Hacker News Reader

Comment by nopinsight | Hacker News Reader

“Claude 3 gets ~60% accuracy on GPQA. It's hard for me to understate how hard these questions are—literal PhDs (in different domains from the questions) [spending over 30 minutes] with access to the internet get 34%.

PhDs in the same domain (also with internet access!) get 65% - 75% accuracy.” — David Rein, first author of the GPQA Benchmark. I added text in […] based on the benchmark paper’s abstract.

https://twitter.com/idavidrein/status/1764675668175094169

GPQA: A Graduate-Level Google-Proof Q&A Benchmark https://arxiv.org/abs/2311.12022

View on HN

I really wanted to read the questions, but they make it hard because they don't want the plaintext to be visible on the internet. Below is a link toa python script I wrote, that downloads the password protected zip and creates a decently formatted html document with all the questions and answers. Should only require python3. Pipe the output to a file of your choice.

https://pastebin.com/REV5ezhv

thank you for not just posting the questions and answers. now we just have to hope that a nascent agi model can't run that script and feed it back to itself for training purposes.

Thanks for that. I did have to append .encode("utf-8") to the strings in the print statements before it would let me pipe the output to an .htm file under Windows, but other than that it worked great.

(Edit: better to use the original script but set PYTHONUTF8=1 before running)

This doesn't pass the sniff test for me. Not sure if these models are memorizing the answers or something else, but it's simply not the case that they're as capable as a domain expert (yet.)

I do not have a PhD, but in areas I do have expertise, you really don't have to push these models that hard to before they start to break down and emit incomplete or wrong analysis.

Have you tried the Opus model specifically?

They claim the model was grounded with a 25-shot Chain-of-Thought (CoT) prompt.

The idea is that they aren't as capable as a domain expert, but they are more capable than an expert from a different domain.

E.g., if you ask a chemistry PhD something about genetics or astrophysics, you're more likely to get a correct answer from the model. Which is pretty interesting IMO.

What's to say this isn't just a demonstration of memorization capabilities? For example, rephrasing the logic of the question or even just simple randomizing the order of the multiple choice answers to these questions often dramatically impacts performance. For example, every model in the Claude 3 family repeats the memorized solution to the lion, goat, wolf riddle regardless of how I modify the riddle.

If the answers were Googleable, presumably smart humans with Internet access wouldn't do barely better than chance?

GPT-4 used to have the same issue with this puzzle early on but they've fixed since then (the fix was like mid 2023).

The fix is to train it on this puzzle and variants of it, meaning it memorized this pattern. It still fail similar puzzles if given in a different structure, until they feed it that structure as well.

LLMs is more like programming than human intelligence, they need to program in the solution to these riddles very much like we did expert systems in the past. The main new thing we get here is natural language compatibility, but other than that the programming seems to be the same or weaker than old programming of expert systems. The other big thing is that there is already a ton of solutions on the web coded in natural language, such as all the tutorials etc, so you get all of those programs for free.

But other than that these LLMs seems to have exactly the same problems and limitations and strengths as expert systems. They don't generalize in a flexible enough manner to solve problems like a human.

Not sure, but I tried using GPT4 in advent of code, and it was absolutely no good.

absolutely? it got a couple of the early ones, didn't it?

it's an interesting benchmark, i had to look at the source questions myself.

i feel like there's some theory missing here. something along the lines of "when do you cross the line from translating or painting with related sequences and filling in the gaps to abstract reasoning, or is the idea of such a line silly?"

Reply on news.ycombinator.com