<p><a target="_blank" href="https://twitter.com/AdamMarblestone">Adam Marblestone</a> is CEO of <a target="_blank" href="https://www.convergentresearch.org/">Convergent Research</a>. He’s had a very interesting past life: he was a research scientist at Google Deepmind on their neuroscience team and has worked on everything from brain-computer interfaces to quantum computing to nanotech and even formal mathematics.</p><p>In this episode, we discuss how the brain learns so much from so little, what the AI field can learn from neuroscience, and the answer to Ilya’s question: how does the genome encode abstract reward functions? Turns out, they’re all the same question.</p><p>Watch on <a target="_blank" href="https://youtu.be/_9V_Hbe-N1A">YouTube</a>; read the <a target="_blank" href="https://www.dwarkesh.com/p/adam-marblestone">transcript</a>.</p><p><strong>Sponsors</strong></p><p>* <a target="_blank" href="https://gemini.google.com">Gemini 3 Pro</a> recently helped me run an experiment to test multi-agent scaling: basically, if you have a fixed budget of compute, what is the optimal way to split it up across agents? Gemini was my colleague throughout the process — honestly, I couldn’t have investigated this question without it. Try Gemini 3 Pro today <a target="_blank" href="https://gemini.google.com">gemini.google.com</a></p><p>* <a target="_blank" href="https://labelbox.com/dwarkesh">Labelbox</a> helps you train agents to do economically-valuable, real-world tasks. Labelbox’s network of subject-matter experts ensures you get hyper-realistic RL environments, and their custom tooling lets you generate the highest-quality training data possible from those environments. Learn more at <a target="_blank" href="https://labelbox.com/dwarkesh">labelbox.com/dwarkesh</a></p><p>To sponsor a future episode, visit <a target="_blank" href="https://www.dwarkesh.com/advertise">dwarkesh.com/advertise</a>.</p><p><strong>Timestamps</strong></p><p>(00:00:00) – The brain’s secret sauce is the reward functions, not the architecture</p><p>(00:22:20) – Amortized inference and what the genome actually stores</p><p>(00:42:42) – Model-based vs model-free RL in the brain</p><p>(00:50:31) – Is biological hardware a limitation or an advantage?</p><p>(01:03:59) – Why a map of the human brain is important</p><p>(01:23:28) – What value will automating math have?</p><p>(01:38:18) – Architecture of the brain</p><p><strong>Further reading</strong></p><p><a target="_blank" href="https://www.lesswrong.com/s/HzcM2dkCq7fwXBej8">Intro to Brain-Like-AGI Safety</a> - Steven Byrnes’s theory of the learning vs steering subsystem; referenced throughout the episode.</p><p><a target="_blank" href="https://www.abriefhistoryofintelligence.com/book"><em>A Brief History of Intelligence</em></a> - Great book by Max Bennett on connections between neuroscience and AI</p><p>Adam’s <a target="_blank" href="https://longitudinal.blog/">blog</a>, and Convergent Research’s <a target="_blank" href="https://www.essentialtechnology.blog/">blog on essential technologies</a>.</p><p><a target="_blank" href="http://yann.lecun.com/exdb/publis/pdf/lecun-06.pdf">A Tutorial on Energy-Based Learning</a> by Yann LeCun</p><p><a target="_blank" href="https://arxiv.org/abs/1907.06374">What Does It Mean to Understand a Neural Network?</a> - Kording & Lillicrap</p><p><a target="_blank" href="https://www.e11.bio/">E11 Bio</a> and their brain connectomics approach</p><p>Sam Gershman on <a target="_blank" href="https://gershmanlab.com/pubs/GershmanUchida19.pdf">what dopamine is doing in the brain</a></p><p><a target="_blank" href="https://www.reddit.com/r/reinforcementlearning/comments/9pwy2f/wbe_and_drl_a_middle_way_of_imitation_learning/">Gwern’s proposal</a> on training models on the brain’s hidden states</p> <br/><br/>Get full access to Dwarkesh Podcast at <a href="https://www.dwarkesh.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">www.dwarkesh.com/subscribe</a>