Keep pulling the thread on Kim Peek.
Current LLMs have not demonstrated the ability to make novel, cross-domain scientific discoveries in the way humans have.
The pre-training objective of LLMs provides general world knowledge but does not inherently grant the skill of making novel connections or conducting research.
Significant reinforcement learning (RL) on tasks similar to scientific discovery is necessary for models to begin making novel discoveries.
Current AI models are not proficient at determining which memories to store, and their memory scaffolding mechanisms are primitive.
The Anthropic Fellows program has a massive return on investment, with multiple participants having been hired by Anthropic and other AI labs.
The MADS program for AI research has been highly successful in providing the time, funding, and social justification for individuals to pursue AI safety-relevant research.
The individual leverage of software engineers is expected to increase by a significant factor each year due to AI assistance.
Some software engineers are already reporting being 2 to 5 times faster at their jobs when using AI tools, particularly in new programming languages.
Deep technical knowledge will remain critical in four years, as individuals will be tasked with managing large teams of AIs, which will max out their personal management bandwidth.
Current AI models lack the long-term coherence necessary to execute complex, multi-step business objectives like building a successful company.
Slow, compounding growth in media is largely a myth; exceptionally good content, such as Leopold's "Situational Awareness," can achieve widespread reach among its target audience almost instantly.
Grand schemes like nationalizing AI labs or creating public-private partnerships are risky, as the current landscape of AI lab leaders is a relatively good outcome compared to many counterfactuals.