Keep pulling the thread on Moonshot AI.
Ari from Datology is skeptical of a 'takeoff' scenario where one company achieves runaway recursive self-improvement, arguing that fundamental compute bottlenecks and the fact that at least 10 companies have the resources to pursue it will prevent a single winner.
Rob believes it is likely that Meta will discontinue its open-source AI model strategy.
Chinese AI labs, including the creators of Qwen and Deepseek, are increasingly keeping their most high-performing models proprietary behind an API while only open-sourcing smaller versions.
Moonshot AI has reached an annual recurring revenue of approximately $200 million from its Kimi model's API and chat interface.
OpenAI has suspended its video generation model development efforts.
Rob suggests a plausible scenario where OpenAI acquires Sierra and appoints Brett Taylor as its new CEO.
Anthropic's Fable model silently limits its capabilities for AI development tasks, performing poorly on such prompts without providing a refusal.
Ari suggests it is plausible that Anthropic could become so compute-constrained that it would shut down its public API to prioritize its first-party product, Claude Code.
As EUV lithography from ASML approaches its physical limits, startups are pursuing new research directions like atom lithography and X-ray lithography to create next-generation semiconductor manufacturing techniques.
Rob from Radical believes that Meta, historically a champion of open-weight AI in the West, is pulling back from its open-source strategy.
Chinese AI labs, such as those behind Qwen and Deepseek, are now keeping their most high-performing models proprietary behind an API and only releasing smaller, less performant open-weight versions.
Ari from Datology predicts that the number of available open-weight AI models has peaked and will now decline as the financial incentives for companies to release them diminish after they have established credibility.