What it would take to convince me
Hey Elliot, you’re still an AI hater right? What would it take to convince you?
Firstly, please call me Mr. Morris. Secondly, you say “still” as though my hatred for AI is related to it’s effectiveness and not to how its existence shatters the democratic principle of laws applying equally to all the human participants of a nation state.
Okay fine, if you really wanted to convince me that heavily LLM driven software development practices (because that’s what you must mean to justify … *gestures broadly*) are now an effective way of producing valuable software, you’re going to need to build and sell a software product to a specific set of conditions.
I’m going to preface this by saying this is what would convince me, it is not a fair test, and I’m a picky bastard. I’m talking about being convinced to change my mind, which means I need to really be certain that specific suspicions of mine are not at play. I’ll explain what I mean as we go.
I’m also going to try and make these relatively falsifiable. No “Produce real value” or any other metaphysically adjacent requirements. Still, I’m sure anyone attempting to actually test me on these will find it frustratingly impossible to agree if any given example “counts”. Too bad.
#1 – Displace the market leader
If LLM development practices are the future, and are an improvement, it stands to reason that eventually a company driven by them will disrupt and replace.
This only makes sense if we are in at least a semi-rational market environment, and development practices actually, you know, matter when it comes to organizational success. If either of them are untrue then nothing we do means anything anyway, and this whole LLM kerfuffle is a massive waste of time and money.
#2 – Do so with a product unrelated to AI
I’m not sure I’ve ever seen a properly off the deep end AI booster ever not be developing an AI product. It’s normally some sort of LLM orchestration frame, or context management system.
It goes without saying this is deeply suspicious, you can’t manifest value solely by eating your own tail.
#3 – Do so in a “traditional” market domain
Let’s take that thinking a step further. To be truly convinced, I’m going to need to be sure you’re not operating within an irrational market environment.
This means “new” domains are out, I can’t take the risk, they could be pump and dump bubbles. Any hype driven domains like crypto, metaverse or anything to do with speculative investment are also way out.
Here are some examples of domains that would satisfy me : consumer or server operating systems, GIS software, FEA/MBD simulators, accountancy/tax software, office suite software, CAD software. You know … stuff that’s actually important.
#4 – Your “engineers” never edit source code
This one may seem wild, and I agree, it’s not fair. However, I’m more aware than anyone how often a small number of motivated engineers can keep a ship moving forward by stealth, despite destructive and baffling behaviors by leadership.
I need to ensure that this is not happening to be convinced. Therefore, no engineers editing the code, ever. I’d quite like to weaken this requirement as even I grant that there’s a huge gap between full-vibe and LLM assisted practices, but I can’t think of a better way to control against engineers compulsively trying to produce working products.
#5 – Make a real profit for three strait years.
Finally, actually make some goddamn money. Three years seems like a good amount of time to prove that you’ve got something sustainable and aren’t riding a bubble. The length of time should also guard against the greenfield problem, as I’ll need to be shown that you can maintain a product to the satisfaction of your customers, whether via traditional maintenance or newfangled AI driven total rewrites.
This has to be real profit too. No weird accounting tricks, no counting investment as profit. Customers need to freely choose to give you their money, ideally they will also be happy about doing so. This also means that you need to be doing this in a non-subsidized token environment, unless those subsidies can be expected to remain available in the long term. I’m confident enough that three years is long enough for that to resolve, although the market is extremely good at staying irrational so perhaps this is too optimistic.
These would convince me. They are not proofs. I can imagine LLM driven development being useful without any given product having fulfilled these conditions, especially the traditional market domain one, which has regulatory capture issues to contend with. I also don’t think an LLM driven company that fulfills these conditions has proven anything, not formally, these are mostly market based and market success doesn’t necessarily equate to value.
Nonetheless, I am a human living in a market economy so they hold a large psychological sway, and these conditions being met would cause me to have to re-evaluate my “LLM systems cannot produce useful value by definition” ideology.
You shouldn’t really care about convincing me though, I don’t matter. If you’re really confident you’ll achieve all these things and keep it to yourself, why bother making a fuss, you’ll have already won.