That said, the results also show the limitations of working with current AI models, because, unlike a human, they can’t actually explain the process by which they are making decisions. For example, some models made very different suggestions from each other, which the researchers say means they were exploring different regions of the space of possible sequences. But we don’t really know whether this is the case, or whether each model had mathematical reasons to dislike the other’s suggestions.
This is one of several cases in the paper where researchers tried to explain what the model was doing based on its outputs. In at least one case, the software redesigned an entire structural element (an alpha helix), where the isoleucine it replaced was located, for reasons we don’t even risk speculating about.
It’s a good reminder that, at this point, these software packages are tools: they let us do things that wouldn’t be possible otherwise, but they don’t really help us understand it all. We have yet to reason through events using neural networks inside our skulls.
This is not necessarily the case; We can place more emphasis on uncovering the inner workings of this software while developing it to gain some insight into the decision making process. But right now, I think there’s an emphasis (quite rightly) on getting something that works.
An amazing achievement, but is it useful?
Overall, this is a surprising work. These proteins have to interact with each other, interact with ribosomal RNA, transfer RNA, messenger RNA, the mounting proteins made by the ribosome – as well as all the normal proteins on the large subunit. Each of them have taken billions of years to develop the ability to work with each other. The fact that we could make such radical changes to the system in the course of a few years is astonishing.
<a href