Training AI Models to reason is currently very expensive. You require a lot of data, tons of compute in Reinforcement Learning, and more. And the reasoning infrastructure is not reusable.
On top of all this, we don't really have a way to improve targeted performance and personalize intelligence within the systems.
Over the last year, I've been looking into latent space reasoning as a way to solve this. By doing this, we can create a different reasoning layer that is reusable across models. We created a simple layer for 50 cents, and it already improves performance. We're working with a few people across major AI Labs at exploring this, but I also wanted to open source because intelligence deserves to be open. To that end, our startup has even opened up a small monthly prize pool for top contributors to the repo.
Would love to have you in there. Here is a report we did breaking down the core design philosophy here-- https://www.artificialintelligencemadesimple.com/p/how-to-te...