Felipe O. Carvalho

27.8K posts

Felipe O. Carvalho banner
Felipe O. Carvalho

Felipe O. Carvalho

@_Felipe

@ApacheArrow / Databases / Compilers. (past @SDFLabs, VoDa, @Spotify). Rust/C++/TLA⁺🇧🇷 → 🇸🇪 → 🌎 https://t.co/vxbdByfADI

Katılım Haziran 2008
2.4K Takip Edilen5K Takipçiler
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
That’s why the fundamental duality between compilers and LLMs is not about determinism.
English
0
0
1
78
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
For many practical reasons, LLMs are non-deterministic, but they could be 100% deterministic by returning the same inadequate answer every time a question is asked. And a compiler can be deterministic but full of bugs.
English
1
0
2
81
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
Compilers and LLMs are programs that expand/decompress compressed input. Compilers deal with LOSSLESSLY-compressed forms of the final code while LLMs inputs are LOSSY versions of the final code.
English
1
0
3
196
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
Never call a module "utils" or "common". Go into debt if you have to.
English
138
52
1.7K
110.7K
Gabor Cselle
Gabor Cselle@gabor·
Does anyone even read code anymore?
English
28
0
18
2.7K
Eric Zhang
Eric Zhang@ekzhang1·
fun systems math: all of the world’s LLM API traffic (~100T tokens/day?) could be served on a single computer in your basement running Nginx (assuming a token is ~4 bytes and you have 50G Ethernet lol)
English
19
7
786
122.9K
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
@sqs I noticed Claude Code doing this one day and thought it was very clever (good!). It was weird that it asked for git stash permission right after all tests passed.
English
0
0
1
98
Quinn Slack
Quinn Slack@sqs·
Useful technique in Amp with GPT-5.5 (deep mode) to make small independent commits from a big uncommitted changeset: "git stash the uncommitted (including untracked) changes, then reapply this refactor that you performed on HEAD, then git stash pop and fix merge conflicts"
English
6
2
93
6.6K
kitze
kitze@thekitze·
remember when pretext changed the way we browse
English
4
1
27
5.2K
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
@slimjimmy The classical example is parser generators vs hand-written parsers Hand-written parsers are more work at first, but doesn’t constrain you when you need to deviate from the theory, do complex error recovery etc.
English
1
0
2
100
Felipe O. Carvalho
Felipe O. Carvalho@_Felipe·
@slimjimmy Both right at the same time. It’s always hard to decide between highbrow algorithms and the lowbrow alternative: brute-force, slow but practical, often more legible (not always), flexible.
English
1
0
16
1.2K
immortal
immortal@immortaldip·
@finbarrtimbers wait, aren't we at stage that it's now silicon, power engineering problem ?
English
1
0
1
991
finbarr
finbarr@finbarrtimbers·
Working in ML starts out as a math problem and very rapidly becomes a distributed systems problem
English
23
94
1.3K
51.5K
alon turing
alon turing@chaumian·
why do ML ppl refer to anything with indices as a tensor
English
25
0
110
29.5K
derek guy
derek guy@dieworkwear·
Why do people write tweets like this? Where every sentence gets a new line. Sometimes a line might have two sentences. Like this one. But generally speaking, every sentence has a new line, making a tweet look like a long block of text that no one reads. Worse still, such tweets are often repetitive and winding, hammering on the same point over and over again. The writing is often very bad.
English
968
231
7.4K
430.3K