Re:Can GPT-4 *Actually* Write Code?
by nwienert on Friday March 17, 2023 @12:13PM
(#35199047)
Even giving this point, GPT has two attributes I haven’t seen improve with scaling much at all:
1. The whole bullshitting thing. If I bullshitted even 1% as much as GPT I’d be fired within a few months.
2. Precision. Ask it to count, or do simple math. It seems to really struggle here, and despite us feeding it nearly all of humanity’s text, it isn’t getting better. Programming is extremely precise in a way painting isn’t. GPT is a painter, not a programmer.
Final note:
9 years ago I worked on an ML startup with a few people, some of which went on and now are at OpenAI. We saw the absolutely incredible progress ML was making and had many discussion with the brightest in the field.
I was much like you, I was convinced that ML would he automating basically everything within 5 years. I mean the rate of progress felt super-exponential and there was hard proof of amazing results. Especially things like radiology for example which seemed to almost already be there.
Turned out to be dramatically wrong. My 5 year predications were laughable. It did well on some things we didn’t expect (fuzzy stuff like language and drawing) but much worse in the precise areas. It still is struggling with precision, and radiology in the most straightforward cases.
I think my calibration on this is much better now. I realize it has limits, where they are defined, and I see now it doesn’t scale in some ways. It doesn’t produce novelties in the way humans do while also being a lot less conservative/precise. It may turn out that solving both of those things at once is the extremely difficult last 10% that is actually 90% of the work (and they feel like polarities that may train against each other - be very precise, also don’t lie, but still be creative - have you noticed the more you push GPT the more “distracted” it gets? I see limits in many ways as it’s scaled that don’t look good for this).
Look at Self Driving Cars. 7 years ago I predicted they’d be here within a year or two. Had to be. They still aren’t! Why? The last 10% is really, really hard. It’s also critical to success. We really seem to have trouble reconciling these two, both underestimating the difficulty, and the need.
Stuff like your “a year or two” note is a big red flag for me here, especially when the claim is so broad. Yes, it will integrate with things in a way that avoids a lot of complexity developers deal with, but needless to say I’d love to see a concrete bet for “in two years it will have (insert exact large effect)” so I can profit off the hype.
1. The whole bullshitting thing. If I bullshitted even 1% as much as GPT I’d be fired within a few months.
2. Precision. Ask it to count, or do simple math. It seems to really struggle here, and despite us feeding it nearly all of humanity’s text, it isn’t getting better. Programming is extremely precise in a way painting isn’t. GPT is a painter, not a programmer.
Final note:
9 years ago I worked on an ML startup with a few people, some of which went on and now are at OpenAI. We saw the absolutely incredible progress ML was making and had many discussion with the brightest in the field.
I was much like you, I was convinced that ML would he automating basically everything within 5 years. I mean the rate of progress felt super-exponential and there was hard proof of amazing results. Especially things like radiology for example which seemed to almost already be there.
Turned out to be dramatically wrong. My 5 year predications were laughable. It did well on some things we didn’t expect (fuzzy stuff like language and drawing) but much worse in the precise areas. It still is struggling with precision, and radiology in the most straightforward cases.
I think my calibration on this is much better now. I realize it has limits, where they are defined, and I see now it doesn’t scale in some ways. It doesn’t produce novelties in the way humans do while also being a lot less conservative/precise. It may turn out that solving both of those things at once is the extremely difficult last 10% that is actually 90% of the work (and they feel like polarities that may train against each other - be very precise, also don’t lie, but still be creative - have you noticed the more you push GPT the more “distracted” it gets? I see limits in many ways as it’s scaled that don’t look good for this).
Look at Self Driving Cars. 7 years ago I predicted they’d be here within a year or two. Had to be. They still aren’t! Why? The last 10% is really, really hard. It’s also critical to success. We really seem to have trouble reconciling these two, both underestimating the difficulty, and the need.
Stuff like your “a year or two” note is a big red flag for me here, especially when the claim is so broad. Yes, it will integrate with things in a way that avoids a lot of complexity developers deal with, but needless to say I’d love to see a concrete bet for “in two years it will have (insert exact large effect)” so I can profit off the hype.