Keep pulling the thread on Liam Fedus.
Liam Fedus predicts that AI systems will soon be able to write complete software repositories, identify bugs, and refactor code, but this capability will not automatically generalize to other domains like biology.
Periodic Labs is an AI foundation lab focused on applying AI to atoms, material sciences, and chemistry.
John Schulman of OpenAI successfully argued that the initial product for GPT-4 should be a general-purpose chatbot, which led to the development of ChatGPT.
Periodic Labs uses large language models as an orchestration layer to direct experiments and coordinate other specialized, low-latency neural nets designed for atomic systems.
Periodic Labs is initially focusing its business model on being a software provider that offers an intelligence system and control plane for experiments, rather than discovering and selling its own materials.
Liam Fedus believes there is an opportunity to accelerate development in the physical world by one to two orders of magnitude using AI.
For companies like Periodic Labs that combine physical labs with AI, compute costs are the primary capital expense and can be larger than the cost of physical infrastructure.
The feedback loop for AI improving AI research is slower than for software engineering because experiments require significant time and GPU resources to evaluate.
Periodic Labs currently uses off-the-shelf, commoditized robotics and is not focused on innovating in robotics hardware.
Liam Fedus believes that the availability of reliable, general-purpose robotic systems will be a massive accelerator for spinning up new scientific labs.
Liam Fedus, co-founder of Periodic Labs, was previously the VP of post-training at OpenAI and worked at Google Brain.
At Google, Liam Fedus's work focused on model sparsity to enable more efficient serving of models at scale.