@YingjunWu I have many thoughts on this. The shortest version is that I don’t think post-training by itself would sell; it’s just differentiation you can use to layer onto your inference stack
Congrats @devvret_rishi! I asked you to join my group project in college because I knew you were brilliant, and here you are 10 years later launching big things at @predibase 🤓 🔥 🚀. Can I say I spotted your talent early?
Google's missing a massive opportunity: Make Gemma fine-tuning in AI Studio as seamless as @predibase. Imagine thousands of specialized models with optimized inference costs. With their infrastructure, they should be dominating this space already.
Starting today, you can build your own custom AI models without collecting labeled data – with the first end-to-end reinforcement fine-tuning platform.
DeepSeek-R1 showed the world how RL can solve challenging reasoning tasks, and now we’ve baked these capabilities into an intuitive platform so anyone can leverage self-improving models for their use cases. RFT guides a model with reward functions and unlabeled data in a new interactive experience that’s a game-changer for tasks like code generation, complex RAG and more.
Our early results with the platform have blown us away. We used it ourselves to build a 32B-param model to write CuDA code 3x more accurately and faster than OpenAI’s o1 or DeepSeek-R1, that we’re open sourcing today. And we’ve already seeing great traction with early customers.
I’m very excited to launch the first version of this experience and give you an early look at what we’re building. Check out how it works in the thread below and let’s shape the future of custom AI together.
With Snowflake and @predibase you can now train small task-specific LLMs that outperform commercial models at a fraction of the cost—all within your Data Cloud.
By combining Predibase’s world-class LLM fine-tuning and inference capabilities with the Snowflake Data Cloud, teams can train their own LLMs directly on their data and deploy using Snowflake’s GPU architecture.
Learn more in this webinar and demo:
https://t.co/hSk6fa0Tze
I like the standard LLM fine-tuning tools. So I wan't sure I'd like Predibase
Then I used it. Their UI is really nice... especially the data visibility part.
I'm going to use this more and more
So I'm pumped that @predibase just offered free compute credits to all participants in the LLM course/conference/woodstock at https://t.co/GxfOjesmGs
Off the shelf is good, fine-tuned is better. Here's everything you need to #efficiently fine-tune @Google's #Gemma7B for your task. 🚀
Best of all, you can do it with @ludwig_ai, the open-source declarative framework for building custom #LLMs.🔥
https://t.co/nN5BM2dTbA
@ItakGol@theallinpod@Jason would love to hear an update on the future of smaller fine-tuned open source models (vs single large walled gardens) in the enterprise on the next pod
@chamath@Hustle@socialcapital@chamath people find fine-tuning challenging, but it doesn't have to be. We made a platform for engineers to fine-tune LLMs, and you can try it out for free (A100s included!) at https://t.co/2Jg79UY5IU.
Our underlying tech is open-source at https://t.co/STvAxaSKAM (10k stars).