Can AI Improve Itself? [Chris Lu, Robert Lange, Cong Lu]

SPONSOR MESSAGES: *** CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. Check out their super fast DeepSeek R1 hosting! https://centml.ai/pricing/ A conversation with three researchers from the Sakana AI ecosystem — Chris Lu, Robert Lange, and Cong Lu — about using AI to discover better AI training algorithms. The centrepiece is DiscoPOP, a framework where language models discover preference optimization algorithms by writing and evaluating loss function code through an evolutionary process. The guests explain how LLMs serve as unusually powerful mutation operators, drawing on their vast pretraining corpus to make structured, creative explorations rather than random perturbations. The conversation ranges from meta-learning and black-box optimization to the AI Scientist project for automated open-ended scientific discovery, the role of non-convexity in discovered loss functions, whether LLM creativity is bounded by interpolation within the training distribution, the future of human-AI collaboration in research, and the Go-Explore algorithm for archive-based open-ended exploration. --- REFERENCES: company: [00:02:14] Sakana AI https://sakana.ai/blog/ paper: [00:02:15] Discovering Preference Optimization Algorithms with and for LLMs https://arxiv.org/abs/2406.08414 [00:13:00] LLMs as Evolution Strategies https://arxiv.org/abs/2402.18381 [01:04:05] First return then explore (Go-Explore) https://arxiv.org/abs/2004.12919 [01:20:00] The AI Scientist https://arxiv.org/abs/2408.06292 --- LINKS: Full Transcript: https://app.rescript.info/share/d391552d59e346d7e042f36f86fc7155 Download PDF transcript: https://app.rescript.info/api/public/sessions/fe6439d98b2c8694/pdf [01:21:15] The AI Scientist announcement, Chris Lu et al. https://x.com/_chris_lu_/status/1823180922767970668