How Do AI Models Actually Think? [Dr. Laura Ruis]

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. https://centml.ai/pricing/ Laura Ruis, a PhD student at University College London supervised by Tim Rocktäschel and Ed Grefenstette, and part-time researcher at Cohere, joins Tim to explore one of the most contested questions in AI: are large language models actually reasoning, or just doing sophisticated pattern matching? The conversation centres on Laura's recent paper on procedural knowledge in pretraining, which uses influence functions to trace how models draw on their training data when producing reasoning traces. The findings are striking — rather than memorizing specific examples, models appear to synthesise knowledge from many different documents in a generalizable way. Even more surprising, code in the training data has an outsized and poorly understood influence on reasoning performance across all domains. From there, the discussion ranges across formal versus approximate reasoning, the role of grounding and embodiment in understanding, Wittgenstein's language games, tensor product representations, and the thorny question of whether agency can emerge from next-token prediction. Laura brings a refreshingly precise perspective to the symbolic-subsymbolic debate, drawing on her earlier work on compositional generalization benchmarks and systematic generalization. The final act ventures into AI safety territory — what it means for a system to be an agent, whether LLMs are better understood as simulators or agent models, and how society might need to adapt to increasingly capable AI systems that were never explicitly designed to be agentic. --- REFERENCES: paper: [00:01:10] Procedural Knowledge in Pretraining Drives Reasoning in LLMs https://arxiv.org/abs/2411.12580 [00:03:50] EK-FAC Influence Functions in Large Language Models https://arxiv.org/abs/2308.03296 [00:19:55] ARC: Abstraction and Reasoning Corpus https://arxiv.org/abs/1911.01547 [00:24:20] Systematic Generalization in Neural Nets https://www.nature.com/articles/s41586-023-06668-3 [00:31:30] Tensor Product Representations https://www.sciencedirect.com/science/article/abs/pii/000437029090007M [00:46:10] Formal Causal Definition of Agency https://arxiv.org/pdf/2208.08345v2 [01:00:40] Language Models as Agent Models https://arxiv.org/abs/2212.01681 [01:05:30] AI Safety and Governance https://arxiv.org/abs/2310.17688 [01:13:35] Pragmatic Understanding in LLMs https://arxiv.org/abs/2210.14986 --- LINKS: Full Transcript: https://app.rescript.info/share/30eacf710e51ce08645c03f266a1469e Download PDF transcript: https://app.rescript.info/api/public/sessions/25c8ef3ea172fe4f/pdf Laura Ruis: https://x.com/LauraRuis https://lauraruis.github.io/