Skip to content

Latest commit

 

History

History
46 lines (37 loc) · 3.29 KB

README.md

File metadata and controls

46 lines (37 loc) · 3.29 KB
Raw

RecursiveLearningAI

Really quick-and-dirty example of AI recursive learning (video in action:https://x.com/jconorgrogan/status/1853943122046144829) ; The TLDR is that we can make AI "Learn" via smart, recursive workflows that rejigger the context window. Goes as follows:

  1. Digest user query. Send towards two paths: 2a. Answer query 2b. Have one LLM create test cases for that query
  2. We then combine 2a and 2b, passing the llm best answer through those test cases
  3. Have a judge LLM review those responses for logic errors. Ideally the testing step is broken down to the most basic components and the judge just reviews pass/fail based on simple criteria
  4. Another LLM with a new context window collects those learnings (if there are any fails)
  5. This LLM then creates a new prompt which amends the original user prompt with the learnings of the LLMs iterations, and sends that through to the model as part of a new context window
  6. This process loops back to 2a until all logic tests are passed, breaking the loop
  7. The answer is cleaned up by another LLM returned to the user

This image is the result of two recursive iterations. The initial starting prompt was, "Construct a sentence where none of the words in that sentence are in the Bible", which o1preview constantly struggles with. As you can see, by the end the "user query" had changed a lot, with learnings

image

Insert your OpenAI key to the "Proofofconcept.html" run and try it.

Future exploration ideas

Example #1: [Not yet in production]

  1. Input prompt, have llm break down goal to the most basic components
  2. test those components, individually, each with a context wiped new pull
  3. have an llm summarize the results; if all pass then pass full response to user. if fail, then synthesize learnings
  4. another llm takes these learnings and then constructs a new prompt with new context, amending the original user prompt
  5. this happens recursively again and again until the test are all passed
  6. Have an LLM then synthesize the best possible answer given original goal
  7. Return to user

Example #2; introspective/entropix-esque [Not yet in production]

  1. User query
  2. LLM responds
  3. Overseer LLM #1 Check branches of that response to assess alternate answers possibilities
  4. Overseer LLM #2 triggers completion for those "alternative universe" tokens
  5. context compressed and sent to an LLM evaluates all COT output from unique branches
  6. LLM responds given best answer
  7. you can even loop this if you want to, eg if there are branches in best answer from the entropy/varentropy profile
  8. Return that answer to user, via an LLM that is focused on response formatting/user objective

Example #3 Introspective pruning- [See 9.11or9.9Solver html file for proof-of-concept which has an LLM replace any numbers like 9.11 (which fire calandar-related activations) with short math equations (9+0.11)]

  1. User querys
  2. LLM responds
  3. LLM reviews neuron activations, if any fire in domains that are counter to the user goal, seek to clamp those down
  4. Alter context/prompt with tokens that still compress the user goal, but in a way that uses tokens that minimize neuron activation for activity that could negatively impact the output