4 comments

  • ssivark57 minutes ago
    Ha, interesting. I wasn&#x27;t aware of Sutton&#x27;s blog post, but if I might make a shameless plug, we demonstrated [1] exactly this problem (see section 4.4.3), and how multi-step world models (using diffusion models as the substrate) could be one potential answer.<p>Since then, I have come to like temporally-abstract models more and more. Rolling out in time -- either step-by-step or many steps at once -- suffers from the <i>tyranny of the specific</i>. For long horizon planning with agents, I care (often only approximately) about where I can end up, and seldom about <i>exactly when</i> I end up there. Successor features, GVFs, Forward-Backward representations, and the like seem like they have an elegant approach for structuring thinking at a &quot;high level&quot;, instead of generating exponentially large search trees by rolling out microscopic world models.<p>[1] <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2410.05364" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2410.05364</a> (funnily, from around the same time &#x2F; few months after Sutton&#x27;s blog post)
  • mxwsn44 minutes ago
    This is the same reasoning behind why Yann Lecun thought test-time scaling would not work for LLMs: compounding error.<p>Instead, the more tokens LLMs use, the better their performance on many tasks. LLMs can self-correct, evidenced by the power of getting models to question themselves by emitting &quot;Wait,&quot; in S1. <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2501.19393" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2501.19393</a>
  • gnabgib38 minutes ago
    (2024)
  • nttylock45 minutes ago
    [flagged]