OpenAI today released a new model, GPT-6, the model called Astra. And for sure you'll see a lot of YouTube videos, reviews and articles about this model and its capabilities. What I want to talk about here today is only one of the things which is changing, and it's the way the model is reasoning. The model is not reasoning in the form of chain of thought, the way previous versions of the models were reasoning. Chain of thought works in the following way: the model, before answering the question, creates a kind of a plan for itself, normally in English, and it goes through these steps one after another, then it performs some tasks, maybe uses different tools, and then you get an answer.
And the good thing about chain of thought is that you can actually read what the model is thinking, and then, for security experts and for different types of tasks where you need to understand where the model went wrong, at which stage, what was not working, it can be used to understand what's happening. Now, with the new way Astra is thinking, the problem is that it's no longer using chain of thought. It's using another technique called recurrent depth, and there you do not really see in English what's happening.
It works through these kinds of steps inside itself, and you only see the final result. The result of this, first of all, is that it doesn't use that many tokens, and if you are charged by token, then the consumption of tokens, tokens, to remind you, are short elements of words, normally of around four letters, goes down: the model is no longer using as many tokens as the previous models used to. At the same time, the problem for safety experts is that now it's really a black box.
We don't know what happened inside, we just have the amazing result. And you can imagine, if the result is not as good as you want, or if the model has done some strange thing which you're not expecting, then it becomes very difficult to understand where this is all coming from. So, for sure you'll see a lot of reviews about Astra, and it looks like it may even be AGI. Maybe we are even there, because out of all the tests Astra is doing, a lot of them are at the maximum or almost the maximum, like 99 percent, which means that potentially we're achieving something close to AGI, and OpenAI has actually already mentioned that they're expecting AGI to come this year.
Thanks for watching and see you in the next video.
