ChatGPT on the visual cortex, Monty, and the BEAM

Research and Theory
, , ,
1

I just had a (seemingly :-) instructive conversation with ChatGPT. Here are some extracts…

Q: I’d like to know more about capabilities of the visual cortex. My subjective impression is that it integrates input from my eyes, building up a virtual image that my cortex can examine at leisure. Discuss.

A: Your subjective impression is actually very close to how neuroscientists used to talk about vision—and also close to how we still talk about it informally—yet modern neuroscience both supports and sharply revises that picture. Let’s unpack where your intuition is right, where it breaks down, and what has replaced it.

…

Q: Discuss how this view aligns with Thousand Brains–style models and, in particular, how camera images could (should?) be pre-processed for use by Monty’s Learning Modules.

A: Great question—this is almost exactly where the modern neuroscience view of vision and the Thousand Brains Project (TBP) line up most cleanly, and it has direct implications for how Monty should treat camera input.

I’ll structure this in four layers:

  1. Conceptual alignment (visual cortex ↔ Thousand Brains)
  2. What not to do with camera images
  3. What camera pre-processing should produce
  4. Concrete implications for Monty Learning Modules (LMs)

…

Q: Discuss how this maps onto your BEAM / Elixir message fabric.

A: This maps extremely well onto a BEAM / Elixir message fabric—arguably better than onto almost any other mainstream runtime. I’ll walk through this in a way that ties neuroscience → Thousand Brains → Monty → BEAM, and then get concrete about message structure, routing, and scaling.

I’ll organize this into six sections:

  1. Why BEAM is a natural substrate for Thousand Brains
  2. Processes as cortical columns / Learning Modules
  3. Messages as sensory evidence and predictions
  4. Routing, addressing, and “voting”
  5. Time, uncertainty, and failure
  6. What this suggests: you should not centralize

…

I found these closing comments to be pretty interesting:

A distilled architectural mantra

For Monty on BEAM:

Processes are beliefs. Messages are evidence. Consensus is behavior.

Or, more provocatively:

If you can print the world state, you’ve centralized too much.

1 Like