> The next generation of personal computers should give their users the most advanced computer languages.
> But someday, according to Minsky, the most useful programs for personal computers will be based on artificial-intelligence programs that write programs of their own. The idea is for an ordinary person — not a programmer — to describe what he wants a program to do in informal terms, perhaps simply by showing the program-writing program some examples.
> Then it will write a computer program to do what was described — a process that will be much cheaper than hiring a professional programmer.
I mean, I'm a programmer but also I had GPT make some asn.1 to DER code for me. Which was cheaper than hiring myself to build from scratch or re-read the spec and code up the tools. It was even easier than researching and testing an external package.
A long time ago I developed a simple SNMP tool in C++. The ASN.1 BER encoder/decoder was very annoying to develop. A very bizarre specification. If GPT can help a lot in a case like that, I'm very impressed and see great value in it.
I've found gpt to excel in esoteric and well defined but novel situations even more than common ones, and I have a suspicion there's just a lot less ambiguity or controversy in the training material for esoteric subjects.
There very likely is. But I had a very narrow, single use case here. And, as I said "It was even easier than researching and testing an external package."
Why spend time researching and vetting some dependency hell when GPT made 90% of what I needed in a few minutes.
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.
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.
> 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?
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.
> 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.
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…
Compilers are translators, but generally from one formal language to another. Speaking English to GPT4 and getting a fully functional program is a whole other beast. Sure, I guess you could abstract it away and call an LLM a compiler as well.
> But someday, according to Minsky, the most useful programs for personal computers will be based on artificial-intelligence programs that write programs of their own. The idea is for an ordinary person — not a programmer — to describe what he wants a program to do in informal terms, perhaps simply by showing the program-writing program some examples.
> Then it will write a computer program to do what was described — a process that will be much cheaper than hiring a professional programmer.