Why I was (kinda) wrong about AI

I am a reasonable person, and I can admit when I’m wrong. Which is why this partial climbdown on AI is hard for me to navigate. My position has been:

1. These companies are dogshit
2. These products are ethically compromised
3. This technology is a dead end

I still stand by 1 and 2, but I am seeing a way forward on 3.

In particular, I see a value for them in software development. I still think that Nilay Patel’s take about Software Brain stands.

But in both a professional environment and even at home, I can see undeniable value in being able to generate code that you understand but did not write. I’m seeing patterns emerge for steering it through loops that are easy to make predictable and create easily reviewable outputs. A way to structure work that is faster than writing the code yourself. Oddly, this actually seems to be more useful on larger, older projects than fresh ones. I suspect that this is because it’s operating in an environment where, good or bad, the overwhelming majority of decisions made were made by humans, and the earlier in a project something is written, the more impactful it is, if only because everything else has to work around it.

I’m still developing thinking on this, but it really does seem to be a game-changer. And I feel quite conflicted and even uncomfortable saying that out loud[^1].

Ah, so you’re an AI guy now

No, wait! No! Nope! Not at all. In fact, this revelation has pissed me off all the more. Think of it this way - this could have been good thing. This could have just been an organic evolution in the way that engineers write software. It didn’t have to be power-hungry, world-threateningly awful. It could have been responsibly trained by research labs and small companies making cool plugins for IDEs that run on laptops or an in-office server with a couple GPUs in it.

I really hope that people come around to this point of view, but my worry is that the AI industry had been desperately searching for something to make their scam seem legitimate and now that they’ve found a valuable use for it, they’re going to use that as leverage and squeeze. In this case, I expect that focus of the global engineering community will switch to optimising local agents that run well enough that we can eschew the giants. Enterprise will be training their own agents to host on their own infrastructure. I really don’t see the AI companies surviving another couple of years, but the technology will endure.

I also recommend reading AI as a Normal Technology, which has been around for a while, but I’ve just come around to this way of thinking.

[^1] well, out loud on the internet.