This model convergence is quite perplexing. Possibly related to recent results on subliminal learning? Basically deeper knowledge correlations transfer when training via distillation. As the amount of data online from LLMs increases, it’s possible this makes them converge to some
@karpathy
-
LLMs Beyond Chatbots: Rethinking AI Model Perception
By
–
Like! My personal top tweets are those that most people find confusing/wrong first and then obvious/self-evident a few years later. LLMs were at one point seen (and still are) as chatbot, like a better Eliza or so, hence the chat bubbles, which lacks imagination by about 1000X.
-
Using Cursor and Claude for AI-Assisted Development Workflow
By
–
I use CC from Cursor and I assumed most do as well (?).
I end up with a mixed thing where Cursor is the UI layer for reading the code, manual edits, tab completion and chunk edits, and CC for larger changes, architecting, Q&A. Still rapidly evolving though… -
Better Model for Analysis Claude Haiku Update
By
–
"Using a better model for analysis" I didn't realize I was using haiku all this time, no idea when claude code snuck this one in rofl.
-
Real-Time Video Diffusion Model MirageLSD Transforms Video Generation
By
–
Diffusion video models but now – **realtime**!
— Andrej Karpathy (@karpathy) 17 juillet 2025
Simple video filters are real-time but can only do basic re-coloring and styles. Video diffusion models (Veo and friends) are magic, but they take many seconds/minutes to generate. MirageLSD is real-time magic. Unlike simple video… https://t.co/34nBZ7Q1zRDiffusion video models but now – **realtime**! Simple video filters are real-time but can only do basic re-coloring and styles. diffusion models (Veo and friends) are magic, but they take many seconds/minutes to generate. MirageLSD is real-time magic. Unlike simple video
-
Deep case analysis outperforms aggregate statistics dashboards
By
–
I always learn a lot more from in-depth analysis of few random cases over dashboards of aggregate statistics across all cases. Both projections can be helpful but the latter is disproportionately pervasive.
-
Reinforcement Learning Field of View Limitations and Policy Optimization
By
–
Yep I think RL is misleading in that it restricts field of view. Eg like you mentioned you can imagine review/reflect doing a lot more – building tools for later use, or actively tuning the distribution for what to try next (instead of just sampling from policy independently as
-
Scaling Reinforcement Learning: Intermediate Gains and Future Limitations
By
–
Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue to yield more intermediate gains, but I also don't expect it to be the full story. RL is basically "hey this happened to go well (/poorly), let me slightly
-
Vision VAE as Ultimate Tokenizer: Unicode Efficiency
By
–
Very cool work direction but also fair question.
I wonder if ultimately is a little vision patch VAE the ultimate "tokenizer"? Unicode + UTF-8 is just too high description length.