Rémy

Rust is fucking disk hungry that's kind of crazy. I guess there are tricks to deal with this?

Given the current war on software from project Glasswing and friends, I'm starting to get a little bit paranoid. What if we started to isolate web applications per view with super-minimal rights? Could this be done?

"Fair point" GAWD SOMEONE PLEASE TEACH THOSE LLMs TO STFU AND DO WHAT THEY ARE TOLD Sorry about that

What's your favorite threat to give to a LLM? At te moment: "... or I'll have you talk with a complotist for 1 trillion tokens"

No RCE in Python today? Come on, what are you doing guys?

👀💎 A Google engineer paired with Gemini 4 Argon in an iterative profiling and kernel-optimization loop to pinpoint the exact vLLM scheduling bottlenecks and land all 8 upstream PRs in days. Full PR list on GitHub: https://goo.gle/4zNtANl

Was today the worst cyber security day of all times? It somehow was very smooth. And yet very real. I just don't know. Almost like of a major city got nuked but everyone was ready for it?

Me: invents algorithm Claude: ah yes, the famous XXX algorithm Me:

Even better than computer use: VM use. My agent has been coding for Windows in a VM that I've not opened ONCE. It even downloaded the ISO on its own. This is fantastic 🥲

Elon Musk is destroying the night sky, pushing us to the brink of Kessler syndrome and then deciding who this benefits to. Wonderful.

wow, i had my doubts, but this is true in every single LLM i can find

The trend is inverting so much: before DeepSeek was slow and larger models were fast, but now DeepSeek is quite fast while larger models struuuuggle like crazy

I may or may not have created 3 ghost tickets in Linear to get the issue number 1337

The Gmail API rate limit is surpisingly steep. It's as if Google didn't want people to get off Gmail or something.

Where DeepSeek v4.1 Flash feels great is on the fact that it's pretty good at avoiding collapse by focusing stronger on some parts. It WILL collapse when you ask for something muti-parametric it can't handle, but that won't prevent the tool call from working, etc. Pretty good honestly...

Am I the only one remembering research they did in the past as a conversation with a LLM, knowing that said research was done years before LLMs were invented?

I should have learned the lesson by now but LLMs are absolutely unable to scope properly user stories and tickets and even less to order them properly. That's really something I need to stop doing.

Not sure what's happening but DeepSeek v4.1 Flash is getting A LOT faster these days. Is it new hosting techniques or is it people going for another model?

Tired: interpreted languages are slow Wired: DeepSeek, rewrite this function and all its hot-path dependencies in Rust, it's too slow

The great thing with AI is that you can take old packages you made, polish them, document them and re-upload them in a much more accessible way My 7-years-old LiveZip module, which allows to stream .zip files without building them locally (yes it's possible).

Flash models can implement back-end features fairly easily while they struggle to do visual things. Back-end snobs owe an apology to front-end developers 😅

Feels like GPT Sol and bigger has the edge on web design?

Is DeepSeek the only AI whose logo isn't an asshole?

If pkill could stop matching itself it would be fucking great

TIL: whenever you register a certificate with Let's Encrypt (or anything else), the domain immediately becomes public and within 1 second a hoard of bots will be pondering your server for .env files

Google's internal process to approve felonies must be longer than Anthropic's. On the other hand, Facebook's management must be disappointed at their engineers, given that felony is assumed

Does anyone know why the Windows Settings dialogs are getting more complex every time I open them?

With all the horrors that the fly is experiencing, did anyone try to get it to fly a drone?

An escape hatch handed to an agent is not a guard. -- Claude

If you’re also fascinated by what people are doing with the fly connectome (and wondering what the fly thinks about all this) read “Lena,” by @qntm.org a very short story written like a Wikipedia article about the near future that may be waiting for us. qntm.org/mmacevedo

Noooo don't be our competitor you'll kill us all ahaha and I am very genuine

Someone realized their model was overpriced 😅

In other news, GLM 3.5 Flash seems to be smart enough AND has vision AND is 1/3rd of the price of DeepSeek v4 Flash?

It's funny that base 64 is basically transparent to LLMs

What I love about modern packaging is that you'll receive a dictionary of disclaimers that you're never going to read nor use, but on the other hand they will drop you a think slice of paper with 3 hieroglyphs to explain how to use the product. Awkward sense of priorities

YES PLEASE FINALLY, Fable is gonna stop bailing out on me when it's looking at dependency code 🥲🥲🥲

Nothing is harder to find than your own repositories in GitHub... I don't even understand

Philosophers are actually the kind of people the best suited for AI engineering. It's been 2600 years those guys are fighting over engineering the proper prompts and it's paying off only now. They got language skills and logics. On hindsight, they are better armed than most developers 🤓

Me: can you validate this model of a Peltier module cooling water? Claude: THIS IS AGAINST MY POLICY I mean, what the fuck?

To be fair it's not even true. It's organized around Nash equilibriums, which are a massive pain in the ass. And the only way to get out of them is a strong and sustainable government. Which you'd also want to keep democratic. We're so doomed.

Person of Interest, a 2011 show, is funny because it's completely unrealistic. The plot is essentially: some dude created Palantir, but he's using it for good