4 ms·

GPT, Copilot, etc. are glorified auto-complete tools. There is nothing wrong with using them, and they are useful. They still need a programmer to give proper

GPT, Copilot, etc. are glorified auto-complete tools. There is nothing wrong with using them, and they are useful. They still need a programmer to give proper instructions, though. I don't think an end user can prompt them to write anything remotely sophisticated the current technology. When these tools are able to detect ambiguity in specifications and start asking the right questions to clarify those ambiguities, they will also start being useful for non-programmers.

serf 3y ago
chatgpt will easily create complete autohotkey scripts with no expert knowledge other than "make an autohotkey script that.." and just good rudimentary step-by-step explanation of what the script must do. that has pretty broad applications for all sorts of non-programmers. yeah, you're right if the goalpost is "please make a HIPAA compliant DBS from scratch for use in California", but most non-programmers aren't really interested in specification compliant programs that meet meticulous measures of rigor; they just want things that simplify their routine when cashing out a customer, organizing their pirated movies, perusing porn sites, whatever it is that people do. Programming as a metaphor for telling a machine what to do was a great gatekeeper for people for many years, chatgpt and the like are reducing the 'fee to entry' immensely for non-technical people, and I say that as someone who is witnessing non-technical people in my life accomplish things they would have never tried previously.
- worrycue 3y ago
  > chatgpt will easily create complete autohotkey scripts with no expert knowledge other than "make an autohotkey script that.." and just good rudimentary step-by-step explanation of what the script must do. How much of this is thanks to hundreds of tutorials / question & answers on stack overflow about these tasks? As AI supplant the sites that host these QnA and professional tutorial writers - even amateur writers might feel miffed that their efforts will be regurgitated by AI with no credit given - how will AI keep up the quality of their responses?
  - anonylizard 3y ago
    Easy, upvoted responses from the AI will be fed back into AI. If humans can train themselves without AI, then so can AI train without humans.
    - worrycue 3y ago
      Can today’s AI even provide a solution if it has never seen a single example of such a solution?
      - zornthewise 3y ago
        Sure, the best chess/go players on this planet are not humans anymore and those games are full of novel situations with no known solutions.
        OJFord 3y ago
        That seems pretty different though, they have relatively few relatively straightforward fixed rules, you can just compute possible outcomes for moves.
        ben_w 3y ago
        The branching factor of Go prevents brute force calculation of the best possible moves. It isn't even obvious how to write a function that evaluates how strong any given board position is. The AI had to learn functions to do that.
        worrycue 3y ago
        Stock Fish and AlphaGo aren’t LLMs - which is the type of AI ChatGPT is.
        anonylizard 3y ago
        Deepmind is trying to combine "Alphazero" style algorithms with LLMs to build Gemini.
- criddell 3y ago
  > chatgpt will easily create complete autohotkey scrtips I’ve used it to do this very thing and while it was helpful, it really struggled with sticking to either v1 or v2 of the language. It felt like playing whack-a-mole. I would tell it some line was incorrect syntax for v2 and it would apologize, fix that line and break something else. The bones of the script were good, but I had to do quite a bit of editing. I don’t feel all that negative about it though because I understand that it will only get better.
ben_w 3y ago
The main limitations of ChatGPT 3.5 for a non-programmer is that non-programmers don't know the right questions to ask; that 10% of functions it produces are buggy in ways the AI can't fix by itself; and that the source code for a useful program is longer than the context window, so it will for example generate a Vector class with "mul(b)" method and later try to call "multiply(b)". I'm told GPT-4 is much better at bug fixes, and there are various options for longer context windows. Knowing what question to ask, though? The AI seems to contain the knowledge to all the questions, as if I ask them for a list of questions I get a useful response; and now someone has to answer the meta-question of "how do you want to present those questions to a user?" as there are many ways…
- blharr 3y ago
  As far as the questions to ask, I really like what Bing has done in including some relevant questions that can be selected to follow up.

serf 3y ago

chatgpt will easily create complete autohotkey scripts with no expert knowledge other than "make an autohotkey script that.." and just good rudimentary step-by-step explanation of what the script must do. that has pretty broad applications for all sorts of non-programmers. yeah, you're right if the goalpost is "please make a HIPAA compliant DBS from scratch for use in California", but most non-programmers aren't really interested in specification compliant programs that meet meticulous measures of rigor; they just want things that simplify their routine when cashing out a customer, organizing their pirated movies, perusing porn sites, whatever it is that people do. Programming as a metaphor for telling a machine what to do was a great gatekeeper for people for many years, chatgpt and the like are reducing the 'fee to entry' immensely for non-technical people, and I say that as someone who is witnessing non-technical people in my life accomplish things they would have never tried previously.

worrycue 3y ago
> chatgpt will easily create complete autohotkey scripts with no expert knowledge other than "make an autohotkey script that.." and just good rudimentary step-by-step explanation of what the script must do. How much of this is thanks to hundreds of tutorials / question & answers on stack overflow about these tasks? As AI supplant the sites that host these QnA and professional tutorial writers - even amateur writers might feel miffed that their efforts will be regurgitated by AI with no credit given - how will AI keep up the quality of their responses?
- anonylizard 3y ago
  Easy, upvoted responses from the AI will be fed back into AI. If humans can train themselves without AI, then so can AI train without humans.
  - worrycue 3y ago
    Can today’s AI even provide a solution if it has never seen a single example of such a solution?
    - zornthewise 3y ago
      Sure, the best chess/go players on this planet are not humans anymore and those games are full of novel situations with no known solutions.
      - OJFord 3y ago
        That seems pretty different though, they have relatively few relatively straightforward fixed rules, you can just compute possible outcomes for moves.
        ben_w 3y ago
        The branching factor of Go prevents brute force calculation of the best possible moves. It isn't even obvious how to write a function that evaluates how strong any given board position is. The AI had to learn functions to do that.
      - worrycue 3y ago
        Stock Fish and AlphaGo aren’t LLMs - which is the type of AI ChatGPT is.
        anonylizard 3y ago
        Deepmind is trying to combine "Alphazero" style algorithms with LLMs to build Gemini.
criddell 3y ago
> chatgpt will easily create complete autohotkey scrtips I’ve used it to do this very thing and while it was helpful, it really struggled with sticking to either v1 or v2 of the language. It felt like playing whack-a-mole. I would tell it some line was incorrect syntax for v2 and it would apologize, fix that line and break something else. The bones of the script were good, but I had to do quite a bit of editing. I don’t feel all that negative about it though because I understand that it will only get better.

ben_w 3y ago

The main limitations of ChatGPT 3.5 for a non-programmer is that non-programmers don't know the right questions to ask; that 10% of functions it produces are buggy in ways the AI can't fix by itself; and that the source code for a useful program is longer than the context window, so it will for example generate a Vector class with "mul(b)" method and later try to call "multiply(b)". I'm told GPT-4 is much better at bug fixes, and there are various options for longer context windows. Knowing what question to ask, though? The AI seems to contain the knowledge to all the questions, as if I ask them for a list of questions I get a useful response; and now someone has to answer the meta-question of "how do you want to present those questions to a user?" as there are many ways…

blharr 3y ago
As far as the questions to ask, I really like what Bing has done in including some relevant questions that can be selected to follow up.