If you use any-llm, you can give this a try by using our existing support for https://t.co/hQDtNS7vGK as a provider. 🚀

If you live in Pittsburgh, take a second to do some reading on the different candidates for tomorrow before you vote

Well the @firefox shake to summarize feature is magic. Instead of scrolling through everything, shake your phone and get straight to the recipe. (This recipe is incredible btw)

When we bought a Subaru, suddenly I realized all my neighbors Subarus. We had kids, and suddenly we noticed all the children at the park. I get similar vibes with Mozilla. After joining I started seeing that many of my favorite projects online are sponsored by Mozilla in some way

https://t.co/4abdqMFCFr 🤝 llamafile We’re adopting llamafile to advance open, local, privacy-first AI. We’re refreshing the codebase, modernizing foundations, and shaping the roadmap with the community. Read more here: https://t.co/WBg6pDAzyO

Lots of back and forth in terms of what to do to replace the OpenAI completions API. We know that we need to migrate off of it eventually, but what do we migrate to? Anthropics API, the OpenAI Responses API, or something else?

Evaluate agents by their function calls, not their answers. The medical triaging agent I'm building succeeds when it: - Escalates emergencies immediately - Fetches medication data before attempting refills - Realizes Adderall is a controlled substance and routes to a

Everyone who’s building real agents knows @karpathy is right.

I don't know what labs are doing to these poor LLMs during RL but they are mortally terrified of exceptions, in any infinitesimally likely case. Exceptions are a normal part of life and healthy dev process. Sign my LLM welfare petition for improved rewards in cases of exceptions.

AI was supposed to save us from UIs like this. Which is why we bet on a pure natural language interface for Replit’s agent builder. Visualization is there for debugging and understanding, not building.

I gave a talk today at The Curve on the state of open models. Here are the slides, recording soon. Topics include: Chinese ecosystem, reflections on DeepSeek, the demise of Llama, who will fill the U.S. market, what local models do, ATOM project & ai2, and more topics

🚀 We’ve just launched our new https://t.co/4abdqMGauZ website! We’ve given our website a full makeover, making it easier than ever to explore our work and connect with what we’re building. 👀 Check it out and explore what’s coming next:

Fine-tuning as a service in 2025? What did they see?

I’m tired of being “absolutely right!” when coding with an agent

Some of my best inspiration hits while I'm on a bike ride or a walk. Now I can have cursor implement my idea for me before I even get back home!? https://t.co/7mDA9JhL2r exciting.

779 ⭐️ Any-LLM is a unified Python SDK that provides a single interface to communicate with multiple LLM providers 🤖 while leveraging their official SDKs for better compatibility and type safety ⌨️. https://t.co/WhHBHou2G7 #starhistory #GitHub #OpenSource

I used gpt-oss:20b to help me out with generating food recipes today, and it's comforting to think that I can do this with this exact LLM for the rest of my life

https://t.co/9xlwTRYFvJ

Using a language model is like learning how to interact with a coworker. Even apart from capability, you learn how to optimize your interactions. Use an open-weight model like gpt-oss or mistral and you never have to worry about that coworker being replaced without your ✅

Source: transformers code: https://t.co/aAMfoFxOSp Sliding window, MoE, and RoPE are pretty standard. Attention sinks are interesting, I don't know any major models that do this. Thanks to @abertsch72 for sharing observations and prompting me to look too.