ramon156

I spent a lot of my time on the internet because I was lonely in my teens and these connections were exciting. I learned software dev because I was a nerd that loved math and loved the concept of building software, having complete control on how it's done. That was the concept I signed up for.

Software dev today is sitting in front of a prompt bar, sometimes checking the IDE, sometimes some DevOps, but nothing about it excites me anymore. Even hobby projects are dull because they're done in mere days. Choosing not to use LLMs has little to no reward imo.

I'm slowly moving away from software dev because the rose tinted glasses I saw the job as are shattering

It hurts my heart too :) I thought it was something I had to push through, I was wrong.

I notice a lot more people moving to a career where they manage, as opposed to do the programming. I see why.

I'm slowly phasing out of my software role, I don't even have that much experience yet. I just cannot sit behind a PC for 40 hrs a week talking to a chat bot. I would rather do something useful with the time i've been given.

next time just write the post yourself. I don't need an entire paragraph to explain that deser = serde but reshuffled

I'm glad online randos know better than people who get paid to do so. I wonder why theyre on the paid side..

can you try this again but by hand? I like the idea but the current state is unusable

the issue is, what will we use? companies are benchmaxxing, so how do you make a bench that cannot be cheated?

Human curated benches aren't accurate enough

it feels yucky to read LLM generated words on a Google blog, but i guess it's not too distracting. cool write-up though

Your clarification makes sense. The distinction between overall benchmark performance and why Terminal-Bench is an outlier is important

If I'm totally honest, I lost my interest in programming. There's not really much money to be made, I'm not efficient enough for the work force and the people I end up working with are mere shells of people. Not everyone, but they end up leaving the company anyway.

I'd rather use my time in a field that is rewarding. Somewhere I can find like-minded people.

thank you Claude

IT's the infamous "OneDrive in 10 lines of code (SFTP)"

While technically correct, it's not the same thing

how come opus 5.5 did so well? I'm surprised

managers that love notion are the cruft you're speaking of

so jev isn't even slightly deterministic? this is worse than an NLU and somehow more expensive. interesting times we live in.

round(stat * 1.05); // rounded down round(stat / 1.05); // rounded down again

that would be my first guess.

The only LLM that does this wel (in my opinion) is gemini. if i ask claude to fix comments, it very clearly fails in doing so. Gemini actually seems to understand the scope of a task

I'd rather read it's slop, it actually adds something to the conversation.

I'd say "imagine if half the posts on HN aren't worth your time" but this is genuinely how it feels nowadays. I'm not sorry

i can see the comment now, so this seems undone?

The more I see about how Jev works, the less interested I get. Jev would be cool as a local model for e.g. Home Assistant. Projects like this are at best misleading, this one also happens to be broken.

if you can be replaced by an algorithm, how useful were you really?

But the evidence is not there...

I've implemented tree-sitter in pi before, and while it works, I have no real proof it saves me tokens, or is more accurate. I think a better implementation is a model that's trained for AST's, not just "use tool, see what happens".

I'd love to do research on this when I have the time.

This sounds good but so far all claims just sound like marketing terms. I'd love to see real proof. e.g. "RLCD" and "parallel sampling" have nothing to back it up.

also "70-500ms vs 3-329 seconds" are apples-to-oranges unless the LLM baseline is doing comparable work (e.g., long chain-of-thought). If Jev is skipping generation entirely for a narrow structured task, of course it's faster.

Nonetheless i want this to be true, so I'm looking forward to Jev

Edit: I really have to say that I like their manifesto https://typesafe.ai/manifesto

Thank you Claude, very cool!

name some examples. easy to use is opinioned,

i quite liked using strix. last time i tried it, deepseek was a mess and bloated the context with nonsense. that was ~5 months ago, i wonder how it performs now

not in this context, though?

they arent being called racist, they are racist. not a small difference there.

how is a filesystem supposed to prevent you from doing that?

i would love a deterministic program that can confidently make plans for lower-cost models like deepseek. ofcourse the LLM part wont be deterministic, but its a lot easier to measure quality like this. you could argue an AGENTS.md is this, but from experience its not enough to make non-frontiers act have a high success rate.