved closed-loop driving by 48.9% over the best checkpoint without it.

   NVIDIA says its deployment recipe achieves 160-190 ms pixel-to-trajectory latency on DRIVE AGX Thor. It also released AlpaJudge, a model for evaluating reasoning quality. All performance figures come from NVIDIA and its partners.

6. 06

   NVIDIA unveils GR00T 2, an open robot model it says runs fully on-board

   NVIDIA research lead Jim Fan announced GR00T 2 on X, calling it NVIDIA's most advanced open "System One" robotics model, able to run untethered on a robot's onboard compute. NVIDIA is also releasing its agentic framework, simulation and annotation engines, hardware design and about 100K hours of training data; X-Adapt recipes are due in December.

   Read the full report

   According to Jim Fan's post, GR00T 2 handles long-horizon tasks chained over 20+ steps and 5+ minutes, such as running chemistry lab experiments or assembling model airplanes; X-Adapt learns from human wearables and simulation and deploys zero-shot to diverse real robots; and the model integrates natively with frontier agents for physical reasoning, error recovery and safety monitoring.

   He says users own every weight, with the agentic framework, simulation engine, annotation engine, hardware design and training data available now and X-Adapt recipes coming in December. These capabilities are NVIDIA's claims and have not been independently evaluated.

7. 07

   MBZUAI's IFM releases PAN 2, an open world model that simulates the outcomes of actions

   MBZUAI's Institute of Foundation Models (IFM) released PAN 2, which it calls an open foundation world model for physical intelligence: given an observation and specified actions, it predicts and renders the "next world" as video so agents can simulate outcomes before acting.

   Read the full report

   According to IFM's project page, PAN 2 has two parts: a multimodal LLM handles "conceptual reasoning" about what happens and when, and a 29B latent diffusion model trained from scratch handles "perceptual reasoning," filling in appearance and motion. Users can control the timing and order of actions and see different outcomes.

   IFM says the model emphasizes physical laws, timing control and spatial consistency as the camera moves. Demo videos were generated by PAN 2; the results are IFM's own claims and have not been independently evaluated.

8. 08

   Geodesic releases NovaAtom-Lite, claiming it beats AlphaFold 3 on antibody-antigen prediction

   Geodesic Intelligence released NovaAtom-Lite, an all-atom structure prediction model for proteins, nucleic acids and the small molecules they bind. The company says it scores 79.07% on FoldBench antibody-antigen tasks, versus 75.58% for IsoDDE and 47.90% for AlphaFold 3. It is live via the Geodesic API and NovaDDE.

   Read the full report

   According to Geodesic, NovaAtom-Lite takes protein, DNA and RNA sequences plus small-molecule ligands and predicts every heavy atom with confidence scores; its training cutoff is September 30, 2021, the same as AlphaFold 3. On FoldBench protein-ligand tasks it reports 77.06% versus 75.99% for IsoDDE.

   Baseline scores come from different reports with differing sample counts, and the company itself says gaps of a point or two should be read as ties; it does not lead on some tasks such as protein-DNA. The model is offered via API only, with no open weights.

9. 09

   WorkflowArena benchmark: best agent orchestration setup fully resolves only about 37% of tasks

   Einsia AI, with Tsinghua University and others, released WorkflowArena, a benchmark of 145 long-horizon tasks across 8 professional domains that tests how well models autonomously orchestrate multi-agent workflows. The best configuration, Claude Opus 5.5 in Claude Code at max effort, fully resolves only 36.95% of tasks on average, and no configuration leads every domain.

   Read the full report

   The paper tests 25 configurations of 9 models on tasks spanning software engineering, formal math, law, finance and 3D production, scored by hidden verifiers with every delegation recorded. Claude models in Claude Code delegate less often but in larger groups, while GPT models in Codex delegate more often in smaller units.

   In a fixed-executor transfer test, workflows written by Opus raised four weaker execution models by about 6.4 points on average. The authors note each configuration was run once and some domains are graded by an LLM judge. The benchmark code is open source on GitHub.