Prompts for Where Experience Goes
The human-side prompts that developed and steered the draft, in conversation order. Assistant responses, button replies, and interface metadata are omitted. The conversation began with Parts 1 and 2; see Prompts for Who Teaches the Weights? and Prompts for Learning About What You Do Not Do.
Part 3
Don’t commit yet. Draft Part 3 (ladder of stores, learned update rules, benchmarks, corrections to The Dynamics That Matter). Thanks!
Going deeper through the RL essays
Can we go deeper technically if we connect to the two prior essays on RL (four Substack articles)? notes/how-can-a-reward-train-a-neural-network.md and notes/how-does-a-policy-improve.md
Review
Go ahead, review parts 2 and 3, then commit. Thanks!
Links
Yes, add inline links to papers at first mention across all three parts, like How Does a Policy Improve?. I thought footnotes might be good but the definition-highlight-style links are better.
Compositionality of programs
“What Should a Program Inherit?” also touches on compositionality of programs, so maybe we could draw a connection to the continual learning’s desideratum of compositionality of skills?
Retiring The Dynamics That Matter
Should we delete “The Dynamics That Matter” altogether, or just mark them as superseded / discouraged? Do we have any (recent) references to it?
In any case, I think part 3 and the forward mentions should drop referring to “The Dynamics that Matter”? How significant are the older references excluding “Solving the Loop”? What’s your opinion of the “Solving the Loop”’s quality?
Go ahead: fold Part 3’s Revisiting section into the body without self-reference, drop the mentions in Parts 1 and 2, The Given and the Found, Which AGI? and Where Does Seeing Become Thinking?, replace the Solving the Loop link with Schlag et al., and add the superseded note to The Dynamics That Matter.
What Breadth Reaches
What’s your assessment of “What Breadth Reaches”, does it need / would it benefit from a significant update? Do we make recent use of it?
Yes: revise §2–§3 and do a light editorial pass on What Breadth Reaches, fix the Golems link, add a dated revision note, update Part 3 and Which AGI? accordingly, and update Substack in place without email. Thanks!
Memory capacity
How do the following resources impact our articles? [a discussion on X headed by two posts, Alex Wang’s and Nicolas Zucchet’s], the former post presenting this article: https://alexwang.ai/posts/associative-memory/ , and the latter post mentioning in particular https://arxiv.org/html/2610.03529v1
Augmented examples
Continuing review of part 3 of the continual learning series, how could we add more color to the claim “Fine-tuning needs many augmented examples to learn the implications of a fact, and those examples are the inferences that in-context learning draws for free.”?
Maybe making it a box below the paragraph would be better? And also mention there sth like “see also Lampinen et al. cited in part 1”
The boundary
What do you think of the closing section “The boundary”? This sounds like a point related to the “What Breadth Reaches” essay? It seems to be running together two points: one that an autonomous learner will need to construct its teaching conditions rather than be handed them (“active meta-learning” to coin a term); another point is the hope that by developing denser representations specific to a particular project, the agent will be able to make a discovery jump not available to the base/original general-purpose model.
Objectives and open-endedness
Do we ever bring up “Why Greatness Cannot be Planned: The Myth of the Objective” and the associated research program?
Draft the Part 2 passage and also a sentence for Part 3’s boundary section on novelty search / OMNI as constructed conditions; verify sources first.
Editorial pass
Part 3 looks good! Do a final editorial pass, then update the prompts file and add an entry to the index.
Recurrent state and the recall criterion
From the freshness check for the architecture series, 26 October 2026.
First re-check the 3 Oct dossier for model releases and new papers since then, and fold in the associative-memory thread.
Yes, revise What Breadth Reaches §3 and Where Experience Goes with the 2609.04434 result and the recall-criterion caveat, then commit everything.
On-Policy Self-Training
From a check of two October 2026 papers against the recent essays, 26 October 2026.
Can you check how these two papers interact with our recent essays, suggest updates? https://arxiv.org/html/2610.31066v1 “Dual Exam: Controlled Evaluation of Multimodal Agents and World Models”, https://arxiv.org/html/2610.32770v1 “On-Policy Self-Training: Data-Free Self-Improvement in Large Language Models”
Apply the OPST edits to Who Teaches the Weights? and Where Experience Goes (sections 2a and 2b of the note), and regenerate the HTML.
Language Models Need Sleep
From a check of Google’s Sleep paper against the series, 28 October 2026.
Google just posted https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/ but the paper is from a few months ago: https://arxiv.org/html/2606.03979v2 . Are we covering it in the continual learning series?
Yes, add to Part 3.