3 ms·

On the other hand there's the consensus bias where you believe that everyone else is like you, when in fact they are not (which is why I said this is something

On the other hand there's the consensus bias where you believe that everyone else is like you, when in fact they are not (which is why I said this is something I've had to learn to accept). Sure, for most distributions, most people will be statistically normal, by definition. On the other hand, I know that e.g. state standard test scores when I was a kid said I was 99th percentile academically (and there was some kind of IQ test for the gifted program? I remember I qualified in all 3 areas they tested me), so I have some evidence to hypothesize that I might be an outlier along other psychometrics too. And I'll go ahead and guess that many people here are also not the median in many ways. I don't think that this makes me somehow immune to lung cancer from smoking, but I might suspect that when I read about others' use of AI and how it rots their minds, they're perhaps not using it to learn Lean (the process of learning has mostly been the AI spits out crap proofs that seem way more complicated than they ought to be, and then I spend some time trying to find ways to simplify it. We have a back-and-forth about different encodings, or ideas for syntax macros, etc. Then I can compare with e.g. Mathlib's solution) or checking their understanding as they work through some notes. I don't know what most people use it for or what they get addicted to, but I believe it might be different from what I do. With regards to ads, I live somewhere where outdoor advertising is banned, I block ads on my computer, I don't watch TV or movies or "news," and I don't install whatever apps, so there aren't a whole lot of channels for ads to reach me. I suspect again that this is not the typical experience. That said I do view ads as fraud-adjacent and damaging to markets, so fully support strongly regulating or outright banning paid advertising.

runarberg 8d ago
This comes across as an athlete excusing their smoking behavior because their superior physical abilities makes them immune to the bad effects of smoking. I assure you a 99th percentile in 100 m dash may have believed that smoking did not affect them the same way it does an average person, but they would be wrong. And so are you about your AI usage.
- ndriscoll 8d ago
  You keep assuming that LLM usage is automatically harmful with literally zero reason to believe that, and several anecdotes from people on how they use it positively. There's simply no reason to even draw a weak analogy to smoking. You'd do much better to draw an analogy to e.g. methylphenidate or amphetamine, which certainly you can abuse and get addicted to, and might even be harmful for the average person (even carrying the possibility of psychosis!), but which carry cognitive benefits with minimal side effects for others. I'll give another concrete use: the other day I scanned a storybook with my phone and had codex OCR it and compare with a list I've started of words and sound patterns my kid knows how to read so that it could identify irregular words or new sounds for us to go over before we start the story. Please, concretely, let me know what the adverse effects were there.
  - runarberg 8d ago
    I keep saying it because all the evidence (while still not conclusive) is pointing in that direction, just like they were with tobacco in the 1950s (It took another two decades for the evidence of tobacco to become conclusive; and two decades still for regulators to catch up). We do know that LLMs can cause psychosis, we do know it can cause people to behave in an unhealthy manner, etc. We are seeing more and more evidence that AI usages in education is at least correlated with bad grades. We are seeing people exhibit addictive behavior around AI. There is plenty of reasons to suspect that as we gather more evidence that they will become conclusive about the addictiveness and the unhealthyness of these products. For me this is obvious.
    - ndriscoll 8d ago
      Evidence doesn't point in that direction though, even if we establish correlations. Like the education effect is pretty obvious: most people will use it to cheat on their homework and avoid learning. Some people will use it to learn new topics or with more depth, and give them additional targeted probing questions to answer (basically, give them more homework). Average achievement will probably drop. The high achievers will probably do even better. Boiling it down to "it harms your brain" when people are already telling you how to use it to learn more effectively is just silly. Most people are happy to rot their brains. They'll take any opportunity to do it. If you're an outlier, you're probably aware of that fact on some level, and it doesn't make sense for you to make decisions looking at the median who behave completely differently. You have to look at the mechanisms behind the data and interpret under what situations they could plausibly apply to you.

runarberg 8d ago

This comes across as an athlete excusing their smoking behavior because their superior physical abilities makes them immune to the bad effects of smoking. I assure you a 99th percentile in 100 m dash may have believed that smoking did not affect them the same way it does an average person, but they would be wrong. And so are you about your AI usage.

ndriscoll 8d ago
You keep assuming that LLM usage is automatically harmful with literally zero reason to believe that, and several anecdotes from people on how they use it positively. There's simply no reason to even draw a weak analogy to smoking. You'd do much better to draw an analogy to e.g. methylphenidate or amphetamine, which certainly you can abuse and get addicted to, and might even be harmful for the average person (even carrying the possibility of psychosis!), but which carry cognitive benefits with minimal side effects for others. I'll give another concrete use: the other day I scanned a storybook with my phone and had codex OCR it and compare with a list I've started of words and sound patterns my kid knows how to read so that it could identify irregular words or new sounds for us to go over before we start the story. Please, concretely, let me know what the adverse effects were there.
- runarberg 8d ago
  I keep saying it because all the evidence (while still not conclusive) is pointing in that direction, just like they were with tobacco in the 1950s (It took another two decades for the evidence of tobacco to become conclusive; and two decades still for regulators to catch up). We do know that LLMs can cause psychosis, we do know it can cause people to behave in an unhealthy manner, etc. We are seeing more and more evidence that AI usages in education is at least correlated with bad grades. We are seeing people exhibit addictive behavior around AI. There is plenty of reasons to suspect that as we gather more evidence that they will become conclusive about the addictiveness and the unhealthyness of these products. For me this is obvious.
  - ndriscoll 8d ago
    Evidence doesn't point in that direction though, even if we establish correlations. Like the education effect is pretty obvious: most people will use it to cheat on their homework and avoid learning. Some people will use it to learn new topics or with more depth, and give them additional targeted probing questions to answer (basically, give them more homework). Average achievement will probably drop. The high achievers will probably do even better. Boiling it down to "it harms your brain" when people are already telling you how to use it to learn more effectively is just silly. Most people are happy to rot their brains. They'll take any opportunity to do it. If you're an outlier, you're probably aware of that fact on some level, and it doesn't make sense for you to make decisions looking at the median who behave completely differently. You have to look at the mechanisms behind the data and interpret under what situations they could plausibly apply to you.