I actually think the surprise is warranted. The model didn't just "get something wrong." It produced specific, confidently stated, errors that look like factual knowledge but are actually side-effects of pattern recognition. The whole point of the article is that this type of error is subtle enough that a non-expert might take it as truth (and to point out just how silly some of those confidently stated answers are for a laugh). Understanding the patterns and why the model behaves like this is useful so it can be improved. I don't think you can discredit the entire technology because it's not 100% perfect.
Understanding the patterns and why the model behaves like this is useful so it can be improved.
The answer to why is simple and obvious --- the model is probabilistic. Improving it substantially would necessitate a total redesign.
I don't think you can discredit the entire technology because it's not 100% perfect.
So dice rolling as a decision making tool shouldn't be discredited in your opinion?
Would you fly on a plane designed using similar "technology"? Anything better than double 3s says it's safe enough.
I expect the legal system to ultimately have a lot of input on the use of this. I don't expect "But AI said it was OK" is going to fly in court once people have been harmed. And once this starts happening, the enthusiasm for this "new technology" will start to wane ---after $trillions have been wasted.
I totally agree. Even though the author states over and over that an LLM is just a pattern-matching machine, he still anthropomorphizes the responses. To say that an LLM "makes mistakes" is granting it a consciousness it doesn't possess.
Why does anyone still find this surprising?
It's disappointing that we (as a society) are investing such huge sums of money in this "amazing" new technology.
I actually think the surprise is warranted. The model didn't just "get something wrong." It produced specific, confidently stated, errors that look like factual knowledge but are actually side-effects of pattern recognition. The whole point of the article is that this type of error is subtle enough that a non-expert might take it as truth (and to point out just how silly some of those confidently stated answers are for a laugh). Understanding the patterns and why the model behaves like this is useful so it can be improved. I don't think you can discredit the entire technology because it's not 100% perfect.
Understanding the patterns and why the model behaves like this is useful so it can be improved.
The answer to why is simple and obvious --- the model is probabilistic. Improving it substantially would necessitate a total redesign.
I don't think you can discredit the entire technology because it's not 100% perfect.
So dice rolling as a decision making tool shouldn't be discredited in your opinion?
Would you fly on a plane designed using similar "technology"? Anything better than double 3s says it's safe enough.
I expect the legal system to ultimately have a lot of input on the use of this. I don't expect "But AI said it was OK" is going to fly in court once people have been harmed. And once this starts happening, the enthusiasm for this "new technology" will start to wane ---after $trillions have been wasted.
I totally agree. Even though the author states over and over that an LLM is just a pattern-matching machine, he still anthropomorphizes the responses. To say that an LLM "makes mistakes" is granting it a consciousness it doesn't possess.