How many people doing this job do we need compared to the tens of millions of people building programs by hand we have today?

How much better would software be if a team of five individuals could produce the same output as a team of 500 individuals?

How much terrible software is out there? How much terrible FOSS software is out there?

How much amazing and humanistic software could be created if the costs for production were drastically reduced and the demands of the market were less of a factor?

How many companies will want new or bigger apps that they can finally afford if it requires fewer developers?

It's possible everything will speed up, and since the competitors are also speeding up some kinds of arms race on steroids will take place where not only are we all much more productive but we are also not cutting back on workers. I find it hard to believe but I concur it is possible.

If you had an AI capable of writing the code you mean to without errors, the demand for software developers would be even larger than what it is today.

Deciding what problem to solve and how to solve it is hard enough to sustain more than the current population of developers. (But if you had an AI capable of deciding those, then yeah, that would be problem.)

Anyway, the current crop of highly hyped AIs are obviously structurally unable to do either. And it's not clear at all how far any AI is from solving any those (what usually means we are many decades away).

If we had a specific quantity of software that was needed, that might be a good argument. But to me, it looks like the quantity of software we want is unlimited. (Or at least, the quantity of stuff that we want software to do.) To the degree that GPT enables the same software to be written with fewer programmers, to that degree we'll write more software, not have fewer programmers.

You might also consider buying your son a Github Copilot license, it's $10/mo and I would easily pay $50/mo for the value it provides.

How does it compare to ChatGPT in terms of code quality?

Github Copilot uses a different model than chatgpt, with a tokenizer more fit for code rather than plain english language.

It's a substantial improvement when actually coding on the fly, since copilot will also analyze code in your codebase. It just makes coding the BS parts less tedious.

I think you get copilot because you don't end up having to cut and paste stuff into a browser.

Have to say I'm very happy with copilot, it's definitely worth the 10 bucks a month.

I've got both running, and I find GPT4 more useful. Unless I'm using it wrong, Copilot is "better autocomplete" and saves some typing. GPT4 can help you reason about code, and has helpful suggestions about debugging. I'm probably going to bin Copilot.

This. It's pretty obvious that any software that can be written by machine should be. (It's almost tautological, eh?) Otherwise you're doing what David Graeber famously called "Bullshit Jobs": pointless work that only serves to establish your position in the economic structure.

The immediate issues are: who decides the boundaries of (publicly) acceptable thought -and- who profits?

I think you touch on the deeper and even more challenging issue:

> I'm now trying to teach my son, an avid programmer, how to work with chatGPT. He's 13, so he's got a lot of opportunity to really integrate it into his career, but he also has a lot of blind spots where chatGPT can really lead him down some blind alleys and knife him.

When the talking computers can answer any (allowed) question, the educational challenge becomes building human minds that can recognize and select the right options from the unbounded "white hole" of content that the machines can produce.

Now, the word "right" in the previous sentence is doing a hellofa lot of work. Deciding the right course of action is an open-ended intelligence test. It's evolution, eh?

> It's pretty obvious that any software that can be written by machine should be.

This is such a good succinct, and I think objective, statement. It strikes me as a great lense through which to look at recent developments.

It does not imply that "all software should be written by a machine". But simply that anything that a machine can do just as well as a human, should probably be done by a machine rather than a human. And all additional value on top of that machine-produced material will continue to come from things that only humans can uniquely do well. And of course, there will continue to be many things in that latter category.

For many of us in this field, this means less busywork and more time spent working at higher levels of abstraction. More time curating, crafting, integrating, strategising. Less time working on the nuts and bolts.

And for those who love to hand-craft the nuts and bolts... I think that opportunity will still be there too. Because handcrafting will become a rarer skill, and there will still be areas where it is the correct approach.

Those of us who used to handcraft nuts and bolts but now delegate this to the machines, will still benefit from our understanding of how the nuts and bolts used to be made.

And those who grow up having never made a nut or bolt - perhaps not even knowing what nuts and bolts even are, will benefit from starting their path much earlier on in life working at a higher level of abstraction.

“Can answer” is doing a lot of work, too. I can write you a program that “can answer“ any question you throw at it. I can assure you the answer won’t be correct in most cases. This is a hyperbole for ChatGPT, but my point is that designing a system that can answer correctly, rather than can answer, is a far more worthwhile effort than retooling humans to be lie detectors, IMO.

We can see from how hard it is to find consensus on trivial topics (are eggs good for you?) amongst experts who study these matters day after day. And the public, who does not, is left confused. A great deal of the problems we face are too complex for the human mind to be able to decide which of several reasonable-sounding options is correct. This is especially true when there is no rationalization for answers given.

Look at our media system. How many people on both sides are stuck in “other side bad” simply because the talking heads on television networks tell them that, or someone on the internet wrote something. A lot of the content that drives them to conclusions isn’t even true. We are very poorly equipped to be vetting answers from a fallible machine.

Maybe I’m too old school, but I’d rather we learn how to solve the more complex problems so that we can do it and understand it rather than take guesses at which of a number of generated options is correct.

> designing a system that can answer correctly

I think the solution is obvious: connect the computers to empirical feedback devices and make them scientists. Humans hallucinate too if they spend too much time in a sensory deprivation tank. Give the computers sensory inputs and they will naturally become scientists.

> find consensus on trivial topics (are eggs good for you?) amongst experts who study these matters day after day

Leaving aside the question of how reliable current soft science really are, this is exactly the problem that these machines can help with, once they have the data. Consider the modified form of the question: Are eggs good for me? At this stage in my life? At this time of day? Given what I ate for dinner last night? And millions of intricate details about my medical history, and my family, and DNA, and "exposome", etc. (I worked very briefly for a medical sensing startup, our chief Doctor would wax lyrical about the possibilities for personalized medicine and nutrition-- once we have the data.)

> the public, who does not [study these matters], is left confused

To some extent, being able to do things without understanding how they are carried out "under the hood" is a measure of the advancement of civilization, eh? (But please don't mistake me as arguing in favor of ignorance!)

> when there is no rationalization for answers given.

We can ask the computers to show their reasoning (soon, if not now), we can ask them to summarize the current state of knowledge, including the grey areas and disputes, eh?

> How many people ... are stuck ...

I think it's clear that these machines will rapidly become perfect salesmen and perfect therapists. It's less clear to me what we will do with that.

> We are very poorly equipped to be vetting answers from a fallible machine.

I don't think the machines will be fallible once we connect them to sensory systems, but I do think that lots of people will try to do silly "Dr. Evil" things like try to mass-program their subjects/citizens. And I think lots of people will let them raise their children, that will probably have mixed results.

> I’d rather we learn how to solve the more complex problems so that we can do it and understand it rather than take guesses at which of a number of generated options is correct.

Selecting from the options the computer generates is the only complex problem left, see "Augmenting Human Intellect: A Conceptual Framework" SRI Summary Report AFOSR-3223 by Douglas C. Engelbart, October 1962 https://dougengelbart.org/pubs/augment-3906.html

well that’s weirdly graphic

A long tradition, see footgun.

You are referring to "lead him down a blind alley and knife him".

ChatGPT has answered some specific code questions with code that is not only wrong, but horribly wrong. Things like presenting anti-patterns as the answer, halucinating APIs that don't exist and suggesting code using them, or code that is subtly wrong. These sorts of things go well beyond leading you down a dark alley, hence my addition.