OpenAI trained a powerful model to write solutions that a weaker model can verify and understand. The technique keeps performance high while making answers legible to humans, tested on math problems. It doesn't yet work well for images.
So here is a new paper from the OpenAI lab on creating even smarter AIs than we previously had. But surprisingly, this paper also shows that smarts are not everything. Now wait, wait, wait. How do we know that they are really smart? Well, we give them a bunch of difficult questions and see if it can answer them correctly. Then we record these answers and compare many AI techniques and make a leaderboard with the best ones. This is from one of our earlier papers called Gaussian Material Synthesis, where we taught an AI to synthesize virtual materials. And it is one of the early works in neural rendering, ray tracing with an AI, if you will. Now, if we look into the paper and, you know, ask for a little help, we get this. Now I have two questions for you fellow scholars. One, is this explanation correct? Well, as far as I see, yes, it is. Now question number two. Is it useful? Well, not quite. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. And that is what this new OpenAI paper is about. These AIs are trained for correctness. Of course they are. They compete on correctness. But this also leads to situations where its solutions are correct. But they are not always that easy to understand. So let's try to train these models to be a bit more understandable too. And the results are pretty amazing. Let's see what it has done. We have this mathematical question as an input. And here is what the AI. Correct? Yes. Useful? Nope. Okay, now our little AI should undergo a legibility training. After round one, we get an actual explanation of the solution. Fantastic. And after round five, we even get a solution that is a tiny bit clearer and a lot better structured. But how? How can we tell an AI to be more understandable? How do we teach it? Well, they propose that we play a little game. Okay, okay. Not that game. This game requires not hiders and seekers. This game requires Einstein and a kid. Yes, you heard it right. The prover model will be Einstein. This is the one that can solve difficult problems. And however, the goal is to hand it over to the verifier, the kid, if you will. So why do all that? Well, the goal is if we have a problem that is difficult to solve, even for Einstein, but when it solved it, the solution should be easy to verify, even for a kid. So Einstein solves, kid verifies. And hold on to your papers, fellow scholars, because this kid can be a up to a thousand times less capable than the big model. And the game still works. That is insane. Loving it. So if it produces solutions that the kid can follow. That's great. But wait. What if Einstein is wrong and just hands over the wrong solution to the kid? How would the kid know? Or in other words, how do we prepare the kids against lies? Well, of course we lie to them. We want kids to be able to see through these lies. And the fantastic thing is that they can actually do that. Glorious. And now hold on to your papers, fellow scholars. Yes again, because here is the main result and it is stunning. Let me try to explain it. First, it says if we generate an initial model, answers might be easy to understand. However, and here comes the trouble. If we improve this model to be smarter, it gets smarter. Fantastic. But the price for being so smart is being less understandable. Huge performance comes at the cost of poor understandability. And now with the new technique, we can make a model smarter without paying this huge legibility tax on it. Absolutely fantastic work that is going to have a meaningful difference in all of our lives. What a time to be alive. And they could have decided to keep this to themselves. But the full research paper with all the technical details to do this is available to all of us for free. So I would like to send a huge thank you for that now. Note that the technique still has its limitations. It works well in the domain of language and and within that especially mathematics, which is highly formalized. However, it is unclear how this could work on images. So what do you think? What would you fellow scanners use this for? Let me know in the comments below. We need new tools for the era of LLMs and weights and biases. Now has weave a lightweight toolkit to confidently iterate on LLM applications using use traces to debug how data flows through each step of your app and use evaluations to measure your progress. It is the best. Try it out now at WNB Me papersllm or click the link in the description below.
