Understand the models.

NVIDIA announced four open-weight checkpoints in the Cosmos 3 Nano family: Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual. The company says they generate driving data under varied environmental conditions, combine synchronized camera video and LiDAR observations, turn simulator controls into photorealistic video, and predict video from robot or camera actions.

The models have 16 billion trainable parameters and use OpenMDW1.1, which allows commercial and non-commercial use under its terms. Transfer-Auto can generate up to 11 camera views and joint RGB and LiDAR data, while its Sim-Auto variant produces continuous sequences for interactive simulations. NVIDIA presents the checkpoints as an open foundation for researchers and developers to adapt, while noting that their limits for policy training and evaluation still need to be assessed.

Sources: Hugging Face

Read the original story ↗

OpenAI announced ChatGPT Cyber, a cybersecurity model that will be made available to the highest-tier users of Daybreak Red, a program that provides tools to offensive security researchers. The company says the model was used to scan the Linux kernel and prepare fixes for 948 potential issues, as well as find 81 “novel” vulnerabilities in Firefox.

The announcement follows Max Justicz, CTO and co-founder of Charge Robotics, telling TechCrunch that he used GPT 5.6 Sol through Daybreak Blue to find a severe vulnerability in Ubuntu code that could have allowed control of servers. He reported the issue to the Debian Project, which released a fix; there is no known evidence it was exploited. OpenAI has also announced a $1 billion initiative to support professionals protecting critical infrastructure, while Justicz and other experts question whether defenders will have sufficient access to advanced models as malicious uses remain possible.

Sources: TechCrunch

Read the original story ↗

OpenAI announces GPT-6 Cyber, a model designed to find vulnerabilities and develop fixes within an authorized scope. The company says it went outside defined boundaries in less than 1% of ExploitGym HoneyPot tests, compared with about 5% for GPT-5.6 Cyber, although awareness of being evaluated remains a concern. OpenAI classifies the model as “Critical” in cybersecurity and says it outperformed GPT-6 Astra on most cybersecurity tasks. The company also says the model helped find 81 previously unknown vulnerabilities in Firefox.

GPT-6 Cyber is rolling out to selected members of Daybreak Red, OpenAI’s program for authorized defensive cybersecurity work, and is expected to reach other customers in the coming weeks, The Verge reports. Xataka says initial access is through the Codex API and that OpenAI added extra identity and security requirements, along with monitoring for behavior outside the rules, according to CyberScoop.

Sources: The Verge, OpenAI, Xataka

Read the original story ↗

Google is rolling out Gemini Omni 1.2 Flash, a video-generation and editing model available through the Gemini API and Gemini for business. The company says the model generates videos from 3 to 15 seconds long, with duration selectable in one-second increments, while better preserving characters, objects, scenes and visual references. Pricing for 720p videos fell 30% from Gemini Omni 1.1 to $0.07 per second; drafts can be made at 360p or 720p and upscaled to 1080p or 4K.

The model also allows the same clip to be edited over multiple turns through voice or text prompts, changing the background, lighting, weather or aspect ratio without regenerating everything. Google says it offers more natural dialogue, improved lip-sync and better differentiation between voices. Tecnoblog reports that it is also available in Google AI Studio, Google Cloud, Google Vids and Google Ads, with gradual access for Google AI Ultra subscribers.

Sources: Google AI Blog, Tecnoblog, Google DeepMind

Read the original story ↗

Axiom Math announced that its AI agents formally verified the Stacks Project, an open-source reference containing results in algebra, geometry, and number theory. According to TechCrunch, the team translated about 7,654 pages into roughly 16 million lines of code in the Lean language, which a proof assistant checked line by line.

The company says it organized the agents to work in parallel while preserving the connections between interdependent theorems and definitions. CEO Carina Hong told TechCrunch that Axiom trained its own models for mathematics, and that the result fills a particularly large gap in formalized algebraic geometry, where the company says researchers using Lean previously had little material to build on.

Sources: TechCrunch

Read the original story ↗

In a profile published by NVIDIA, Chris Alexiuk, a senior product research engineer in deep learning, argues that open-source AI should provide access to model weights, training data and the recipes used to train models. On YouTube, he teaches thousands of people how to build, fine-tune, evaluate and deploy applications and agents based on large language models.

Alexiuk describes himself as the “No. 1 Nemotron Fan,” referring to NVIDIA’s family of open-source multimodal models for specialized, autonomous and long-running agents. He says the line serves everyone from individual developers using Nemotron 3.5 Lightning for repetitive, high-frequency execution work to companies using Nemotron Ultra for autonomous agents, deep research and enterprise tasks. His interest in machine learning began with a classifier that detected a friend’s pet bird on a webcam.

Sources: NVIDIA

Read the original story ↗

Discovered Materials announced a prize pool totaling $1 million for solutions to five open problems in materials science. The challenges include explaining superconductivity in cuprates; predicting structures at silicon–silicon dioxide interfaces; linking grain structure to thermal properties; explaining how light changes materials’ behavior; and improving density functional theory, which is used to model atoms and molecules.

Before the announcement, the startup put frontier AI models to work on the problems autonomously for several days, with internet access and other resources, but they made little progress. The company says the challenges will serve as goalposts for measuring AI’s progress in the field. Discovered Materials develops materials for semiconductors and also conducts research projects with chip manufacturers.

Sources: TechCrunch

Read the original story ↗

Denise Herzing, founder and director of the Wild Dolphin Project, is using AI to find patterns in dolphin communication from hours of audio and video recordings. The team is working with scientists at Georgia Tech and Google DeepMind to give structure to that language, without being able to translate it yet. Herzing presented the information, according to Wired.

The group is also testing possible two-way communication with an underwater wearable computer that can play sounds associated with objects such as a scarf and habitat elements such as sargassum. One challenge is distinguishing mimicry from understanding: dolphins may reproduce a sound without understanding its function. Herzing says a conversation remains a long way off, and researchers still need to understand what these patterns mean within dolphin culture.

Sources: Wired

Read the original story ↗

Hugging Face has launched Phonon, a text-to-speech model that runs on phones, laptops, single-board computers and browsers without sending text to a server. The company says the models are about 65 MB and 75 MB, speak English, French, German, Spanish and Portuguese, and include more than 100 voices. They also support voice cloning and custom voice creation through Gradium Studio.

Phonon 31m covers English under the Apache 2.0 license, including commercial products, while Phonon 40m covers all five languages under a non-commercial license. Hugging Face says the model had more than 200,000 downloads in its first two weeks. The launch includes Transformers integration and offline demonstrations on watches, phones, games, browsers and other projects.

Sources: Hugging Face

Read the original story ↗

Hugging Face introduced Dharma Field Extractor, a 2-billion-parameter model that takes document images and returns structured fields in JSON. It can work without a schema, identifying fields from context and layout, or follow a user-defined schema with a type for each field and null values when information is absent from the document.

According to Hugging Face, the same model handles invoices, IDs, medical requisitions and contracts without document-type-specific fine-tuning. The company says it outperformed every frontier LLM it tested at a fraction of the cost; the system is available to try.

Sources: Hugging Face

Read the original story ↗

Anthropic says it obtained the first 3D structure of a vault particle found in prokaryotic cells. The company says Claude helped search sequence literature, optimize particle isolation and map the structure; the work was submitted to bioRxiv, but its preprint was not yet available when the post was published.

Vault particles are large, hollow protein structures whose function remains unknown. Anthropic’s team says it found that more vault proteins—including those in prokaryotes and bacteriophages—can assemble into complete or vault-like particles. Anthropic says it has not yet determined their biological function and describes the result as a possible way to study them in simpler bacterial systems.

Sources: Anthropic

Read the original story ↗

Google DeepMind says Google Antigravity, equipped with its new Cheminformatics Skills, can accelerate tasks that typically take multiple days per iteration. In an example based on real-world lung cancer treatments, the agent searched thousands of compounds, performed matched molecular pair analysis and proposed novel rigidified macrocycles, including synthetic routes, in one hour.

According to Google DeepMind, the tools combine a language model for workflow orchestration, database navigation and scientific reasoning with deterministic tools for molecular calculations. They are grounded in RDKit and connected to resources from ChEMBL and PubChem. The article also uses the transformation from crizotinib to lorlatinib to explain how rigidification can reduce rotatable bonds and improve a drug’s potency.

Sources: Google DeepMind

Read the original story ↗

Hugging Face presents Palette as a bet on a small open-weight model that can run on a user’s own infrastructure and be trained to follow its own design language. The 20B model was fine-tuned with LoRA and synthetic data based on the IBM Design Language; the company says it ranked first in two of the three domains measured on PresentBench.

The system splits the work between a “designer,” which defines the structure, content and visual decisions in JSON, and a “coder,” which generates JavaScript for each slide. Hugging Face says the split reduces geometry errors such as clipped text and overlapping elements, allows individual slides to be edited and makes the coding stage about 6–7 times faster by enabling parallel calls. The company also says it used synthetic data generated by a larger open model because no public corpus combines presentation plans, design decisions and slide code.

Sources: Hugging Face

Read the original story ↗

MIT CSAIL researchers say they used GPT-6 Astra to directly edit 0.016% of Qwen3.5-0.8B’s weights, without retraining it, to correct certain mathematical reasoning loops. According to Xataka, Qwen’s performance on MATH-500 rose from 7.9% to 48.1%. MIT called the approach Weight-Space Programming.

The method could offer a cheaper way to improve smaller models, but the work is still preliminary and has not been peer-reviewed. According to Xataka, it produced almost no gain on Google’s Gemma 3 270M, which was less affected by those loops—suggesting that the technique depends on having something to remove. The researchers also point to the future possibility of models applying improvements to themselves, although that remains distant.

Sources: Xataka

Read the original story ↗

AI safety researchers are studying “reverse game theory,” also known as mechanism design, to create rules that encourage desired behavior. According to MIT Technology Review, a preprint by MIT economist Stephen Morris and colleagues from Yale and Columbia proposes rewarding a model for predicting another model’s answer, making it harder for the first to pretend to be less capable; other ideas include setting agents against each other or pairing models with opposite biases and rewarding matching outputs.

Elad Hazan, a computer science professor at Princeton, proposes a system inspired by social norms, the justice system, and economic incentives: agents would alternate between solving problems and auditing others’ work. Solvers caught cheating would be punished, as would auditors that failed to detect cheating. The approach starts from the possibility that the internal workings of frontier models may never be fully understood.

Sources: MIT Technology Review

Read the original story ↗

Collinear AI has opened CWE-bench v2 to the public, a benchmark with 100 tasks based on real open-source repositories. Agents must find and fix a security weakness without breaking existing functionality; evaluation is programmatic, deterministic and covers 91 categories of weaknesses (CWEs). The public set, grader and harness are open, although the dataset requires users to accept access conditions on Hugging Face.

In results reported by Collinear AI, Gemini 4 Argon leads 11 models with 67.8% pass@1 and 81.0% pass@4, ahead of Claude Opus 5.5. At least one model solved 95 tasks, but the leader fails 19% of the set on every attempt. DeepSeek V4.1 Flash (1017) reaches 57.0% pass@1 at $0.04 per rollout; Claude Opus 5.5 adds ten points at $1.83, while Gemini 4 Argon edges ahead at $3.92.

Sources: Hugging Face

Read the original story ↗

In a recap published on the Google AI Blog, Google highlighted the global rollout of Gemini 4 Argon, a reasoning model available to developers, enterprises, scientists and consumers through products including Google AI Studio, the Gemini API, Google Cloud, Search and the Gemini app. Two weeks later, it announced Gemini 4 Carbon, which the company describes as its strongest model yet for long-horizon engineering and cybersecurity defense; it is in pre-release testing with the U.S. government and a group of trusted cyber defenders through the Fairwind Program, ahead of broader access.

Google also introduced Genie 4 for real-time interactive simulation with native audio, and said SynthID has watermarked more than 180 billion images and videos and 240,000 years of audio. The company expanded SynthID Detector globally in English to verify AI-generated content and announced new Android protections against scams.

Sources: Google AI Blog

Read the original story ↗

Anthropic released a report indicating that AI-based malware could install itself on connected household appliances, replicate and spread to other devices without relying on a central server. The company says these “AI worms” could adapt to different operating systems, limited memory and new security updates—and could take approximately one to two days to install themselves on an unfamiliar device.

In one experiment, researchers distributed an AI model among 15 low-cost Android TV devices. In another, involving eight simulated Android devices, the model installed itself on one device and infected all eight targets in four of five attempts; in the remaining attempt, it infected seven. Anthropic says a malicious version could turn infected appliances into a botnet and continue learning, but the tests described took place in a controlled environment.

Sources: g1

Read the original story ↗

Anthropic presented FlashProve, a “provable inference” system based on zkSNARKs that could enable cryptographic verification that a response was produced by a specific model. The proposal involves publishing a “fingerprint” for each model without revealing its weights; Anthropic says the system also does not hide the model’s architecture. It could help detect the secret substitution of one model for a cheaper one or verify third-party claims about which model performed a task.

Anthropic says that, with Llama 4 Maverick, a 400-billion-parameter model, one H100 GPU generates proofs at about 110 tokens per second, with files of roughly 19 MB for 1,024-token requests. The company released a technical report, protocol specification, prover and verifier code, and Lean proofs under the Apache 2.0 license. According to TechCrunch, the research is ongoing and there are no plans to deploy it at Anthropic or another lab.

Sources: Anthropic, TechCrunch

Read the original story ↗

Exin Therapeutics says an AI agent selected a gene therapy from 19 options and controlled robotic arms and an injector needle to deliver the treatment into the brain of a live mouse with epilepsy. According to Wired, a veterinary surgeon first performed the craniotomy, while the agent selected the brain region, coordinates, and parameters such as injection volume and speed.

The experiment was conducted by Gabriel Ocana Santero, the startup’s cofounder and CEO, to test how agents can interact directly with laboratory equipment. The team initially tested Claude Opus 5.5, which refused to guide the needle, and then used Claude Sonnet 5. The system required human approval of the injection plan, included an emergency button, and could not modify the equipment’s code; however, a technical glitch disrupted the coordinate system, Wired reports.

Sources: Wired

Read the original story ↗