Today was my last day at @jetbrains. It’s very bitter sweet for me because I had a very rare chance to build the llm team (which became Mellum) from ground up. Probably will never have such an experience again. When I joined, except from local completion, there were practically nobody who knew how to train models. We needed to learn everything ourselves from first principles.
The first big pre-training run was launched by essentially three people: @ivanbondyrev, Marko, and me. Now the team is still super small but one of the few who do the whole model cycle from pre-training data and arch to post-training and evals. They were the locomotive of the company transformation from making good code editors to being a big player in the AI era.
Wishing them big luck! Hope they’ll surprise everyone with a new release in a good way.
Now gonna enjoy my two days of being unemployed. Stay tuned!
P.S. a photo of me pushing a start button of Mellum2 pre-training in an Uzbek restaurant. Marko’s babysitting load-balancing loss.
There is probably nothing wrong with the soverign ai positioning and a lab post-training Kimi. Both have their own place: see Mistral and Cursor Composer. But it's kinda either this or that?