HackerTrans
トップ新着トレンドコメント過去質問紹介求人

stephantul

598 カルマ登録 2 年前
NLP engineer

投稿

From Chesterton's fence to Chesterton's gap

stephantul.github.io
86 ポイント·投稿者 stephantul·24 日前·56 コメント

Why scikit learn's fit transform is probably not for you

stephantul.github.io
1 ポイント·投稿者 stephantul·2 か月前·0 コメント

Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep

github.com
8 ポイント·投稿者 stephantul·2 か月前·0 コメント

Show HN: Semble – Fast code search for agents with near-transformer accuracy

github.com
7 ポイント·投稿者 stephantul·3 か月前·0 コメント

Show HN: Skeletoken, a Python package for editing model tokenizers

github.com
1 ポイント·投稿者 stephantul·5 か月前·0 コメント

Show HN: PyNIFE. 400-900× speedup for embedding-based retrieval pipelines

github.com
2 ポイント·投稿者 stephantul·8 か月前·0 コメント

Show HN: Skeletoken, a Package for Editing Tokenizers

github.com
1 ポイント·投稿者 stephantul·10 か月前·0 コメント

コメント

stephantul
·一昨日·議論
Why do all this work and then let an ai write the blog post.
stephantul
·9 日前·議論
Ok, thanks!
stephantul
·9 日前·議論
It’s personal ad, basically. The author is trying to get a job as an evaluator somewhere and is hoping that putting 1000$ on the line will get them enough publicity to land them an interview/get a job somewhere.
stephantul
·9 日前·議論
Is it even legal to publish excerpts of books like this? Or does this fall under some kind of exemption/fair use clause?
stephantul
·14 日前·議論
Thanks for introducing me to the article! I’ve experienced this myself but didn’t know it had a name.
stephantul
·19 日前·議論
A non-autoregressive transformer trained with a classification objective.
stephantul
·20 日前·議論
The second part of this comment is not what I expected. I also don’t think it is true. I got bit by a CORS error at work recently that passed by Claude, copilot, and another senior engineer.
stephantul
·23 日前·議論
We’ve been on the receiving end of this complaint with Semble. I think it is a valid complaint, but constructing a benchmark for this kind of thing is just very difficult and expensive because of the (harness) x (model) x (mcp/cli) combination.

With traditional ml/tooling, not showing benchmarks was usually a red flag. But for llm tooling, I’m not so sure.
stephantul
·24 日前·議論
Ha thanks, it was pink a while ago
stephantul
·24 日前·議論
Extreme programming in a nutshell. I like doing this to features: build it, then take it down and rebuild but better.
stephantul
·24 日前·議論
Puppy slush automatically pushed through vents into your codebase
stephantul
·24 日前·議論
Thanks! This is very similar indeed. Related: I see a lot of “drive-by” PRs by agents, who obviously have no intent of ever maintaining the code they wrote.
stephantul
·24 日前·議論
I’m not sure I share your view of PRs. I still see submitting PRs as something that puts pressure on maintainers. Even incorrect PRs take time to verify and review.

I also don’t see how this differs between the “gap” and the “fence” part of the metaphor. Whether someone submits a rewrite/removal (fence) or a new feature (gap) for PR review, it’s still going to cost me attention.
stephantul
·先月·議論
It’s an interesting question: I’d say this is more of a vulnerability creator than the actual vulnerability.

Similar to how using very difficult technologies makes you more likely to create code with vulnerabilities: the technologies are not the vulnerability, but it’s easier to cause them.
stephantul
·先月·議論
This paper oversells on the title. Like, what is chronos, which embedding model was used, which reranker, how was the reranking done, why is chronos much better than claude code
stephantul
·先月·議論
Sure, the whole premise is exactly that proof of work reduces the value of scraping, while having negligible impact on users. If the data is so valuable that bot operators are willing to pay 10s of cpu, then other measures are necessary.

Nevertheless even for these high value cases, you can still argue that it disincentivizes the business model, it becomes less efficient.
stephantul
·先月·議論
Because it destroys the economics of scraping. It’s too expensive with proof of work, or at least not as economically viable
stephantul
·先月·議論
I was also at the event and was pretty disappointed. Most of the talks were pretty low on information. I was at the “build” stage, which supposedly was the technical stage, but the talks there didn’t really go into technical specifics.

The papyrus talk was awesome though.
stephantul
·2 か月前·議論
It was directed at the parent who implied that we didn’t think about this.

I agree with your point about the evals and how you can get discontinuities: good search can be worse than bad search when agents can do many searches. We’re working on it
stephantul
·2 か月前·議論
It's not probabilistic, and exact matches will always be preferred over non-exact. So if you search for a function name this will surface it.