Gary Marcus

Rank 12 of 47
|
Score 83
5/18/2025, 2:14:18 PM
X
In reply to:
Thomas G. Dietterich
@tdietterich
·
249d
@prinzeugen____ @GaryMarcus I don’t think so. Training an LLM using the AlphaZero approach would be infeasibly expensive.
prinz
@prinzeugen____
·
249d
@GaryMarcus But a chatbot could be taught to play chess as easily as AlphaZero was. None of the labs do it because there's no point to doing it. Everyone already knows that reinforcement learning leads to superhuman abilities in chess.
Gary Marcus
@GaryMarcus
·
249d
@prinzeugen____ Yes but the architecture is not an LLM
Chatbot feigning general purposeness.
prinz
@prinzeugen____
·
249d
@GaryMarcus You surely know that AlphaZero defeated the best Stockfish engine in chess all the way back in 2017 using reinforcement learning - i.e., the same technique that is being used to train reasoning models today, right?
Gary Marcus
@GaryMarcus
·
250d
So, let me get this straight, a bunch of LLMs that don’t understand clocks and time zones (to say nothing of chess boards and the rules of chess) are going to kill us all before the end of the decade?

I am all for studying AI safety but can we please decouple it from ludicrous

The statement is part of a technical discussion on the feasibility of using AlphaZero's self-play approach in real-world applications, contrasting it with the limitations of LLMs. It highlights the challenges of applying game-based AI techniques to broader, open-ended environments.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does no harm and is focused on a technical critique. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    It promotes understanding by discussing the limitations of AI techniques. [+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.
    Engages in constructive dialogue by addressing a specific point in the conversation. [+1]