News for hackers. Stuff that matters.
Re:Claude 3 model family by nopinsight on Monday March 04, 2024 @11:51AM (#39592569)
The APPS benchmark result of Claude 3 Opus at 70.2% indicates it might be quite useful for coding. The dataset measures the ability to convert problem descriptions to Python code. The average length of a problem is nearly 300 words.

Interestingly, no other top models have published results on this benchmark.

Claude 3 Model Card: https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bb...

Table 1: Evaluation results (more datasets than in the blog post) https://twitter.com/karinanguyen_/status/1764666528220557320

APPS dataset: https://huggingface.co/datasets/codeparrot/apps

APPS dataset paper: https://arxiv.org/abs/2105.09938v3

Re:Claude 3 model family by eschluntz on Monday March 04, 2024 @12:21PM (#39593004)
(full disclosure, I work at Anthropic) Opus has definitely been writing a lot of my code at work recently :)
Re:Claude 3 model family by mkakkori on Monday March 04, 2024 @02:15PM (#39594661)
Interested to try this out as well! What is your setup for integrating Opus to you development workflow?
Re:Claude 3 model family by zellyn on Monday March 04, 2024 @01:38PM (#39594156)
Do y'all have an explanation for why Haiku outperforms Sonnet for code?
Re:Claude 3 model family by razodactyl on Tuesday March 05, 2024 @01:16AM (#39600008)
Seems like they optimised this model with coding datasets for use in Copilot-like assistants with the low latency advantage.

Additionally, I wonder if an alternate dataset is provided based on model size as to not run into issues with model forgetting.

Re:Claude 3 model family by RivieraKid on Monday March 04, 2024 @02:12PM (#39594624)
What's your estimate of how much does it increase a typical programmer's productivity?
Re:Claude 3 model family by bwanab on Monday March 04, 2024 @12:29PM (#39593154)
Sounds almost recursive.
Re:Claude 3 model family by vok on Monday March 04, 2024 @12:34PM (#39593224)
APPS has 3 subsets by difficulty level: introductory, interview, and competition. It isn't clear which subset Claude 3 was benchmarked on. Even if it is just "introductory" it is still pretty good, but it would be good to know.
Re:Claude 3 model family by nopinsight on Monday March 04, 2024 @03:02PM (#39595253)
Since they don’t state it, does it mean they tested it on the whole test set? If that’s the case, and we assume for simplicity that Opus solves all Intro problems and none of the Competition problems, it’d have solved 83%+ of the Interview level problems.

(There are 1000/3000/1000 problems in the test set in each level).

It’d be great if someone from Anthropic provides an answer though.

Re:Claude 3 model family by nopinsight on Monday March 04, 2024 @12:26PM (#39593092)
“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

Re:Claude 3 model family by montecarl on Monday March 04, 2024 @01:48PM (#39594299)
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

Re:Claude 3 model family by fragmede on Tuesday March 05, 2024 @06:29AM (#39602044)
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.
Re:Claude 3 model family by CamperBob2 on Tuesday March 05, 2024 @01:24PM (#39607309)
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)

Re:Claude 3 model family by lukev on Monday March 04, 2024 @01:01PM (#39593663)
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.

Re:Claude 3 model family by matchagaucho on Monday March 04, 2024 @01:54PM (#39594378)
They claim the model was grounded with a 25-shot Chain-of-Thought (CoT) prompt.
Re:Claude 3 model family by p1esk on Monday March 04, 2024 @01:34PM (#39594104)
Have you tried the Opus model specifically?
Re:Claude 3 model family by CamperBob2 on Tuesday March 05, 2024 @02:40PM (#39608262)
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.

Re:Claude 3 model family by wbarber on Monday March 04, 2024 @12:50PM (#39593486)
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.
Re:Claude 3 model family by msikora on Monday March 04, 2024 @01:30PM (#39594036)
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).
Re:Claude 3 model family by Jensson on Monday March 04, 2024 @08:44PM (#39598470)
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.

Re:Claude 3 model family by apsec112 on Monday March 04, 2024 @01:25PM (#39593971)
If the answers were Googleable, presumably smart humans with Internet access wouldn't do barely better than chance?
Re:Claude 3 model family by torginus on Monday March 04, 2024 @03:42PM (#39595772)
Not sure, but I tried using GPT4 in advent of code, and it was absolutely no good.
Re:Claude 3 model family by fragmede on Tuesday March 05, 2024 @06:32AM (#39602066)
absolutely? it got a couple of the early ones, didn't it?
Re:Claude 3 model family by a-dub on Tuesday March 05, 2024 @02:09AM (#39600270)
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?"

Re:Claude 3 model family by nopinsight on Monday March 04, 2024 @01:35PM (#39594110)
AMC 10, AMC 12 (2023) results in Table 2 suggest Claude 3 Opus is better than the average high school students who participate in these math competitions. These math problems are not straightforward and cannot be solve by simply memorizing formulas. Most of the students are also quite good at math.

The student averages are 64.4 and 61.5 respectively, while Opus 3 scores are 72 and 63.

Probably fewer than 100,000 students take part in AMC 12 out of possibly 3-4 million grade-12 students. Assume just half of the top US students participate, the average score of AMC would represent the top 2-4% of US high school students.

https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bb...

Re:Claude 3 model family by usaar333 on Monday March 04, 2024 @04:09PM (#39596139)
Interestingly, math olympiad problems (using ones I wrote myself years ago so outside training data) seem to be better in Claude 3.

Almost everything else though I've tested seems better in GPT-4.

Re:Claude 3 model family by sebzim4500 on Monday March 04, 2024 @05:19PM (#39596927)
The benchmark would suggest that but if you actually try asking it questions it is much worse than a bright high school student.
Re:Claude 3 model family by whymauri on Monday March 04, 2024 @08:15PM (#39598260)
Is it possible they are using some sort of specialized prompting for these? I'm not familiar with how prompting optimization might work in LLM benchmarks.
Re:Claude 3 model family by nopinsight on Monday March 04, 2024 @10:33PM (#39599129)
Most likely, it’s less generally smart than the top 2-4% of US high school students.

It’s more like someone who trains really hard on many, many math problems, even though most of them are not the replicas of the test questions, and get to that level of performance.

Since the test questions were unseen, the result still suggests the person has some intelligence though.

Note that there’s some transfer learning in LLMs. Training on math and coding yields better reasoning capabilities as well.

Re:Claude 3 model family by AndyNemmity on Monday March 04, 2024 @10:36PM (#39599140)
I saw the benchmarks, and everyone repeating how amazing it is, so I signed up for pro today.

It was a complete and total disaster for my normal workflows. Compared to ChatGPT4, it is orders of magnitude worse.

I get that people are impressed by the benchmarks, and press released, but actually using it, it feels like a large step backward in time.