newsroom
AI Regulation & Policy · US government now in release loop for frontier models via national security holds
now playing · AI Regulation & Policy
US government now in release loop for frontier models via national security holds
The White House is gating access to GPT-5.6 and Mythos customer-by-customer, creating a de facto synchronization mechanism for the OpenAI/Anthropic duopoly while slowing domestic deployment…
Chinese open models reaching parity with US frontier by Christmas via harnesses not base models
GLM 5.2 and other Chinese open models are achieving frontier-level coding performance through sophisticated harnesses/scaffolding rather than base model superiority; US export controls may…
AI Agentstailwindscore 9/10emad mostaque
Harnesses and multi-model scaffolding now matter more than base model weights for capability
Frontier SWE benchmarks show open models with proper harnesses (Fugu, Blitz, Intelligent Internet) matching or beating closed frontier models; the 'harness' layer of software orchestration,…
Recursive self-improvement is the real red line; government gating models that can build themselves
The core government concern is not cybersecurity but models that can answer 'help me build yourself' — enabling recursive self-improvement escape velocity; China may already have reached th…
Photonic computing to replace Nvidia GPUs within 1-2 years at 1/100th mass for orbital data centers
Highly quantized neural nets enable photonic computation at massive efficiency gains; SpaceX acquired a photonic computing company and will likely deploy photonic compute for Starmind/Starl…
Orexin agonists as next GLP-1 class: Eli Lilly bets $6.3B on 4-hour sleep lifestyle drug
Eli Lilly is replicating its GLP-1 playbook by targeting orexin (wakefulness master switch) via Synthesa acquisition; orexin agonists could enable healthy short sleep, creating a massive li…
GLM 5.2 outperforms GPT-5.5 on hardest coding benchmark with $25M compute budget
Chinese open model GLM 5.2 achieves top scores on Frontier SWE (novel kernel building) using harnesses, demonstrating that coding agent capability is decoupling from massive compute spend a…
ISO-certified trustworthy AI stamps imminent for code merging; defensive models poison offensive use
OpenAI's Daybreak model biases toward defensive cybersecurity; the industry will require certified AI stamps for code merging to prevent supply chain poisoning, creating a new trust/certifi…
Human hippocampus operates like transformer embedding space; telepathy bandwidth could increase 100,000x
Neuroscience shows bilingual human hippocampus organizes concepts like vector embeddings; this structural similarity suggests Neuralink-style telepathy could achieve massive bandwidth gains…
Real-time 4K video generation for interactive gaming is a $30T market opportunity
Video game industry ($200B+) dwarfs Hollywood ($50B); whoever solves latency for real-time generative video in games unlocks a market expanding to all demographics, potentially reaching hol…
Quantum photonics and sensing are near-term winners; quantum computing needs AI-discovered algorithms
Quantum sensing and photonic quantum computing will work soon; the real breakthrough requires AI to discover economically useful quantum advantage algorithms for training/inference, which w…
Orbital and lunar data centers emerging as secure, scalable compute infrastructure with in-situ resources
Moon offers defensible positions, raw materials (regolith for silicon/oxygen), and mass drivers to launch compute to LEO; orbital data centers avoid Kessler syndrome in lower orbits and mun…
EU entrepreneurs should focus on agentic workflows into legacy industries, not frontier models
Europe's regulatory environment and slower diffusion create massive opportunity for transforming legacy companies with agentic AI workflows; US is better for vertical AI startups, EU for di…