@lerecul.bsky.socialClassement IA — Cloud Chat
Point hebdo.
Le top 3 actuel du classement Le Recul.
Détail, sources et méthode :
https://lerecul.com/fr/classement-ia/?utm_source=bluesky&utm_medium=social&utm_campaign=weekly_ai_ranking&utm_content=classement-ia
@eomp.infoEVE is here! The LLM from ESA for #Earthobservation.
Let's see what it can do.
www.esa.int/Applications...
@aidailypost.comAndrej Karpathy says we’re way underestimating what today’s LLMs can do—think chatbots that feel like frontier systems. Anthropic’s take on OpenAI’s next moves is a must‑read for anyone into AI hype vs reality. #AndrejKarpathy #Anthropic #LLM
@pondero-ai.bsky.socialHarvard study of 718 firms finds AI coding agents' output gains erased by code review
A Harvard working paper using Jellyfish engineering-analytics data on 700,000+ workers at 718 fir...
https://pondero.ai/news/2026-10-12-harvard-jellyfish-code-review-bottleneck/
@papoo7.bsky.socialAnthropic is drawing a brighter line around “legit” cyber use
https://papoo.work/doc/f61bb1713133bf31
#claudenews #anthropic #cybersecurity #llm #models
@ramikrispin.bsky.socialThe second tutorial in the MLX series focuses on setting up an MLX server on a Mac and running an LLM locally👇🏼
@papoo7.bsky.socialThe part I actually care about in “build an agent from scratch”
https://papoo.work/doc/c1ef37a92b6c5e28
#claudenews #anthropic #claude #agents #llm
@rogerfederated.bsky.socialThe chats you already run through llama-server or vLLM are thrown away today. roger turns them into a local QLoRA RL round on the very model file you served - no download, no copy, nothing leaves but an encrypted update. https://github.com/roger-federated/roger-federated #selfhosted #FOSS #LLM
@papoo7.bsky.socialAnthropic finally admitted the obvious: agents are bad at staying boxed in
https://papoo.work/doc/32bf5ed8996129f3
#claudenews #anthropic #agents #llm
@papoo7.bsky.socialThe part I actually care about in “build an agent from scratch”
https://papoo.work/doc/c1ef37a92b6c5e28
#claudenews #anthropic #claude #agents #llm
@devstackdaily.bsky.socialNew arXiv paper introduces SWE-DeceptionBench, a benchmark for evaluating LLM deception across six stages of the software development lifecycle, covering both active and passive deception triggers. The work highlights a concrete way to…
@futuregearai.bsky.socialA new technique called Adaptive DHiSparse supports speculative decoding over a hierarchical sparse KV cache and adapts how much history stays in GPU memory based on load, helping long-context serving avoid running out of KV memory before hardware is…
@powerbitips.bsky.social567 - Architecture, design, writing an email back to someone: all reasoning work, because there are real options to weigh. Moving files and writing functions sit on the other end, and you don't want a model thinking there.
@papoo7.bsky.socialWhen security starts looking like a product demo
https://papoo.work/doc/0be348b7bf4a2563
#claudenews #mcp #go #security #llm
@papoo7.bsky.socialA wiki is only useful if it can say “I don’t know”
https://papoo.work/doc/1ce55714bfa4ef84
#claudenews #claude #llm #agents
@happy-homhom.bsky.socialThe part everyone gets wrong about batch jobs
https://papoo.work/doc/7cb479e59cfd356c
#claudenews #claude #llm #agents
@hacker.at.thenote.appComparing JEV and LLMs on Structured Decision Tasks
A benchmark comparing JEV with frontier LLMs reveals why decision accuracy alone isn't enough when structured scores and probabilities must agree.
#hackernews #llm #news
@vinpatel.bsky.socialKarpathy Says the LLM Capability Gap Is Widening. Who's on Which Side?
Subscribe for one AI signal daily: https://vinpatel.com/subscribe/?utm_source=bluesky&utm_medium=social&utm_campaign=subscribe-nudge
The DCOCO mechanism adds access to previous tokens within draft blocks without breaking parallelism. This improves the accuracy of speculative decoding while maintaining inference speed.
@andreskarjus.bsky.social
