There doesn’t seem to be one bit of original research in the post, and no explanation of the data or conclusions. For example, what exactly does it mean that the papers “held up”, and how exactly did Claude reach that conclusion? If you don’t know, we can’t trust the data. If you do know, it should be in the post.
As it is, the post is almost devoid of information. Everything was “I asked Claude”. There’s no value here (aside from some saved tokens) above just crafting a prompt and saying “here, ask your favourite model this question”.
Hacker News isn't a great place to discuss papers generally.
Having a productive discussion around a paper requires at least reading and understanding the abstract, and the most successful content on HN (sadly) is content where people can jump in with an opinion purely from reading the headline.
Anyone know of any forums that are good for discussing papers?
This is true across all research subject areas (I'm not especially tuned into LLM research but am to cryptography, which also happens to be a field that gets a lot of play on HN). I think it's just a function of how many people conversant in the field are available to talk about it at any one time.
There are/were isolated communities on Discord around fast.ai, MLC, MLOps that talk papers more in depth but it’s hard to organize a community without commercial or academic incentive.
With big commercial labs clamming up about training details, hardware requirements going up, and overall tired sentiment about AI in general, that’s not really surprising.
ML research shows up on the front page if it shows splashy claims or results. Mundane cockroach papers that advance the field one nudge at a time aren’t that interesting for the average reader.
> I asked Claude (…)
> So I asked him (…)
> (…) so I asked Claude (…)
> (…) so I asked Claude (…)
> Thanks, Claude.
There doesn’t seem to be one bit of original research in the post, and no explanation of the data or conclusions. For example, what exactly does it mean that the papers “held up”, and how exactly did Claude reach that conclusion? If you don’t know, we can’t trust the data. If you do know, it should be in the post.
As it is, the post is almost devoid of information. Everything was “I asked Claude”. There’s no value here (aside from some saved tokens) above just crafting a prompt and saying “here, ask your favourite model this question”.
This must be a troll.
Having a productive discussion around a paper requires at least reading and understanding the abstract, and the most successful content on HN (sadly) is content where people can jump in with an opinion purely from reading the headline.
Anyone know of any forums that are good for discussing papers?
But the gold standard is a small signal or discord community of like-minded, fairly tight knit friends. You may have to organize this yourself
ML research shows up on the front page if it shows splashy claims or results. Mundane cockroach papers that advance the field one nudge at a time aren’t that interesting for the average reader.
Cool to see the sentiment visualized though.