I'm almost 50, and have been writing code professionally since the late 90s. I can pretty much see projects in my head, and know exactly what to build. I also get paid pretty well for what I do. You'd think I'd be the prototype for anti-AI.
I'm not.
I can build anything, but often struggle with getting bogged down with all the basic work. I love AI for speed running through all the boring stuff and getting to the good parts.
I liken AI development to a developer somewhere between junior and mid-level, someone I can given a paragraph or two of thought out instructions and have them bang out an hour of work. (The potential for then stunting the growth of actual juniors into tomorrow's senior developers is a serious concern, but a separate problem to solve)
I love AI for speed running through all the boring stuff and getting to the good parts.
In some cases, especially with the more senior devs in my org, fear of the good parts is why they're against AI. Devs often want the inherent safety of the boring, easy stuff for a while. AI changes the job to be a constant struggle with hard problems. That isn't necessarily a good thing. If you're actually senior by virtue of time rather than skill, you can only take on a limited number of challenging things one after another before you get exhausted.
Companies need to realise that AI to go faster is great, but there's still a cognitive impact on the people. A little respite from the hardcore stuff is genuinely useful sometimes. Taking all of that away will be bad for people.
That said, some devs hate the boring easy bits and will thrive. As with everything, individuals need to be managed as individuals.
That makes me think of https://store.steampowered.com/app/2262930/Bombe/ which is a version of Minesweeper where instead of clicking on squares you define (parametric!) rules that propagate information around the board automatically. Your own rules skip all the easy parts for you. As a result, every challenge you get is by definition a problem that you've never considered before. It's fun, but also exhausting.
There's also the fact that, while you're coding the easy stuff, your mind is thinking about the hard stuff, looking things up, seeing how they articulate. If you're spending 100% of your time on hard stuff, you might be hurting these preliminaries.
> Devs often want the inherent safety of the boring, easy stuff for a while.
Part of the problem is that in big orgs, you need to show consistent progress in order to not get put on some PIP and kicked out of the company. There are performance review cycles and you have to show something continuously.
That ONLY works if you have boring, easy work. It's easy to deliver consistent progress on that.
Interesting and difficult work is nice only if you are trusted to try your best and given the freedom to fail. That's the nature of hard problems; progress in those domains is very sudden and Poissonian and not consistent by nature. If you're going to be judged on your ability to be sub-Poissonian and consistent, and get put on a PIP for not succeeding at it one review cycle (and possibly risking income that you use to put a roof over your head or feed your family) it's not worth the career risk to try difficult things.
Not saying this is the way I think, it's just the reality of how things often work in big orgs, and one of the reasons I dislike many big orgs.
> Devs often want the inherent safety of the boring, easy stuff for a while
That's matches my experience. In my first job, every time a new webapp project has been starting it was fun. Not because of challenges or design, but simply because of the trivial stuff done for n-th time - user accounts, login, password reset, admin panel. Probably should have been automated at this point, but we got away with reinventing the wheel each time.
>AI changes the job to be a constant struggle with hard problems
I find this hilarious. From what I've seen watching people do it, it changes the job from deep thought and figuring out a good design to pulling a lever on a slot machine and hoping something good pops out.
The studies that show diminished critical thinking have matched what i saw anecdotally pairing with people who vibe coded. It replaced deep critical thinking with a kind of faith based gambler's mentality ("maybe if i tell it to think really hard it'll do it right next time...").
The only times ive seen a notable productivity improvement is when it was something not novel that didnt particularly matter if what popped out was shit - e.g. a proof of concept, ad hoc app, something that would naturally either work or fail obviously, etc. The buzz people get from these gamblers' highs when it works seems to make them happier than if they didnt use it at all though.
> AI changes the job to be a constant struggle with hard problems.
Very true. I think AI (especially Claude Code) forced me to actually think hard about the problem at hand before implementing the solution. And more importantly, write down my thoughts before they fleet away from my feeble mind. A discipline that I wished I had before.
This is an interesting take on it. I've been using Cursor at work and notice that days where I have it generate all the "easy parts", I end up feeling much more mentally exhausted than if I spend the day writing everything by hand.
Yes I can see why some devs might prefer the safety of the boring and the familiar stuff. But the employers aren’t going to care about that. They’re going to hire and retain the devs who are more productive in the new era.
On the flip side, there have been lots of times where I personally didn’t have a lot of time to deeply research a topic (read papers, build prototypes of different ideas, etc) due to lack of time and resources. If all of the boring stuff is gone, and building prototypes is also 3x faster maybe what will end up happening is we can now use all of this free time to try lots of different ideas because the cost of exploration has been lowered.
I think you're describing things we already knew long before this era of AI. Less code is better code, and the vast majority of bugs come from the devs who "hate the boring easy bits".
I disagree that this has anything to do with people needing a break. All code eventually has to be reviewed. Regardless of who or what wrote it, writing too much of it is the problem. It's also worth considering how much more code could be eliminated if the business more critically planned what they think they want.
These tensions have existed even before computers and in all professions.
That's crazy to me. I solve problems. I'm not a janitor or tradesman, you bring me in to envision and orchestrate solutions that bring bottom line value. I live to crack hard nuts, if I never have to bother with rigging again I'll be so happy.
> In some cases, especially with the more senior devs in my org, fear of the good parts is why they're against AI. Devs often want the inherent safety of the boring, easy stuff for a while. AI changes the job to be a constant struggle with hard problems. That isn't necessarily a good thing. If you're actually senior by virtue of time rather than skill, you can only take on a limited number of challenging things one after another before you get exhausted.
The issue of senior-juniors has always been a problem; AI simply means they're losing their hiding spots.
I'm just slightly younger than you, but have the exact same sentiment. Hell, even moreso maybe, because what I realized is that "writing code to implement interesting ideas" is not really what I enjoy - it's coming up with the interesting ideas and experimenting with them. I couldn't care less about writing the code - and I only did it because I had to...if I wanted to see my idea come to life.
AI has also been a really good brainstorming partner - especially if you prompt it to disable sycophancy. It will tell you straight up when you are over-engineering something.
It's also wonderful at debugging.
So I just talk to my computer, brainstorm architectures and approaches, create a spec, then let it implement it. If it was a bad idea, we iterate. The iteration loop is so fast that it doesn't matter.
Did you end up regretting a design choice, but normally you'd live with it because so much code would have to be changed? Not with Agentic coding tools - they are great at implementing changes throughout the entire codebase.
And its so easy to branch out to technologies you're not an expert in, and still be really effective as you gain that expertise.
I honestly couldn't be happier than I am right now. And the tools get better every week, sometimes a couple times a week.
> I can build anything, but often struggle with getting bogged down with all the basic work. I love AI for speed running through all the boring stuff and getting to the good parts.
I'm in the same boat (granted, 10 years less) but can't really relate with this. By the time any part becomes boring, I start to automate/generalize it, which is very challenging to do well. That leaves me so little boring work that I speed run through it faster by typing it myself than I could prompt it.
The parts in the middle – non-trivial but not big picture – in my experience are the parts where writing the code myself constantly uncovers better ways to improve both the big picture and the automation/generalization. Because of that, there are almost no lines of code that I write that I feel I want to offload. Almost every line of code either improves the future of the software or my skills as a developer.
But perhaps I've been lucky enough to work in the same place for long. If I couldn't bring my code with me and had to constantly start from scratch, I might have a different opinion.
> By the time any part becomes boring, I start to automate/generalize it, which is very challenging to do well. That leaves me so little boring work that I speed run through it faster by typing it myself than I could prompt it.
The two aren't mutually exclusive. You can use AI to build your tooling. (Unless it's of sufficient complexity or value that you need to do the work yourself)
I have a couple of niche areas of non-coding interest where I'm using AI to code. It is so amazing to write rust and just add `todo!(...)` through out the boiler plate. The AI is miserable at implementing domain knowledge in those niche areas, but now I can focus on describing the domain knowledge (in real rust code because I can't describe it precisely enough in English + pseudo code), and then say "fill in the todos, write some tests make sure it compiles, and passes linting", verify the tests check things properly and I'm done.
I've struggled heavily trying to figure out how to get it to write the exactly correct 10 lines of code that I need for a particularly niche problem, and so I've kind of given up on that, but getting it to write the 100 lines of code around those magic 10 lines saves me so much trouble, and opens me up to so many more projects.
I find it best as a "personal assistant," that I can use to give me information -sometimes, highly focused- at a moment's notice.
> The potential for then stunting the growth of actual juniors into tomorrow's senior developers is a serious concern
I think it's a very real problem. I am watching young folks being frozen out of the industry, at the very beginning of their careers. It is pretty awful.
I suspect that the executives know that AI isn't yet ready to replace senior-levels, but they are confident that it will, soon, so they aren't concerned that there aren't any more seniors being crafted from youngsters.
Exactly. I tend to like Hotz, but by his description, every developer is also "a compiler", so it's a useless argument.
My life quality (as a startup cofounder wearing many different hats across the whole stack) would drop significantly if Cursor-like tools [1] were taken away from me, because it takes me a lot of mental effort to push myself to do the boring task, which leads to procrastination, which leads to delays, which leads to frustration. Being able to offload such tasks to AI is incredibly valuable, and since I've been in this space from "day 1", I think I have a very good grasp on what type of task I can trust it to do correctly. Here are some examples:
- Add logging throughout some code
- Turn a set of function calls that have gotten too deep into a nice class with clean interfaces
- Build a Streamlit dashboard that shows some basic stats from some table in the database
- Rewrite this LLM prompt to fix any typos and inconsistencies - yeah, "compiling" English instructions into English code also works great!
- Write all the "create index" lines for this SQL table, so that <insert a bunch of search usecases> perform well.
[1] I'm actually currently back to Copilot Chat, but it doesn't really matter that much.
That's one of the thing that I wouldn't delegate to LLM. Logging is like a report of things that happens. And just like a report, I need relevant information and the most useful information.
...
A lot of these use cases actually describes the what. But the most important questions is always the why. Why is it important to you? Or to the user? That's when things have a purpose and not be just toys.
Exactly. If you know how the whole thing works end to end, AI makes you incredibly dangerous. Anyone who specializes or never really learned how everything works is at a huge disadvantage.
However. There's also good news. AI is also an amazing tool for learning.
So what I see AI doing is simply separating people who want to put effort forth and those who don't.
> However. There's also good news. AI is also an amazing tool for learning.
Absolutely. For example, I've been learning Autodesk Fusion, and after establishing a small foundation through traditional learning techniques, I've been able to turbocharge my learning by asking precise questions to AI.
I image this really sucks for those whose business model relied on gatekeeping knowledge. (like training companies)
I agree. Having AI write your entire application is equivalent to having it write a song for your band, or even generate audio so you don't have to do record it yourself.
If you aren't talented enough to write or record your own music, you aren't really a musician.
If you have a quick question about music theory and you want a quick answer, AI can be a benefit.
> I can pretty much see projects in my head, and know exactly what to build.
This is where AI actually helps - you have a very precise vision of what you want, but perhaps you've forgotten about the specific names of certain API methods, etc. Maybe you don't want to implement all the cases by hand. Often validating the output can take just seconds when you know what it is you're looking for.
The other part of making the output do what you want is the ability to write a prompt that captures the most essential constraints of your vision. I've noticed the ability to write and articulate ideas well in natural language terms is the actual bottleneck for most developers. It takes just as much practice communicating your ideas as it does anything else to get good at it.
Yes, unfortunately the boring parts are what junior devs used to do so the senior devs could work on the good stuff. Now that AI is doing the boring stuff nobody has to hire those pesky jr developers anymore. Yay?
The problem is that junior developers are what we make senior developers with— so in 15 years, this is going to be yet another thing that the US used to be really good at, but is no longer capable of doing, just like many important trades in manufacturing. The manufacturers were all only concerned with their own immediate profit and made the basic sustainability of their workforce, let alone the health of the trades that supported their industries, a problem for everyone else to take care of. Well, everyone else did the same thing.
It’s yet another place where we know our own capacity as a society is shrinking and hoping that ??? (Ai? Robots? Fusion?) will fix it before it’s too late. I never thought programming would join elder-care in this category though, that came as a surprise.
This resonates with me. I'm also around the same age and have the same amount of experience.
I love AI and use it for both personal and work tasks for two reasons:
1. It's a way to bounce around ideas without (as much) bias as a human. This is indispensable because it gives you a fast feedback mechanism and validates a path.
2. It saves me typing and time. I give it one-shot, "basic work" to do and it's able to do accomplish at least 80% of what I'd say is complete. Although it may not be 100% it's still a net positive given the amount of time it saves me.
It's not lost on me that I'm effectively being trained to always add guardrails, be very specific about the instructions, and always check the work of AI.
Most of my work is for my employer, but the bigger point is that you wouldn't be able to tell my "AI work" from my other work because I primarily use it for the boring stuff that is labor-intensive, while I work on the actual business cases. (Most of my work doesn't fall under the category of "web application", but rather, backend and background-processing intensive work that just happens to have an HTML front-end)
Yes, I wrote my own DSL, complete with BDD testing framework, to write my blog with. In Rust!
GET /hello/:world
|> jq: `{ world: .params.world }`
|> handlebars: `<p>hello, {{world}}</p>`
describe "hello, world"
it "calls the route"
when calling GET /hello/world
then status is 200
and output equals `<p>hello, world</p>`
100% agree. I am interested in seeing how this will change how I work. I'm finding that I'm now more concerned with how I can keep the AI busy and how I can keep the quality of outputs high. I believe it has a lot to do with how my projects are structured and documented. There are also some menial issues (e.g. structuring projects to avoid merge conflicts becoming bottlenecks)
I expect that in a year my relationship with AI will be more like a TL working mostly at the requirements and task definition layer managing the work of several agents across parallel workstreams. I expect new development toolchains to start reflecting this too with less emphasis on IDEs and more emphasis on efficient task and project management.
I think the "missed growth" of junior devs is overblown though. Did the widespread adoption of higher-level really hurt the careers of developers missing out on the days when we had to do explicit memory management? We're just shifting the skillset and removing the unnecessary overhead. We could argue endlessly about technical depth being important, but in my experience this hasn't ever been truly necessary to succeed in your career. We'll mitigate these issues the same way we do with higher-level languages - by first focusing on the properties and invariants of the solutions outside-in.
An important skill for software developers is the ability to reason about what the effects of their coce will be, over all possible conditions and inputs, as opposed to trial and error limited to specific inputs, or (as is the case with non-deterministic LLMs) limited to single executions. This skill is independent of whether you are coding in assembly or are using higher-level languages and tooling. Using LLMs exactly doesn’t train that skill, because the effective unpredictability of their results largely prevents any but the most vague logical reasoning about the connection between the prompt and the specific output.
>> developer somewhere between junior and mid-level,
Analogies to humans don't work that well. AI is super-human in some respects while also lacking the ability to continually work toward a goal over long periods of time. AI can do very little on its own - just short / scoped / supervised tasks.
However, sometimes the situation is reversed, AI is the teacher who provides some examples on how to do things or provides hints on how to explore a new area and knows how others have approached similar things. Then, sometimes, AI is an astute code reviewer, typically providing valuable feedback.
Anyway, I've stopped trying anthropomorphize AI and simply try to reason about it based on working with it. That means combinations of direct ChatGPT usage with copy / paste / amend type workflows, async style / full PR style usage, one-shot "hail Mary" type throw away PRs just to establish an initial direction as well as PR reviews of my own code. I'm using AI all the time, but never anything like how I would work with another human.
> I love AI for speed running through all the boring stuff and getting to the good parts.
But the issue is some of that speedrunning sometimes takes so much time, it becomes inefficient. It's slowly improving (gpt5 is incredible), but sometimes it get stuck on really mundane issue, and regress endlessly unless I intervene. And I am talking about straightforwars functional code.
I've used many different ones, and find the result pretty similar. I've used Copilot in VS Code, Chat GPT stand-alone, Warp.dev's baked in tools, etc. Often it's a matter of what kind of work I'm doing, since it's rarely single-mode.
> I can pretty much see projects in my head, and know exactly what to build.
I think you’re the best case support for AI coding. You know clearly what you want, so you know clearly what you don’t want. So if you had decent verbal dexterity you could prompt the AI model and manage to accomplish what you intended.
A lot of programming problems / programmer contexts don’t match that situation. Which is the problem with universalizing the potency of AI / benefits of AI coding.
same. 50’s, coding since the 90’s, make more $$ than I can spend in 3 lifetimes. one constant thing in my 3 decade career has been that absolute best people I worked with all had one common thread - absolute laziness. sounds strange but every other trait of a great SWE has not been universal except laziness.
the laziness manifest itself into productivity as crazy as this sounds. how? lazy people find a way to automate repetitive tasks. what I have learned from these over the years is that anything you do twice has to find a way to be automated as third time is around the corner :)
what does this have to do with AI? the AI has taken automation to another level allowing us to automate so much of our work that was not previously possible. I found myriad of ways to use AI and several of my best (lazy) co-workers have as well. I cannot imagine doing my work anymore without it, not because of any “magic” but because my lazy ass will be able to do all the things that I have automated out
Yes! Once I've figured out that this problem is best solved using parser combinators, and that I have a good idea of how to model the transformation, I'm so glad I can delegate work to the LLM code gen and focus on the specification, test cases, corner cases, etc;
I don't think thats contrary to the article's claim: the current tools are so bad and tedious to use for repetitive work that AI is helpful with a huge amount of it.
That‘s exactly why i like AI too. I even let them play roles like „junior dev“, „product owner“ or „devops engineer“ and orchestrate them, to play together as a team - with guidance from me (usually the „solution architect“ or „investor“)! This „team“ achieves in weeks what we usually needed months for - for 2.40€/h*role!
I have a similar relation to AI with programming -- and my sense is very many HN readers do as well, evidenced not least by the terrific experience report from antirez [1]. Yet it is rare to see such honest and open statements even here. Instead, HN is full of endless anti-AI submissions on the front page where the discussion underneath is just an echo chamber of ill-substantiated attacks on 'AI hype' and where anything else is down-voted.
It's what is, to me, so bizarre about the present moment: certainly investment is exceptionally high in AI (and of course use), but the dominant position in the media is precisely such a strange 'anti-AI hype' that positions itself as a brave minority position. Obviously, OpenAI/Altman have made some unfortunate statements in self-promotion, but otherwise I genuinely can't think of something I've read that expresses the position attacked by the anti-AI-ers -- even talk of 'AGI' etc comes from the AI-critical camp.
In a sense, the world seems divided into three: obvious self-promotion from AI companies that nobody takes seriously, ever-increasingly fervent 'AI critique', and the people who, mostly silent, have found modern AI with all its warts to be an incomparably useful tool across various dimensions of their life and work. I hope the third camp becomes more vocal so that open conversations about the ways people have found AI to be useful or not can be the norm not the exception.
It’s hard to see how the current rate of progress is compatible with, 30 years from now, it being good business sense to pay human professionals six figure salaries. Opinions then split: the easiest option is pure denial, to assume that the current rate of progress doesn’t exist. Next easiest is to assume that progress will halt soon, then that we will be granted the lifestyle of well paid professionals when unsupervised AI can do our job for cheaper, then that Altman will at least deign to feed us.
antirez's success doesn't mean anything since antirez is far, far above average as a developer and is able to wield AI very effectively because of his talents and knowledge. Most developers are not remotely close to antirez's skill level (half are below average, by definition) and are likely going to suffer the problems that skeptics are already seeing.
> developer somewhere between junior and mid-level
Why the insistence on anthropomorphizing what is just a tool? It has no agency, does not 'think' in any meaningful manner, it is just pattern matching on a vast corpus of training data. That's not to say it can't be very useful - as you seem to have found - but it is still just a tool.
This isn’t necessarily anthropomorphizing, as from a company’s point of view, or uncharitably even from a tech leads point of view, developers are also just tools. The point being made is that LLMs (supposedly) fulfill those developer roles.
It's less about what the tool is, and more about the kind of work we often assign to less experienced developers. Pattern matching in meatspace is still pattern matching.
This is a more extreme example of the general hacker news group think about AI.
Geohot is easily a 99.999 percentile developer, and yet he can’t seem to reconcile that the other 99.999 percent are doing something much more basic than he can ever comprehend.
It’s some kind of expert paradox, if everyone was as smart and capable as the experts, then they wouldn’t be experts.
I have come across many developers that behave like the AI. Can’t explain codebases they’ve built, can’t maintain consistency.
It’s like a aerospace engineer not believing that the person that designs the toys in an Kinder egg doesn’t know how fluid sims work.
I’m surprised to see this. From my perspective, reading comments and seeing which posts rise to the top, HN as a whole seems pretty bullish on the tech as whole…
I think there might also just be a vocal minority and/or some astroturfing hyping AI coding around hackernews. I personally keep trying all the latest shit and always stop using it, because it actively slows me down.
It’s changing over time. When copilot came out a few years ago, people were very against it due to it being trained oh GitHub codebase. Now there’s more support around it.
We don't want to admit it but HN has similar characteristics as many other platforms; echo chamber / group think. You see it over and over again.
HN participants (generally speaking) are against: AI, Crypto, HFT. I've worked in 2/3 of these industries so have first hand experience. My basic summary is that the average commenter here has a lot of misinformation on these topics (as a insider).
It also seems wildly inconsistent for someone who founded a self driving company. If models can't write code, which is a very constrained problem space, how are they supposed to drive?
I don’t know. He covers this pretty early in the post.
> The only reason it works for many common programming workflows is because they are common. The minute you try to do new things, you need to be as verbose as the underlying language.
"The best model of a programming AI is a compiler... You give it a prompt, which is “the code”, and it outputs a compiled version of that code. Sometimes you’ll use it interactively, giving updates to the prompt after it has returned code, but you find that, like most IDEs, this doesn’t work all that well and you are often better off adjusting the original prompt and “recompiling”."
I think his excellency in his own trade limited his vision for the 99% who just want to get by in the job. How many dev even deal with compiler directly these days? They write some code, fix some red underlines, then push, pray and wait for pipeline pass. LLMs will be gods in this process, and you can even beg another one if your current one does not work best.
First, the assertion that the best model of "AI coding" is that it is a compiler. Compilers deterministically map a formal language to another under a spec. LLM coding tools are search-based program synthesizers that retrieve, generate, and iteratively edit code under constraints (tests/types/linters/CI). That’s why they can fix issues end-to-end on real repos (e.g., SWE-bench Verified), something a compiler doesn’t do. Benchmarks now show top agents/models resolving large fractions of real GitHub issues, which is evidence of synthesis + tool use, not compilation.
Second, that the "programming language is English". Serious workflows aren’t "just English." They use repo context, unit tests, typed APIs, JSON/function-calling schemas, diffs, and editor tools. The "prompt" is often code + tests + spec, with English as glue. The author attacks the weakest interface, not how people actually ship with these tools.
Third, non-determinism isn't disqualifying. Plenty of effective engineering tools are stochastic (fuzzers, search/optimization, SAT/SMT with heuristics). Determinism comes from external specs: unit/integration tests, type systems, property-based tests, CI gates.
False dichotomy: "LLMs are popular only because languages/libraries are bad."
Languages are improving (e.g. Rust, Typescript), yet LLMs still help because the real bottlenecks are API lookup, cross-repo reading, boilerplate, migrations, test writing, and refactors, the areas where retrieval and synthesis shine. These are complementary forces, not substitutes.
Finally, no constructive alternatives are offered. "Build better compilers/languages" is fine but modern teams already get value by pairing those with AI: spec-first prompts, test-gated edits, typed SDK scaffolds, auto-generated tests, CI-verified refactors, and repo-aware agents.
A much better way to think about AI coding and LLMs is that they aren’t compilers. They’re probabilistic code synthesizers guided by your constraints (types, tests, CI). Treat them like a junior pair-programmer wired into your repo, search, and toolchain. But not like a magical English compiler.
Author here. I agree with this comment, but if I wrote more like this my blog post would get less traction.
"LLM coding tools are search-based program synthesizers," in my mind this is what compilers are. I think most compilers do far too little search and opt for heuristics instead, often because they don't have an integrated runtime environment, but it's the same idea.
"Plenty of effective engineering tools are stochastic," sure but while a SAT solver might use randomness and that might adjust your time to solve, it doesn't change the correctness of the result. And for something like a fuzzer, that's a test, which are always more of a best effort thing. I haven't seen a fuzzer deployed in prod.
"Determinism comes from external specs and tests," my dream is a language where I can specify what it does instead of how it does it. Like the concept of Halide's schedule but more generic. The computer can spend its time figuring out the how. And I think this is the kind of tools AI will deliver. Maybe it'll be with LLMs, maybe it'll be something else, but the key is that you need a fairly rigorous spec and that spec itself is the programming. The spec can even be constraint based instead of needing to specify all behavior.
I'm not at all against AI, and if you are using it at a level described in this post, like a tool, aware of its strengths and limitations, I think it can be a great addition to a workflow. I'm against the idea that it's a magical English compiler, which is what I see in public discourse.
I think the key insight I walked away with from this whole thread, for me, was:
A compiler takes source and maps it to some output. Regardless of the compiler detail, this is an atomic operation; you end up with source (unmodified) and an artifact.
These “agent workflows” are distinctly different.
The process of mapping prompt to an output is the same; but these agent workflows are destructive; they modify the source.
Free reign over the entire code base; They modify the tests. The spec, the implementation.
It seems like this is a concept people are still struggling with; if your specification is poorly defined, and is dynamically updated during the compilation process, the results are more than just non deterministic.
Over time, the specification becomes non deterministic.
Thats why unsupervised agents go “off the rails”; not because the specification cant be executed, but because over time the spec drifts.
People knock "English as a programming language", but in my opinion this is the whole value of AI programming: by the time you've expressed your design and constraints well enough that an LLM can understand it, then anyone can understand it, and you end up with a codebase that's way more maintainable than what we're used to.
The problem of course is when people throw away the prompt and keep the code, like the code is somehow valuable. This would be like everyone checking in their binaries and throwing away their source code every time, while arguments rage on HN about whether compilers are useful. (Meanwhile, compiler vendors compete on their ability to disassemble and alter binaries in response to partial code snippets.)
The right way to do AI programming is: English defines the program, generated code is exactly as valuable as compiler output is, i.e. it's the actual artifact that does the thing, so in one sense it's the whole point, but iterating on it or studying it in detail is a waste of time, except occasionally when debugging. It's going to take a while, but eventually this will be the only way anybody writes code. (Note: I may be biased, as I've built an AI programming tool.)
If you can explain what needs to be done to a junior programmer in less time than it takes to do it yourself, you can benefit from AI. But, it does require totally rethinking the programming workflow and tooling.
I don't think that a prompt can be a valuable object, similar to how code used to be. Unless Mira Murati is successful at scaling her approach to deterministic inference, a prompt is fragile and transient. And even if she is successful, LLM updates make a prompt much less useful over longer time horizons.
I think that the only useful objects to keep right now are DSPy programs together with well-crafted examples, with examples being the most valuable because they are transferable across models and architectures.
I also noticed several people in the thread comparing coding assistants to junior programmers. I disagree. The only parallel is that they will do what you tell them to. Otherwise, a coding assistant can hold an entire codebase in context, reason across patterns, and generate boilerplate faster than any human. That capability has no human analogue. And unlike a junior, they have no agency, so the comparison breaks down on multiple fronts.
I think this gets to a fundamental problem with the way the AI labs have been selling and hyping AI. People keep on saying that the AI is actually thinking and it's not just pattern matching. Well, as someone that uses AI tools and develops AI tools, my tools are much more useful when I treat the AI as a pattern matching next-token predictor than an actual intelligence. If I accidentally slip too many details into the context, all of a sudden the AI fails to generalize. That sounds like pattern matching and next token prediction to me.
> This isn’t to say “AI” technology won’t lead to some extremely good tools. But I argue this comes from increased amounts of search and optimization and patterns to crib from, not from any magic “the AI is doing the coding”
* I can tell claude code to crank out some basic crud api and it will crank it out in a minute saving me an hour or so.
* I need an implementation of an algorithm that has been coded a million times on github, I ask the AI to do it and it cranks out a correct working implementation.
If I only use the AI in its wheelhouse it works very well, otherwise it sucks.
I think this comes down to levels of intelligence. Not knowledge, I mean intelligence. We often underestimate the amount of thinking/reasoning that goes into a certain task. Sometimes the AI can surprise you and do something very thoughtful, this often feels like magic.
Both CRUD and boilerplate are arguably a tooling issue. But there are also a bunch of things only AI will let you do.
My tests with full trace level logging enabled can get very verbose. It takes serious time for a human to parse where in the 100 lines of text the relevant part is.
Just telling an AI: "Run the tests and identify the root cause" works well enough, that nowadays it is always my first step.
Yesterday I used Opus to diagnose an error and it sent me down a bunch of wrong paths, suggested fixes with bad unintended consequences. If I hadn’t carefully guided it and corrected it, not only would the bug not have been fixed but it would have made things worse. My point being that if you don’t have the domain knowledge and put on your thinking cap then you will likely be screwed. But this is not how AI is sold.
Ofcourse, there is some truth in what you say. But business is desperate for new tech where they can redefine the order (who is big and who is small). There are floating billions which chase short term returns. Fund managers will be fired if they are not jumping on the new fad in the town. CIO's and CEO's will be fired if they are not jumping on AI. It's just nuclear arms race. It's good for none. but the other guy is on it, so you need to be too.
Think about this. Before there were cars on roads, people were just as much happy. Cars came, cities were redesigned for cars with buildings miles apart, and commuting miles became the new norm. You can no longer say cars are useless because the context around them has changed to make the cars a basic need.
AI does same thing. It changes the context in which we work. Everyone expects you use AI (and cars). It becomes a basic need, though a forced one.
To go further, hardly anything produced by science or technology is a basic need for humans. The context got twisted, making them basic needs. Tech solutions create the problems which they claim to solve. The problem did not exist before the solution came around. That's core driving force of business.
That METR study gets a lot of traction for its headline; and I doubt many people read the whole thing—it was long—but the data showed a 50% speed up for the one dev with the most experience with Cursor/AI, suggesting a learning curve and also wild statistical variation on a small sample set. An errata later suggested another dev who did not have a speedup had not represented their experience correctly, but still strongly draws into question the significance of the findings.
The specific time sucks measured in the study are addressable with improved technology like faster LLMs and improved methodology like running parallel agents—the study was done in March running Claude 3.7 and before Claude Code.
We also should value the perception of having worked 20% less even if you actually spent more time. Time flies when you’re having fun!
It was such a great study imo and rightfully deserves a lot of attention since it's a great alternative to twitter vibes.
But to stop here and say AI coding doesn't work will just not hold up well. We have a sample size of 16 developers for 250 tasks.
Here is my data point: adding AI to your engineering team really is like adding a new member. At first it slows everybody down: explain this, explain that, stupid mistake made that the experienced dev would have avoided. But over the months you learn how to get the most out of them: they speak fluently all programming languages, don't mind writing 20 tests or coming up with good industry ideas to your problems.
Over time I have come to appreciate that the market is rather smart and even hype has its place. You often have to cross a line to know you crossed it, especially with something as fuzzy and quickly changing as technological progress hitting the messy real world. But then again, we need the heretics to stay smart and not follow an emperor without clothes.
Just putting this up as a reference the next time this comes up on HN. The study data shows that the median task is 1.5hrs and the 15 minutes that developers think they saved was actually more than that many minutes less researching, testing, and writing code and the 15 minutes longer they actually spent was more idle (5mins) and waiting for AI (5 mins) and reviewing and prompting dominating their work over actually writing code.
The person who showed a speed-up indicated over a week of prior experience with cursor while all others under a week.
AI coding is the one thing that got my back to programming. I got to the point in life, when my ability to focus is reducing, and I prefer to send the remaining energy elsewhere. I kind of gave up on programming, just doing architecture and occasionally doing very small programming tasks. It all changed when I discovered Claude Code and saw that the way it works, is kind of how I work. I also use a lot of grep to find my way through a new codebase, I also debug stuff by adding logs to see the context, I also rely on automated tests to tell me something is broken. I'm still very good at reading code, I'm good at architecture, and with these tools, I feel I can safely delegate the boring bits of writing code and debugging trivial things to AI. Yes, it's slower than if I focused on the task myself, but the point is that I'd not be able to focus on the task myself.
I feel the same way. After 30+ years coding, I know what I want to build, but finding the time to work on it and the focus is harder than it used to be.
I feel like many opinion pieces on AI coding are written from the perspective of highly experienced software engineers, often from a kind of ivory tower. (even that study they cite was based on “16 experienced open-source developers.”)
But for people who aren’t seasoned coders, these tools are incredibly valuable. I have some coding experience myself, but it’s never been my profession (I’m a visual artist). Now, I can accomplish in an afternoon what would otherwise take me days. Two months ago I left my job to work solo on my video game, and even though my budget is limited, I still make sure to keep Claude and ChatGPT. Also, being able to write something on my phone at 1 a.m. when I can’t sleep, send it to Codex, and then test it first thing in the morning at my computer feels pretty magical. It also takes away the worry of “what if this isn’t the best way to do it?” since refactoring parts of my codebase is now so easy. It helps not just with writing code, but also with removing the fear that might otherwise keep me from sitting down and getting the work done.
That said, I understand that much of the criticism is really aimed at the marketing hype around these tools and the broader “this will replace the engineers at your company” narrative.
Readit News
I'm not.
I can build anything, but often struggle with getting bogged down with all the basic work. I love AI for speed running through all the boring stuff and getting to the good parts.
I liken AI development to a developer somewhere between junior and mid-level, someone I can given a paragraph or two of thought out instructions and have them bang out an hour of work. (The potential for then stunting the growth of actual juniors into tomorrow's senior developers is a serious concern, but a separate problem to solve)
In some cases, especially with the more senior devs in my org, fear of the good parts is why they're against AI. Devs often want the inherent safety of the boring, easy stuff for a while. AI changes the job to be a constant struggle with hard problems. That isn't necessarily a good thing. If you're actually senior by virtue of time rather than skill, you can only take on a limited number of challenging things one after another before you get exhausted.
Companies need to realise that AI to go faster is great, but there's still a cognitive impact on the people. A little respite from the hardcore stuff is genuinely useful sometimes. Taking all of that away will be bad for people.
That said, some devs hate the boring easy bits and will thrive. As with everything, individuals need to be managed as individuals.
There's also the fact that, while you're coding the easy stuff, your mind is thinking about the hard stuff, looking things up, seeing how they articulate. If you're spending 100% of your time on hard stuff, you might be hurting these preliminaries.
Part of the problem is that in big orgs, you need to show consistent progress in order to not get put on some PIP and kicked out of the company. There are performance review cycles and you have to show something continuously.
That ONLY works if you have boring, easy work. It's easy to deliver consistent progress on that.
Interesting and difficult work is nice only if you are trusted to try your best and given the freedom to fail. That's the nature of hard problems; progress in those domains is very sudden and Poissonian and not consistent by nature. If you're going to be judged on your ability to be sub-Poissonian and consistent, and get put on a PIP for not succeeding at it one review cycle (and possibly risking income that you use to put a roof over your head or feed your family) it's not worth the career risk to try difficult things.
Not saying this is the way I think, it's just the reality of how things often work in big orgs, and one of the reasons I dislike many big orgs.
That's matches my experience. In my first job, every time a new webapp project has been starting it was fun. Not because of challenges or design, but simply because of the trivial stuff done for n-th time - user accounts, login, password reset, admin panel. Probably should have been automated at this point, but we got away with reinventing the wheel each time.
I find this hilarious. From what I've seen watching people do it, it changes the job from deep thought and figuring out a good design to pulling a lever on a slot machine and hoping something good pops out.
The studies that show diminished critical thinking have matched what i saw anecdotally pairing with people who vibe coded. It replaced deep critical thinking with a kind of faith based gambler's mentality ("maybe if i tell it to think really hard it'll do it right next time...").
The only times ive seen a notable productivity improvement is when it was something not novel that didnt particularly matter if what popped out was shit - e.g. a proof of concept, ad hoc app, something that would naturally either work or fail obviously, etc. The buzz people get from these gamblers' highs when it works seems to make them happier than if they didnt use it at all though.
Very true. I think AI (especially Claude Code) forced me to actually think hard about the problem at hand before implementing the solution. And more importantly, write down my thoughts before they fleet away from my feeble mind. A discipline that I wished I had before.
On the flip side, there have been lots of times where I personally didn’t have a lot of time to deeply research a topic (read papers, build prototypes of different ideas, etc) due to lack of time and resources. If all of the boring stuff is gone, and building prototypes is also 3x faster maybe what will end up happening is we can now use all of this free time to try lots of different ideas because the cost of exploration has been lowered.
I disagree that this has anything to do with people needing a break. All code eventually has to be reviewed. Regardless of who or what wrote it, writing too much of it is the problem. It's also worth considering how much more code could be eliminated if the business more critically planned what they think they want.
These tensions have existed even before computers and in all professions.
The issue of senior-juniors has always been a problem; AI simply means they're losing their hiding spots.
AI has also been a really good brainstorming partner - especially if you prompt it to disable sycophancy. It will tell you straight up when you are over-engineering something.
It's also wonderful at debugging.
So I just talk to my computer, brainstorm architectures and approaches, create a spec, then let it implement it. If it was a bad idea, we iterate. The iteration loop is so fast that it doesn't matter.
Did you end up regretting a design choice, but normally you'd live with it because so much code would have to be changed? Not with Agentic coding tools - they are great at implementing changes throughout the entire codebase.
And its so easy to branch out to technologies you're not an expert in, and still be really effective as you gain that expertise.
I honestly couldn't be happier than I am right now. And the tools get better every week, sometimes a couple times a week.
Can you share more about how you are prompting to disable sycophancy?
I'm in the same boat (granted, 10 years less) but can't really relate with this. By the time any part becomes boring, I start to automate/generalize it, which is very challenging to do well. That leaves me so little boring work that I speed run through it faster by typing it myself than I could prompt it.
The parts in the middle – non-trivial but not big picture – in my experience are the parts where writing the code myself constantly uncovers better ways to improve both the big picture and the automation/generalization. Because of that, there are almost no lines of code that I write that I feel I want to offload. Almost every line of code either improves the future of the software or my skills as a developer.
But perhaps I've been lucky enough to work in the same place for long. If I couldn't bring my code with me and had to constantly start from scratch, I might have a different opinion.
The two aren't mutually exclusive. You can use AI to build your tooling. (Unless it's of sufficient complexity or value that you need to do the work yourself)
I've struggled heavily trying to figure out how to get it to write the exactly correct 10 lines of code that I need for a particularly niche problem, and so I've kind of given up on that, but getting it to write the 100 lines of code around those magic 10 lines saves me so much trouble, and opens me up to so many more projects.
I find it best as a "personal assistant," that I can use to give me information -sometimes, highly focused- at a moment's notice.
> The potential for then stunting the growth of actual juniors into tomorrow's senior developers is a serious concern
I think it's a very real problem. I am watching young folks being frozen out of the industry, at the very beginning of their careers. It is pretty awful.
I suspect that the executives know that AI isn't yet ready to replace senior-levels, but they are confident that it will, soon, so they aren't concerned that there aren't any more seniors being crafted from youngsters.
Would suck, if they bet wrong, though…
My life quality (as a startup cofounder wearing many different hats across the whole stack) would drop significantly if Cursor-like tools [1] were taken away from me, because it takes me a lot of mental effort to push myself to do the boring task, which leads to procrastination, which leads to delays, which leads to frustration. Being able to offload such tasks to AI is incredibly valuable, and since I've been in this space from "day 1", I think I have a very good grasp on what type of task I can trust it to do correctly. Here are some examples:
- Add logging throughout some code
- Turn a set of function calls that have gotten too deep into a nice class with clean interfaces
- Build a Streamlit dashboard that shows some basic stats from some table in the database
- Rewrite this LLM prompt to fix any typos and inconsistencies - yeah, "compiling" English instructions into English code also works great!
- Write all the "create index" lines for this SQL table, so that <insert a bunch of search usecases> perform well.
[1] I'm actually currently back to Copilot Chat, but it doesn't really matter that much.
That's one of the thing that I wouldn't delegate to LLM. Logging is like a report of things that happens. And just like a report, I need relevant information and the most useful information.
...
A lot of these use cases actually describes the what. But the most important questions is always the why. Why is it important to you? Or to the user? That's when things have a purpose and not be just toys.
However. There's also good news. AI is also an amazing tool for learning.
So what I see AI doing is simply separating people who want to put effort forth and those who don't.
Absolutely. For example, I've been learning Autodesk Fusion, and after establishing a small foundation through traditional learning techniques, I've been able to turbocharge my learning by asking precise questions to AI.
I image this really sucks for those whose business model relied on gatekeeping knowledge. (like training companies)
If you aren't talented enough to write or record your own music, you aren't really a musician.
If you have a quick question about music theory and you want a quick answer, AI can be a benefit.
This is where AI actually helps - you have a very precise vision of what you want, but perhaps you've forgotten about the specific names of certain API methods, etc. Maybe you don't want to implement all the cases by hand. Often validating the output can take just seconds when you know what it is you're looking for.
The other part of making the output do what you want is the ability to write a prompt that captures the most essential constraints of your vision. I've noticed the ability to write and articulate ideas well in natural language terms is the actual bottleneck for most developers. It takes just as much practice communicating your ideas as it does anything else to get good at it.
The problem is that junior developers are what we make senior developers with— so in 15 years, this is going to be yet another thing that the US used to be really good at, but is no longer capable of doing, just like many important trades in manufacturing. The manufacturers were all only concerned with their own immediate profit and made the basic sustainability of their workforce, let alone the health of the trades that supported their industries, a problem for everyone else to take care of. Well, everyone else did the same thing.
In 15 years senior developers will not be needed as well. Anyway no company is obliged to worry about 15 years timescale
I love AI and use it for both personal and work tasks for two reasons:
1. It's a way to bounce around ideas without (as much) bias as a human. This is indispensable because it gives you a fast feedback mechanism and validates a path.
2. It saves me typing and time. I give it one-shot, "basic work" to do and it's able to do accomplish at least 80% of what I'd say is complete. Although it may not be 100% it's still a net positive given the amount of time it saves me.
It's not lost on me that I'm effectively being trained to always add guardrails, be very specific about the instructions, and always check the work of AI.
Shhh, WIP blog post (on webpipe powered blog)
https://williamcotton.com/articles/introducing-web-pipe
Yes, I wrote my own DSL, complete with BDD testing framework, to write my blog with. In Rust!
My blog source code written in webpipe:http://github.com/williamcotton/williamcotton.com
Deleted Comment
I expect that in a year my relationship with AI will be more like a TL working mostly at the requirements and task definition layer managing the work of several agents across parallel workstreams. I expect new development toolchains to start reflecting this too with less emphasis on IDEs and more emphasis on efficient task and project management.
I think the "missed growth" of junior devs is overblown though. Did the widespread adoption of higher-level really hurt the careers of developers missing out on the days when we had to do explicit memory management? We're just shifting the skillset and removing the unnecessary overhead. We could argue endlessly about technical depth being important, but in my experience this hasn't ever been truly necessary to succeed in your career. We'll mitigate these issues the same way we do with higher-level languages - by first focusing on the properties and invariants of the solutions outside-in.
Analogies to humans don't work that well. AI is super-human in some respects while also lacking the ability to continually work toward a goal over long periods of time. AI can do very little on its own - just short / scoped / supervised tasks.
However, sometimes the situation is reversed, AI is the teacher who provides some examples on how to do things or provides hints on how to explore a new area and knows how others have approached similar things. Then, sometimes, AI is an astute code reviewer, typically providing valuable feedback.
Anyway, I've stopped trying anthropomorphize AI and simply try to reason about it based on working with it. That means combinations of direct ChatGPT usage with copy / paste / amend type workflows, async style / full PR style usage, one-shot "hail Mary" type throw away PRs just to establish an initial direction as well as PR reviews of my own code. I'm using AI all the time, but never anything like how I would work with another human.
But the issue is some of that speedrunning sometimes takes so much time, it becomes inefficient. It's slowly improving (gpt5 is incredible), but sometimes it get stuck on really mundane issue, and regress endlessly unless I intervene. And I am talking about straightforwars functional code.
I think you’re the best case support for AI coding. You know clearly what you want, so you know clearly what you don’t want. So if you had decent verbal dexterity you could prompt the AI model and manage to accomplish what you intended.
A lot of programming problems / programmer contexts don’t match that situation. Which is the problem with universalizing the potency of AI / benefits of AI coding.
the laziness manifest itself into productivity as crazy as this sounds. how? lazy people find a way to automate repetitive tasks. what I have learned from these over the years is that anything you do twice has to find a way to be automated as third time is around the corner :)
what does this have to do with AI? the AI has taken automation to another level allowing us to automate so much of our work that was not previously possible. I found myriad of ways to use AI and several of my best (lazy) co-workers have as well. I cannot imagine doing my work anymore without it, not because of any “magic” but because my lazy ass will be able to do all the things that I have automated out
This persona driven workflow is so weird to me. Feels like stuck in old ways.
It's what is, to me, so bizarre about the present moment: certainly investment is exceptionally high in AI (and of course use), but the dominant position in the media is precisely such a strange 'anti-AI hype' that positions itself as a brave minority position. Obviously, OpenAI/Altman have made some unfortunate statements in self-promotion, but otherwise I genuinely can't think of something I've read that expresses the position attacked by the anti-AI-ers -- even talk of 'AGI' etc comes from the AI-critical camp.
In a sense, the world seems divided into three: obvious self-promotion from AI companies that nobody takes seriously, ever-increasingly fervent 'AI critique', and the people who, mostly silent, have found modern AI with all its warts to be an incomparably useful tool across various dimensions of their life and work. I hope the third camp becomes more vocal so that open conversations about the ways people have found AI to be useful or not can be the norm not the exception.
[1] https://antirez.com/news/154
Why the insistence on anthropomorphizing what is just a tool? It has no agency, does not 'think' in any meaningful manner, it is just pattern matching on a vast corpus of training data. That's not to say it can't be very useful - as you seem to have found - but it is still just a tool.
Geohot is easily a 99.999 percentile developer, and yet he can’t seem to reconcile that the other 99.999 percent are doing something much more basic than he can ever comprehend.
It’s some kind of expert paradox, if everyone was as smart and capable as the experts, then they wouldn’t be experts.
I have come across many developers that behave like the AI. Can’t explain codebases they’ve built, can’t maintain consistency.
It’s like a aerospace engineer not believing that the person that designs the toys in an Kinder egg doesn’t know how fluid sims work.
I’m surprised to see this. From my perspective, reading comments and seeing which posts rise to the top, HN as a whole seems pretty bullish on the tech as whole…
HN participants (generally speaking) are against: AI, Crypto, HFT. I've worked in 2/3 of these industries so have first hand experience. My basic summary is that the average commenter here has a lot of misinformation on these topics (as a insider).
But don't worry. The company puts them somewhere they can't do any damage. Most of them become managers.
> The only reason it works for many common programming workflows is because they are common. The minute you try to do new things, you need to be as verbose as the underlying language.
Dead Comment
Really?
"The best model of a programming AI is a compiler... You give it a prompt, which is “the code”, and it outputs a compiled version of that code. Sometimes you’ll use it interactively, giving updates to the prompt after it has returned code, but you find that, like most IDEs, this doesn’t work all that well and you are often better off adjusting the original prompt and “recompiling”."
Really?
If you do not understand his point, you need to read more about our field.
I think his excellency in his own trade limited his vision for the 99% who just want to get by in the job. How many dev even deal with compiler directly these days? They write some code, fix some red underlines, then push, pray and wait for pipeline pass. LLMs will be gods in this process, and you can even beg another one if your current one does not work best.
Deleted Comment
First, the assertion that the best model of "AI coding" is that it is a compiler. Compilers deterministically map a formal language to another under a spec. LLM coding tools are search-based program synthesizers that retrieve, generate, and iteratively edit code under constraints (tests/types/linters/CI). That’s why they can fix issues end-to-end on real repos (e.g., SWE-bench Verified), something a compiler doesn’t do. Benchmarks now show top agents/models resolving large fractions of real GitHub issues, which is evidence of synthesis + tool use, not compilation.
Second, that the "programming language is English". Serious workflows aren’t "just English." They use repo context, unit tests, typed APIs, JSON/function-calling schemas, diffs, and editor tools. The "prompt" is often code + tests + spec, with English as glue. The author attacks the weakest interface, not how people actually ship with these tools.
Third, non-determinism isn't disqualifying. Plenty of effective engineering tools are stochastic (fuzzers, search/optimization, SAT/SMT with heuristics). Determinism comes from external specs: unit/integration tests, type systems, property-based tests, CI gates.
False dichotomy: "LLMs are popular only because languages/libraries are bad." Languages are improving (e.g. Rust, Typescript), yet LLMs still help because the real bottlenecks are API lookup, cross-repo reading, boilerplate, migrations, test writing, and refactors, the areas where retrieval and synthesis shine. These are complementary forces, not substitutes.
Finally, no constructive alternatives are offered. "Build better compilers/languages" is fine but modern teams already get value by pairing those with AI: spec-first prompts, test-gated edits, typed SDK scaffolds, auto-generated tests, CI-verified refactors, and repo-aware agents.
A much better way to think about AI coding and LLMs is that they aren’t compilers. They’re probabilistic code synthesizers guided by your constraints (types, tests, CI). Treat them like a junior pair-programmer wired into your repo, search, and toolchain. But not like a magical English compiler.
"LLM coding tools are search-based program synthesizers," in my mind this is what compilers are. I think most compilers do far too little search and opt for heuristics instead, often because they don't have an integrated runtime environment, but it's the same idea.
"Plenty of effective engineering tools are stochastic," sure but while a SAT solver might use randomness and that might adjust your time to solve, it doesn't change the correctness of the result. And for something like a fuzzer, that's a test, which are always more of a best effort thing. I haven't seen a fuzzer deployed in prod.
"Determinism comes from external specs and tests," my dream is a language where I can specify what it does instead of how it does it. Like the concept of Halide's schedule but more generic. The computer can spend its time figuring out the how. And I think this is the kind of tools AI will deliver. Maybe it'll be with LLMs, maybe it'll be something else, but the key is that you need a fairly rigorous spec and that spec itself is the programming. The spec can even be constraint based instead of needing to specify all behavior.
I'm not at all against AI, and if you are using it at a level described in this post, like a tool, aware of its strengths and limitations, I think it can be a great addition to a workflow. I'm against the idea that it's a magical English compiler, which is what I see in public discourse.
A compiler takes source and maps it to some output. Regardless of the compiler detail, this is an atomic operation; you end up with source (unmodified) and an artifact.
These “agent workflows” are distinctly different.
The process of mapping prompt to an output is the same; but these agent workflows are destructive; they modify the source.
Free reign over the entire code base; They modify the tests. The spec, the implementation.
It seems like this is a concept people are still struggling with; if your specification is poorly defined, and is dynamically updated during the compilation process, the results are more than just non deterministic.
Over time, the specification becomes non deterministic.
Thats why unsupervised agents go “off the rails”; not because the specification cant be executed, but because over time the spec drifts.
That doesnt happen with compilers.
in your HN comment: “I agree with this comment, but if I wrote more like this my blog post would get less traction.”
Seems like you also not care about the truth.
The problem of course is when people throw away the prompt and keep the code, like the code is somehow valuable. This would be like everyone checking in their binaries and throwing away their source code every time, while arguments rage on HN about whether compilers are useful. (Meanwhile, compiler vendors compete on their ability to disassemble and alter binaries in response to partial code snippets.)
The right way to do AI programming is: English defines the program, generated code is exactly as valuable as compiler output is, i.e. it's the actual artifact that does the thing, so in one sense it's the whole point, but iterating on it or studying it in detail is a waste of time, except occasionally when debugging. It's going to take a while, but eventually this will be the only way anybody writes code. (Note: I may be biased, as I've built an AI programming tool.)
If you can explain what needs to be done to a junior programmer in less time than it takes to do it yourself, you can benefit from AI. But, it does require totally rethinking the programming workflow and tooling.
I think that the only useful objects to keep right now are DSPy programs together with well-crafted examples, with examples being the most valuable because they are transferable across models and architectures.
I also noticed several people in the thread comparing coding assistants to junior programmers. I disagree. The only parallel is that they will do what you tell them to. Otherwise, a coding assistant can hold an entire codebase in context, reason across patterns, and generate boilerplate faster than any human. That capability has no human analogue. And unlike a junior, they have no agency, so the comparison breaks down on multiple fronts.
> This isn’t to say “AI” technology won’t lead to some extremely good tools. But I argue this comes from increased amounts of search and optimization and patterns to crib from, not from any magic “the AI is doing the coding”
* I can tell claude code to crank out some basic crud api and it will crank it out in a minute saving me an hour or so.
* I need an implementation of an algorithm that has been coded a million times on github, I ask the AI to do it and it cranks out a correct working implementation.
If I only use the AI in its wheelhouse it works very well, otherwise it sucks.
My tests with full trace level logging enabled can get very verbose. It takes serious time for a human to parse where in the 100 lines of text the relevant part is.
Just telling an AI: "Run the tests and identify the root cause" works well enough, that nowadays it is always my first step.
Think about this. Before there were cars on roads, people were just as much happy. Cars came, cities were redesigned for cars with buildings miles apart, and commuting miles became the new norm. You can no longer say cars are useless because the context around them has changed to make the cars a basic need.
AI does same thing. It changes the context in which we work. Everyone expects you use AI (and cars). It becomes a basic need, though a forced one.
To go further, hardly anything produced by science or technology is a basic need for humans. The context got twisted, making them basic needs. Tech solutions create the problems which they claim to solve. The problem did not exist before the solution came around. That's core driving force of business.
The specific time sucks measured in the study are addressable with improved technology like faster LLMs and improved methodology like running parallel agents—the study was done in March running Claude 3.7 and before Claude Code.
We also should value the perception of having worked 20% less even if you actually spent more time. Time flies when you’re having fun!
But to stop here and say AI coding doesn't work will just not hold up well. We have a sample size of 16 developers for 250 tasks.
Here is my data point: adding AI to your engineering team really is like adding a new member. At first it slows everybody down: explain this, explain that, stupid mistake made that the experienced dev would have avoided. But over the months you learn how to get the most out of them: they speak fluently all programming languages, don't mind writing 20 tests or coming up with good industry ideas to your problems.
Over time I have come to appreciate that the market is rather smart and even hype has its place. You often have to cross a line to know you crossed it, especially with something as fuzzy and quickly changing as technological progress hitting the messy real world. But then again, we need the heretics to stay smart and not follow an emperor without clothes.
The person who showed a speed-up indicated over a week of prior experience with cursor while all others under a week.
But for people who aren’t seasoned coders, these tools are incredibly valuable. I have some coding experience myself, but it’s never been my profession (I’m a visual artist). Now, I can accomplish in an afternoon what would otherwise take me days. Two months ago I left my job to work solo on my video game, and even though my budget is limited, I still make sure to keep Claude and ChatGPT. Also, being able to write something on my phone at 1 a.m. when I can’t sleep, send it to Codex, and then test it first thing in the morning at my computer feels pretty magical. It also takes away the worry of “what if this isn’t the best way to do it?” since refactoring parts of my codebase is now so easy. It helps not just with writing code, but also with removing the fear that might otherwise keep me from sitting down and getting the work done.
That said, I understand that much of the criticism is really aimed at the marketing hype around these tools and the broader “this will replace the engineers at your company” narrative.