What if your MCP servers could provide interfaces, not just tools & data? We're designing MCP Apps with OpenAI and MCP UI, an official MCP standard for how MCP servers can provide interactive interfaces: https://t.co/8TnhGj2SWu

Deepwiki now faces a challenger: Google wiki. But I got addicted to Deepwiki MCP as feeds directly the agent. https://t.co/htz8NRf5yy Announcement: https://t.co/khGenEE4Zg Also they plan to integrate it into Gemini CLI, so expect an API. #AI #Deepseek #Google

MCP Specification 2025-11-25 Release Candidate: We have frozen the spec for the upcoming release of MCP. Upcoming additions are Tasks, enabling long running operations, and Client ID Metadata replacing Dynamic Client Registration as well as a lot of small ergonomics. Take a look

Use Colab directly in vscode, amazing from Google. https://t.co/8fpGlFPAR6 https://t.co/YQfY6maZJ7 #AI #Google #Colab

Claude Code 2.0.37 killed CPU 100% in Devcontainer after using --teleport.

Owasp Released AIVSS a new method to assess risk for AI instead of CVSS scoring. Scores seem lower, which seem more logic in threat modeling. The topic is still a bit fuzzy and suffering from a lot of hype. https://t.co/BK8pd0j6MF https://t.co/wl5btcGGR8 #AI #Security

Sonnet 4.5 have major flaws, when following plans. It manage always to create havoc, despite solid specs, SDK, docs all prepared and laid out. The least it do, is telling you all done while it's 40% completion. Bad for complex long tasks.

Helm 4.0 is out https://t.co/GABj8rMXpe #Kubernetes #K8s #Helm

More than sure Claude Code Web have a "loop" detection. To detect when the model is unable to complete a task and start redoing/undoing stuck in sort of loop, unable to really finish the task.

Claude Code cloud died trying to fix some tests. It was no more responding until Codex provided it with clear direction. It was quite stalling, refusing to do anything! And the genius Sonnet 4.5 marked them skipped and said all done 5 times.

#Codex since few days starting using like crazy inline python scripts with very bad results. This is strange. Never noticed this before. Any one noticed this?

How to see that Kimi K2 is built on top of Deepseek: https://t.co/fslCSQWj1j @Yuchenj_UW

Nvidia latest GB300 is kicking! A full cluster system can get up to 1M t/s with Llama 3 70B which is not a small dense model or MoE. That's 62k t/s per node!! https://t.co/1qH3hHbR5H #AI #NVIDIA #INFERENCE

DeepWiki is one of my favorite tools to explore and understand code on Github. But I found they had an MCP open and free! That's a real steal. Thanks #Cognition https://t.co/gm7ilrKVNi #AI #MCP

MCP/Tools can't be mapped to API 1:1. May sound easy shortcut for some API lovers, but AI models can use more efficiently API's thru wrappers. Doesnt' mean CURL is the solution! You need to help the model and steer it a bit. #AI #MCP

No one looking at #llama.ccp output? Or updating it? The Web UI had been there since a while! #AI

Great rex from Hugging face over Smol llm training: https://t.co/0G8eAcbfat Thanks for sharing @huggingface #AI

Surprised to see AMD so close in the high end vs B200. TCO is close too for AMD. MI355is close. (inference here as training is another story). https://t.co/NbXPBgCP2v #AI #Inference

VLLM support sleep mode to allow faster models switching, which is practical in constrained GPU env while testing or to save on GPU when deploying multiple models. https://t.co/oFKTzcQcjE #AI #VLLM

Router from VLLM similar to hugging face chat using models routing with ArchGW model Curious how this scale with models that are more capable here. https://t.co/JtYK0x07Fr