Shirayuki2
new account of Shirayuki, lost old password
User ID: 4180
I'm sympathetic to the idea that if you're doing anything large or complex, a sufficiently detailed spec can boil down to writing pseudo-code in English, and that verification and optimization of generated code for anything non-trivial is still quite difficult.
I do find the recent "SWE is over!!!" boosterism pretty annoying as well.
However, I will say that if your criticism is on the level of getting syntax errors, and you aren't doing anything extremely out of distribution (Brainfuck, Malbolge, things like that), there is probably something you are doing wrong.
Provide the agent with tests (or have it write them itself with oversight) and I would be extremely surprised if you are still getting non-compiling code from any of the large models/harnesses released in the last few months.
I am curious how you evaluate these life trajectories from the perspective of AI timelines.
I know you're pretty bullish on AI; even if you don't subscribe to short 5-10 year timelines, it would seem to me that if you believe in transformational AI, there would be a very high probability you would see some sort of strong AI by the time you are 60 or 70.
It would seem to me, then, that the optimal strategy is to front-load your life trajectory by not having children early. If AI goes poorly, then at least you managed to enjoy your 20s and perhaps your 30s without having to deal with the suffering of your children, while if it goes well it's unlikely that your age would be a barrier to having children, and indeed it would be much easier to raise children in such a world.
My impression is that the SOTA agents (Claude Code or Codex) can now, if given a sufficiently detailed specification and sufficient tools to validate their work, provide code that fully satisfies that spec for the vast majority of moderately long coding tasks now. This is a big jump in capabilities and obviously very powerful.
This does come with a lot of caveats that imo make the claims of 100% AI written code misleading at best; the creation of "sufficiently detailed" specs (which is famously difficult) fed to the agent and validation of the generated code is still dependent on engineering skill, and one-shotted code will generally be of worse quality than code written for the same purpose hand-crafted by a skilled engineer.
In the short term I think this is actually bullish for SWE, coding agent output is still bottlenecked by engineering skill and I don't believe we've yet saturated the demand for software. In the mid-long term it's unclear if there will still be value in having engineers in the loop, but imo end-to-end automation of software engineering is/requires AGI so my job isn't high up on my list of concerns in that scenario.
IMO a lot of this effect is that you can get a lot more out of agents if you already know what you're doing, and a lot of people assume that the median person knows a lot more about their own niche than they actually do (see XKCD 2501). Right now you still need a lot of specialist knowledge to make the most out of your agent.
Even in your original post, you need to have enough knowledge to understand when to use io_uring, when you should drop down to assembly, and understand when and why you might want a JIT for your app. It's not like my mother is doing any of that even if I give her access to Claude Code lol.
As I mentioned in my other post I am quite unsure whether the reverse centaur phase will last, but at the point where someone completely non-technical is outputting the same amount and quality of software as a SWE, requirements and all, I think this is already AGI and either everyone or no-one is cooked.
It's really not worth worrying about that scenario too much imo, at least from an employment perspective.
More options
Context Copy link