huggingface transformers remains extremely useful for getting started but has bloated beyond recognition and is no longer helpful for learning someone should make nano-transformers. implement all the models from scratch without all the dependencies. would be amazing
CODE
-
Box’s Agentic Architecture Journey with LangGraph
By
–
Hear from @Box
's CTO Ben Kus in his LangChain Interrupt talk on how Box is building agentic applications Box rebuilt their AI with agentic architecture after simple approaches failed. They’re now starting to standardize on LangGraph. Watch full video here: -
Uber’s AI Developer Agents Built with LangGraph Save 21000 Hours
By
–
How @Uber used LangGraph to build AI developer agents that generate thousands of daily code fixes and saved 21,000+ hours — serving an organization of 5,000 developers working with hundreds of millions of lines of code. Watch their full session here: https://
youtube.com/watch?v=Bugs0d
VcNI8/?utm_medium=social&utm_source=twitter&utm_campaign=q2-2025_interrupt-2025_co
… -
o3 Price Drop Now Available in Cursor IDE
By
–
The o3 price drop is now reflected in Cursor! o3 is a fantastic model, and we're excited that it can now be used as a daily driver.
-
Cursor AI and Anthropic Team Conversation on AI Tools
By
–
Listen to a conversation between the team at @cursor_ai and Anthropic:
-
Kontext API Issues Under Investigation
By
–
yeah, we're having issues with Kontext's APIs. we're on it
-
Found AI dev agent that builds real full-stack apps on AWS GCP
By
–
Found an AI dev agent that isn't just gluing together a Next.js frontend and a Supabase backend. @leapdotnew is building real apps. Full-stack, production-grade, deployed to your own cloud on AWS or GCP, with clean infrastructure as code and no vendor lock-in. Most agents out
-
GPT Researcher Integrates LangChain’s Model Context Protocol
By
–
GPT Researcher MCP GPT Researcher now leverages LangChain's Model Context Protocol adapters for intelligent tool selection and research. Seamlessly combines MCP with web search capabilities for comprehensive data gathering. Check out the integration docs!
-
Multimodal Reasoning Emerges from Online RL Training
By
–
We also found that multimodal reasoning capabilities emerge naturally after online RL (even when RL is only done on math & code text data), as long as you start from an initial multimodal model.
-
Distilled Models Gain from Online RL Training Beyond Initial Distillation
By
–
We found that a model distilled on reasoning traces from a larger model still benefits a lot from additional online RL training. In particular, we trained Magistral Small by distilling it from Magistral Medium and running additional RL.