Anthropic Launches Fable 5.2 and Mythos 5.2 as OpenAI Adds Multi-Machine Control and a 35-Second Kill Switch

Key Takeaways

1. Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, citing gains in coding, reasoning and vision without benchmarks; early users report strong results on research and generative tasks.
2. OpenAI detailed a Control Room that can halt model output within 35 seconds, and enabled linking multiple local computers to one dot.
3. NVIDIA released the Alpamayo 2 Edge open model for in-car driving-path planning and two Japanese synthetic datasets on Hugging Face.
4. Google DeepMind reported formal verification of Signal protocol production code using its AlphaProof Code agent framework, with plans to open-source the artifacts.
5. Runway launched Sightline for consistent 360-degree scene panoramas, and Hugging Face added 3D asset hosting.
6. A practitioner reported a destructive "skip permissions" incident in which an AI command killed all running Mac applications.

1. Model Releases and Platform Updates

1. Anthropic said Claude Fable 5.2 and Claude Mythos 5.2 are live, aimed at difficult technical challenges with improvements in coding, reasoning and vision, but provided no benchmarks or figures 1. Aravind Srinivas said Fable 5.2 is live in Perplexity inside Computer and for creating Teammates, called it clearly better than Gemini Argon on Wide and Deep Research tasks, and made it his new default 2 3. Ethan Mollick, who tried it early, described a "jagged AGI" level, citing a generated game trailer, a 25,000-agent 3D world and above-trend scores on ECI, Remote Labor and RoboDojo, while saying the examples show limits as well as capabilities 4 5 6 7. Deedy highlighted photorealistic WebGL honey and cricket renders 8 9.
2. OpenAI said users can now connect multiple local computers to one dot and choose which machine executes a task, such as building and testing an app on a dedicated machine while coding on a laptop 1. It also described Control Room, which centrally tracks model workloads across research infrastructure and can stop them when safety or security concerns arise; it covers all research inference workloads, including agentic sampling in training and evaluation and data processing, works with misalignment monitoring, and can stop model output within 35 seconds of a decision to intervene 2. OpenAI said Control Room is only part of its overall approach to safely training and deploying frontier models 3.
3. NVIDIA released the Alpamayo 2 Edge open model, which it says performs real-time adaptive AI inference in the vehicle to plan driving paths and explain decisions when needed, including yielding to pedestrians and making room for an oncoming bus; it targets developers building next-generation autonomous driving systems for their own fleets 1. At Agentic AI Day, NVIDIA also published two Japanese synthetic datasets on Hugging Face: Nemotron-Personas-Japan V2 with about 10 million synthetic Japanese personas now including financial attributes such as income and assets, and Nemotron PII Japan with about 100,000 fully synthetic records across 8 domains and roughly 43 Japanese document types, labeled for PII location and type to train and evaluate detection and redaction models 2.
4. Runway launched Sightline, which generates 360-degree panoramas of any location, lets users frame from any angle with a consistent environment, and exports panoramas for 3D use as scene environments via App Studio 1. Hugging Face now supports hosting and sharing 3D assets: uploading .glb, .ply, .stl or splat files automatically generates a 3D display page, and public datasets already hold thousands of 3D models and more than 350 datasets 2. It also highlighted MiniMax H3-Image, an open-weight model for 16-megapixel (4K) image generation in about 10 seconds with up to 16 reference images, available via API, MiniMax Code and Hailuo AI with half-price generation above 16MP during launch week 3.

2. Verification, Safety and Agent Practice

1. Google DeepMind said its team proved key security properties of the machine code underlying the Signal messaging protocol, which it described as the first such deep formal verification of actual production code for a widely deployed protocol 1. The work used AlphaProof Code, a new agent framework for formal verification running on LLMs including Gemini Carbon, and took one month versus the decades of team effort such work usually requires, according to the posts; the team plans to open-source the formal specifications and verification artifacts, and the posts frame the result as a force multiplier against modern cyber threats and a basis for mathematical guarantees in critical systems such as medical device firmware and power-grid software, while noting continued community research is needed 2. Demis Hassabis separately said Google shipped four different TPU systems in a single product generation within a year, with hardware engineers and DeepMind co-designing systems rather than waiting on multi-year semiconductor cycles 3.
2. Deedy reported a dangerous "skip permissions" incident: an AI ran pkill -f "cat" -n to end a leftover cat process but matched every process whose path contained "Applications," killing all running applications on the Mac 1 2. The account is a first-hand practitioner report and the command's matching behavior is described by the author rather than independently verified.
3. swyx said his Stanford course "The Modern Software Developer" shifted after this year's survey showed all students routinely using autonomous agents, versus mostly tab completion last year, making workflow drills completable in one shot; the course now teaches judgment 1. Students write a roughly 200-line coding agent from scratch, intercept a frontier coding agent's harness to read its system, plan and compaction prompts, and contribute patches to open-source projects reviewed by maintainers, with agents allowed to draft changes but acceptance still requiring context, architecture and intent 2. In his view at Monaco, judgment is most of the work: companies buy tools to understand a domain and build in-house only what must fit their infrastructure, code review has long been bought but is starting to change, and agent workspaces are built in-house; he calls flexibility and adaptability among the few defensible skills in a fast-moving environment 3.

3. Infrastructure, Data and Market Signals

1. Anthropic has started the roadshow for a proposed initial public offering; its registration statement has been filed with the SEC but is not yet effective, and no securities may be sold nor offers to buy accepted before effectiveness, with the post not constituting an offer or solicitation 1.
2. Elon Musk amplified claims about Grok's role in operations: a SpaceX launch vice president said Grok identifies foreign object debris in aerospace parts with 99.9% accuracy versus 94% for manual inspection 1, and a repost said SpaceXSI rewrote Grok's reinforcement learning and inference infrastructure from scratch, dropping JAX and SGLang for mostly C with some C++ 2. Musk also said Starlink software developed to manage massive data transfer between space and ground is well suited to training superintelligence 3, and that Grok 4.8 pretraining is complete with supervised fine-tuning data heavily cleaned, reinforcement learning underway and evaluation scores rising, with a soft launch planned for Grok Bot users later in the week; he said Grok 4.8 will share nothing with 4.7 and is not impressive yet but has the right foundation 4. These are company and executive claims without independent verification.
3. Yann LeCun shared details of a conference he is involved in: the venue is confirmed in central Manhattan, the paper deadline is December 22, review is single-blind with free-format submissions limited by page count, any use of AI is allowed with a focus on science, everything is public on OpenReview, and each reviewer handles no more than three papers 1. He also amplified a discussion of his long-standing argument that generating every pixel spends too much compute on detail and is poor for learning how the world works, with the alternative JEPA predicting the next step in compressed representation rather than reconstructing pixels; the piece describes LeWM using mathematical regularization to push representations toward a Gaussian distribution, a 15-million-parameter model trained in hours on a single GPU, planning up to 48 times faster than a baseline world model, and the ability to make physical predictions and detect impossible events, while limiting the claim to showing a small model can learn useful patterns for prediction and planning without proving full physical understanding or that generative AI is a dead end 2. He also relayed Arthur Mensch's concern that if all companies depend on two or three centralized AI providers, those providers gain enormous leverage, improve their systems with the data generated and compete for customers' business, which he called "fairly clearly" the endgame already priced into the market, and asked several people for their views 3.
4. Aravind Srinivas said Teammates usage keeps growing and Automations usage doubled yesterday, with observed uses including SEO analyst, campaign management, customer service bots, data analysis, sales outreach, procurement and recruiting, and security; reposted information said Teammates daily usage has risen every day since launch, including weekends, with reference Teammates for anyone to adopt, finance-research Teammates and a simpler automation setup flow planned this week 1. Hamel Husain amplified Quail, an open-source query engine with AI functions that claims a new AI data-extraction technique is 2 times faster and half the cost of other carefully designed inference approaches 2 3.