Anyone who was affected by the FAIR layoffs, feel free to DM. We are a fortune 500 company on east coast currently growing our AI tech devision needing all sorts of expertise across AI/ML, post training, data, eval, infra and SW.
This is interesting. Assuming 10GB of the 12GB is for the model, usually the model deployed is at INT4 , so this is a 20B model. Taking LLM as an example, 20B is the size that starts to having a taste of frontier model. With AI5, Tesla can definitely push the model size to another 2X~5X.
The rest of the year is all about fixing the long tail. FSD is solved, just matters of iterations and time.
Hey all, I just open sourced this project I did in the past few weeks.
Introducing UCIP: The open-source protocol for seamless, multi-device AI experiences! Say goodbye to fragmented context—capture user details like mood, location & habits for smarter LLMs. MIT licensed. Devs, join the revolution: star, fork & contribute! 🚀
github.com/JianbangZ/UCIP
In the earnings call, Tesla team said diff parameter set for diff road conditions are needed, then the head of AI talked about MoE (Mixture of Experts) right away. So instead manual parameter tuning, Tesla is using MoE to automatically let the neural net to pick the best "expert". MoE is rather common tech in modern Large Language Model. With MoE, model can have a deisgn of predefined any # of experts, during training, by feeding into diverse dataset, the model itself figures out the "expertise" for each expert automatically. At the running time, one or multiple experts are fired to map it's strength to the road conditions it senses at real-time.
