remilouf

5 ms·

1.
by remilouf 4mo ago
Rémi here, really cool! It kind of turned into a rabbit hole on my end: https://github.com/rlouf/sigil
2.
by remilouf 4mo ago
Of course: https://github.com/rlouf/sigil
3.
by remilouf 4mo ago
It was indeed inspired from my IRC days :)
4.
by remilouf 4mo ago
Original author here, the project has evolved quite a bit since then, you can follow here if that interests you: https://github.com/rlouf/sigil (The $7k was sarcasm)
5.
by remilouf 4mo ago
Author here. Sorry my writing is tedious. Next time I’ll use AI to make it more readable.
6.
Tool calls that execute 100% of the time(blog.dottxt.ai)
5 points by remilouf 5mo ago|0 comments
7.
by remilouf 5mo ago
> Ironically LLMs solve the MxN problem he's complaining about Enlighten me please
8.
by remilouf 5mo ago
Ooops sorry
9.
by remilouf 5mo ago
Author here. You're right, it's not a hard problem, but a particularly annoying one.
10.
The M×N problem of tool calling and open-source models(thetypicalset.com)
159 points by remilouf 5mo ago|50 comments
11.
by remilouf 6mo ago
I haven't always done this, and the knowledge base used to visibly degrade over time. Reviewing a PR does not take a long time, maybe a few minutes, and this compounds over time.
12.
I run my company from Emacs(thetypicalset.com)
20 points by remilouf 6mo ago|1 comments
13.
Every AI Integration Is Held Together with Parsing Logic and Prayer(blog.dottxt.co)
2 points by remilouf 1y ago|0 comments
14.
by remilouf 2y ago
This is actually pretty funny.
15.
by remilouf 2y ago
That’d be a pretty inefficient way to generate bullshit at scale
16.
by remilouf 2y ago
LLM evaluations are very sensitive to the details of the prompt's structure. This post shows how using structured generation reduces the results' variance and the ranking shifts.
17.
by remilouf 2y ago
Looks like it’s quite the opposite: http://blog.dottxt.co/performance-gsm8k.html
18.
by remilouf 3y ago
What do you mean by "semantic dimension"?
19.
by remilouf 3y ago
That whole structured generation line of work looks promising. I hope someone else takes this and runs evaluations on other benchmarks. Curious to see if the results translate!
20.
by remilouf 3y ago
Awesome work! I am really impressed by how much structured generation improves model performance.
21.
Tokenization Is a Problem for LLMs(blog.dottxt.co)
2 points by remilouf 3y ago|1 comments
22.
by remilouf 3y ago
This article presents a way to make structured generation with LLMs much faster than standard generation, but what I find most interesting is how it highlights the issues that tokenization entails towards the end.
23.
by remilouf 3y ago
We already support regex-guided generation in the library, and could easily make an API to serve this as well if that's a feature people want!
24.
Show HN: Vectorize OpenAI API calls (like NumPy)(github.com)
3 points by remilouf 3y ago|0 comments
25.
by remilouf 3y ago
It is currently limited by the time it takes to build the index. There are obvious optimizations we can apply to this, however in a production setting it does not matter much since you only need to build the index once for each (schema, voc
26.
by remilouf 3y ago
You mean nested JSON? It's totally possible.
27.
by remilouf 3y ago
You can definitely let the model improvise by defining `weapon` as `Union[Weapon, str]` if that's what you're asking.
28.
by remilouf 3y ago
It's limited to 300 logit biases at a time. Knowing GPT4's vocabulary is ~100k tokens it's not nearly enough to get reliable guided generation. Although it could work in some cases, and another advantage of this work is that
29.
by remilouf 3y ago
Sorry for misrepresenting your work. Thank you for correcting me and the explanation. Will take a closer look.
30.
by remilouf 3y ago
Regex-Gen is implemented in all generality in the library (minus some constructs that we still have to add). JSON is merely an application. You can read https://blog.normalcomputing.ai/posts/2023-07-27-regex-guide... f

More ›