5 ms·

Am I the only one who's completely exhausted with these blog posts and twitter threads that just throw some chat logs into the ether paired with a sloppy, biase

Am I the only one who's completely exhausted with these blog posts and twitter threads that just throw some chat logs into the ether paired with a sloppy, biased analysis that misses important details and obvious refutations? Then as a bonus they'll leave out parts of the chat log too just to make it even harder to make sense of. There's been an endless stream of these for months now and I just can't any more. I'm sick of sifting these people's cognitive diarrhea for useful things. Any sufficiently advanced LLM contains several ad-hoc, unspecified, bug-ridden, slow implementations of more capable if narrowly trained transformer models.

namuol 4y ago
FWIW, the “HellaSwag” benchmark from the GPT-4 paper consists of many such “riddles” and it achieves ~95% accuracy (essentially par with humans). The next best model (LLM or otherwise) was around 85%. It is pretty remarkable. GPT-4 paper: https://arxiv.org/pdf/2303.08774.pdf https://arxiv.org/pdf/2303.08774.pdf Original HellaSwag paper from 2019: https://arxiv.org/pdf/1905.07830.pdf https://arxiv.org/pdf/1905.07830.pdf
- mtlmtlmtlmtl 4y ago
  Sure, papers are great. Happy to read papers. I just have no patience left for this endless stream of garbage from Twitter. It's so pointless.
  - namuol 4y ago
    If you’re an insider it’s frustrating to see a lot of “bad science” out there among laypeople but that’s what assimilation looks like for new technologies. But I couldn’t agree more about Twitter, in general… (I don’t consider myself an “insider” by any stretch, to be clear)
- YeGoblynQueenne 4y ago
  Every time someone on HN invokes an LLM's scoring on a benchmark as if it meant something, I'll post a link to this paper: What Will it Take to Fix Benchmarking in Natural Language Understanding? https://aclanthology.org/2021.naacl-main.385/ https://aclanthology.org/2021.naacl-main.385/ To summarise, it's been quite some time now that language models have been beating NLU benchmarks left and right, but that still doesn't tell us anything about their true capabilities, because those benchmarks don't measure what they purport to be measuring. Or, to rephrase in a way that shows how much I love pattern matching: you keep saying this word, "benchmark". I don't think it means, what you think it means.
  - namuol 4y ago
    Thank you for the perspective; I'm an ML outsider so I'm not familiar with the debates going on within the field. That said, I was didn't mean to suggest that these benchmarks were measuring anything other than ...well, the performance of the benchmark. I was trying to give the parent commenter something less "subjective" to read that I think connects with the somewhat squishier Twitter discussion.
    - YeGoblynQueenne 4y ago
      I think that's a great motivation. But now you can also read the paper I linked above and be more informed about the discussion around benchmarks for NLU (Natural Language Understanding). You can also follow the references in that paper. That will then help you know how to contribute to discussions, in particular when it comes to pointing interlocutors to results, without increasing the amount of noise. We sorely need that right now, so please consider becoming better informed.
  - sebzim4500 4y ago
    That paper makes some very fair criticisms, I think that GPT-4's performance in exams (bar exam, LSAT, AP bio, etc.) is far more impressive than its ability to beat benchmarks. My concern is that fragments of those exams might be found in the training set, and it is hard to see how to account for that. Even if we use exams ostensibily written more recently than the training data was collected, exam boards do tend to reuse questions (sometimes verbatim, sometimes slightly modified).
    - YeGoblynQueenne 4y ago
      This was posted on HN today (not by me!): https://news.ycombinator.com/item?id=35245626 https://news.ycombinator.com/item?id=35245626 It talks about bar exams and contamination, so spot on :)
YeGoblynQueenne 4y ago
>> Am I the only one who's completely exhausted with these blog posts and twitter threads that just throw some chat logs into the ether paired with a sloppy, biased analysis that misses important details and obvious refutations? No no, you're not alone. It just kills me how many people seem to be so ready to abandon critical thinking and go straight for the results that confirm their beliefs. But of course, with large language models, there's so much behaviour that just poking them to see what falls off is never going to work, if one wants to understand what they do. And why do we even need to do any poking? We know how they work. Because "we" made them [1]. Have people completely lost the ability to reason about a complex system, without getting lost in the complexity? Have people lost the ability to reason about a system's behaviour from the principles of its design? When a piece of software does something unexpected, what would most programmers do? Stare at it and try to guess why it did that? Keep generating results and come up with a fancy theory? Wouldn't they just look at the damn code, debug it maybe, and try to find out why it's doing that? Has everyone simply lost the ability to draw a straight line from a system's internal structure, to its external behaviour? What the effing 'ell is wrong with people? It's like someone's gone and replaced all the engineers and the scientists with homeopaths and dowsers. I just checked the twitter bio of the guy who started the Twitter thread. He's an astronomer. Is that how astronomers study the cosmos? They just stare hard at the output of their instruments and try to imagine why it looks that way? Note: that's astro_nomer_, not astrologist. ______________ [1] Well, OpenAI has made that one, but they didn't invent the idea of modelling language by taking statistics over a corpus [2]. And they're not the first to make language models, either, not by a very, very long shot. So that's "we" as in humans, who have done some language modelling before. Hands up! No? Well, I guess you know who you are. [2] Haha no. You know who did? No, that wasn't the one. This is the one: https://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf https://people.math.harvard.edu/~ctm/home/text/others/shanno... Scroll down to page 7. See the section titled "3. The series of approximations to English"? Yeah, Shannon invented statistical language modelling. Or maybe it wasn't even he, and he's just showing how to do it right. That was in 1948, btw. So much for language models being a shiny new thing that will keep improving because it's only now getting started.
- sebzim4500 4y ago

namuol 4y ago

FWIW, the “HellaSwag” benchmark from the GPT-4 paper consists of many such “riddles” and it achieves ~95% accuracy (essentially par with humans). The next best model (LLM or otherwise) was around 85%. It is pretty remarkable. GPT-4 paper: https://arxiv.org/pdf/2303.08774.pdf https://arxiv.org/pdf/2303.08774.pdf Original HellaSwag paper from 2019: https://arxiv.org/pdf/1905.07830.pdf https://arxiv.org/pdf/1905.07830.pdf

mtlmtlmtlmtl 4y ago
Sure, papers are great. Happy to read papers. I just have no patience left for this endless stream of garbage from Twitter. It's so pointless.
- namuol 4y ago
  If you’re an insider it’s frustrating to see a lot of “bad science” out there among laypeople but that’s what assimilation looks like for new technologies. But I couldn’t agree more about Twitter, in general… (I don’t consider myself an “insider” by any stretch, to be clear)
YeGoblynQueenne 4y ago
Every time someone on HN invokes an LLM's scoring on a benchmark as if it meant something, I'll post a link to this paper: What Will it Take to Fix Benchmarking in Natural Language Understanding? https://aclanthology.org/2021.naacl-main.385/ https://aclanthology.org/2021.naacl-main.385/ To summarise, it's been quite some time now that language models have been beating NLU benchmarks left and right, but that still doesn't tell us anything about their true capabilities, because those benchmarks don't measure what they purport to be measuring. Or, to rephrase in a way that shows how much I love pattern matching: you keep saying this word, "benchmark". I don't think it means, what you think it means.
- namuol 4y ago
  Thank you for the perspective; I'm an ML outsider so I'm not familiar with the debates going on within the field. That said, I was didn't mean to suggest that these benchmarks were measuring anything other than ...well, the performance of the benchmark. I was trying to give the parent commenter something less "subjective" to read that I think connects with the somewhat squishier Twitter discussion.
  - YeGoblynQueenne 4y ago
    I think that's a great motivation. But now you can also read the paper I linked above and be more informed about the discussion around benchmarks for NLU (Natural Language Understanding). You can also follow the references in that paper. That will then help you know how to contribute to discussions, in particular when it comes to pointing interlocutors to results, without increasing the amount of noise. We sorely need that right now, so please consider becoming better informed.
- sebzim4500 4y ago
  That paper makes some very fair criticisms, I think that GPT-4's performance in exams (bar exam, LSAT, AP bio, etc.) is far more impressive than its ability to beat benchmarks. My concern is that fragments of those exams might be found in the training set, and it is hard to see how to account for that. Even if we use exams ostensibily written more recently than the training data was collected, exam boards do tend to reuse questions (sometimes verbatim, sometimes slightly modified).
  - YeGoblynQueenne 4y ago
    This was posted on HN today (not by me!): https://news.ycombinator.com/item?id=35245626 https://news.ycombinator.com/item?id=35245626 It talks about bar exams and contamination, so spot on :)

YeGoblynQueenne 4y ago

>> Am I the only one who's completely exhausted with these blog posts and twitter threads that just throw some chat logs into the ether paired with a sloppy, biased analysis that misses important details and obvious refutations? No no, you're not alone. It just kills me how many people seem to be so ready to abandon critical thinking and go straight for the results that confirm their beliefs. But of course, with large language models, there's so much behaviour that just poking them to see what falls off is never going to work, if one wants to understand what they do. And why do we even need to do any poking? We know how they work. Because "we" made them [1]. Have people completely lost the ability to reason about a complex system, without getting lost in the complexity? Have people lost the ability to reason about a system's behaviour from the principles of its design? When a piece of software does something unexpected, what would most programmers do? Stare at it and try to guess why it did that? Keep generating results and come up with a fancy theory? Wouldn't they just look at the damn code, debug it maybe, and try to find out why it's doing that? Has everyone simply lost the ability to draw a straight line from a system's internal structure, to its external behaviour? What the effing 'ell is wrong with people? It's like someone's gone and replaced all the engineers and the scientists with homeopaths and dowsers. I just checked the twitter bio of the guy who started the Twitter thread. He's an astronomer. Is that how astronomers study the cosmos? They just stare hard at the output of their instruments and try to imagine why it looks that way? Note: that's astro_nomer_, not astrologist. ______________ [1] Well, OpenAI has made that one, but they didn't invent the idea of modelling language by taking statistics over a corpus [2]. And they're not the first to make language models, either, not by a very, very long shot. So that's "we" as in humans, who have done some language modelling before. Hands up! No? Well, I guess you know who you are. [2] Haha no. You know who did? No, that wasn't the one. This is the one: https://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf https://people.math.harvard.edu/~ctm/home/text/others/shanno... Scroll down to page 7. See the section titled "3. The series of approximations to English"? Yeah, Shannon invented statistical language modelling. Or maybe it wasn't even he, and he's just showing how to do it right. That was in 1948, btw. So much for language models being a shiny new thing that will keep improving because it's only now getting started.

sebzim4500 4y ago