Gary Marcus

Rank 13 of 47
|
Score 83
9/14/2024, 4:53:50 PM
X
In reply to:
Cameron Williams
@Wasgo
·
486d
@GaryMarcus So the most likely thing is the model has a new internal ending condition, similar but different to the prior models using “answering” as the end condition.
Cameron Williams
@Wasgo
·
486d
@GaryMarcus There’s some evidence, but also reasoned speculation.

We know CoT is a new model because they confirmed it in a tweet.

And based on everything we know, there’s nothing smart enough to know when to stop the thinking and the amount of thinking varies wildly based on the answer.
Gary Marcus
@GaryMarcus
·
486d
@Wasgo My understanding: from the output protocols there is a cycle and a stop condition, and those not as far as i can tell governed by LLM; rather those symbolic operations govern how the LLM is invoked and when it stops.

is there evidence to the contrary?
Cameron Williams
@Wasgo
·
486d
@GaryMarcus I have to find the tweet but I don’t think it is.

The way all the existing LLMs models worked was they were primed to generative an answer then stop.

This new CoT model is primed to generate output according to a new structure they’re not disclosing.
Gary Marcus
@GaryMarcus
·
486d
@Wasgo its not a pure LLM; there is as far I can tell a symbolic control cycle. no?
Cameron Williams
@Wasgo
·
486d
@GaryMarcus I think it still is a pure LLM though.

Based on their answers yesterday, the changes made to do CoT output happened through changes to the model.

It sounds like once the hidden generative output happens, it then uses a different generative model to summarize it and answer.

The statement is part of a technical discussion about the behavior and internal mechanisms of a new model, likely related to machine learning or artificial intelligence. It engages substantively with the topic, contributing to the understanding of how the model operates and its stopping conditions. Therefore, it constitutes public discourse.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not cause harm and is focused on technical aspects, thus adhering to the principle of doing no harm with words and actions. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement aims to promote understanding by discussing the technical details of the model, which aligns with promoting understanding, empathy, and compassion. [+1]
  3. Principle 4:
    I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
    The statement engages in constructive dialogue without personal attacks or ad hominem arguments, adhering to the principle of constructive criticism and dialogue. [+1]