Reinforcement Learning. Large Language Models like GPT-4 can’t “think”, they just produce text. You can simulate thinking by having a model prompt itself through a technique called “Chain of Thought”, and get it to produce better reasoning chains by giving it (virtual) “rewards” or positive reinforcement when it gets to correct results, similar to how the human brain works as kids. This relatively simple “Reinforcement Learning” produces smarter and smarter models. This is how R1 works, and — again since it’s open source — shows that LLMs + Chain of Thought + Reinforcement Learning = a very viable path to achieving AGI.

🌎 Geopolitics. Various regions are approaching the “AI race” differently. The new administration in the US seems to hold a “throw as much money on AI as possible” strategy. China seems to be funding really smart research teams looking for efficient solutions, and Europe seems to be focused on regulation and safety.

Full link here: https://timkellogg.me/blog/2025/01/25/r1