There's still something off with AI
One year after my article There's something off in the "dev with AI or die" narrative, the question I still get asked most about it is: "Have you changed your mind? Your dev job is toast. It must be." Let's ignore the fact that most people can't see beyond "Are you pro- or anti-AI? ANSWER!" and that I was already saying I use all these tools all day, every day. Have I changed my mind?
Now I've never been discussing the capabilities of LLMs and other AIs and even less predict. I'll admit that Opus 5.5 and 6.1 Sol are really good. But fundamentally, my position is the same: two truths coexist. The hype and the claims are absolutely overblown, AND we still haven't adapted socially to the changes that Sonnet 4-level capabilities brought a year ago. And while I don't question the capabilities of current models, I definitely question what we're doing with them.
The "I have Opus too" paradox and C-levels losing their minds
Not a day goes by without some "big shot" calling out their employees for not moving fast enough. The latest in the news was Proton. But a lot of us are living with this every day. It has become the latest toxic management behavior: whenever a senior dev pushes back, managers try to do it themselves in Claude. If they fail, they ask a prompting genius junior to see whether it can be done anyway. I've had to say "I have Opus too" (and, by the way, I burn more tokens than you) to people implying that I could just prompt "fix app, make better."
This is the dehumanization dream I talked about in my original article, now on steroids. If only we could tell engineers to do exactly as asked, without having them ask questions in return, express doubts, or flat-out refuse to build your genius feature… First of all, we're apparently still here. I won't dwell on this one, but the fact is that senior devs are still in demand.
But it's the second part that gets interesting. Let's imagine you really could prompt "Make a great, secure, private email service hosted in Europe. Purple." and get a working Proton. Then anybody with Opus could do exactly that, couldn't they, Mr. Genius CEO? AGI, ASI, whatever-BS-AI or not, if it really takes just a two-sentence prompt to conjure up a business, you'd better have the best two sentences in the world. Otherwise, you're the CEO of nothing, because, well, we all have Opus.
That's exactly what we're seeing with the current slopageddon wave. With the barrier to entry for coding a game or an app so low, surely some solo genius with a great idea but no coding skills will emerge with the best thing ever? Well, they haven't. We have a bajillion games popping up, indistinguishable from one another. Yet we don't really have the next McMillen or ConcernedApe. We haven't seen the "I made MyHouse.wad in a week" moment. What we certainly have is a lot of people who also have Opus and have all prompted more or less the same thing. At the very least it proves that the productivity gain are no shortcut for taste and execution. But it has me question it more: what is productivity and how do we measure gains on it, if all we ship are shitty product and decoherent system?
This is, IMO, how most overblown claims about dehumanization fall apart. We all have access to the same models anyway, so what's your edge? And what you thinks your job is more relevant than mine? Judgment? You just dismissed mine…
Yes, but frontier models haven't stopped improving, and small models haven't won!
I never expected to see a solution to the Navier–Stokes problem, I'll give you that. I studied it enough as a student to have a good understanding of why it's hard. It's mind-blowing, really. Yeeeeet… if you read how it unfolded, it's not all that surprising. Finding a counterexample under specific constraints is a task for which LLMs can be great tools, especially with expert guidance and north of $10 million worth of tokens.
But here again, dehumanization hits a wall and the great scientific achievement didn't instantly materialize into real world benefits, even academic ones. The mathematicians were quick to point out that they don't really know what to do with millions of lines of proof that nobody reads and nobody understands. Funnily enough, I was finishing Octome when the news came out. It's a novel in which the big AI has decided to go completely silent and just think to itself.
"Still, Opus 5.5 is too good, and small models aren't catching up!" Well, as far as LLMs are concerned, a Qwen 3.8 27B that you can run on a laptop falls somewhere between Sonnet 4.6 and Opus 4.6 for everyday coding. Those are models from a year ago. Had we waited one year and dropped them in our IDE, people would have been gone mad all the same. Some devs quit their jobs because they thought Opus 4.6 was better than they were. I'm using a good mix of Opus 5.5, 6.1 sol, and DeepSeek v4.1 Flash, and DeepSeek is such a workhorse. It's not that small, but people have found ways to run it on less than 1 GB of RAM. Not to mention models like Jev. I mean, how cool are real-time UIs?
Now, I'm no expert at ranking and evaluating models. What I can tell from my daily work, though, is that LLMs continue to be very LLM-ish and that a lot of bottlenecks are still here—and no, this is not review. But when quite a few high profile AI scientists sayed LLM were a dead end, they were probably pointing at what the sloppagedon is producing.
But are you pro- or anti-AI? ANSWER!
At the very least, I use quite a lot of it. At the very least, I wish it had rolled out differently.
I still think we're far from having adapted to what Sonnet 4.x brought. I'm still in the early stages of experimenting with having my designer/PO start the agent thread and taking over when needed. I'm still experimenting with designing directly in the real app rather than in a Figma file that drifts out of sync. I don't have a pipeline for using it to estimate the blast radius of changes across teams, etc. There so many UIs to rework beyond a chatbox, there are so many things we can do off screen that could benefit real UX design, not a fullblown self-managed dev VPS. Yes, a model like Jev requires significantly more work upfront, but has great potential.
So many processes should have been reworked by now. Yet we're mostly just collectively waiting for the next model to see whether "it will finally one-shot it" this time.
Had it rolled out in a more conventional way—where, say, a university released the core intelligence as public research and companies fine-tuned it on business data (or whatever…)—we would have had more time to adapt. Time to see where it works, where it doesn't, where it needs improvement, what new workflows and interfaces we need to develop, etc., and ultimately where it's worth training a better model. Maybe also more time to reflect on the fact that, just as we realized that logic and computational power aren't the whole of intelligence, PAC learning also has its limits, which we would do well to acknowledge more accurately.
But there's so much money flowing into it right now that it's genuinely hard to compare, judge, and adapt. I mean, pour trillions into anything and it'll look amazing at first glance, and sure, it'll move fast. But the end results are disappointing right now. Where are the mind-blowing products we should be seeing pop up everywhere? While we could be refining interfaces and ways of working together, we've instead spawned a battle royale of "let's all prompt the same thing and see who can raise billions." I should be excited about Git getting a healthy dose of innovation, but I'm put off by the fact that so many people are doing their own thing with seemingly zero cooperation.
I still think AI is one of the most fascinating scientific subjects right now. I still think Probably Approximately Correct is one of the greatest books ever. I still think my skills are relevant, and I see more evidence of that every day. What I doubt, though, is whether I want to participate in this tsunami of software. Is more software really what we need at this point? Do I need to waste more time, energy and resources to throw one more turd at the wall?
At the very least, I think my use of AI should be to ship less, more focused code and features.