[runarberg]

I’m not an expert on AI but I was under the impression that AI is just a fancy way to fit data with massive computative power. So if you wanted AI to do science for us it would basically just return the result we expected in a fancy and unexpected way, and even then we might be overfitting. If an AI would return new science we didn’t expect—I don’t see how but lets imagine—we would reject it as a miss.

Take the Event Horizon Telescope as an example here. It requires a massive amount of computational modeling to create those images from a huge dataset. Humans did their science to figure out how the supermassive black holes would look like and created their models (AI if you will) based on this knowledge. Even so they are not free from criticism that they might be overfitting onto a conformation bias and many scientists are waiting for an independent confirmatory observation before they believe the image. Now if you replace the humans in that equation with more models (AGI if you will) how do you even know what you are looking at is even sensical?

I’m not a fan of the LHC++ project either, but I think you might have too much faith in AI and AGI.