3 ms·
This is one of those conversations, as an AI researcher, that I tend to not like. There's two major voices and both are ridiculous. Neither should we completely
by godelski 4y ago
This is one of those conversations, as an AI researcher, that I tend to not like. There's two major voices and both are ridiculous. Neither should we completely rely on these tools and neither should we throw them out. It is a cat and mouse game, and yes, the mouse will always win. We're on hacker news, and everyone knows this same game exists for hackers and the advantage bias is similar. Doesn't mean we shouldn't stop trying to defend against the hackers. But ML requires a lot of nuance. We've already been using automated tools to ban people and we all know how frustrating that is. There have been several subreddits that would ban anyone that used the term "vaccines" because they wanted to protect their subreddits from anti-vaxxers but these backfire and also autoban people defending vaccines.
The problem isn't using these systems, the problem is that there is no nuance. No way to reach a human who can see the absurdity. But this has been a progression and even bureaucratic institutions do this with real humans. This is a frustration we all experience and we shouldn't make this worse. Use the tools, but always maintain a way to reach a human who can account for system failures.
An important issue with ML system failures is that when they fail, they fail bad. Anyone who has used GPT should be well aware of this and have experienced it (if you haven't, you haven't used it much, are incredibly lucky, or have been duped). It is also the reason AlphaGo lost that one game, and it played REALLY poorly. These higher dimensional spaces are rather complex and don't fit your intuitions. You'll have a better idea of how to stroll through flat land than you will through 10D land, let alone 2562563 (an image) land. The curse of dimensionality is that any intuition you have of geometrical meaning goes out the window.