5 ms·

I don't know that his argument is based on essentialism since he goes on to describe algorithms which he considers artificially intelligent: "There are programs

I don't know that his argument is based on essentialism since he goes on to describe algorithms which he considers artificially intelligent: "There are programs that can look at a photo of some magnified cells and tell you, with greater likelihood of being right than any human doctor, whether it’s cancerous or not,”. And he puts it in contrast to his previous statement about LLMs not being artifically intelligent. It seems to me he wants people to look past the impressive facade of correct grammar and see that a lot of the LLMs often spew out bullshit. Personally I've seen them perform intelligence tasks, generate correct code and so forth. But I think many people ascribe to them a general intelligence that they don't have, which is why you see them quoted increasingly often online as if they were knowledgeable or correct-by-default.

vidarh 3y ago
I think he's seriously contradicting himself there in that the algorithms he argue are AI certainly do not meet his stated standard, and it seems like he's drawing a line where he refuses to use the term about anything that might be seen as intelligent that doesn't meet his standard for intelligence, but is fine with using it for techniques where there's no chance of confusion. I think if he wasn't making an essentialist argument about LLMs he worded himself exceedingly poorly. Arguing against the danger of blindly believing LLMs is good, but at the same time he's destroying his own credibility on the subject by exaggerating how bad they are vs. focusing on the dangers of blind trust. This is especially notable when he then goes on to point to measurable success as a metric for other AI tech - he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust.
- simiones 3y ago
  > This is especially notable when he then goes on to point to measurable success as a metric for other AI tech - he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust. The examples he gives are technologies where you can point the AI to a problem and trust the results will work as intended. You don't need to check every individual output of those AIs, you can just trust that it will be better than the output you'd get from a human expert (even though it's of course not 100%, and a combination of AI + human expert may be better). What is the equivalent for LLMs?
  - vidarh 3y ago
    > The examples he gives are technologies where you can point the AI to a problem and trust the results will work as intended. No, they're technologies where AI's have been tuned and tested and validated for there domain so that we after that process know that they will work as intended sufficiently often to measurably be a net benefit, just as you yourself go on to describe. They'll still make mistakes, just like you yourself point out. But they've been tested to ensure they make few enough mistakes to be worthwhile. > What is the equivalent for LLMs? The equivalent for LLMs is exactly the same process. To quote myself: > he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust. In other words: Test them on your use case, and validate their performance on that use case. Don't assume. We wouldn't take the performance of any domain specific model on trust, and there's no more reason to take the performance of an LLM on trust for a domain we've not tested it thoroughly for, and quite possibly fine tuned it for. The difference is that LLMs do well enough to convince some people of the idea they can skip the testing and validation step, and that point - that people are prone to give them a level of trust that they should not be given - is valid. But extending that to dismissing them as "bullshit generators" as he did is equally ridiculous.

vidarh 3y ago

I think he's seriously contradicting himself there in that the algorithms he argue are AI certainly do not meet his stated standard, and it seems like he's drawing a line where he refuses to use the term about anything that might be seen as intelligent that doesn't meet his standard for intelligence, but is fine with using it for techniques where there's no chance of confusion. I think if he wasn't making an essentialist argument about LLMs he worded himself exceedingly poorly. Arguing against the danger of blindly believing LLMs is good, but at the same time he's destroying his own credibility on the subject by exaggerating how bad they are vs. focusing on the dangers of blind trust. This is especially notable when he then goes on to point to measurable success as a metric for other AI tech - he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust.

simiones 3y ago
> This is especially notable when he then goes on to point to measurable success as a metric for other AI tech - he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust. The examples he gives are technologies where you can point the AI to a problem and trust the results will work as intended. You don't need to check every individual output of those AIs, you can just trust that it will be better than the output you'd get from a human expert (even though it's of course not 100%, and a combination of AI + human expert may be better). What is the equivalent for LLMs?
- vidarh 3y ago
  > The examples he gives are technologies where you can point the AI to a problem and trust the results will work as intended. No, they're technologies where AI's have been tuned and tested and validated for there domain so that we after that process know that they will work as intended sufficiently often to measurably be a net benefit, just as you yourself go on to describe. They'll still make mistakes, just like you yourself point out. But they've been tested to ensure they make few enough mistakes to be worthwhile. > What is the equivalent for LLMs? The equivalent for LLMs is exactly the same process. To quote myself: > he'd have made a far better point if he argued that LLMs also need to be evaluated carefully on a case by case basis like these other systems, rather than be taken on trust. In other words: Test them on your use case, and validate their performance on that use case. Don't assume. We wouldn't take the performance of any domain specific model on trust, and there's no more reason to take the performance of an LLM on trust for a domain we've not tested it thoroughly for, and quite possibly fine tuned it for. The difference is that LLMs do well enough to convince some people of the idea they can skip the testing and validation step, and that point - that people are prone to give them a level of trust that they should not be given - is valid. But extending that to dismissing them as "bullshit generators" as he did is equally ridiculous.