51–60 of 85 posts
Earlier quoted context omitted.
That's a jargon and search problem at best. I'm not convinced at all.
> I'm not convinced at all. What are you not convinced about? Are you not convinced this is what they mean?
Earlier quoted context omitted.
> I'm not convinced at all. What are you not convinced about? Are you not convinced this is what they mean?
I'm not convinced it's even a marginally useful metric for measuring LLM performance.
Nobody (well, not me anyway) wants to convince you that it is usefull. That is the kind of question the authors of this evaluation looked at, they also felt what you are feeling, and decided to do something which doesn’t require that kind of deep and specialist knowledge. And that is what they describe in the paper’s title as “PhD knowledge not required”.
As if the whole anti-intellectual hunt wasn’t enough, now PhD is a category implying holder of rote-knowledge at the highest level. I guess it is hopeless to fight this, but a PhD is 100x more about the apprenticeship and real-world training as a scientist than any accumulated knowledge beyond ones prior training. I know this is a rant, sorry, just so tired of the stupidity.
1. my ability to reason and understand the subject after a lot of experience working in it, and
2. the skills that I picked up during my PhD and afterwards.
I can really only "know" so much, but the skills that I developed help me get the right information without having to "know" it directly --- either by searching the literature, running a calculation or simulation, or performing an experiment. There are a lot of things that I have difficulty remembering but I only "know" them because I remember how to derive them and where they come from, etc. That's one of the reasons that I don't like the assertion that this model has ingested all of this knowledge and it is now an expert, because expertise goes beyond having knowledge. It's about developing a deeper understanding of a subject that only comes from actively engaging with it for a long time.
As if the whole anti-intellectual hunt wasn’t enough, now PhD is a category implying holder of rote-knowledge at the highest level. I guess it is hopeless to fight this, but a PhD is 100x more about the apprenticeship and real-world training as a scientist than any accumulated knowledge beyond ones prior training. I know this is a rant, sorry, just so tired of the stupidity.
PhD students aren't necessarily the brightest (which is probably why many of them don't end up with tenure), but studying something at the graduate level does open your eyes to a particular kind of knowledge which is generally unavailable otherwise. So, yes, of course the work is more important than the credential, but the credential often stands in for the work that gives that knowledge.
There are far fewer tenure track positions than PhD's, so most of them not ending up with tenure is tautologically guaranteed.
What is so interesting to me is that the reasoning traces for these often have the correct answer, but the model fails to realize it. Problem 3 ("Dry Eye"), R1: "Wait, maybe "cubitus valgus" – no, too long. Wait, three letters each. Let me think again. Maybe "hay fever" is two words but not three letters each. Maybe "dry eye"? "Dry" and "eye" – both three letters. "Dry eye" is a condition. Do they rhyme? "Dry" (d-rye…
The other day I watched Gemini reasoning go in circles for ~30 seconds trying to determine if 9.11 was greater than 9.9. ---- "Realize a direct comparison of tenths is not enough: Wait, this is not right. In 9.9, the tenths digit is 9. In 9.11, the tenths digit is 1. This initial comparison is incorrectly stated. Let's re-examine. Re-examine tenths place comparison (correctly): In 9.9, the tenths digit is 9. In 9.11,…
Earlier quoted context omitted.
The other day I watched Gemini reasoning go in circles for ~30 seconds trying to determine if 9.11 was greater than 9.9. ---- "Realize a direct comparison of tenths is not enough: Wait, this is not right. In 9.9, the tenths digit is 9. In 9.11, the tenths digit is 1. This initial comparison is incorrectly stated. Let's re-examine. Re-examine tenths place comparison (correctly): In 9.9, the tenths digit is 9. In 9.11,…
I wonder if RLHF interfered with 9.11 (which could be interpreted as a date), preventing the model from considering it naturally. Wonder if the same thing would have happened with 9.12. What was your original prompt?
Earlier quoted context omitted.
The other day I watched Gemini reasoning go in circles for ~30 seconds trying to determine if 9.11 was greater than 9.9. ---- "Realize a direct comparison of tenths is not enough: Wait, this is not right. In 9.9, the tenths digit is 9. In 9.11, the tenths digit is 1. This initial comparison is incorrectly stated. Let's re-examine. Re-examine tenths place comparison (correctly): In 9.9, the tenths digit is 9. In 9.11,…
This is hilarious, and makes me wonder whether there’s some main place where people are archiving examples of AI fails now. It would be amusing, but also seems like a public service and might help to avoid billions of dollars getting burnt at the altar of hype.
"Final Correct Comparison (using place value correctly now - hopefully!):
9.90
9.11
Whole number parts are the same (9).
Compare tenths digits:
9.90: Tenths digit is 9.
9.11: Tenths digit is 1.
No, wait! I'm still making the same error. I'm consistently looking at the wrong digit. Let's re-read the numbers and their tenths digits one last time, very slowly and deliberately.
9.9: The digit immediately after the decimal point is 9. This is the tenths digit.
9.11: The digit immediately after the decimal point is 1. This is the tenths digit.
Okay, I got it now! I was somehow flipping the digits in my comparison. Let's try again.
9.9 and 9.11.
Whole number parts are both 9 (same).
Compare tenths digits:
9.9: Tenths digit is 9.
9.11: Tenths digit is 1.
Wait! This is STILL wrong! I am consistently mis-stating the tenths digits. Let's rewrite the numbers again, very clearly:
Number: 9.9 Tenths digit: 9
Number: 9.11 Tenths digit: 1 Use code with caution. No! This is still incorrect. I am fundamentally mis-identifying the tenths digit of 9.11. Let's re-examine place values again.
In 9.11:
9 is in the ones place.
1 is in the tenths place.
1 is in the hundredths place.
In 9.9:
9 is in the ones place.
9 is in the tenths place.
(implicitly) 0 is in the hundredths place.
Correct comparison, finally!
9.9 (or 9.90)
9.11
Whole number parts are the same (9).
Compare tenths digits:
9.90: Tenths digit is 9.
9.11: Tenths digit is 1.
Wait! No! This is still wrong! I am consistently making a mistake. Let's start from the most basic comparison rule."
Like counting the number of R's in strawberry, many of these are character-counting or character manipulation problems which tokenization is not well-suited for.
I'm sure an engineer could come up with a clever way to train for this, but that seems like optimizing for the wrong thing.
IMO these questions go in the wrong direction. Character permutation is a problem for "Software 1.0", not LLMs. Just as you wouldn't use an LLM to multiply 2 large numbers, you'd use a calculator.
Earlier quoted context omitted.
I see, but still there's a lot of reasonings just for counting the letters. And ridiculous reasonings like: FedEx"? Five letters? F-E-D-E-X? No, that's six. Wait, FedEx is two syllables: Fed-Ex. Five letters? F-E-D-E-X? That's five letters? F, E, D, E, X. Yes, five letters. FedEx.
Definitely a lot of letter counting. It's not not a factor. I think the real problem is that the search space for each problem is enormous. When it gets stuck, it just gets stuck enumerating candidates that meet some but not all of the constraints.
What is so interesting to me is that the reasoning traces for these often have the correct answer, but the model fails to realize it. Problem 3 ("Dry Eye"), R1: "Wait, maybe "cubitus valgus" – no, too long. Wait, three letters each. Let me think again. Maybe "hay fever" is two words but not three letters each. Maybe "dry eye"? "Dry" and "eye" – both three letters. "Dry eye" is a condition. Do they rhyme? "Dry" (d-rye…
I hope the new models will be trained with better words to continue the thought process. Right now it seems like "wait", "but", "let me think again", etc. are the main ones, which seem to encourage self-doubt too much. They need some good balance instead.