In my opinion, the future of software development looks like this:
1) Human developer instructs A.I. bots to build large volumes of code (like Second)
2) Human developer uses A.I. code assistants in the IDE to make "low level" changes or additions
I do not believe that developers or engineers will be replaced. The world needs more software, and there's not enough of us to build it all. Quite frankly, we need the help of bots to meet the demand.
AI changes this, however. The scenario you are describing assumes that AI will get stuck, as I said. Otherwise I don't see why we would still need humans to supervise it. But if it can write good enough software without expensive human beings making corrections left and right, it will be able to do a lot of the jobs that we're trying to make more efficient (or even possible) with all this software we're writing. E.g. you don't really need GitHub (Jira, etc.) if you don't have a lot of developers cooperating. And this is true for a lot of white collar professions that we're creating these heaps of web apps for, that generate all that demand that you are mentioning.
If we get stuck (and, BTW, have some time to figure out how to live with AI) then sure, we'll have programmers (and non-programmers with good analytical skills) command and supervise AI coders. Everyone and their colleague will turn into a tech-lead/product manager. If.
That's basically my opinion, yeah sure this looks impressive, until you realise that it's basically recreating Wordpress with less modules.
Generating code is far from enough to replace developers at this point, it's the earliest step, even Frontpage itself is 30 years old at this point.
Instead of looking for layoffs, look for the jobs and payment transactions never happening
Note: this doesn't bother me
If it is successful, it increases jobs in the field, while shifting skill demands to higher levels of analysis/abstraction. Just like every change improvement in how you program computers since "physically configure wiring or switches".
Possibly some individuals won't get hired in the field that otherwise would have, but the size of the field developing software isn't likely to shrink because of it.
I feel like the early days of programming might've been actually easier to automate. There was often well designed specs, cleaner design and architectural patterns to follow etc. Not in all cases of course however now, I'm seeing people hoping to god that ChatGPT can write a client for the worlds shittiest XML API because no one else sure as hell wants to do it.
So what happens now? We just generate more shit and then use moar AI (tm) to work with moar shit (tm), and evolve that pattern fast and faster?
Yesterday I was using co-pilot and it was suggesting a bunch of obvious auto-completes etc, which is fine, that's why I use it, then it hit me, most people are probably going to start to not bother with libraries, clean APIs and specs because soon you will be just brute forcing your way to a solution using text to code and on we go. Maybe the answer will really just be to keep evolving generative coding AIs to keep up with it, let's see how it goes.
OTOH, so much for the idea that people will guide the hand of AI to create better code. As this will unfold, there will be ever more incentives to remove (most) humans from the loop. And if past can be used to predict the future, what we have seen so far is that when AI gets reasonably competent at something it will gain superhuman capabilities very quickly. I still keep bringing up how people thought that after alpha go beat Fan Hui (European go champion) it would have taken a huge leap for it to beat Lee Sedol. Because Lee Sedol was so much better than Fan Hui. At least in human terms. It only took a few months of training and tinkering for DeepMind.
Speaking of loops: since AIs are taught from the internet (simple language models or coding specialists), we're creating an unwanted feedback loop here. More AI generated content will likely make teaching future AIs harder. At least with information harvested from the internet after ~2022-ish.
But these are fundamentally different from AI. As I mentioned above, Fred Brooks in his essay (The Mythical Man-month) would argue that all these just decrease the accidental complexity. The incidental complexity still remains. I.e. taking a vague description, an idea, something embedded in the heads of the stakeholders and turning it into some kind of actual computer code, finding and removing inconsistencies and with a minimal number of bugs. Now AI will be able to do it one day. And it seems that day is not that far away in the future.
The hard move is going from a vague set of incomplete requirements (which is always the case) to specific, executable code. We never had a tool that can do this transformation. Until now(ish).
> The world needs more software, and there's not enough of us to build it all
... is a paraphrasing of the following from your top-level text
> Moreover, the world needs more software than there are engineers to build it all.
... which I'm very skeptical about. In what sense does the world "need" more software?
What exactly is it that "the world" really needs this additional software for, and is it really software that's missing, or can the same need be fulfilled via other means (e.g. better software, or improved processes)?
And to follow-up on that, if the world does need "more software", is there some level of software that is sufficient? I.e. is there a steady state of software/world, or at least, software/person, or are we "maximizing paperclips"?
It sends an AI generated letter to a congress person in a seconds. Users can edit it, but frankly the results are amazingly good.
I bring it up because effectively exactly what you described - a shim on top of 4 APIs. (Code is open sourced, link on the site)
The question is how will AI replace what I have?
The AI might get better, but ultimately someone has to mail the letter (I use a different API for that). Someone has to ensure there’s no threats (we do a manual review, combined with the AI review) and someone needs to verify location.
Sure the AI systems of the future could theoretically do all that, but it’s still the same work/services.
Before the AI the service I built required me to manually create templates or users to write the full letter. Frankly, AI opened up this business, not closed it. It was more expensive and a lot more logistical work to manually generate these letters.
Probably not [0]; oh, the output may read nicely, but letters to legislators, even when I was working in a state legislative office in the 1990s, before mass online template generated near-zero-user-effort spam letters were common (but not before organizations sent out template for people to use, or even xerox-and-add-name-and-signature near-zero-effort spam) mostly got read by interns and processed by a heuristic like this (for issues mail, constituent service is different):
(1) Is it from a constituent? No = done and file, Yes = continue.
(2) Is it for a specific, identifiable active bill before the body in which the legislator sits (including the other House) -- see (3), or is it a general issue or one before a different body (if it seems to be maybe about an active bill, but the information is insufficient to identify the bill readily, treat as the second) -- see (4)?
(3) Does it clearly support or oppose the bill? If so, record it in the running tally (creating it if necessary) of constituent support/opposition for that bill; file and done.
(4) Does it address an issue that the legislator has identified as important and has a message on? If so, respond with that message; file and done.
Literally, the ultimate constituent letter on a bill for effect (maximum effect, minimum chance of getting misinterpreted, minimum effort) was a postcard with sufficient address information to identify the sender as a constituent, a bill number, and "Support" or "Oppose".
I don't imagine individual issue letters are getting even that much attention these days, which is why all the serious activist trainings focus on live calls ass the "low effort but low effect" approach and visiting legislators offices as the high effort/high effect approach to influence. Letters and emails might get categorized and mined for propaganda use by legislators, but they aren't moving positions, and every new tool to spam legislators with them reinforces that effect. They are a placebo button, and incorporating AI technology to make a newer, fancier-looking placebo button isn't changing that.
[0] But at least it gets the federal legislative reps (almost) right (it makes the error of using the legal name of a Senator rather than the name they actually use publicly, and hey, if it was going to actually be read, nothing makes a better impression than using the wrong name), for state reps it gives me none of the actual reps, but three of the eight statewide Constitutional officers (the Governor, AG, and Secretary of State).
> Literally, the ultimate constituent letter on a bill for effect (maximum effect, minimum chance of getting misinterpreted, minimum effort) was a postcard with sufficient address information to identify the sender as a constituent, a bill number, and "Support" or "Oppose".
That’s exactly what is provided. First line is a clear description, support / opposed in bold and the policy/law.
We used the billing info to verify they’re in district and provide the return address.
Phone number should also be provided for a call effort.
I was told they evaluate scales of issues, ie 10 letters from people in district might matter, 100 would matter a lot, 1000 would be a burning issue. That’s effectively all I’m trying to provide.
> for state reps it gives me none of the actual reps, but three of the eight statewide Constitutional officers (the Governor, AG, and Secretary of State).
Yeah, I automated what I could, I can manually enter / edit state reps and did for several states. Constitutional officers are mostly included because they can be pulled via api (plus their still parties who are often elected)
I think you answered your own question:
> Sure the AI systems of the future could theoretically do all that, but it’s still the same work/services.
That's it. At the moment, you had to wire together 4 APIs and, as you said, this wouldn't do the job without using (someone else's) AI. So this works as a business/service as long as the AI is not able to handle those 4 APIs by itself (or 3, if the AI is one of them). And it doesn't even have to figure out using these APIs, some company (I'd expect the owner of the AI) can create interfaces.
Not that ChatGPT doesn't know how to use APIs itself. I started a quick experiment, that I didn't have the time to finish yet, making it create a simple GUI app that is connected to a specific SaaS API. Nothing fancy.
At first it would reject my request when I told it to create an app that uses that API saying that it can't create apps. Then I told it to generate code for the main window with the most important details. And it did. Then I told it to create an API wrapper for the API. Then I told it to connect the API wrapper with the gui. Honestly, nothing that it couldn't do if it didn't reject my original request with "I'm sorry, but as an AI language model, I'm not able to generate code for you."
So I don't see how the current state of affairs is too far from me being able to say to a similar AI agent to send a letter to a congress person regarding this and that cause where my opinion is X.
Integrations are more than code (you have to setup physical addresses in this case, transfer money, etc). I do believe an AI will get there, but all that human evaluation will still need to happen.
I personally assume these AI systems will be used as the middleware for a long time. It’ll be hard to enter other spaces because (1) it won’t be something legally the AI company wants to be liable for and (2) it often requires niche semi-documented items.
Much easier for AI companies to let people take the risks. They can then blame that party.
That said, you have exactly the same category of risk with hiring an employee or team. They could leave at any point also. You always have the option to contract in handcuffs or a notice period, but then the same applies to the AI tooling as well.
In the same way this can't end up much worse than cloud vendor lock-in either. Many businesses dream or discuss about being 'cloud agnostic' but realistically most are tightly coupled, especially around cost-of-change and institutional knowledge.
That's one thing. The other thing is that employees who leave only hamper future development not the current business/operation. (Again, unless a large number of employees leave in a short time, which means you have screwed up something.) Once you have a stable organization, you can replace individual employees at a reasonable rate without serious issues. (You actually have to prepare for this as employees will leave, no matter what.)
The big difference with cloud providers is that they really offer something that's complementary and not something that can easily be turned into your product. (I mean except for the SaaS companies that really just repackage the cloud offer into an easier to use product. They are definitely at risk all the time.)
Another risk for building businesses on AI in general is that unlike with software that anyone can build (e.g. for a niche), AI solutions seem to be costly because they require a lot of data and a lot of processing power (at least to train) so you'll end up using someone else's AI. And the real magic is the AI not whatever you build around it so it can be pulled from under you while the AI itself keeps developing and obviously it can make the thin layer built around it irrelevant. I know that AI companies, at least OpenAI will sell the vision that they will be like AWS (amazon) and other similar cloud providers, but I just don't see this as a realistic scenario. It can only happen if AI gets stuck near the level where it is now for a substantial period. Say, at least a decade. Where it's good enough to be included in SaaS (and maybe other) software products but it's not good enough yet to do away with all software products or at least the need for companies building software products.