AI News

1. Ex-OpenAI Researcher Jerry Tworek Launches Core Automation With Anthropic and DeepMind Defectors

Former OpenAI VP of research Jerry Tworek announced Core Automation, a new AI lab whose self-description is "the world's most automated AI lab." The founding team picked off researchers from Anthropic (Rohan Anil) and Google DeepMind (Anmol Gulati, a Gemini research scientist), with The Information reporting the startup is raising $500M-$1B. Tworek's stated thesis is training models that need roughly 100x less data and learn continuously from experience — a direct bet against the scaling-by-pretraining orthodoxy that dominates frontier labs today. Source

2. Jeff Bezos's Project Prometheus Closes In on $10B at a $38B Valuation

Project Prometheus, the AI venture co-founded by Sherjil Ozair and William Guss and backed by Jeff Bezos, is near-closing a $10 billion round at a $38 billion valuation, with JPMorgan and BlackRock named among the investors. The company's stated focus is applying AI to engineering and manufacturing across computers, automobiles, and spacecraft — a vertical-industrial framing that differs from the general-purpose labs chasing AGI. At that valuation, Prometheus would slot in alongside Thinking Machines Lab and SSI as the best-funded new frontier labs less than two years from founding. Source

3. 10x Science Raises $4.8M to Triage AI-Generated Drug Candidates

10x Science, a Stanford-Bertozzi-lab spinout, raised a $4.8M seed led by Initialized Capital (with Y Combinator, Civilization Ventures, and Founder Factor participating) to tackle the bottleneck forming downstream of AI drug discovery: characterizing the flood of candidate molecules AI models produce. The platform combines deterministic chemistry and biology algorithms with AI agents that interpret spectrometry data and produce traceable, regulator-friendly output in minutes rather than months. The pitch lands at a moment when pharma is drowning in AI-proposed candidates but lacks the experimental throughput to evaluate them. Source

4. Ten EU Member States Reject Germany's Push to Deregulate Parts of the AI Act

Austria, Denmark, the Netherlands, Slovakia, Slovenia, Spain, Greece, Portugal, Romania, and Latvia formally opposed a German-led Parliament proposal to carve industrial AI products — machinery, medical devices, toys — out of the AI Act's horizontal framework and into sectoral legislation instead. Germany's coalition, backed by Siemens and Bosch, frames the move as removing "double regulation"; opponents call it deregulation in disguise. EU ambassadors met to broker compromise, with a final trilogue deal scheduled next Tuesday. Source

5. AlphaSignal Publishes a Karpathy-Inspired CLAUDE.md Baseline

AlphaSignal published a 65-line CLAUDE.md template that codifies Andrej Karpathy's January observations about LLM coding failure modes: silent wrong assumptions, overcomplicated code, damage to orthogonal systems, and loss of goal state. The file formalizes rules like "Don't assume. Don't hide confusion. Surface tradeoffs" and installs in under 30 seconds via curl or plugin marketplace. It's a useful reference point for teams still figuring out what kinds of behavioral constraints actually move the needle on Claude Code or Codex CLI output quality. Source

6. India's App Market Grows Fast, With Streaming and AI Capturing Most of the Spend

TechCrunch reports India's non-gaming app market is booming — streaming and AI apps the twin drivers — but global platforms are capturing most of the gains while spending per user stays well below global peers. The data point matters for AI product teams: India is the single largest addressable market by install volume for ChatGPT, Gemini, and Perplexity consumer apps, but monetization still funnels to US headquarters rather than local developers. Source