Jev: A Model that Substract Language from the LLM

Nowadays, so many AI models spring up like mushrooms. Every new one seems to be revolutionary (at least in the overwhelming media coverage). The “best AI model” in some scope don’t even last a half of month.

But truthfully, how many of them are really stunning? Or, just the another approach for media to make us anxious and make profits from us?

Jev’s Philosophy and Its Wisdom

Just yesterday, TypeSafe AI opened its first model named Jev to public tests. And overnight, hundreds of thousands of media pushed it to me. At first, I thought it was just a new company launched a new LLM and nothing really matter. But when I sat down to look at its official website and read its blogs, I realized it’s not that simple as I previously thought.

First of all, Jev is even not a LLM! It’s a model that can only read plain text or JSON and generate valid JSON. And the important thing is, it is extremely specialized in it. It’s 193.6 times faster than other LLMs, and it’s even 444.6 times cheaper. And the output fee for it is ZERO because of its highly structured answers.

And in an era that every company wants to achieve the multimodality of LLMs, TypeSafe AI just builds an AI model that has a very limited modality. But it’s just the most clever point.

The point that interests me most is its machine-friendliness. So far, most LLMs have been trained in a method called RLHF (Reinforcement Learning from Human Feedback), which has led to LLMs that are optimized for human preferences. This has led to models that are superhuman at instruction following, and are what we now call “chat.” Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.

But do you ever think about it: We don’t always need LLMs to chat. We need them to work. Especially work with machines and computers.

So here comes Jev, which is natively used by machines. TypeSafe AI is building it with a new architecture, a new sampler, and a new training algorithm: RLCD (Reinforcement Learning for Calibrated Decisions). Which means Jev is less dependent on humans which are often less stable and likely to make some mistakes. Owing to this innovative training method, Jev’s tool call error rate and structured output error rate are ZERO (according to official data)! Which means this model has no hallucinations. Unbelievable, right?

This reminds me of the OpenAI’s newly released flagship model–Astra for Law, which is also built for ultimate certainty. But obviously, these two models are not in the same racing track.

Meanwhile, Jev is also multithreaded, which means it can run in parallel. Therefore, it boasts such high output speed.

The Prospect of the AI Model

From my perspective, models like Jev that less depend on humans and language and have a high output certainty are what should be achieved in the future. So, I have great expectations for the WM (World Models) just like 杨立昆, a Turing Award laureate who left Meta and pursue full-time research on WM.

The AI Wave in the Eyes of the Public

I know a lot of people get anxious when there are some new AI tools but you even heard of them before. As if others will exceed you with these easily. I don’t think it’s a bad thing, because it’s a solid example for how AI is powerful. But a lot of AI tools, LLMs or something similar, are overrated by the media to catch the public attention. But obviously the most media don’t have the capabilities to test AI models like Vals AI (a platform that tests AI models professionally). What they publish is merely impressions and assumptions from people who have never used these models in real scenarios.

So actually, there’s no need at all for us to chase the wave of AI models. For most people, in most situations most of the time, Doubao is still the AI you need most.