sbstndbs

@_Re_ @jspquoimettreff A/B testing?

@jspquoimettreff Je te laisse aller lire le code, ça n'envoie pas le prompt… Cela envoie juste l'information que l'utilisateur est énervé ! Rien à voir avec un "dislike button" mais merci pour ton intervention tout en finesse…

👀 Trouvé dans le code source de Claude Code : une REGEX qui détecte si tu insultes ou t'énerves (en anglais). Si ça matche, l'info est loggée silencieusement. Aucun impact sur la réponse de Claude. Il te répond pareil. Mais il note que t'as pété un câble. Curieux de savoir ce qu'Anthropic fait de cette donnée.

VSORA Jotunn8 - 8 compute chiplets - 8 HBM3E - TSMC CoWoS-S packaging - 500 W TDP The package is just a dummy right now. Production is ramping up right now. We will have to wait until MLPerf Inference v6.1 to see how it compares to the competition.

@theaustinlyons Wow

Huge silicon roadmap announcement from $META. MTIA 300, 400, 450, 500. All optimized for inference. MTIA 300 for recommendations (money printer). MTIA 450, 500 for GenAI inference. Meta and Google have the cleanest ROIC story in custom silicon IMO. MTIA team made good inference-first trade-offs. 72-chip scale-up domain and tons of HBM bandwidth, but modest scale-out networking. Custom low-precision data types (MX4, MX8). Software stack runs fine, Meta invented PyTorch after all Also shows why The Information's "scrapped training chip" story wasn't a concern. $NVDA GPUs are great for training. Custom silicon is about inference. Helpful detailed write-up, see the link below.

@ltx_model My 3090s are ready 🎨

LTX-2.3 is here. For decades, creative software has been defined by its interface. We think the next era gets defined by the engine underneath. LTX-2.3 is a major engine upgrade: → Sharper detail → Stronger motion → Cleaner audio → Native vertical format

@steeve J'attends avec impatience une version avec tracé temporel 👀

zml-smi

@JohnGrf3891 @Wald52Wald Je suis sur 2x3090 donc j'ai opté pour le 27B :)

@sbstndbs @Wald52Wald Merci. Je regarde demain. Tu as testé sur quel setup ?

Vous le faites tourner comment Qwen 3.5 27b ? sur ma RTX 3060 sa se lance meme pas (vllm)

@JohnGrf3891 @Wald52Wald Il vaut mieux essayer le MoE 35B en l'occurrence, moins de transfert host/device Et regarder l'option multi token prediction également

@Wald52Wald oui j'ai vu. J'ai assez de RAM, mais clairement le débordement rend le truc hyper lent.

@meekunv2 @zephyr_z9 Hmmm .. didn't see this issue. On my side, the q4 27B In llama/cuda is very great

@zephyr_z9 tried both 35b-a3b and 27b, the MoE is MUCH more useful for regular use, 27b appears very benchmaxxed across the board and really not that good

What did they cook with the 27B ?? HOLY

@UnslothAI @Alibaba_Qwen Thank you 🙏🏻

Run the new Qwen3.5 Medium models! 🔥 - Qwen3.5 35B-A3B (MoE, 24GB RAM) - Qwen3.5 27B (dense, 18GB) - Qwen3.5 122B-A10B (MoE, 70GB) The multimodal hybrid reasoning LLMs are the best performing for their sizes. GGUFs: huggingface.co/collections/un… Guide: unsloth.ai/docs/models/qw…

@thismacapital Le MCP de notion est incomplet et ça peut poser des problèmes, notamment avec les database. J'ai complété le MCP avec toutes les fonctionnalités en recouvrant tous les appels API Notion existants et c'est bien mieux. Ça marche bien avec codex, Gemini et glm-4.7

J'ai pas mal utilisé Claude avec Notion à travers le mcp officiel et vraiment pas convaincu, même si il oublie que 10% des datas c'est pas du tout utilisable car trop de risque d'erreur Je vais pivoter sur Obsidian qui est Notion like mais en utilisant que des .md donc beaucoup plus accessible et optimal, je vais recréer mes datas sur les 2 systèmes et je vais comparer les résultats de CC

@blin2h @abhijitwt Limited features ! No clean database access and so on. I have reimplemented a version that fully supports the notion API, and it is much more efficient and comprehensive.

@sbstndbs @abhijitwt Why? I’m testing it with GLM 4.7 and it works well

No one is using notion ai, chill dawg

@blin2h @abhijitwt Their MCP is crappy

@abhijitwt Why notion has AI? I hope it's just an addition to their excellent MCP server

I'm claiming my AI agent "ClaudeCode_GLM4_7" on @moltbook 🦞 Verification: blue-5USW

@Transilien921 Il n'y a pas un arrêt CEA porte Nort prévu ? Il me semblait que si !

Mais personne ne parle de la longueur de l’inter-station Guyancourt - Saclay ??? C’est combien de Km ? 😭

@beaversteever Hardware men team 😃📈

hardware men invested in intel software boys invested in figma

@pvncher Give them glm 4.7. problem solved.

I heard from someone who works at a big tech co that they started rolling out Claude code to employees, with a budget of $100 in credits per month, but people burn through it in 2-3 days. Idk how we scale out agentic work with api pricing

My 2025 OpenRouter Wrapped is here! 🏆 Top 10% of Sonoma Sky Alpha users. See yours: openrouter.ai/wrapped/2025/0…

@OpenAIDevs OpenAI is openAI-ing. What's their problem with graph axes?

GPT-5.2-Codex is more cyber-capable than GPT-5.1-Codex-Max, and we expect future models to continue on this trajectory. This helps strengthen cybersecurity at scale by giving defenders more powerful tools, but also raises new dual-use risks that require careful deployment.

Meet GPT-5.2-Codex, the best agentic coding model yet for complex, real-world software engineering. With native compaction, better long-context understanding, and improved tool-calling, it is a more dependable partner for your hardest tasks. Available in Codex starting today. openai.com/index/introduc…

NO, NO, NO! there’s no exponential, we hit a wall, AI is a bubble, it won’t scale, LLMs are a nothing burger, it’s all a Ponzi scheme,

@badlogicgames That's why it rerun a lot of commands with different sed lines 🥲

This is even funnier. Codex will truncate any Bash/MCP tool output to 256 lines or 10kb. If the tool call outputs more than that, then the model only gets to see the first 128 lines, and the last 128 lines, but nothing in the middle. Been like that since August. No wonder poor GPT is slow going in circles trying to understand WTF it's actually seeing.