Alex Rives, founder of the ESM project, has open-sourced ESMFold2, a new tool designed for protein structure prediction and therapeutic design. This advancement is significant as improved inference compute scaling has demonstrated enhanced success rates in binder design, potentially accelerating drug discovery and development.
Sam Altman of OpenAI and Dario Amodei of Anthropic are said to be moderating their previous warnings about AI's potential to cause widespread job losses. Critics, including Gary Marcus, suggest this shift in rhetoric is a strategic move to protect their companies' valuations in anticipation of upcoming IPOs.
OpenAI, in partnership with Thrive and Crete, has launched self-improving AI tax agents powered by Codex. These agents processed over 7,000 tax returns with up to 97% accuracy, reducing preparation times by one-third and boosting tax firm throughput by 50% by continuously refining models with accountant corrections. This initiative demonstrates AI's significant potential for complex financial tasks.
Cognition, the creator of the AI coding agent Devin, has secured $1 billion in new funding, pushing its pre-money valuation to $25 billion. This significant investment, coupled with a reported $492 million annualized revenue run rate and growing adoption by major companies, highlights strong investor confidence in the burgeoning field of AI software engineers.
Axiom Math's AxiomProver AI has achieved a significant milestone by generating machine-verified Lean proofs for eight arXiv math papers, with five of these proofs subsequently accepted by peer-reviewed journals. This marks the first time an AI system has had its formal proofs published in this manner, demonstrating a major advancement in AI's capability for automated theorem proving and formal mathematics.
Robinhood has introduced "Agentic Trading Beta" (also referred to as "Agentic Accounts"), a new feature enabling users to connect AI agents to dedicated brokerage accounts for automated stock trading. Led by co-founder Vlad Tenev, this initiative marks a significant step towards AI agents performing direct financial actions, including the use of agentic virtual cards for purchases. It brings advanced AI capabilities to automated financial trading, drawing comparisons to existing algorithmic trading systems.
YouTube is implementing more prominent AI disclosure labels on both long-form videos and Shorts to increase transparency for viewers regarding AI-generated content. The platform is also expanding its automatic detection capabilities for realistic AI-generated media, with a full rollout for automatically labeling undisclosed content planned by May 2026, while providing creators with dispute options.
Amazon struck a five-year, $6 billion deal with Snowflake to supply AWS Graviton CPU chips, specifically to meet Snowflake's increasing demand for agentic computing. This major agreement underscores the significant infrastructure investment required for AI agent workloads and the strategic importance of custom silicon in the AI era.
Amazon struck a five-year, $6 billion deal with Snowflake to supply AWS Graviton CPU chips, specifically to meet Snowflake's increasing demand for agentic computing. This major agreement underscores the significant infrastructure investment required for AI agent workloads and the strategic importance of custom silicon in the AI era.
Yasmin Razavi of Spark Capital has been added to the Forbes Midas List, recognizing her highly successful $75 million early investment in AI startup Anthropic, which has since yielded a $3 billion return.
A recent study analyzing five frontier large language models found they disagreed on 67% of 1,000 fact-checking claims, with 34% of these disagreements being substantive. The research highlighted significant inconsistencies, particularly for nuanced verdicts and in domains like legal claims, though models like GPT-5.4 showed higher alignment with peer majorities and different models exhibited varied verdict distribution strategies.
Hugging Face's TRL library now includes a delta weight synchronization method that dramatically reduces the data payload for model updates in asynchronous Reinforcement Learning, cutting transfers from gigabytes to tens of megabytes. This innovation makes large-scale RL training more cost-effective and enables distributed setups where trainers and inference engines can operate independently across different locations.
Alex Rives announced ESMFold 2, an open scientific engine that leverages Cryo-EM data to achieve state-of-the-art performance in protein interaction prediction, particularly for antibodies. This release includes an atlas of 6.8 billion proteins and 1.1 billion predicted structures, advancing protein biology research and therapeutic development.
A new research paper introduces PAST2HARM, an adaptive jailbreak framework that effectively bypasses safety measures in leading multimodal AI models like Gemini Nano, GPT Image 2, and SD XL by exploiting past tense reformulations, leading to the generation of harmful content. This discovery highlights significant vulnerabilities in current multimodal AI safeguards and underscores the urgent need for enhanced safety training.
Google introduced Gemini Omni, a new multimodal model capable of creating high-quality videos from diverse inputs like images, audio, video, and text, with its first iteration, Gemini Omni Flash, rolling out to subscribers and YouTube apps. This marks a significant advancement in generative AI for video content.