We might all be AI engineers now

> But when you start comparing it to the surrounding code, there's a pretty big lack of coherency and it'll happily march down a very bad architectural path.

I had an idea earlier this week about this, but haven’t had a chance to try it. Since the agent can now “see” the whole stack, or at least most of it, by having access to the repos, there’s becoming less of a reason to suspect they won’t be able to take the whole stack into account when proposing a change.

The idea is that it’s like grep: you can call grep by itself, but when a match is found you only see one line per match, not any surrounding context. But that’s what the -A and -B flags are for!

So you could tell the agent that if its proposed solution lies at layer N of the system, it needs to consider at least layers N-1 (dependencies) and N+1 (consumers) to prevent the local optimum problem you mentioned.

The model should avoid writing a pretty solution in the application layer that conceals and does not address a deeper issue below, and it should keep whatever contract it has with higher-level consumers in good standing.

Anyway, I haven’t tried that yet, but hope to next week. Maybe someone else has done something similar and (in)validated it, not sure!

I'm sorry you've had that experience, and I agree there are a good share of "skeptics" who have latched on to anecdata or outdated experience or theorycrafting. I know it must feel like the goalposts are moving, too, when someone who was against AI on technical grounds last year has now discovered ethical qualms previously unevidenced. I spend a lot of time wondering if I've driven myself to my particular views exclusively out of motivated reasoning. (For what it's worth, I also think "motivated reasoning" is underrated - I am not obligated to kick my own ass out of obligation to "The Truth"!)

That said, I _did_ read your comments history (only because you asked!) and - well, I don't know, you seem very reasonable, but I notice you're upset with people talking about "hallucinations" in code generation from Opus 4.6. Now, I have actually spent some time trying to understand these models (as tool or threat) and that means using them in realistic circumstances. I don't like the "H word" very much, because I am an orthodox Dijkstraist and I hold that anthropomorphizing computers and algorithms is always a mistake. But I will say that like you, I have found that in appropriate context (types, tests) I don't get calls to non-existent functions, etc. However, I have seen: incorrect descriptions of numerical algorithms or their parameters, gaslighting and "failed fix loops" due to missing a "copy the compiled artifact to the testing directory" step, and other things which I consider at least "hallucination-adjacent". I am personally much more concerned about "hallucinations" and bad assumptions smuggled in the explanations provided, choice of algorithms and modeling strategies, etc. because I deal with some fairly subtle domain-specific calculations and (mathematical) models. The should-be domain experts a) aren't always and b) tend to be "enthusiasts" who will implicitly trust the talking genius computer.

For what it's worth, my personal concerns don't entirely overlap the questions I raised way above. I think there are a whole host of reasons people might be reluctant or skeptical, especially given the level of vitriol and FUD being thrown around and the fairly explicit push to automate jobs away. I have a lot of aesthetic objections to the entire LLM-generated corpus, but de gustibus...

I understand your concern. The copycat problem is real.

But if you come from a technical background and this is your first time building a product, you'll soon learn that it is so damn hard to get users, especially *paying* ones.

I was there. I built something, shared it, prayed people would notice. The truth is most of the time your product fails. Better explore the problem you are trying to solve first, share your idea if necessary, and collect feedback. You'll have a much clearer picture of what you need to do from there.

> I think these developers should face the fact that those skills and those ways they are reticent to give up are more or less obviated at this point.

Yes. We are this generations highly skilled artisans, facing our own industrial revolution.

Just as the skilled textile workers and weavers of early 19’th century Britain were correct when they argued this new automated product was vastly inferior, it matters not at all. And just as they were also correct, that the government of the day was doing nothing to protect the lives and livelihoods of those who had spent decades mastering a difficult set of professional skills (the middle class of the day), the government of this day will also do nothing.

And it doesn’t end with “IT”; anything that can be turned into a factory process with our new “thinking engines” will be. Perhaps we can do better in society this time around. I am not hopeful.

Your response is definitely on the top 5% of reasonableness from AI skeptics, so I appreciate that :-)

But, if you don't mind me going on a rant: the hallucinations thing. It kind of drives me nuts, because every day someone trots out hallucinations as some epic dunk that proves that AI will never be used in the real world or whatever. I totally hear you and think you are being a lot more reasonable than most (and thank you for that) -- you are saying that AI can get detail-oriented and fiddly math stuff wrong. But as I, my co-workers, and anyone who seriously uses AI in the industry all know, hallucinations are utterly irrelevant to our day-to-day.

My point is that hallucinations are irrelevant because if you use AI seriously for a while you quickly learn what it hallucinates on and what it does not, you build your mental model, and then you spend all your time on the stuff it doesn't hallucinate on, and it adds a fantastic amount of value there, and you are happy, and you ignore the things it is bad at, because why would you use a tool on things it is bad at? Hearing people talk about hallucinations in 2026 sounds to me like someone saying "a hammer will never succeed - I used it to smack a few screws and it NEVER worked!" And then someone added Hammer-doesnt-work-itis to Wikipedia and it got a few citations in Arxiv now it's all people can say when they talk about hammers online, omfg.

So when you say that I should spend more time asking "what do they see that I don't" - I feel quite confident I already know exactly what you see? You see that AI doesn't work in some domains. I quite agree with you that AI doesn't work in some domains. Why is this a surprise? Until 2023 it worked in no domains at all! There is no tool out there that works in every single domain.

But when you see something new, the much more natural question than "what doesn't this work on" is "what does this work on". Because it does work in a lot of domains, and fabulously well at that. Continuously bringing up how it doesn't work in some domain, when everyone is talking about the domains it does work, is just a non-sequitur, like if someone were to hop into a conversation about Rust and talk about how it can't solve your taxes, or a conversation about CSS to say that it isn't turing complete.