Dharmesh Kakadia

Building to give post-training superpower to everyone. @mixtrainai
Past @nuro @zoox @Microsoft @MSFTResearch

[Dharmesh Kakadia @dharmeshkakadia]

"'Three general-purpose models ought to be enough for everybody."

https://t.co/RsiFnB9N0F

[Dharmesh Kakadia @dharmeshkakadia]

@natolambert Congrats! Met recently @latkins and came away very impressed and optimistic about American Open AI! Great fit for both 👏

[Dharmesh Kakadia @dharmeshkakadia]

@GoingBallistic5 @Rewkang @standardbots yeah @evanbeard is great! R2D2 is very creative. Does that make the visit 32 BBY :P ?

[Dharmesh Kakadia @dharmeshkakadia]

@LusciousPear 💯 I think the biggest unlock is allowing rapid creation & deployment of robot models. Working on giving post training superpower to every robot @mixtrainai

[Dharmesh Kakadia @dharmeshkakadia]

Why do you need your own model?

"We’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design"

[Claude @claudeai]

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Its capabilities exceed those of any model we’ve ever made generally available.

[Dharmesh Kakadia @dharmeshkakadia]

@evanbeard Congrats Evan and the entire @standardbots team!

[Dharmesh Kakadia @dharmeshkakadia]

@xxunhuang Congrats Xun and the entire team!

[Dharmesh Kakadia @dharmeshkakadia]

For some reason, lately I have to remind people a lot about https://t.co/W0baTFo75g

[Dharmesh Kakadia @dharmeshkakadia]

@dwarkesh_sp Anything to do with time estimation

[Dharmesh Kakadia @dharmeshkakadia]

@dwarkesh_sp Writing.

[Dharmesh Kakadia @dharmeshkakadia]

@rronak_ @MichaelElabd @QuantumArjun Congrats on the launch!

[Dharmesh Kakadia @dharmeshkakadia]

@_arohan_ I am gonna steal the Big token phrase.

[Dharmesh Kakadia @dharmeshkakadia]

@bernhardsson I think it makes sense if they consider owning the entire stack is worth as muscle in the long run (and not just about bps). They have done things in programming language, compiler etc for the same reason I believe.

[Dharmesh Kakadia @dharmeshkakadia]

@lukas_m_ziegler Post training platform for robotics, so more and more robotics companies can ship task specific models for every use case x embodiment, while still leveraging the power & economics of generalized learning in large VLA/WAM/VAM.

[Dharmesh Kakadia @dharmeshkakadia]

@_advaitpatel @Mind_Robotics 🚀

[Dharmesh Kakadia @dharmeshkakadia]

@eatpraydiehard @madiator This is largely because post training *platforms* aren’t wide spread yet. With right platform partners, post training is not lot more complicated then harnessing around a single model.

[Dharmesh Kakadia @dharmeshkakadia]

@mixtrainai You can't rely on any labs post training offering, because it's not their core focus. You need platform that is independent/agnostic of a model, and brings you frontier grade infra to your problem domains (custom kernels, data curation, eval infra,..)

[Dharmesh Kakadia @dharmeshkakadia]

@alexstauffer_ Hey Alex, would love to chat about post training stuff :)

[Dharmesh Kakadia @dharmeshkakadia]

@mark_k @OpenAI Your post training strategy should be model provider agnostic. Partner with @mixtrainai to get frontier grade platform for your custom models
https://t.co/EavaavgaTY

[Dharmesh Kakadia @dharmeshkakadia]

@mixtrainai You can't rely on any labs post training offering, because it's not their core focus. You need platform that is independent/agnostic of a model, and brings you frontier grade infra to your problem domains (custom kernels, data curation, eval infra,..)

[Dharmesh Kakadia @dharmeshkakadia]

@mixtrainai You can't rely on any labs post training offering, because it's not their core focus. You need platform that is independent/agnostic of a model, and brings you frontier grade infra to your problem domains (custom kernels, data curation, eval infra,..)

[Dharmesh Kakadia @dharmeshkakadia]

If you were looking for a platform (and company) dedicated to post training, we at @mixtrainai are fully focused on helping you build the best model for your task+constraint+evaluation...

[Mark Kretschmann @mark_k]

OpenAI has announced they will be winding down fine tuning. I got the email today. Existing active @OpenAI customers can keep running fine-tuning jobs until January 6, 2027, but after that no new training jobs can be created. Existing fine-tuned models will still run, but only until the underlying base model is eventually deprecated.

I get the argument that newer models follow instructions much better, and that prompts plus RAG cover more use cases than before. But not all of them.

[Dharmesh Kakadia @dharmeshkakadia]

@willbitsky @oyhsu @JacobZietek 3. Task specific models optimized both for inference compute and latency
4. Efficient training env + kernels when co training video/world models with diffusion heads.
5. Data infra for curation and evaluation for physical AI

[Dharmesh Kakadia @dharmeshkakadia]

@willbitsky @oyhsu @JacobZietek Agree with all 10, and it's great to see more than 5 are bottlenecked on "engineering/deployment" and not on research any more. Few more to add:
1. Latent vs pixel space prediction
2. Long horizon context management between autoregressive backbone and diffusion heads