I strongly agree with that last statement—I hate using agents because their code smells awful even if it works. But I have to use them now because otherwise I’m going to wake up one day and be 100% obsolete and never even notice how it happened.
Today I gave a lecture to my undergraduate data structures students about the evolution of CPU and GPU architectures since the late 1970s. The main themes:
- Through the last two decades of the 20th century, Moore’s Law held and ensured that more transistors could be packed into next year’s chips that could run at faster and faster clock speeds. Software floated on a rising tide of hardware performance so writing fast code wasn’t always worth the effort.
- Power consumption doesn’t vary with transistor density but varies with the cube of clock frequency, so by the early 2000s Intel hit a wall and couldn’t push the clock above ~4GHz with normal heat dissipation methods. Multi-core processors were the only way to keep the performance increasing year after year.
- Up to this point the CPU could squeeze out performance increases by parallelizing sequential code through clever scheduling tricks (and compilers could provide an assist by unrolling loops) but with multiple cores software developers could no longer pretend that concurrent programming was only something that academics and HPC clusters cared about.
CS curricula are mostly still stuck in the early 2000s, or at least it feels that way. We teach big-O and use it to show that mergesort or quicksort will beat the pants off of bubble sort, but topics like Amdahl’s Law are buried in an upper-level elective when in fact it is much more directly relevant to the performance of real code, on real present-day workloads, than a typical big-O analysis.
In any case, I used all this as justification for teaching bitonic sort to 2nd and 3rd year undergrads.
My point here is that Simon’s assertion that “code is cheap” feels a lot like the kind of paradigm shift that comes from realizing that in a world with easily accessible massively parallel compute hardware, the things that matter for writing performant software have completely shifted: minimizing branching and data dependencies produces code that looks profoundly different than what most developers are used to. e.g. running 5 linear passes over a column might actually be faster than a single merged pass if those 5 passes touch different memory and the merged pass has to wait to shuffle all that data in and out of the cache because it doesn’t fit.
What all this means for the software development process I can’t say, but the payoff will be tremendous (10-100x, just like with properly parallelized code) for those who can see the new paradigm first and exploit it.
It definitely wasn’t a waste of time! I passed JLPT N1 back in 2014 after ~6 years of mostly Anki-based studying. Did Heisig’s RtK first and then mostly played old Japanese console games that I was familiar with. Never opened a JLPT study guide and passed the test on my first attempt.
Could I speak Japanese at that point? No not really… I even had a Japanese spouse! But we spoke mostly English at home. I could read quite well, but conversation was very challenging.
Then we moved to Japan. Despite not having a job that requires me to speak Japanese, I got enough live exposure just from chatting with people at the gym or in social activities that now, a few years later, I’ve backfilled all that conversational fluency that was missing. No special extra effort required, just living in an environment where I used the language reasonably often.
Anyways, the point is that all the time spent in Anki laid a rock-solid foundation that merely needed activation in the right environment for active fluency to emerge. Of course I no longer do my daily flashcard drills (and I’ve forgotten how to write quite a few kanji as a result) but the work paid off.
Miraikan is one of our favorites, been there like 3-4 times with my son. The current exhibit that turns quantum logic gates into a DJ game is really innovative but they only give you like 5 mins which is barely enough time to figure out WTF is even going on
For all the disparagement of “fact regurgitation” as pedagogical practice, it’s not like there’s some proven better alternative. Higher-order reasoning doesn’t happen without a thorough catalogue of domain knowledge readily accessible in your context window.
Your last statement misses the mark—of course the brain is the root of human intelligence. The error is in assuming that consciousness is the primary learning modality. Or, as you put it, “arguing semantics”.
From my own personal experience, this realization came after finally learning a difficult foreign language after years and years of “wanting” to learn it but making little progress. The shift came when I approached it like learning martial arts rather than mathematics. Nobody would be foolish enough to suggest that you could “think” your way to a black belt, but we mistakenly assume that skills which involve only the organs in our head (eyes, ears, mouth) can be reduced to a thought process.
Your analogy to PHP developers not reading assembly got me thinking.
Early resistance to high-level (i.e. compiled) languages came from assembly programmers who couldn’t imagine that the compiler could generate code that was just as performant as their hand-crafted product. For a while they were right, but improved compiler design and the relentless performance increases in hardware made it so that even an extra 10-20% boost you might get from perfectly hand-crafted assembly was almost never worth the developer time.
There is an obvious parallel here, but it’s not quite the same. The high-level language is effectively a formal spec for the abstract machine which is faithfully translated by the (hopefully bug-free) compiler. Natural language is not a formal spec for anything, and LLM-based agents are not formally verifiable software. So the tradeoffs involved are not only about developer time vs. performance, but also correctness.
For a great many software projects no formal spec exists. The code is the spec, and it gets modified constantly based on user feedback and other requirements that often appear out of nowhere. For many projects, maybe ~80% of the thinking about how the software should work happens after some version of the software exists and is being used to do meaningful work.
Put another way, if you don't know what correct is before you start working then no tradeoff exists.
> Put another way, if you don't know what correct is before you start working then no tradeoff exists.
This goes out the window the first time you get real users, though. Hyrum's Law bites people all the time.
"What sorts of things can you build if you don't have long-term sneaky contracts and dependencies" is a really interesting question and has a HUGE pool of answers that used to be not worth the effort. But it's largely a different pool of software than the ones people get paid for today.
> For many projects, maybe ~80% of the thinking about how the software should work happens after some version of the software exists and is being used to do meaningful work.
Some version of the software exists and now that's your spec. If you don't have a formal copy of that and rigorous testing against that spec, you're gonna get mutations that change unintended things, not just improvements.
Users are generally ok with - or at least understanding of - intentional changes, but now people are talking about no-code-reading workflows, where you just let the agents rewrite stuff on the fly to build new things until all the tests pass again. The in-code tests and the expectations/assumptions about the product that your users have are likely wildly different - they always have been, and there's nothing inherent about LLM-generated code or about code test coverage percentages that change this.
> So the tradeoffs involved are not only about developer time vs. performance, but also correctness.
The "now that producing plausible code is free, verification becomes the bottleneck" people are technically right, of course, but I think they're missing the context that very few projects cared much about correctness to begin with.
The biggest headache I can see right now is just the humans keeping track of all the new code, because it arrives faster than they can digest it.
But I guess "let go of the need to even look at the code" "solves" that problem, for many projects... Strange times!
For example -- someone correct me if I'm wrong -- OpenClaw was itself almost entirely written by AI, and the developer bragged about not reading the code. If anything, in this niche, that actually helped the project's success, rather than harming it.
(In the case of Windows 11 recently.. not so much ;)
> The "now that producing plausible code is free, verification becomes the bottleneck" people are technically right, of course, but I think they're missing the context that very few projects cared much about correctness to begin with.
It's certainly hard to find, in consumer-tech, an example of a product that was displaced in the market by a slower moving competitor due to buggy releases. Infamously, "move fast and break things" has been the rule of the land.
In SaaS and B2B deterministic results becomes much more important. There's still bugs, of course, but showstopper bugs are major business risks. And combinatorial state+logic still makes testing a huge tarpit.
The world didn't spend the last century turning customer service agents and business-process-workers into script-following human-robots for no reason, and big parts of it won't want to reintroduce high levels of randmoness... (That's not even necessarily good for any particular consumer - imagine an insurance company with a "claims agent" that got sweet talked into spending hundreds of millions more on things that were legitimate benefits for their customers, but that management wanted to limit whenever possible on technicalities.)
OK but, I've definitely read the assembly listings my C compiler produced when it wasn't working like I hoped. Even if that's not all that frequent it's something I expect I have to do from time to time and is definitely part of "programming".
It's also important to remember that vibe coders throw away the natural language spec each time they close the context window.
Vibe coding is closer to compiling your code, throwing the source away and asking a friend to give you source that is pretty close to the one you wrote.
Imagine if high level coding worked like: write a first draft, and get assembly. All subsequent high level code is written in a repl and expresses changes to the assembly, or queries the state of the assembly, and is then discarded. only the assembly is checked into version control.
Or the opposite, all applications are just text files with prompts in them and the assembly lives as ravioli in many temp files. It only builds the code that is used. You can extend the prompt while using the application.
People choose to hold non-yield-bearing assets when they believe the returns offered by current investment opportunities are not sufficient to justify the risks.
It is the miracle of modern capital markets that enables almost anyone to quickly and easily invest their savings in productive assets, but of course capital markets aren’t perfect. The availability of “none of the above” options (like gold) that remove savings from the pool of active investment capital is the essential feedback loop that balances risk and return.
