The thing AI lacks right now in my opinion is the ability to self-systemise and derive new internal logic structures. Right now, AIs like ChatGPT only have the system of English and other written languages. It's why it's seemingly good at writing prose and code, and why it fails at mathematics and logic. Once it can alter modality depending on context, we'll be entering some amazing and scary territory.
Anyone actually using these things right now would know how wrong this is. Text to Speech, Speech to Text, Text to Image, image to image, text to text, ALL crossed the threshold of being curiosities to being impressively usable in the last year, and there are communities of people actively using them for useful and impressive things for the first time. In the last year I have all of those models in various creative workflows in my day to day life and they've all been actively more useful than flawed. A bad tech demo isn't going to change the reality on the ground, and all of this usage is very likely to result in more improvements.
Agree these detractors seem to be missing the big picture. An AI winter will never happen again because AI is so incredibly useful in so many areas now.
And yes it still fails in some, we aren’t at GAI yet, but we also likely aren’t far off
I just asked ChatGPT to translate some code form Python to C++ for me, and it did just that (there were a few minor errors, but fixable).
I then asked for some advice how to improve some Python code. Initially it just summarized what the code is doing. Funnily enough, I forgot to copy-paste one function, but ChatGPT figured it out and created a version of it (which, I have to admit, was slightly better than mine). It then help me to refactor the code by using dataclasses and type annotations.
If someone claims we are heading for an AI winter, they did not experiment enough with ChatGPT yet.
It's coding abilities are hit and miss right now. There are times when it suggests code that just doesn't work, or it makes up a library or a function that doesn't exist. When I point out the non-existent library or function it actually agrees with me, and will then continue to fail, likely because the required knowledge just isn't in there.
But these are still early days and LLM-based products will surely get much better.
I used to think Gary Marcus is a voice of reason in the AI field. Now I think he's just a moron shouting hoopla over and over again. Insecurity piling from his work not being relevant any more?
10 comments
[ 2.7 ms ] story [ 27.8 ms ] threadhttps://ai.facebook.com/blog/ai-math-theorem-proving/
https://www.infoq.com/news/2022/09/mit-math-solving-ai/
And yes it still fails in some, we aren’t at GAI yet, but we also likely aren’t far off
I just asked ChatGPT to translate some code form Python to C++ for me, and it did just that (there were a few minor errors, but fixable).
I then asked for some advice how to improve some Python code. Initially it just summarized what the code is doing. Funnily enough, I forgot to copy-paste one function, but ChatGPT figured it out and created a version of it (which, I have to admit, was slightly better than mine). It then help me to refactor the code by using dataclasses and type annotations.
If someone claims we are heading for an AI winter, they did not experiment enough with ChatGPT yet.
But these are still early days and LLM-based products will surely get much better.