Most LLMs can’t see this is the Mona Lisa, but it works if you use o3-pro and squint:
@goodside
-
Propose moving –llm verbosity to man page for LLM
By
–
I get why you’d want –help to stay brief, but whatever verbosity/examples you print on –llm, couldn’t that just be included in the man page? I imagine sticking to the formal conventions of man would help the LLM.
-
AGI belief implies human economy irrelevant within years
By
–
enjoying the implication in his post that if you think the human economy matters for even “several more years” you don’t actually believe in AGI
-
o3 hallucinates due to losing trace visibility
By
–
It helps to keep in mind o3 loses visibility of its own reasoning/tool traces after each turn, and will hallucinate whatever it takes to avoid acknowledging this is happening. IMO this explains almost everything in Transluce’s (great) blog post here: https://
transluce.org/investigating-
o3-truthfulness
… -
What changed your mind on inference scaling progress?
By
–
what changed your mind on this? i mostly assume the opposite (i.e. most recent progress is in inference scaling) so curious what convinced you
-
Leaving out ‘solo female’ makes problem too hard for o3
By
–
Yeah — I picked this as the most impressive of the variations I tried that it gets right If you leave out “solo female” the problem seems to be too hard for o3 — I think part of the trick here is there are fewer female artists on music charts than one would expect
-
Challenge: o3 identifies street from Lorde’s ‘Hammer’ lyrics via SHA1 hashes
By
–
o3 names street x mentioned in the lyrics of song y by solo female artist z, given the last four chars of the SHA1 hashes of x, y, and z. Note the answer for y — Lorde’s “Hammer” — was just released yesterday.
-
ChatGPT o3 solves task requiring web search and multi-hop deduction in 2m54s
By
–
* = the answer to the next question A task meant to require web and computational search guided by multi-hop deduction starting from a partial SHA1, answered in 2m 54s by ChatGPT o3 featuring Tom Cruise and Addison Rae
-
LLM limitations, tool usage, and a specific Gemini 2.5 Pro failure
By
–
I think it would be a worse puzzle. A good LLM knows its limitations and uses tools. (Raw LLMs also can’t compute SHA1, which is why I used it.) In another reply in this thread, Gemini 2.5 Pro is seen failing at this task because it thinks the fifth letter of “Madrid” is d.