A stylized brain or neural network rendered in warm tones against a dark background, suggesting both biological and computational aspects of mind.

Tokens All The Way Down

by OCDevel

5 episodes · updated 30 Sep 2026

Education

Tokens All The Way Down is an introductory twelve-episode course on the questions a language model forces into the open: what consciousness is and whether a machine could have it, what attention is in a person and in a machine, whether anything we do is free, and whether life or the universe has a direction. It starts from the hard problem and works through the leading theories of mind, attention, purpose and freedom, then closes with a bonus episode on attention and how to live. Every episode states each position as its best defenders would, with the people and evidence behind it and the strongest objections, and points to books and papers to read next.

Spotify Apple Podcasts YouTube Download EPUB

1. Is Anyone There?

When chatbots like LaMDA and Bing's Sydney say they have feelings, is anyone there to have them? The chapter explains why a fluent report can't answer that, then traces the hard problem of consciousness from Descartes through Nagel, Jackson, Levine and Chalmers to the main competing answers, none of which has won. Along the way each idea meets its AI analogue, from what it might be like to be a model made of tokens to a text-only model as Mary in her room.

2. A Mind Is What It Does

Physicalism gets its hearing, from Place and Smart's identity theory through Putnam's functionalism, Turing's imitation game and Searle's Chinese room. Block's access/phenomenal split, the phenomenal concept strategy, Chalmers's meta-problem and illusionism then test whether an account of what a brain does leaves anything out. Each idea is paired with a machine analogue, from distillation and the China brain in silicon to a model misdescribing its own arithmetic.

3. Lighting Up the Workspace

Neuroscience's main theories of consciousness are global workspace, higher-order, recurrent processing, predictive processing and Penrose-Hameroff, each set out with its evidence and its open objections. Each is paired with its nearest AI analogue, from activation patching and the Transformer's residual stream as a workspace to feedforward networks and next-token prediction. The Cogitate adversarial test is covered, along with how a panpsychist novel's view of attention compares.

4. The Measure of a Mind

Integrated information theory begins from experience itself and measures consciousness as phi, the integration of a physical system. The episode covers its brain evidence, the perturbational complexity index, its verdict that conventional chips are near-zero, and the major critiques. It then sets the novel's bet about Relay beside the theory.

5. Experience All the Way Down

The case that experience is the inner nature of matter, running from Russell and Eddington to Strawson and Goff, set beside AI models. Then the combination problem: whether small minds can make larger ones, tested against split brains, group agents and the novel's own bet on attention.