HNHacker News
TopNewBestAskShowJobs

LelouBil

2,229 karma · joined March 23, 2022

submissionscomments
LelouBil··on Timestral: Mistral's time-series foundation models
What IRL company is the closest to lumon right now ?
LelouBil··on Show HN: Qwen3-4B with the AI writing tics trained out
Honestly I don't see the em dash as for sure AI. It's just that it's often used in LLM tropes like the "it's not XXX — it's more" and other tagline-esque sentences.
LelouBil··on Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
I feel like everybody forgot that prompt injection is not solved
LelouBil··on There's never been a better time to go into math (by: 3blue1brown) [video]
I think for many fields starting from scratch is useful to gain understanding (well it may be hard in biology's case but you get my point)
LelouBil··on Show HN: Sphinx – A hook that quizzes you on your AI's code before shipping
I feel the same a bit. I'm learning how to use agent harnesses and how to get them to work on problems I have with projects.

But it very often feels like brute forcing the solution, and I'm not a big fan of it.

But maybe I'm using it wrong.

LelouBil··on Jev meets the Trolley Problem
Yeah, that's one of the points of the EU AI Act, I don't have the specifics, but for some critical sectors decision making cannot be automated by AI and you are obligated to get a human explanation if you request it.
LelouBil··on I know why the mathematicians weep
This was a very interesting read, from 2003.

https://karmak.org/archive/2003/12/manna/manna1.htm

LelouBil··on Do Frameworks Matter Anymore?
Even though I'm not a web developer I resonate with this a lot, but in the more generic way of "Does understanding the code even matter anymore ?". And it makes me kind of sad.

I just graduated (and was very lucky to get a junior job), and I'm someone that likes to dig deep when I'm having issues with a software. I care about correctness, I prefer to use strongly typed languages, I never used Haskell professionally, but I do all the advent of code with it. I like reading precise documentation.

But yesterday, my father showed me an app he made using Mistral's Vibe and android studio. He's in IT support, he knows computer science even though he never learned Kotlin or anything, but he was able to "make" an app, and he was excited, and I was happy for him as well, but only happy because he was happy basically.

I used this opportunity to ask him lots of questions about how he approached it, does he care that his only source for "how the code works" is the AI's own explanations. And basically we ended up at "if it's not critical" (it was just a launcher app for his phone) "it's fine".

But then, even though I myself never "vibe coded" an app or anything, I feel like I am always underestimating what AI can do now with all of the projects popping up. It seems almost magical, because I have no idea of the process behind it.

And it makes me think about when (not if) we'll just stop reviewing the AI code collectively, like 90% people stopped verifying compiler outputs except in very special cases.

I don't like the non-deterministic/fuzzy nature of AI coding, and I can't really bring myself to use it more, and I don't like the fact that I'm kind of drifting away from the software development scene.

But maybe I'm thinking about the future too much, I'm selfishly hoping we will hit a bottleneck, and it's not right to be selfish about something like this.

LelouBil··on Thinking will become a hobby
Wow, thanks, it was a great read, I am not sure if I want to go to Australia or not now !

This part made me chuckle about current AI systems and prompt injection

> The Vertabrane system is operating in the same way. It is learning things, not running programs. It acts consciously rather than being 'programmed', and it has a far more rigid moral code than most human beings do. The Vertebrane system never blindly 'executes' a program, so it cannot be taken over.

I would say that the fact that LLMs never execute a program but "learn" it, mean that they can for sure be taken over by very unpredictable means.

LelouBil··on Terence Tao: Math 2.0 [pdf]
When do you think we will stop reviewing the AI's results ? Kind of like mathematicians are not checking results from calculators ?

Will there be a point where AI will just become trustworthy by default ?

The fact that current methods are not deterministic makes me think it'll happen by a new method, but if the success rate for most people is 99%, will it matter that it's not deterministic ?

LelouBil··on QGIS enacts moratorium on AI contributions
Actually I was extrapolating, I should have expanded about it.

But to me, having graduated last year, it doesn't seem far-fetched to think that AIs will also get better in terms of design.

And will conciseness and readability even matter in the long term ?

It's all things I think about more and more often and I'm not sure where I will end up

LelouBil··on QGIS enacts moratorium on AI contributions
AI agents can build a whole app and review PRs better than a lot of people, while having access to more context at once.

Honestly, where did the bottleneck end up now ? It sure isn't simply reviewing, or designing, so where is it ?

LelouBil··on Terence Tao: Math 2.0 [pdf]
I really like this way of putting it
LelouBil··on Play Skyrim as a Minecraft Player
I'm not really sure how I feel about this.

On one hand: It works, the mechanism is interesting (run both games in paralell, minecraft is hidden, IPC to have a minecraft entity for each Skyrim entity and synchronize health and other properties, and to show the minecraft UI it mixins minecraft render code to send the rendered GUI layer to skyrim)

On the other hand: I don't have anyone to say "congratulation" to, apart from maybe the person who prompted this, and it's a very weird feeling.

I am also having issues with people saying "this is only possible because of AI" or "Modding has changed forever".

First of this is possible because of the countless documentation and frameworks for modding both minecraft and Skyrim, these are two of the most modded games ever.

Secondly, the architecture wasn't "Impossible" before AI. It was just very tedious, and honestly I am surprised no one bothered trying before (or maybe it just didn't blow up); But I don't know if I'm being too dismissive here.

And lastly, with all of the big vibecoded projects that work great coming, I have a weird feeling that it isn't the "right way" to do it, it seems like this is the equivalent of brute forcing to have working software, when in my opinion advances like stricter typed languages or better developer experience sould make software development "just work" for humans as well.

I am probably still in denial about many things, very complicated feelings about this whole situation (I started my first real development job a month ago)

LelouBil··on Port of the TypeScript compiler, checker and lsp to Rust, by LLM
They didnt actually spent 400k, that's in "API-priced" tokens, he spent ~$24,047
LelouBil··on Navier–Stokes Lost in Translation
As an outsider to the field that what I understood from the person you are replying to. What's the difference ?
LelouBil··on Newgrounds.com – A community of games, music, and art
I never got to actually be on newgrounds, when I was 8 to 14 years old I didn't understand much english so I was always playing newgrounds games but off sites, it's only later that I understood that most of it was from the same community!
LelouBil··on CS240 AI Cheating Retrospective
I was an assistant teacher at a private engineering school, teaching a toy kernel project.

We had to add a new significant assignment which was the student sitting down next to me and we both look at their code and I ask questions.

Our policy for the project wasn't even anti AI, it was simply "do not submit code you do not understand" so obviously we had to test understanding, but already with 90 students it was hard, I don't know how you would manage something like this with 500

LelouBil··on Singapore govt dating app uses Gale-Shapley stable marriage algorithm
Right, FOMO is a human psychology problem...

...that many businesses exploit for profit.

LelouBil··on America.gov
Honestly just doing a Google search about what you want to do makes a page from this website appear, and it is usually very detailed while having a small "quiz" at the start to filter the information based on your situation. It's not just for life events, but for basically any public French administrative work. You'll have pages about your car, about your job, or if you want to immigrate, and so on.

At the bottom you also have other relevant links (usually to the website of the specific administrations of what you want to do) including reference to the current law that the page is based on (https://legifrance.gouv.fr is also great)

For what I used it for, it was never ambiguous in determining your situation

LelouBil··on Responsible Release of AI-Generated Mathematics
I saw an interview (in French) of Cédric Villani, speaking on behalf of him and other Fiels medalists, who said that in order to advance mathematics we need three things:

- ideas

- students

- problems

Basically, students to bring original ideas to try and solve existing problems and this generates new ideas and possibly new problems for new students to try and solve with new ideas and so on.

And he continued to say that the issue with LLMs in mathematics, is that he's afraid they could run out of problems, and so students wouldn't bother trying, and this could hurt understanding of mathematics as a whole.

It's totally not my field so I'm not sure what to think of it but it seems important

LelouBil··on America.gov
It can be good without resorting to an LLM.

I France service-public.gouv.fr is amazing

https://www.service-public.gouv.fr/particuliers/vosdroits/F1...

LelouBil··on VSCode's SSH Agent Is Bananas (2025)
Before reading I thought of the classic meaning of an SSH Agent, which saves your private keys
LelouBil··on OpenAI is well positioned to fast-follow Jev
Not directly related, but still jev related:

Would it be intesting/useful to use jev to generate a block of text like LLMs do ?

Like asking it to pick the n + 1 word given the starting text (using it's choice primitive), but also asking n + 2,n+3 and so on at the same time.

Would it give coherent or useful results ? Or does the fact that it computes it "all at once" means it cannot make one of it's answer influence the other ones ?

LelouBil··on OpenAI is well positioned to fast-follow Jev
Not an expert at all here, but I saw a comment on the jev post saying that it you constrain an LLM suck that it outputs a valid structure, if the token with the highest probability is not the one that you expected because of the structure (and so you pick the valid lower one) this means the LLM was already confused and your answer is less likely to be correct anyways.

Is this actually true ?

LelouBil··on Type Punning in C and C++
Oooh thanks, I missed that the cast was into a 64 bit integer
LelouBil··on Type Punning in C and C++
I don't understand the example in "Why C and C++ Differ".

Shouldn't the code return 3 in the base case and 4 if the check for 2 was assumed to always hold ?

LelouBil··on JetBrains Air: A System of Products for Agentic Software Development
There's an ACP connector that allows you to use Claude with a subscribtion.

It's not bundled with the IDEs but it's available on github

LelouBil··on AI-generated posters don’t have to be horrible
Of course if the text is made with care and re-read by a human it will be okay even if ai generated.

But I frequently encounter two issues:

1. Careless text generation, that was never read and so the human never prompted the LLm for proofreading and so on, this results in a text/readme/documentation that is using way too much words to describe simple or mildly complex things

2. LinkedIn post/corporate-trying-to-seem-friendly type of text that is always writing like it has to convince you of something (frequent user of the "it's not just X - It's y" trope)

I find that for most AI generated texts, it defaults to my 2nd point, and the human often doesn't read it or is happy with it, whereas I would have been much happier reading the shorter version they would have written instead.

Or just give me the prompt directly

LelouBil··on Data Protection Commission fines Google €403M over processing of location data
They paid all the fines that were confirmed after appeal, you can look at the linked table yourself.
← PreviousPage 2 of 24Next →