Hacker News Reader

Comment by zozbot234 | Hacker News Reader

The summarized chain of thought for this task (linked in the blogpost) is 125 pages. That's an insane scale of reasoning, quite akin to what Anthropic has been teasing with Mythos.

View on HN

That's here for anyone wondering - https://cdn.openai.com/pdf/1625eff6-5ac1-40d8-b1db-5d5cf925d...

I'm disappointed only that the chain of thought needed to be rewritten. Need to train these LLMs to natively communicate in LaTeX research paper format.

I believe they rewrite the chain of thought to protect their IP, i.e. the chain of thought reveals information about how the model works in a manner that may aid replication

So, I'm not really a mathematician, but the first 3-8 pages reads like nonsense and a bunch of unrelated facts. A bit surreal may be, but if this the norm for this kind of thing, I'm amazed it arrives at any useful result at all.

It didn't seem like nonsense to me. (Recently graduated undergrad with a math degree; probably could have gone to PhD). It seemed like the AI was cycling through a bunch of different possible approaches to tackle the problem. Eventually it finds one and makes more progress there until it reaches the solution

Today I generated the equivalent of two LOTR books just to fix three missing rows in my SQL models (and open a PR), so +1

or put differently, you melted x cubic meters of polar ice

Based on some napkin math, that would be about ~100 watt hours of electricity on an H100 cluster, or, roughly the same amount of energy needed to boil a kettle for a cup of tea.

That's an exceptionally fast output you have there...

Mind showing your working out?

Lotr ~1.4k pages 1 page ~1k tok -> ~700k tok 1tok ~0.5J -> 350kJ ~100Wh ~1 cup of tea

Doing anything melts x cubic meters of polar ice.

I note that (though summarized), this is ~100k tokens. Anyone who routinely works with Codex (or any agentic harness really) can tell you how trivial it is to eat up 100k tokens doing complex work. I've personally had plenty of codex 5.5 xhigh sessions where just the pure chain of thought token count in a single turn exceeds 200k (and I assume doesn't go further only due to compaction meta-guidance; the harness will push the model to stay under 256k per turn/thinking block) .

I think the more interesting question is how many tokens were spent all told; the most interesting graph in the article imo is the success rate by log test-time compute: how many tokens are being spent on the right of the graph to hit a winning CoT/solution like this >50% of the time?

AFAIK, extremely long CoT is fairly routine for those math-finetuned private models. Also, is Mythos unusually verbose compared to Opus?

Reply on news.ycombinator.com