Sabitlenmiş Tweet
Guigui
1K posts
Guigui retweetledi
I just finished working on a very important system for my game :
Multi tiles entities.
Before, every entity had a 1 tile by 1 tile footprint, which was enough for small props, small monsters and the player.
English
@BrumeDofus @Huzounet Tu peux faire un bot qui utilise tout comme un humain (lit les pixels de l’écran, déplace la souris, ect) et utilise l’ia pour faire ces choix mais ça va te bouffer une quantité affreuse de token et le mec sera pas rentable.
Français
@BrumeDofus @Huzounet L’IA lit les paquets pour comprendre ce qu’il se passe.
Si les paquets sont offusqués l’ia voit plus rien.
Je suis même pas sûr que ce soit l’IA qui clique sur l’écran, elle envoie très probablement des paquets pour ça aussi.
Français
So just following up yesterdays discussions with @theo here are the results testing various reasoning efforts through @OpenRouter on the @convex evals.
The GPT 5.4 xhigh result was the most surprising to me, so I re-ran it again to check and it got the same result which is inline with what Theo was saying that xhigh is worse than high.
English
Open source projects that desesperatly need contributions will allow a lower trust score, and the ones that need high quality contributions will set a high minimal trust score.
To prove you are trust worthy to contribute on big softwares you need to contribute to small ones 1st
English
We need a « trust score » on developers GitHub accounts.
you submit a PR:
it’s approved => your score goes up
it’s closed because it’s ai slop => the score goes down
And project can decide of a minimal trust score to submit a PR.
Pirat_Nation 🔴@Pirat_Nation
Open-source game engine Godot is drowning in 'AI slop' code contributions: 'I don't know how long we can keep it up' Many submissions contain nonsensical code changes, fabricated test results, and overly verbose descriptions typical of LLM output. Reviewing every new contributor's PR has become extremely time-consuming and demoralizing. Rémi Verschelde stated they may not be able to sustain the current level of manual vetting much longer.
English
@scaling01 I don’t get why 5.3 codex is not present in benchmarks, it went out the same day as opus 4.6.
This makes all this benchmark completely worthless
English
There's yet again, another SWE benchmark
and Opus 4.6 is on top
English
Opus 4.6 has become so fucking slow holy shit.
What have you done to my boy @AnthropicAI ?
English
@EU_Propaganda I don’t want the country that started the 2 more murderous war of human history to have access to weapons able to obliterate all life on earth.
English
I can't believe this mentality still exists in the 21st century. 🤦🤦
@levelsio@levelsio
You absolutely cannot trust Germans with nukes Did we learn anything from history? I'm Dutch so I sure as hell did when we were invaded by them French yes, British yes they can have nukes, they're trustworthy, not Germans They'll start behavior policing the entire EU with nuclear weapons as a threat Absofuckinglutely not
English
You absolutely cannot trust Germans with nukes
Did we learn anything from history? I'm Dutch so I sure as hell did when we were invaded by them
French yes, British yes they can have nukes, they're trustworthy, not Germans
They'll start behavior policing the entire EU with nuclear weapons as a threat
Absofuckinglutely not
New York Post@nypost
Germany may develop its own nuclear weapons with EU allies: ‘Talks are taking place’ #Echobox=1769990057" target="_blank" rel="nofollow noopener">nypost.com/2026/02/01/wor…
English
This is how I code now.
I have a list of prompts that I send to Claude after it's done with the current task. I only use one agent at a time, to track the changes and the current task properly.
I just open a text buffer on Zed to plan my next steps, with the help of autocomplete I create these tasks easily.
English
@mat_augmente « Je ne fais que mon métier », le métier en question : harceler des gens toutes la journée pour les scams
Français
@MarekaDoBrasil Qui utiliserait un logiciel médical critique codé en langage naturel, dont le code produit serait non déterministe et impossible à faire relire et valider par un pair ?
Français
@MarekaDoBrasil Le langage naturel est imprécis, sujet à interprétation, un logiciel sérieux ne pourra jamais être codé en langage naturel.
Tu auras toujours besoin d’une étape intermédiaire, dans un langage compréhensible par un humain et produisant un code machine de manière déterministe.
Français
Débat intéressant. Je ne pense pas qu'un langage (déjà lequel ? Python ? C ? Pourquoi pas l'assembleur ?) est une brique fondamentale de savoir. Coder c'est communiquer avec la machine. Si l'IA était une interface parfaite j'aurais pas de pb à plus jamais rien coder.
farghie@farghie
Danger de l’IA. Découvrir une discipline en faisant l’impasse sur ses primitives finit par se payer. La calculette a 50 ans mais les programmes scolaires perpétuent la division posée. Sans maîtrise des fondamentaux, la pensée est incapable de s’exercer avec ordre.
Français
