Even within the N% that is more genuine coding and system reasoning; system reasoning is
really hard, oftentimes requiring weird leaps of faith, and I also don't see a path for AI to be helpful with that.
Some random recent thing: "We have a workflow engine that's composed of about 18 different services. There's an orchestrator service, some metadata services on the side, and about 14 different services which execute different kinds of jobs which flow through the engine. Right now, there is no restriction on the ordering of jobs when the orchestrator receives a job set; they just all fire off and complete as quickly as possible. But we need ordering; if a job set includes a job of type FooJob, that needs to execute and finish before all the others. More-over, it will produce output that needs to be fed as input to the rest of the jobs."
There's a lot of things that make this hard for humans, and I'm not convinced it would be easier for an AI which has access to every bit of code the humans do.
* How do the services communicate? We could divine pretty quickly: let's say its over kafka topics. Lots of messages being published, to topics that are provided to the applications via environment variables. Its easy to find that out. Its oftentimes harder to figure out "what are the actual topic names?" Ah, we don't have much IaC, and its not documented, so here I go reaching for kubectl to fetch some configmaps. This uncovers a weird web of communication that isn't obvious.
* Coordination is mostly accomplished by speaking to the database. We can divine parts of the schema by reverse engineering the queries; they don't contain type information, because the critical bits of this are in Python, and there's no SQL files that set up the database because the guy who set it up was a maverick and did everything by hand.
* Some of the services communicate with external APIs. I can see some axios calls in this javascript service. There's some function names, environment variable names, and URL paths which hint to what external service they're reaching out to. But, the root URL is provided as an environment variable; and its stored as a secret in k8s in order to co-locate it in the same k8s resource that stores the API key. I, nor the AI, have access to this secret thanks to some new security policy resulting from some new security framework we adopted.
* But, we get it done. We learn that doing this ordering adds 8 minutes to every workflow invocations, which the business deems as unacceptable because reasons. There is genuinely a high cardinality of "levels" you think about when solving this new problem. At the most basic level, and what AI today might be good at: performance optimize the new ordered service like crazy. But that's unlikely to solve the problem holistically; so we explore higher levels. Do we introduce a cache somewhere? Where and how should we introduce it, to maximize coherence of data? Do some of the services _not_ depend on this data, and thus could be ran outside-of-order? Do we return to the business and say that actually what you're asking for isn't possible, when considering the time-value of money and the investment it would take to shave processing time off, and maybe we should address making an extra 8 minutes ok? Can we rewrite or deprecate some of the services which need this data in order to not need it anymore?
* One of the things this ordered workflow step service does is issue about 15,000 API calls to some external service in order to update some external datasource. Well, we're optimizing; and one of the absolute most common things GPT-4 recommends when optimizing services like this is: increase the number of simultaneous requests. I've tried to walk through problems like this with GPT-4, and it loves suggesting that, along with a "but watch out for rate limits!" addendum. Well, the novice engineer and the AI does this; and it works ok; we get the added time down to 4 minutes. But: 5% of invocations of this start failing. Its not tripping a rate limit; we're just seeing pod restarts, and the logs aren't really indicative of what's going on. Can the AI (1) get the data necessary to know what's wrong (remember, k8s access is kind of locked down thanks to that new security framework we adopted), (2) identify that the issue is that we're overwhelming networking resources on the VMs executing this workflow step, and (3) identify that increasing concurrency may not be a scalable solution, and we need to go back to the drawing board? Or, lets say the workflow is running fine; but the developers@mycompany.com email account just got an email from the business partner running this service that they had to increase our billing plan because of the higher/denser usage. They're allowed to do this because of the contract we signed with them. There are no business leaders actively monitoring this account, because its just used to sign up for things like this API. Does the email get forwarded to an appropriate decision maker?
I think the broader opinion I have is: Microsoft paid hundreds of millions of dollars to train GPT-4 [1]. Estimates say that every query, even at the extremely rudimentary level GPT-3 has, is 10x+ the cost of a typical google search. We're at the peak of moores law; compute isn't getting cheaper, and actually coordinating and maintaining the massive data centers it takes to do these things means every iota of compute is getting more expensive. The AI Generalists crowd have to make a compelling case that this specialist training, for every niche there is, is cheaper and higher quality than what it costs a company to train and maintain a human; and the Human has the absolutely insane benefit that the company more-or-less barely trains them, the human's parents, public schools, universities paid for by the human, hobbies, and previous work experience do.
There's also the idea of liability. Humans inherently carry agency, and from that follows liability. Whether that's legal liability, or just your boss chewing you out because you missed a deadline. AI lacks this liability; and having that liability is extremely important when businesses take the risk of investment in some project, person, idea, etc.
Point being, I think we'll see a lot of businesses try to replace more and more people with AIs, whether intentionally or just through the nature of everyone using them being more productive. Those that index high on AI usage will see some really big initial gains in productivity; but over time (and by that I mean, late-20s early-30s) we'll start seeing news articles about "the return of the human organization"; the recognizing that capitalism has more reward functions than just Efficiency, and Adaptability is an extremely important one. More-over, the businesses which index too far into relying on AI will start faltering because they've delegated so much critical thinking to the AI that the humans in the mix start losing their ability to think critically about large problems; and every problem isn't approached from the angle of "how do we solve this", but rather "how do I rephrase this prompt to get the AI to solve it right".
[1] https://www.theverge.com/2023/3/13/23637675/microsoft-chatgp...