Without any funny business (meaning no re-valuation of the debt, which I guess there are strategies for) and assuming an interest rate on the debt of between 3-5%, I figured between 10-20 years before the interest payments eat up most essential services.
I don’t disagree if models stay as capable as they are today. But devils advocate: the point of the saaspocalypse isn’t just that anyone will be able to make their own software, it’s also that the AI will be good enough and interconnected enough to maintain it.
The world these investors are envisioning is not one where a software engineer gives a detailed spec to a model and reviews its output, deploys the resulting files and monitors said application. It’s where Jo-shmo at the law firm can tell the model “give me a new billing system”, and the AI does everything correctly and better than a team of software engineers, in a matter of minutes or hours. And that AI maintains it for them, better than the engineers would have
Imagine that you are a consultant. You get a call that starts with, "Hi, this is Joe Schmoe and Schmoe Law Firm. I need a new billing system. Can you build me one?"
And you respond by saying that you can, but you need to do a _lot_ of work with him to spec this billing system out. You can't just build "a new billing system" without any more details. You tell him that this will take many hours of work between the two of you where you ask him questions, write a spec, get his feedback, and repeat that a number of times.
At this point, he says "wow, that sounds like a ton of of work for me just get started", and he gives up.
AI does not fix any of this, and this is the thing that I think most people will not want to do, and that's why I think this blog post is making a very good point. The amount of work it takes to build a new software system, even with a super competent programmer as a partner, is still quite significant. And it requires thinking about hundreds of tiny little details in a way that drives a lot of people nuts. They will only do it if they _really_ have to do it.
Think payroll. I used to think payroll is relatively simple. Then I spent some time on government of Canada Phoenix pay system (go ahead... Google and weep). And it's... Insane. System has been live for a decade and still regularly gets hit with some weird scenario from some department that nobody foresaw, it wasn't captured in requirements, but upon review by business analysts is a valid scenario. Bob was a CS5 in department of defense and speaks French so gets bilingual bonus and his boss was away for half a day so Bob gets acting cs5 pay and is in public alliance union so these are the dues, and it is second Tuesday of a month and blue moon, but then Bob got moved to department of agriculture and then 3 months later realized that his previous manager at defense didn't put in his promotion on time so now you have to figure out his retro pay for when he was in defense even though everything on his file now has agriculture labour agreement and codes and rates etc etc etc. And this made up example is a fraction of the complex examples.
Clear and comprehensive Requirements are always the tricky bit, at least in business software. Twilight zone covered it perfectly and presciently decades before AI, with genies taking your requests literally and giving you unpredictable and usually negative outcomes.
Or AI won't fix diffusion of responsibility that you see in companies through outsourcing, offshoring or matrix organizations...Or to go through committees to know if they should change shh root access with abc123 as password.
Everything you describe is fantasy, though. It’s not real. It’s not possible to be real. “Give me a new billing system”?? No way is that going to produce a good result for the company or their clients. But the second that Joe Schmo has to start laying out all the ever-evolving requirements for his custom billing system, he will run back to traditional SaaS providers.
At best, if AI is supergenius enough to just intuit everything Joe needs, then the cost of running the AI to constantly maintain a billing system will far exceed the cost of just paying someone for their existing billing system SaaS.
I think the idea is you'd basically have it take a look at your current system, it would learn what features you're actually using at all, it'd check company emails for past and current pain points or stuff you wish was possible or just simpler, it'd Slack everyone in the company asking what their biggest wish and biggest pet peeves are currently, it'd do a small interview with Joe himself presenting the above to see if it's gotten the right idea, create a very detailed spec and then implement it.
Of course both models and tooling will need to be far more powerful for all this, but it doesn't exactly seem sci-fi to me.
Once system is built it could run detailed analysis on its usage and figure out what parts seem to be confusing or slow for users, and simply refine, deploy, keep analyzing, rinse and repeat.
The biggest upside is probably that workers could also simply request features, have Joe sign off on them (would get messy otherwise) and minutes later they actually roll out.
To me anyways most systems are a PITA because they do so much and your own organization only utilizes a small subset. Good systems actually let you turn off stuff you don't use so that users don't even know it's possible and don't have to drown in menu options, but that's still rare enough. And good luck getting dev focus on your specific requests regarding the parts of the system most important to your specific company, since there are a zillion other things and hundreds or thousands of other customers.
Something literally tailored to what you need will surely be the norm eventually. In five years or whatever I'm sure we'll be plenty on our way towards something like that.
But again just like LLM training in general this all requires having something existing to analyze and work off of. So yeah nobody will be going from paper to custom agent-built system.
Other idea:
Stay with SaaS, real devs, real core product, closed source, but each customer can (if they want and pay up) literally skip multi-tenant and being on the same codebase as everyone else, and get an interface to actually customize their own version to their liking.
Remove unneeded features, change UX, UI, add features. Some dev spends tiny amounts of time ensuring nothing gets too crazy, but apart from that it's basically an autonomous fork of the product, continuously tracking main.
It's science fiction, because when you have this capabilities you don't need any payroll system, you just let the AI do the job Joe was doing.
The system you are describing is far more intelligent than Joe, there's no point to use it to solve a operational problem, just point it to the problem Joe is solving.
Or even better to the problem that create the need for Joe role.
doesn't matter if it's just some agent using the system in the end, you do understand you still need an actual system for consistency etc, right?
Just that instead of cirrent "making something possible" it'll mainly be for restricting what's possible, hence forming a stable format.
That will never ever stop being a thing, even if the inputs and possibly outputs are entirely rock n roll it still requires coercion into something suitable for storage and processing.
I was being a little facetious - I really dont think AI systems are there yet. It would probably look more like an interview, and there will be some amount of human-required maintenance and subjectivity for a while.
But I think thats what the investors are envisioning.
reply