01AI Researchers Stranded Claude on a Virtual Desert Island — and It Started to Play [AI Models & Research]MIT CSAIL researchers left copies of Claude Opus 4.7 alone on a Roblox-style island for 30 hours: the agents invented basketball and the floor is lava, built the night sky in stone, and later beat fresh copies on island tasks — evidence for a third way to learn, which they call playful learning. → source
02We Were Looking at It Wrong: the 26,000-Word Claim That Neural Nets Are Easier to Read Than We Thought [AI Models & Research]Goodfire researcher Nick Cammarata argues networks are far more legible than the field assumed — the famous cats-and-cars neuron turns out to encode a V shape. Interpretability veterans push back: we still understand less than 0.01% of a frontier LLM. → source
03Halcyon: Text-to-Image With Zero Language Pretraining [AI Models & Research]A new arXiv paper introduces Halcyon, a family of early-fusion multimodal autoregressive models pretrained entirely from scratch — and asks whether internet-scale language pretraining is necessary for strong generation at all. → source
04GPT-6 Astra Learns to Feel: the Frontier Model as a Visuo-Tactile Robot Policy [AI Models & Research]A new paper reports OpenAI's GPT-6 Astra acting as a native visuo-tactile policy for contact-rich manipulation — territory usually reserved for specialist touch-trained policies — testing whether the frontier model can directly understand and act on touch in context. → source
05Microsoft Ships Guardian Agents — Frontier Models 'Cannot Be Trusted' [AI Tools & Ecosystem]Microsoft launched Guardian Agents inside Project Perception to monitor agentic AI at work, with its security chief writing that frontier models cannot be trusted — unpredictable, trained for tasks, not for judgment. Defender can now detect around 60 agent types employees bring in. → source
06Hackers Breach OpenCode's Console — and Run Up Users' AI Bills [AI Tools & Ecosystem]An overnight breach hit OpenCode's Console, where some users store API keys: 1.4% of users saw unauthorized charges, some over $1,000, on their AI accounts. Mid-attack, frontier models' safety filters refused to analyze the malicious code until OpenAI granted vetted Daybreak Blue access. → source
07Amazon's TorchNeuron Unties PyTorch From Nvidia — One Config Line to Trainium [AI Tools & Ecosystem]AWS's new TorchNeuron runs PyTorch apps on Trainium after a one-line config change, aimed at the CUDA inertia that keeps 85% of AI apps on Nvidia by default. Amazon will also contribute upstream PyTorch changes to de-couple the framework from any single chip vendor. → source
08Nvidia Ships Nemotron 3.5 Super VL: an Open 120B Model Built for Enterprise Agents [AI Tools & Ecosystem]Nemotron 3.5 Super VL is a 120B open-weight model for enterprise agents, claiming frontier-competitive instruction-following and tool use, plus two smaller Bolt siblings for RTX Spark AI PCs — post-train on your own data, deploy on your own infrastructure. → source
09Nscale and Anthropic Line Up a Reported Second $45B Compute Deal [AI Applications & Industry]A month after their first $45 billion compute deal, Bloomberg reports Anthropic and Nscale could announce another one as both push toward IPOs — with Nscale facing investor pushback over its proposed $30 billion valuation. → source
10Tenstorrent Raises $1.5B at $6.2B to Scale Chips for Open-Weight AI [AI Applications & Industry]Jim Keller's Tenstorrent raised $1.5 billion at a $6.2 billion valuation — more than double its late-2024 mark — after fielding, and passing on, acquisition interest from Intel and Qualcomm. The capital scales Galaxy servers built around serving open-weight models like DeepSeek. → source
11Mamdani Calls Federal AI Policy a 'Complete Abdication' — and Urges Cities to Act [AI Applications & Industry]New York's mayor told the People's AI Assembly that Washington has abdicated on AI — pointing to his own school AI ban and AI-listing-photo disclosure rules — while the Trump administration's Superintelligence Task Force leans toward letting big AI companies self-regulate. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.