3 ms·

As a counter-point, I paid for it for three months immediately after its launch, had a lot of fun, and then cancelled my subscription. Yes, GPT4 is very notice

As a counter-point, I paid for it for three months immediately after its launch, had a lot of fun, and then cancelled my subscription. Yes, GPT4 is very noticeably smarter than GPT 3.5 or any of the competitors like LLAMA 2. I rank almost all LLMs at the same level as a high school student, and GPT4 at the level of a graduate student just starting out as a junior employee. However, I don't need a "junior" assistant, I need more of "me", which GPT4 can't currently emulate. If I were to hire a flesh & blood human junior tech, that would be with the expectation that the junior would learn with time and pick up the specific techniques and approaches I prefer to use. Currently, GPT models cost an absurd amount of time, effort, and money to customise, and there's little chance that this would result in the equivalent of a trained employee. These days I don't write much code, and the code that I write I can bang out faster than I can explain what I want to ChatGPT. I've also found that the act of coding is a part of the learning process. I can't really understand something until I've taken it apart myself and put it back together again. I keep telling my customers: I can teach it to you, but if I learn it on your behalf, you'll end up knowing nothing yourself.

JaDogg 3y ago
Yes I have this experience with GPT 3.5, where my code is superior when it actually matters. But I'm not sure if GPT 4 will help me with that. There is a waitlist to pay now. I recently tried to use it write a .scad parser. I ended up writing it myself as code it generated was useless. (This was even in Python and I tried rephrasing and giving it multiple examples) It also loves to import libraries. Which I do not really like I prefer avoid dependencies as much as I can.
zh3 3y ago
Personally I'm on the wait list, mainly because it cracked a problem in seconds that one hour plus on stackoverflow hadn't solved. I didn't go to ChatGPT first because (erroneously) I figured the issue was too esoteric for it to crack. Specifically, following an update to Samba 4.13 our diskless DOS systems failed to map their drives on the server. ChatGPT came up with the answer first time, with a bunch of specific settings in smb.conf, unclear which did the trick but it worked at first attempt (after numerous failed SO-inspired guesses). The day before, a co-worker sent a draft of a Web UI layout that was clearly doable but well beyond my CSS abilities . Four hours later, with copilot assistance, the web page (a complex form) got a big thumbs up. Today, I asked a question about how to solve a problem I'd struggled for some months. No useful answers on SO, lots of net-poking and thinking got me a working solution that seemed pretty cool. And...I asked ChatGPT, first try it gave me CSS that was definitely better than my efforts. On the other hand, in my specific domain of expertise, the answer are either egg-suck or trival. So (for me, personally) the real boost is in areas I have a little knowledge and can figure out the right question to ask an AI.

JaDogg 3y ago

Yes I have this experience with GPT 3.5, where my code is superior when it actually matters. But I'm not sure if GPT 4 will help me with that. There is a waitlist to pay now. I recently tried to use it write a .scad parser. I ended up writing it myself as code it generated was useless. (This was even in Python and I tried rephrasing and giving it multiple examples) It also loves to import libraries. Which I do not really like I prefer avoid dependencies as much as I can.

zh3 3y ago

Personally I'm on the wait list, mainly because it cracked a problem in seconds that one hour plus on stackoverflow hadn't solved. I didn't go to ChatGPT first because (erroneously) I figured the issue was too esoteric for it to crack. Specifically, following an update to Samba 4.13 our diskless DOS systems failed to map their drives on the server. ChatGPT came up with the answer first time, with a bunch of specific settings in smb.conf, unclear which did the trick but it worked at first attempt (after numerous failed SO-inspired guesses). The day before, a co-worker sent a draft of a Web UI layout that was clearly doable but well beyond my CSS abilities . Four hours later, with copilot assistance, the web page (a complex form) got a big thumbs up. Today, I asked a question about how to solve a problem I'd struggled for some months. No useful answers on SO, lots of net-poking and thinking got me a working solution that seemed pretty cool. And...I asked ChatGPT, first try it gave me CSS that was definitely better than my efforts. On the other hand, in my specific domain of expertise, the answer are either egg-suck or trival. So (for me, personally) the real boost is in areas I have a little knowledge and can figure out the right question to ask an AI.