This was also my thought, and I think holds true of the ones with invisible requirements that aren't stated up front and are only captured in tests. Oh, you need to rework your solution to handle requirements nobody mentioned before? Well, me too.
Yeah, it's testing a different thing than what the benchmark claims to test, but it's also accidentally testing something more real-world applicable than a clean benchmark would be, so hey.
(EDIT: That is, if the agent is allowed to see the failed tests and iterate. If not, then yeah, that's just a problem. And either way, the ones with tests that just encode a particular solution's implementation details, thereby demanding that your solution have some rando internal details, are junkier. That's not a situation you'd run into in reality.)
I kinda love that you made this post feel negative enough that a bunch of AI skeptics are enthusiastically agreeing with a post suggesting that the realistic, pragmatic bear case for AI is, uh, 2-3x productivity improvements.
Go to chat.z.ai right now and ask it about what happened in Tiananmen Square. Do you think it's good for the world if software is written by the model that answers that question that way?
No, I know what it is. But it's not as good as Claude Code or Codex. Harnesses matter for coding agents, it's not just the model. There's a reason Microsoft had to make its developers stop using Claude Code, and it's because they all found that Copilot sucked in comparison.
I mean, literally the answer is that nobody knows. Maybe the robots replace us all. Maybe they shift those who remain into being some combination of Product Manager and QA. Maybe there's still a role for a technical overseer even in the medium-long run.
But it sounds like you're really asking about the state of the world today. If so, I don't think that ideal state is like your friend's company (or at least, as it appeared to be to you). It might be possible that you can make that "dark factory" pattern work (StrongDM seems to be doing it), but it would require infrastructure and discipline that I doubt they're mustering. Think about how CD didn't involve taking a sloppy build process with no testing or observability and just going straight to prod -- it required building up a lot of infra and discipline first.
But on the other hand, I don't think the ideal present involves artisan hand-crafting code either. I haven't written a line of code by hand in enough months that it would genuinely feel weird if I were to try to program that way despite decades of having done just that. That era's done with, and moderate normie practices right now today are more about supervising and guiding agents than about chiseling code into clay tablets.
10% of a developer's cost is something like $4000/month. Many companies are complaining at a point that's well, well below that.
(I think they are being irrational, and that the mental model they have of AI costs -- "how much are we spending on tooling for this developer?" -- is going to shift over time to something more sensible, but those kinds of short-sighted companies are the ones that are having cost panics.)
Which is why one of the big problems for the field right now is that a) most code bases still need someone more skilled than a mere robot driver, and b) many developers are not better than that.
In the past, a team of five mid devs and one good one would be fine, because that good one would ride herd on the mid ones. But now those mid ones are slamming out robot code that they're incapable of meaningfully reviewing (because it's better than they are already), and they're just overwhelming the good dev's capacity.
The solution, of course, is to fire them all -- they're worthless now -- but this is not going to happen quickly, and it's probably for the best that it doesn't.
Most code they write is obviously fine. Much of the rest isn't obviously fine, but is in fact fine once you've gone through understanding it. But yes, there's some that still benefits from a human eye.
(For as long as that's true, "software developer" is still a job. It's not clear for how long it will be true.)
Completely disagree. I think this is one of the big wins of agentic engineering. When you look back at your own completed change and realize that you made it too complicated because your initial abstraction was wrong, you have to debate long and hard about whether it's worth going back and redoing the work -- is the abstraction actually that bad? Would you really get a huge win by changing it, enough to justify spending another day on the task?
But with the agent, you know that the change will be relatively quick and easy, so the bar to tell it to shift approaches is much, much lower.
Yeah, but I think there's a difference here: If your coworker puts up code that you don't understand quickly, in most environments people give it an approval, as withholding approval is meant to indicate that there's a problem with the code. It's very rare that you'd actually force them to wait to merge until they've explained the code to your satisfaction.
(There are workplaces where that's the norm, I know -- it tends to be a thing with smaller teams with codebases that everyone understands fully, and much less a thing with larger teams where different people have areas of the code they understand more than others.)
With AI code, though, it's _your code_ and you can't give it a lgtm, you actually need to dig at it until you do fully understand it, fully agree with it, and could justify it to a hostile reviewer. It's a different level of rigor.
Not all engineers apply that rigor, though, which becomes a problem.
I was hoping that the web UI would be better -- I like Anthropic better than OpenAI from a values perspective and want to use their products, but ChatGPT in thinking mode has been just vastly better than claude.ai.So my fingers were crossed that these changes would bring it up to par.
But trying it out... alas, no. Simple factual questions where ChatGPT would go do a quick search and get the facts and report them back to me, get a "Great question! [totally invented bullshit]" from Claude, even with this new model and thinking set to high. I have to explicitly tell it to search to get it to look up basic facts, rather than it recognizing that it needs to do that, like GPT does.
5.2 still had a Codex variant, which this doesn't describe using. It also notably is not using the Codex harness -- it does everything with open source harnesses (which obviously are worse). And while it uses two harnesses with its cheap models, it only uses the worse-performing one of those with GPT 5.2 for cost reasons. (They also don't specify effort/thinking level used for GPT 5.2, but given that it performs worse in their baseline testing than obviously non-SOTA models, I'm guessing it wasn't set to anything high.)
Yeah, they conflate Microsoft's actions (which are not about cost) with a random quote from the "vice president of applied deep learning at Nvidia," who says that compute costs more than people on his team -- but his team isn't using LLMs for software development, they're literally a deep learning team that is burning compute in deep learning development ways.
If people would do even a little bit of math, they'd see that Microsoft can't possibly be paying more for AI than for developers: They have about 80K employees in product development roles. Senior developers probably cost them $400K all-in.
Do they have a $32 billion Claude bill? I suspect they do not.
The big thing on their roadmap is rearchitecting for something that can handle the increased load, though. Like, they're clearly paranoid that if they don't move fast, they're going to be just as busted as Github.
Yeah, it's testing a different thing than what the benchmark claims to test, but it's also accidentally testing something more real-world applicable than a clean benchmark would be, so hey.
(EDIT: That is, if the agent is allowed to see the failed tests and iterate. If not, then yeah, that's just a problem. And either way, the ones with tests that just encode a particular solution's implementation details, thereby demanding that your solution have some rando internal details, are junkier. That's not a situation you'd run into in reality.)