Gary Marcus

Rank 12 of 47
|
Score 83
9/18/2024, 2:13:31 PM
X
In reply to:
Colin Fraser
@colin_fraser
·
491d
@GaryMarcus We don’t know exactly what’s in the training data but there are a trillion 6 digit addition problems, for example. They can’t all be in the training data.
Gary Marcus
@GaryMarcus
·
491d
@colin_fraser We don’t know what is in the training data. And do know there was a lot of undisclosed augmentation in o1.
Colin Fraser
@colin_fraser
·
491d
@GaryMarcus I think this is kind of contrary to that result. These examples don’t appear in the training data.
Gary Marcus
@GaryMarcus
·
491d
@colin_fraser https://arxiv.org/abs/2202.07206
Colin Fraser
@colin_fraser
·
491d
@GaryMarcus This is 4o so no CoT, it’s just blurting out the words. There’s certainly some relevant training data but the probability that these particular addition problems are in the training data is basically zero. Somehow it’s sort of adding.
Gary Marcus
@GaryMarcus
·
491d
@colin_fraser There were multiple previous similar studies and there are obviously some relevant data in the corpus. And for 01 there may be synthetic data as well.

The statement by @colin_fraser engages in a technical discussion about the training data and capabilities of language models, specifically focusing on the use of Reinforcement Learning from Human Feedback (RLHF) and data augmentation. This discussion is part of a broader conversation about the capabilities and limitations of AI models, which is a significant public issue in the field of technology and ethics. The tone is analytical and speculative, aiming to contribute to the understanding of AI training processes.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does not appear to cause harm with words or actions. It is a technical discussion aimed at understanding AI capabilities. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by discussing technical aspects of AI training, which can help others grasp the complexities involved. [+2]
  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. It focuses on the technical aspects of the discussion. [+2]
  4. Principle 7:
    I will uphold the principles of free speech and use my platform responsibly and with integrity.
    The statement upholds the principles of free speech and uses the platform responsibly to discuss a relevant public issue. [+2]