The Jim Rutt Show with Jeff and Viviane

General
1

Jeff and Viviane appear on the Jim Rutt Show and dive into the Thousand Brains Project and Theory.

4 Likes
2

I tried to listen to the podcast, using my MacBook Air (15-inch, M2, 2023) and Chrome (Version 133.0.6943.54 (Official Build) (arm64)). The play icon changed to a pause icon, but the podcast didn’t start up. I was able to view the transcript, but it clearly needs some touch-up editing. Help?

3

Here’s a YouTube link https://youtu.be/KiJ5FGYMPC0

3 Likes
4

I get the impression that a given cortical column might specialize on handling input from a set of nerve or retinal cells that are in close proxiity to each other. In this context, reference frames could be used to handle relationships among the incoming signals within a single column.

This begs the question, however, of how larger-scale relationships are handled. My guess is that columns try first to exchange info with neighboring column in the cortex, taking advantage of the fact that neighboring nerve cells are likely to have neighboring cortical columns.

However, there also needs to be a mechanism for contacting and exchanging info with more distant columns. For example, some events could be broadcast widely, letting columns which notice simultaneous events form connections.

I could imagine ways that (say) Elixir PIDs could be used to implement this sort of thing, but the details are still TBD… One possibility is that, if multiple columns report an event (nearly) simultaneously, shared higher-level columns could “reflect” IDs back down as suggestions for collaboration.

5

Jeff’s comments on ways to organize historical knowledge reminded me of a friend who had two bookcases full of history books, organized chronologically…

6

One possible advantage to the TBT approach is that (the data stored by) topically-related sets of Learning Modules could be extracted from the overall data store and given to other data stores as a shortcut around training. This is rather in contrast to the situation with LLMs and such.

In Monty, I’d like to see a way to keep IDs from imported model(s) from colliding with IDs from existing ones (e.g., a UUID). There would also need to be a way for related models to “link up” with imported ones. One possibility might involve the use of keywords and such…