Joel Lehman

I'm a machine learning researcher. The unifying theme behind my research is open-endedness -- the study of processes of creative divergence, like biological evolution, science, art, and markets. Recently I'm interested in the tension between open-endedness and safety -- both in the context of AI and in human society. My main technical expertise is in evolutionary computation, reinforcement learning, LLMs, and artificial life. The Myth of the Objective This book challenges conventional thinking about achievement, arguing that innovative discoveries often come from serendipity rather than direct objective-based approaches. We need a field that pursues what matters (flourishing) rather than what is only instrumental (progress); and that pursuit needs to be rooted in a philosophy that meets the complexity of the modern world (metamodernism) — rather than a naive (modern) or cynical (post-modern) philosophy. Machine learning may have a blindspot for unknown unknowns (Knightian uncertainty);

We need a field that pursues what matters (flourishing) rather than what is only instrumental (progress); and that pursuit needs to be rooted in a philosophy that meets the complexity of the modern world (metamodernism) — rather than a naive (modern) or cynical (post-modern) philosophy. Machine learning may have a blindspot for unknown unknowns (Knightian uncertainty); biological evolution thrives on Knightian uncertainty, and perhaps our algorithms can as well. Base model LLMs are powerful pattern-completion engines. With minimal prompting they can interactively breed text (like poetry, or qu

related reading

• Open-endedness: The last grand challenge you’ve never heard of – O’Reillyoreilly.com
• Lluminatejoelsimon.net
• LessWronglesswrong.com
• LLM Daydreaming · Gwern.netgwern.net
• Guardian Angels: LLM Personalization for Productivity and Security · Gwern.netgwern.net
• Open-endedness: The last grand challenge you’ve never heard of – O’Reillyoreilly.com
• GenAI Handbookgenai-handbook.github.io
• Andrej Karpathy — AGI is still a decade awaydwarkesh.com
• Unfalsifiable stories of doom | Mechanize, Inc.mechanize.work
• Animals vs Ghosts – karpathykarpathy.bearblog.dev
• Taking LLMs Seriously (As Language Models) — LessWronglesswrong.com
• Best explanations of how LLMs work | Roman Vorushinvorushin.github.io