3 ms·
We know how a hammer works. We don't command the hammer to "analyze the nails you see in front of you and find the optimal way to hammer them all in". With AI,
by typon 5y ago
We know how a hammer works. We don't command the hammer to "analyze the nails you see in front of you and find the optimal way to hammer them all in".
With AI, there's a qualitative difference. We know exactly how the AI is built, but we have no idea what the Gigabytes of FP32 numbers that make up the weight tensors actually encode. We can actually ask the AI to "find the optimal structure of a protein" without knowing how it finds the optimal. Currently, it requires a lot of massaging of data and cajoling the weight tensors to be "just right" in order to get good answers, but in the future we can imagine AI giving very good answers that are "out of our league" (in terms of understanding as to how or why) that are nonetheless optimal according to some metric we can measure.
- mattkrause 5y agoI'd argue that the prize is mostly for identifying the question and determining constitutes a satisfactory answer. In other words, it's for the result, rather than the data. The work itself can be fairly mechanized. Even when it isn't, most of it is often done by an army of techs and grad students, rather than the nominal "prize winner".
- typon 5y agoDon't disagree there. I just think it's incorrect to compare (modern) AI to hammers.
- ianhorn 5y agoWe used to not understand how steam engines worked either, even once we got them working. At least not past the most macro level. Understanding at the micro level came later.
- DharmaPolice 5y agoIf someone identified a chemical compound which cured Alzheimers they'd presumably still get awards, even if they couldn't fully explain why/how the treatment worked. I don't see the AI solutions being too different from that. Now, an AI that undertook an independent program of research - that would obviously be different.