Robert Ta and Jonathan compare TypeSafe's Jev System 1 model to OpenAI's 5.6-trillion-parameter model. Jev skips text generation entirely, outputting structured data directly, which cuts costs and latency by orders of magnitude. It's built for parallel decisions like identifying which business units are affected by a change.
Jonathan, what are your thoughts on the example here on the left of using TypeSafe Jev System 1 model architecture versus a typical LLM frontier model where you're trying to ask questions like, is revenue currently impacted? What business is impacted? Integration issues present? These types of things that, you know, would drive decision-making.
Yeah, I mean, I think the output says it all. 5.6 tera is the cheapest and fastest model from OpenAI right now. And you compare it to the TypeSafe model, TypeSafe is getting the advantage of not having to render text through its processing and specifically trained to like, can ask a whole bunch of questions in parallel. Terra is not really optimized for that use case. So it's orders of magnitude in cost and speed. And that's the whole fun of this, this JEV approach, right?
It makes sense. We've had this problem of like, okay, we throw in the system prompt and throw in evals with the dataset to say, you need to have a contract for how you output stuff when the input's like this. For a lot of our AI production use cases, that has been the meta for everybody. And so this is a complete game changer where structure comes with it. You don't have to parse through stuff and then use a bunch of server processing to do that. The experience on the user side, because of the performance gains, are a lot better, not to mention the more efficacy that you get with Jev's specific fine-tuning around the decision-making probabilistic confidence scoring use case.
