The Sequence Radar #906: Last Week in AI: Open Models, Intelligent Robots, and the Price of Conviction
NVIDIA's letter, Gemini Robotics, Kimi release and more.
TL;DR
- The AI race is progressing from spectacular demonstrations to addressing complex issues like distribution, embodiment, ownership, and economic returns.
- An industry letter, backed by Jensen Huang and 25 companies, advocates for open-weight AI models as crucial for American leadership, arguing against premature restrictions.
- Moonshot AI released Kimi K3, a large mixture-of-experts model with a one-million-token context window, highlighting the growing significance of open models.
- Google DeepMind's Gemini Robotics 2 expands AI capabilities into controlling the physical world, focusing on planning, coordination, and manipulation.
- The collapse of Leopold Aschenbrenner’s Situational Awareness fund illustrates the financial risks of concentrated AI positions, even with a correct long-term thesis.
- Big Tech earnings show investors are looking for a clear link between AI capital expenditure and customer revenue, with Microsoft and Amazon demonstrating success.
- Meta faced scrutiny as infrastructure spending accelerated faster than market confidence in near-term AI monetization.
- The market is now focused on who can translate intelligence into durable economics while maintaining control over models, machines, and capital.
- New AI research explores agentic AI in scientific computing, verifiable research agents, visual prompt engineering for video models, and efficient safety classifiers.
- Recent AI tech releases include Gemini Robotics 2, LFM2.5 encoder models, MAI-Cyber-1-Flash for code vulnerabilities, and DeepSeek-V4-Flash API.
- Significant AI news includes calls for open AI models, an AI hedge fund's portfolio sale after losses, Anthropic's security incidents, and major funding and partnerships for AI companies like Recursive, Safe Superintelligence, and Moonshot AI.