Alex

@pbeyssac Je viens de le faire

@Fax3l Ok mais faut m'écrire car de votre côté c'est fermé.

Pour info, l'attaque est toujours en cours, l'attaquant se permet même de relancer. *Ne cliquez pas* Quand je vois le nombre de gens pourtant pas nés de la dernière pluie qui ont cliqué sur le lien d'attaque, je suis effaré. On peut critiquer l'ANTS, on n'est pas rendus.

@dnhkng Was the vocabulary shared/multilingual for each individual training and did it fully cover all 8 languages?

1/7 I fed the same sentences in 8 languages through 4 different LLMs and extracted the hidden states at every layer. The result: in the reasoning layers, language disappears. The models organize thought by meaning, not by language. Then I tested code and math. Same thing. 🧵

@karpathy @KenWattana This reminds me the first time I talk to gemini and it asked me “how’s the wether in <exact city where i was> “

@KenWattana yes exactly! a bit like i'm being manipulated in some creepy way. "please like me, look how much i know about you, we are good friends".

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

@theo @BlazeisCoding Ahhahha microslop

@BlazeisCoding It means you’re probably using windows and your codebase has file names/paths that are too long for windows

Hii @theo , What's his error supposed to mean?

1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy

@steipete @ebukagaus Agreed. Though « submit yolo » is very fun

@ebukagaus Not sure if you are joking or not, but there'a big difference between a thoughtful conversation with an agent, then it writing code, then testing it and "yo fix issue x and submit yolo"

Are you a Grifter??

Seeing a lot of hype around ANE reverse engineering breakthroughs lately. Cool progress, but we benchmarked jina-v5-nano (12-layer 768d transformer) on M3 Ultra: • MLX GPU: 4ms • ANE hybrid (baked weights): 49ms • ANE hybrid (dynamic weights): 236ms ANE's value is low-power inference. It is not necessarily faster than MLX on desktop. MLX GPU does in 4ms what ANE does in 236ms on the same chip. Unified memory doesn't help when ANE must shuttle data through IOSurface kernel calls while GPU reads it directly via Metal with zero copy. People claiming ANE beats MLX for transformer inference have simply never used MLX.

😂

incredible that we built all this RAG and vector database stuff and it turns out that grep from 1973 works better than all that

Google has identified an iOS exploit kit named Coruna. 5 full exploit chains, 23 vulnerabilities, documentation in native English, modular architecture. Full professionalism. It must have cost millions of dollars. Who built it? Google doesn’t say, but the evidence points to US government tools. The kit also contains components previously used in a cyber operation that Russia attributed to the NSA. Coruna traveled. First, an anonymous “company client”, then used by a Russian cyber espionage group, which hid the code on Ukrainian websites inside a visitor-counter script, delivering it only to selected users from a specific geolocation. Later a financially motivated actor “operating from China” deployed it (infecting over 42,000 devices). The malware added to the ready-made kit was lower quality than the original suggesting the tools were acquired and modified by someone else. One US government subcontractor, Peter Williams, just received a 7-year prison sentence for selling tools to Russian broker Operation Zero. The US government spent millions on a tool that now steals cryptocurrency. A good return on investment, just not for themselves. One more detail: Coruna did not attack devices with Lockdown Mode enabled.​​​​​​​​​​​​​​​​ cloud.google.com/blog/topics/th…

@powl_d @oooooooorion @ropau_ai 400k token dans le contexte c est 160 milliards d elements pour que le systeme d attention choisisse ce qu il se passe. Sans « tricher » avec des flop16 c est 300 gb + de ram. Donc l attention est diluée…

Franchement c’est une facon de faire pour du dev solo sur Claude Code ou Cursor. Mais nous on build Orchestria : 5 à 10 agents (Opus, Sonnet, Gemini…) qui se challengent ( On désactive actuellement des agents et on dit a Orchestria de se limiter), se review, font de la QA visuelle avec Playwright et se corrigent entre eux. Un run complet c’est 200 à 400k tokens easy. Avec 1M de contexte tout tient en mémoire, zéro troncature, zéro résumé foireux, zéro perte d’info. L’empoisonnement c’est un problème de dev solo en mode chat. En multi-agents avec quality gates intégrés, ils se démontent entre eux. El famoso /clear c’est bien quand tu codes solo. 1M de contexte, ça change la game quand tu manages une équipe d'agents

On a débloqué l’option 1M context sur Opus 4.6 Windows pour @ropau_ai 🚀 Objectif : aller plus vite, scaler plus fort. Pas juste regarder mon argent brûler en 4K 😅🔥

@bluetouff A croire qu’il aurait ete dresser pour surtout ne pas paraitre woke

@bluetouff 😂 mais wtf Bluetooth noooon

J'ai vraiment trop hâte que Grok soit utilisé par le Pentagone 😂

@robertgraham state = table[state][column[c]] This so smart i ll now have to live with that knowledge and probably try to put this everywhere

Here was my answer to the problem, a faster implementation of the 'wc' program that the interviewer wouldn't be able to understand how it works.

@mfranz_on Python -m venv venv

A new toy is arrived. Sooo what should I install first?😂

@KieranCrown @ipla03 Dude!’ Thanks!

@ipla03 That 13 mini bug is so annoying. For anyone seeing it go into Display > Screen zoom set to large and then back to normal 👌🏻

That moment when you spin up these three guys to check if your UI is responsive enough

@SanhEstPasMoi “Senior Slop Janitor” 😂

On my timeline: - “Engineers haven’t written code in 3 months.” - “SaaS apocalypse.” - “PMs are obsolete.” - “If you’re not running 10 coding agents in parallel, you’re ngmi.” - Crab costumes In our repo: - “You are absolutely right!!! Great catch!!!” - Agent deletes half the file and rewrites it right away. - Hallucinated API endpoints. - 47 new tests. 2 are actually testing something non-superficial. - README .md README_SUMMARY .md FINAL_SUMMARY_V2_ACTUAL_FINAL .md Goal Architect. But in the meantime… Senior Slop Janitor. If you feel that tension, we’re hiring builders in NYC.