Company@X

Signal of the Day

OpenAI has temporarily paused reinforcement learning (RL) training on its latest deployment-bound models for two weeks to harden its research environments, run red-teaming, and expand monitoring against cyber-critical risks.

Key Announcements

OpenAI · Source OpenAI announced a voluntary two-week pause on reinforcement learning (RL) training for its deployment-bound models to harden and red-team its research environments. The pause keeps their largest planned frontier RL runs on hold while smaller-scale runs validate new security measures. To support safer development, the lab introduced stronger workload and network isolation, continuous security testing, and expanded multi-stage monitoring to detect concerning agent behaviors quickly.

Etched · Source ASIC developer Etched has raised $700 million in new funding at a $21 billion valuation, led by Jane Street with participation from top-tier venture firms like Sequoia, a16z, Kleiner Perkins, and Blackstone. Crucially, the hardware startup has already shipped its first customer server rack to Jane Street, demonstrating significant execution following their successful A0 tape-out earlier this year. The massive capital injection is positioned to help fuel the production of their specialized Transformer-only chips designed to maximize intelligence-per-watt inference.

Anthropic · Source Anthropic demonstrated that Claude can autonomously design novel protein binders from scratch, successfully targeting 14 out of 15 proteins when guided by an expert prompt. In laboratory testing conducted with Adaptyv Bio and Twist Bioscience, Claude’s binders achieved a success rate of 22% to 35%, drastically outperforming the typical industry success rate of 10% to 15%. While high-affinity binders represent only the initial phase of drug development, Anthropic is actively teaching Claude to run the entire pharmaceutical development process end-to-end.

Google DeepMind · Source Google DeepMind, alongside academic collaborators and its Gemini-powered coding agent AlphaEvolve, announced a new record for matrix multiplication complexity, reducing the theoretical upper limit of omega to ω < 2.371177. Lowering this exponent represents a major theoretical breakthrough because matrix multiplication is the core computational bottleneck of modern AI training and inference. The achievement highlights how LLM-powered coding agents are actively driving fundamental breakthroughs in complexity theory and computer science.

Rillet · Source Modern finance platform Rillet raised a $100 million Series C at a $1 billion valuation led by ICONIQ Capital, representing its third funding round in 14 months and bringing its capital to over $200 million. CFOs are increasingly adopting the platform to replace legacy ERPs like Oracle Fusion, SAP, and Workday, driven by AI agent usage that is expanding at 70% month-over-month. Alongside the funding, Rillet announced a strategic alliance with EY and has officially partnered with more than half of the top 20 CPA firms.

Moving Atoms · Source Spun out of TikTok and Oxford, YC startup Moving Atoms launched its first video world model, Atom 1, designed specifically for training physical robotics. The model is trained on internet-scale video and conditioned on robot actions to accurately predict real-world physical interactions. Atom 1 has already taken the top spot on DeepMind’s Physics-IQ benchmark, notably outperforming NVIDIA’s Cosmos 3 hardware model.

Also Noted

• Exa (Source): Partnered with Firefox to power search for an upcoming AI browser built to protect user choice on the open internet.
• Hugging Face (Source): Crossed 3 million models on the Hub, highlighting an ICML reproduction challenge where human-agent teams verified 2,226 research papers in public logs.
• Tesla (Source): Updated its Tesla Semi fleet page, showcasing lower operating costs, 500 miles of loaded range, and 1.2 MW Megacharging capable of recovering 60% range in 30 minutes.
• SpaceX (Source): Guided the Starship vehicle back to calm waters near Christmas Island after a 24-day ocean recovery effort to perform engineering evaluations.
• Cursor (Source): Shared the technical architecture of “Origin,” their custom Git storage layer designed to scale like a database with improved performance and reliability.

📊 I can compile this into a beautifully formatted PDF report if you’d like to share this daily digest with your wider team or executives.