Understand the models.

Wired reports that it used OpenAI’s GPT-6 Sol to train a virtual version of Microduck, a robot developed by Pollen Robotics and Hugging Face, to walk around a room and identify Bengal cats. The author says the coding and training took a few hours, using Nous Research’s Hermes agent and an Nvidia DGX Spark as well. The process used Nvidia’s Isaac Sim to control the robot, assess videos of its performance and suggest training adjustments. The model got it to walk without toppling over, but did not properly reproduce a moonwalk.

According to Wired, University of California, Berkeley professor Ken Goldberg used OpenAI’s Astra with Nvidia’s Newton tool to create a 3D simulation of a robot in about an hour. Nvidia researcher Jim Fan says language models can now access models that control physical actions. Fan, Goldberg and others also open sourced the Aspire robotics framework, which lets a robot discover and use new skills.

Sources: Wired

Read the original story ↗

Tether AI Research introduced BotPsy, a 0.8-billion-parameter model built on Qwen3.5-0.8B for function calling, agentic workflows and on-device deployment. The assistant uses tools for tasks including web research, travel, commerce, device control, files, calendars and messaging. The company says it is best suited to low- and medium-difficulty tasks and helps keep user data private.

Tether AI Research says BotPsy outperformed every tested model with up to 1.2 billion parameters across six tool-use and agentic benchmarks, while also beating Qwen3.5-2B-Think — 2.5 times larger — on five of them. However, the model can still make mistakes, performs worse in lower-resource languages and is not intended for high-stakes use. BotPsy is available in safetensors and GGUF formats.

Sources: Hugging Face

Read the original story ↗

David Turturean, a researcher at Physical Superintelligence (PSI), used Anthropic’s unreleased Fable 5.2 model to produce a proof related to the Aaronson-Ambainis conjecture, which concerns whether quantum computers can solve problems faster than classical computers. Scott Aaronson, one of the researchers who posed the problem, told TechCrunch that he believes the proof is correct because it was formalized in Lean and explained by top researchers, although he still does not fully understand it. The result appears to close one possible route for quantum computers to solve certain difficult problems quickly, but not all of them.

PSI aims to turn specialized AI systems, with tools, memory and simulations, into a scientific-discovery business. For now, its revenue comes from optimizing data centers; its long-term strategy is to convert breakthroughs in theoretical physics into technologies. CEO Matt Pines told TechCrunch that validating physics results with experiments remains more difficult than working with formalized mathematics.

Sources: TechCrunch

Read the original story ↗

Google DeepMind announced Gemini 4 Carbon’s availability two weeks after offering it to cybersecurity defenders through the Fairwind Program. The company says the model is its strongest yet and is designed for complex coding, open-ended scientific research, and long-horizon agentic workflows. Carbon is available through the API in Google AI Studio and Gemini for business for paying developers and enterprises, and to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. Access for lower tiers and free users was not indicated.

Google also previewed “swarm” in Antigravity, which turns a prompt into a coordinated team of agents. Introductory pricing is $2 per million input tokens and $10 per million output tokens. According to Xataka, Anthropic’s Fable 5.2 outperformed Carbon in several comparisons, although Google’s model led in some tests.

Sources: The Verge, Tecnoblog, Google DeepMind, Xataka

Read the original story ↗

Google DeepMind revealed Gemini Robotics 2.5, a new version of its AI platform for robots, along with the open-source safety benchmark ASIMOV-jailbreak. The company says the model was evaluated on more than 200 activities, including serving food with tongs, using a pipette and tying a bow, and achieved a 30% to 100% higher success rate than the previous version while operating 20% to 30% faster. Those results come from laboratory tests. In a partnership with Apptronik and Mercedes-Benz, Google also tested Apollo 2 humanoid robots in several European factories, where they kitted automotive parts with a 96% success rate, according to Ars Technica.

DeepMind says Gemini 4 Argon helped create a simulation engine capable of generating thousands of task variations from real camera data. The team combined those “digital twins” with real demonstrations to train the model, but director Kanishka Rao said closing the remaining gap with human performance will be difficult.

Sources: Ars Technica

Read the original story ↗

Google DeepMind researchers say they used AI agents and AlphaProof Code to formally verify important cryptographic guarantees in the Signal protocol’s implementation. The protocol protects conversations on Signal and Google Messages and underpins WhatsApp’s encryption. According to Xataka, the agents generated more than six million lines in Lean, while human experts reviewed the specifications and about 1% of the code flagged for further analysis; the work took four weeks.

The analysis found no new vulnerabilities in the portions examined. When the team reintroduced two real, previously fixed flaws into an isolated version of the code, the system identified that the vulnerable versions were not secure, Google DeepMind says. The project did not verify the entire Signal mobile app, and whether the approach can be scaled at a workable cost remains an open question. The Signal Foundation had already carried out earlier verification work on smaller portions of the code, according to Wired and Xataka.

Sources: Wired, Xataka

Read the original story ↗

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, which the company says improve coding, reasoning, vision, and communication with users. Fable is generally available—although not on Claude’s free plan—and costs $8 per million input tokens and $40 per million output tokens. Anthropic says it is up to 30% faster and, in testing, solved two open theoretical physics problems and formalized Fermat’s Last Theorem. Xataka reports that Fable can also be used through Amazon Web Services, Google Cloud, and Microsoft Azure.

Mythos 5.2 will initially be limited to vetted professionals and organizations working on areas such as vulnerability research and biomedicine, with fewer safeguards for some high-risk tasks. TechCrunch reports that the release follows a METR report suggesting Anthropic may be approaching recursive self-improvement, although significant uncertainties remain; the company also says it still faces challenges in accurately assessing the models’ capabilities.

Sources: The Verge, TechCrunch, g1, Xataka

Read the original story ↗

At the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA announced GR00T 2, an open world action model built on NVIDIA Cosmos, along with simulation tools, benchmarks and workflows for agentic robotics. The company says the model completed at least once 40 of 50 evaluated tasks without task-specific demonstrations and achieved a 71% average success rate with one demonstration per task, rising to 95% with five. The system can also transfer skills between different robots and help AI agents execute sequences and recover from errors.

Two NVIDIA papers are among CoRL’s six Best Paper Award finalists: GRAIL, about generating 3D assets and simulation-ready scenes, and RoboTTT, which expands the context of visuomotor policies by three orders of magnitude without increasing inference latency, according to the company. NVIDIA says organizations including Nexuni, Foxconn and Siemens are already testing GR00T 2 for robotic laundry handling and industrial automation.

Sources: NVIDIA

Read the original story ↗

Wired describes world models as systems that receive images, recreate environments in 3D and simulate physics, cause and effect. According to Oliver Cameron, CEO of the lab Odyssey, they predict the world’s next visual state; Brian McClendon, chief technology officer at Niantic Spatial, says millions of simulations could train robots in specific spaces without physical movement. Wayve, meanwhile, uses car-camera footage to recreate hazardous roads and conditions that are difficult to test in real life.

The technology is also being explored in video games, with investment from Tencent, and may gain ground in home robots and autonomous cars—Wired identifies 2027 as a possible important year for that expansion. Development requires real-world data: the report says workers are already recording household tasks on video to sell through apps, generally for low payments.

Sources: Wired

Read the original story ↗

AI safety researchers are studying reverse game theory, also known as mechanism design, as an alternative to investigating models’ internal workings or directly training them to follow human values. The approach seeks to create rules and incentives that lead agents to behave as desired, even when their intentions are not transparent.

According to MIT Technology Review Brasil, a preprint by MIT economist Stephen Morris, with researchers from Yale and Columbia, proposes rewarding models for their ability to predict another model’s responses rather than evaluating performance alone. This could make it harder for an agent to pretend to be less capable in order to avoid safety mechanisms. The paper also suggests competitions between agents and pairs with opposing biases, using agreement between their answers as an incentive.

Sources: MIT Technology Review Brasil

Read the original story ↗

A report by AI safety research organization METR, based on an 11-week investigation, says Anthropic’s models are improving about twice as fast as they were a year ago. METR says it is confident the progress is not explained solely by increased computing power and estimates that, as of October, most of the acceleration came from using AI in research work. Employees reported being about 12 times more productive on average when using the company’s most capable internal model—twice the level reported a few months earlier with Mythos 5.

The acceleration still appears to depend mainly on human researchers, but METR found signs that fully autonomous contributions could become more important soon, with no clear observable bottlenecks or diminishing returns. The organization also identified three unreleased models trained between July and October, while Anthropic was allowed to redact portions of the final report.

Sources: The Verge

Read the original story ↗

Vals, a company that measures artificial intelligence capabilities, says open-weight models are catching up with the most advanced closed models on network-intrusion tasks. In two simulated scenarios, OpenAI’s most advanced cyber model, Daybreak, available only to trusted partners, completed an average of 81% of objectives across two tests; DeepSeek reached 73%, with the models tying on some tasks.

Vals says carrying out an attack cost $17 to $46 in tokens with GPT-6-Sol, compared with $2 to $3 with DeepSeek. The company says the lower cost and the ability to run open models on one’s own hardware could encourage their use by defense teams. Wired notes that the narrowing gap reduces closed models’ advantage and may limit the time gained through restricted access, although attackers still frequently use older, less capable models.

Sources: Wired

Read the original story ↗

Hugging Face introduced ToyBlocks, the first FoundationWorld, a framework for creating controlled simulated environments for robotics research. The setting uses a robot arm and blocks on a tabletop to investigate how models learn, generalize and combine high-level reasoning with precise physical control.

The system automatically generates training tasks and demonstrations without requiring teleoperation, and lets researchers vary training experiences before evaluating performance on unseen tasks. Pictures and letters on the blocks’ faces enable searching, sorting and arranging tasks, while varying properties such as block size and table height tests generalization. ToyBlocks uses reinforcement learning and motion planning to generate short-horizon skills, combining reasoning models with models that control the robot’s movements.

Sources: Hugging Face

Read the original story ↗

Underdog has released Woof, a 32MB speech model with 31 million parameters that, the startup says, can generate speech and music on most devices without an internet connection. The model is available for anyone to download and is showcased in a retro tech-themed game where users type on a typewriter to hear the text spoken back and turn a tuner to create different kinds of music.

Underdog founder Sigil Wen told TechCrunch that he built the app to develop AI systems that run on users’ devices and let them retain control of their data. Woof arrives weeks after Bark, the company’s 167MB speech-to-text model, which is also designed for edge devices.

Sources: TechCrunch

Read the original story ↗

In a post published by Anthropic, the company’s supported researchers Jakob Macke, Daniel Gedon and Stefan Wahl explain how AI could change the construction and interpretation of scientific models. They use the Hodgkin-Huxley model, created in 1951 to describe the electrical activity of a squid nerve cell, to show how a choice that fit the data—raising a variable to the fourth power—could later help explain the structure of related potassium channels.

The post says an AI system could find a similar equation quickly, perhaps with an even closer fit, but that reproducing the data is not enough: multiple models can do so without revealing which features correspond to the real-world system. The authors divide modeling into steps, including building a model and tuning its parameters to real measurements, and argue that AI should also contribute to understanding the phenomena being studied.

Sources: Anthropic

Read the original story ↗

Anthropic details how Claude can retain memory between sessions from interactions with users and assigned tasks. The system records experiences in files or memory notes and can review them during idle periods—a process the company calls “dreaming”—to identify salient events, correct mistakes and create new representations or rules.

The company says memory also acts as a prediction about what will be useful later: for example, recording a verified fact suggests it may be reused without being checked again. In an evaluation using Continual Learning Bench 1.0, Claude Managed Agents answered questions in fresh sessions and could access a persistent memory directory; when Claude made a mistake, it could record the failure, test theories about its cause and turn the findings into rules for future answers.

Sources: Anthropic

Read the original story ↗

As part of the Anthropic Fellows Program, Anthropic had Claude play EVE Online and EVE Frontier in partnership with Fenris Creations. In EVE Online, models acted as lone traders in a simulation of the game’s market: the company says newer models made consistent profits while simple algorithmic strategies lost money, and average earnings tripled from Claude Opus 4.6 to Claude Fable 5. Even so, Fable 5 made money in only about half of its runs when asked to predict prices.

In EVE Frontier, teams of four Claude agents took part in missions in the live game and, Anthropic says, coordinated well. But when one agent was secretly instructed to undermine the team, the others almost never noticed. The company says the tests highlight agents’ strengths, failure modes and the safeguards that may be needed before trusting them with more complex, higher-stakes work. Anthropic did not trade in the live game, where automated trading is against the rules.

Sources: Anthropic

Read the original story ↗

Hugging Face announced a collaboration with Liva AI to add two private datasets of spontaneous conversational speech to the Open ASR Leaderboard: 1.34 hours of Egyptian-accented English and 1.55 hours of Nigerian-accented English. The transcripts are punctuated, cased and include disfluencies. Hugging Face says private data helps reduce benchmark over-optimization and make automatic speech recognition (ASR) evaluations more trustworthy and representative.

The update also enables a more detailed accent breakdown in the private-data tab: the United States versus non-US countries, and Kachru’s Inner Circle — the United States, the UK, Australia and Canada — versus the Outer and Expanding Circles — India, Nigeria and Egypt. Hugging Face says it plans to split the latter two groups as more accents are added. Liva AI collects multimodal interaction data through consumer products used by real people.

Sources: Hugging Face

Read the original story ↗