undefined | Better HN

0 pointsadastra221y ago0 comments

Again, you need to define AGI. According to the classical definition of the term by the people who coined it, the example you gave proves that it is AGI. You were able to specify a problem using a fully generic interface (chat) that the AI was not specifically trained on, and it got an answer close to the right answer. The process it used to generate that answer is general intelligence.

0 comments

jqpabc1231y ago

Semantics aside, it is definitely lacking and leaves much to be desired in terms of "intelligence".

Either it didn't understand the question or it can't count --- take your pick. Any results come at a high cost and are simply not reliable.

adastra22OP1y ago

What ChatGPT actually sees when you input that question is the output of the tokenizer:

[5299, 1991, 18151, 553, 306, 290, 2195, 392, 491, 33465, 69029]

This happens to be 11 tokens, but I think that's a coincidence. Token 491 is "int" and token 33465 is "elligence", but ChatGPT doesn't actually see the letters.

How can you expect it to count, given those limitations? It had to guess how many letters each token represented. It got close, but not exact.

This is an artificial example pretty much maximally designed for ChatGPT to screw up.

jqpabc1231y ago

This is an artificial example pretty much maximally designed for ChatGPT to screw up.

This is an extremely simple example that ChatGPT screws up maximally.

If it can't be trusted for a simple, obvious example, why should it be trusted in less obvious cases were accuracy and reliability are important.

Answer --- it shouldn't. Maximal cost with minimal reliability.

1 more reply

alphan0n1y ago

Seems like it can count just fine: https://news.ycombinator.com/item?id=42391312

Maybe you’re using it wrong.

j / k navigate · click thread line to collapse

0 comments

jqpabc1231y ago

Semantics aside, it is definitely lacking and leaves much to be desired in terms of "intelligence".

Either it didn't understand the question or it can't count --- take your pick. Any results come at a high cost and are simply not reliable.

adastra22OP1y ago

What ChatGPT actually sees when you input that question is the output of the tokenizer:

[5299, 1991, 18151, 553, 306, 290, 2195, 392, 491, 33465, 69029]

This happens to be 11 tokens, but I think that's a coincidence. Token 491 is "int" and token 33465 is "elligence", but ChatGPT doesn't actually see the letters.

How can you expect it to count, given those limitations? It had to guess how many letters each token represented. It got close, but not exact.

This is an artificial example pretty much maximally designed for ChatGPT to screw up.

jqpabc1231y ago

This is an artificial example pretty much maximally designed for ChatGPT to screw up.

This is an extremely simple example that ChatGPT screws up maximally.

If it can't be trusted for a simple, obvious example, why should it be trusted in less obvious cases were accuracy and reliability are important.

Answer --- it shouldn't. Maximal cost with minimal reliability.

1 more reply

alphan0n1y ago

Seems like it can count just fine: https://news.ycombinator.com/item?id=42391312

Maybe you’re using it wrong.

j / k navigate · click thread line to collapse