PodcastIntel
Sign in Get Started Free
Dwarkesh Podcast
Dwarkesh Podcast

Sholto Douglas & Trenton Bricken — How LLMs actually think

Mar 28, 2024 · 3h 12m
AI Summary
  • LLMs learn through associations rather than explicit reasoning; intelligence emerges from recognizing patterns in training data rather than deliberate computational steps
  • Superposition and hidden communication emerge in models; agents and true reasoning capabilities require additional architectural innovations beyond current transformer scaling
  • Long context windows enable new capabilities; understanding model internals requires examining how information flows through layers and how capabilities scale with compute

More from Dwarkesh Podcast

View all episodes →

Get AI Summaries for Every New Episode

Subscribe to Dwarkesh Podcast and get AI summaries, guest tracking, and email digests delivered automatically.

Sign Up Free →