I’m going to take a big chance here and make predictions about GenAI’s future. Yeah, I know, you’re feeling overloaded on this stuff and me too, but it seems to have sucked the air out of all the other conversations in my profession. I would so like to return to arguing about Functional Programming or Integration Testing. This is risky and there’s a pretty good chance that I’m completely wrong. But I’ll try to entertain while prognosticating.

Reverse Centaurs · That’s the title of a Cory Doctorow essay, which I think is spot on. I’m pretty sure anyone who’s read even this far would enjoy it and it’s not long, and it’d help understand this. Go have a look, I’ll wait.

Hallucinations won’t get fixed · I have one good and one excellent argument to support this prediction. Good first: While my understanding of LLMs is not that deep, it doesn’t have to be to understand that it’s really difficult (as in, we don’t know how) to connect the model’s machinations to our underlying reality, so as to fact-check.

The above is my non-expert intuition at work. But then there’s Why Language Models Hallucinate, three authors from OpenAI and one from Georgia Tech, which seems to show that hallucinations are an inevitable result of current training practices.

And here’s the excellent argument: If there were a way to eliminate the hallucinations, somebody already would have. An army of smart, experienced people people, backed by effectively infinite funds, have been hunting this white whale for years now without much success. My conclusion is, don’t hold your breath waiting.

Maybe there’ll be a surprise breakthrough next Tuesday. Could happen, but I’d be really surprised.

(When it comes to LLMs and code, the picture is different; see below.)

The mass layoffs won’t happen · The central goal of GenAI is the elimination of tens of millions of knowledge workers. That’s the only path to the profits that can cover the costs of training and running those models.

To support this scenario the AI has to run in Cory’s “reverse centaur” mode, where the models do the work and the humans tend them. This allows the production of several times more work per human, generally of lower quality, with inevitable hallucinations. There are two problems here: First, that at least some of the output is workslop, whose cleanup costs eat away at the productivity wins. Second, that the lower quality hurts your customers and your business goes downhill.

I just don’t see it. Yeah, I know, every CEO is being told that this will work and they’ll be heroes to their shareholders. But the data we have so far keeps refusing to support those productivity claims.

OK then, remove the “reverse” and run in centaur mode, where smart humans use AI tools judiciously to improve productivity and quality. Which might be a good idea for some people in some jobs. But in that scenario neither the output boost nor the quality gain get you to where you can dismiss enough millions of knowledge workers to afford the AI bills.

The financial damage will be huge · Back to Cory, with The real (economic) AI apocalypse is nigh. It’s good, well worth reading, but at this point pretty well conventional wisdom as seen by everyone who isn’t either peddling a GenAI product or (especially) fundraising to build one.

To pile on a bit, I’m seeing things every week like for example this: The AI boom is unsustainable unless tech spending goes ‘parabolic,’ Deutsche Bank warns: ‘This is highly unlikely’.

The aggregate investment is ludicrous. The only people who are actually making money are the ones selling the gold-mining equipment to the peddlers. Like they say, “If something cannot go on forever, it will stop.” Where by “forever”, in the case of GenAI, I mean “sometime in 2026, probably”.

… But the economy won’t collapse · Cory forecasts existential disaster, but I’m less worried. Those most hurt when the bubble collapses will be the investing classes who, generally speaking, can afford it. Yeah, if the S&P 500 drops by a third, the screaming will shake the heavens, but I honestly don’t see it hitting as hard as 2008 and don’t see how the big-picture economy falls apart. That work that the genAI shills say would be automated away is still gonna have to be done, right?

The software profession will change, but not that much · Here’s where I get in trouble, because a big chunk of my professional peers, including people I admire, see GenAI-boosted coding as pure poison: “In a kind of nihilistic symmetry, their dream of the perfect slave machine drains the life of those who use it as well as those who turn the gears.” (The title of that essay is “I Am An AI Hater.”)

I’m not a hater. I argued above that LLMs generating human discourse have no way to check their output for consistency with reality. But if it’s code, “reality” is approximated by what will compile and build and pass the tests. The agent-based systems iteratively generate code, reality-check it, and don’t show it to you until it passes. One consequence is that the quality of help you get from the model should depend on the quality of your test framework. Which warms my testing-fanatic heart.

So, my first specific prediction: Generated code will be a routine thing in the toolkit, going forward from here. It’s pretty obvious that LLMs are better at predicting code sequences than human language.

In Revenge of the junior developer, Steve Yegge says, more or less, “Resistance is useless. You will be assimilated.” But he’s wrong; there are going to be places where we put the models to work, and others where we won’t. We don’t know which places those are and aren’t, but I have (weaker) predictions; let’s be honest and just say “guesses”.

Where I suspect generated code will likely appear:

• Application logic: “Depreciate the values in the AMOUNT field of the INSTALLED table forward ten years and write the NAME field and the depreciated value into a CSV.” Or “Look at JIRA ticket 248975 and create a fix.”

  (By the way, this is a high proportion of what actual real-world programmers do every day.)

• Glorified StackOverflow-style lookups like I did in My First GenAI Code.

• Drafting code that needs to run against interfaces too big and complex to hold in your head, like for example the Android and AWS APIs (“When I shake the phone, grab the location from GPS and drop it in the INCOMING S3 bucket”). Or CSS (“Render that against a faded indigo background flush right, and hold it steady while scrolling so the text slides around it”).

• SQL. This feels like a no-brainer. So much klunky syntax and so many moving pieces.

Where I suspect LLM output won’t help much.

• Interaction design. I mean, c’mon, it requires predicting how humans understand and behave.

• Low level infrastructure code, the kind I’ve spent my whole life on, where you care a whole lot about about conserving memory and finding sublinear algorithms and shrinking code paths and having good benchmarks.

Here are areas where I don’t have a prediction but would like to know whether and how LLM fits in (or not).

• Help with testing: Writing unit and integration tests, keeping an eye on coverage, creating a bunch of BDD tests from a verbal description of what a function is going to do.

• Infrastructure as code: CI/CD, Terraform and peers, all that stuff. There are so many ways to get it wrong.

• Bad old-school concurrency that uses explicit mutexes and java.lang.Thread where you have to understand language memory models and suchlike.

The real reason not to use GenAI · Because it’s being sold by a panoply of grifters and chancers and financial engineers who know that the world where their dreams come true will be generally shitty, and they don’t care.

(Not to mention the environmental costs and the poor folk in the poor countries where the QA and safety work is outsourced.)

Final prediction: After the air goes out of the assholes’ bubble, we won’t have to live in the world they imagine. Thank goodness.