Developing our position on AI
recurse.com
recurse.com
I did a lot of AI assisted coding this week and I felt, if anything, it wasn't faster but it led to higher quality.
I would go through discussions about how to do something, it would give me a code sample, I would change it a bit to "make it mine", ask if I got it right, get feedback, etc. Sometimes it would use features of the language or the libraries I didn't know about before so I learned a lot. With all the rubber ducking I thought through things in a lot of depth and asked a lot of specific questions and usually got good answers -- I checked a lot of things against the docs. It would help a lot if it could give me specific links to the docs and also specific links to code in my IDE.
If there is some library that I'm not sure how to use I will load up the source code into a fresh copy of the IDE and start asking questions in that IDE, not the one with my code. Given that it can take a lot of time to dig through code and understand it, having an unreliable oracle can really speed things up. So I don't see it as a way to gets things done quickly, but like pairing with somebody who has very different strengths and weaknesses from me, and like pair programming, you get better quality. This week I walked away with an implementation that I was really happy with and I learned more than if I'd done all the work myself.
Just a super quick test: "what are 3 obscure but useful features in python functools. Link to doc for each."
GPT 4o gave good links with each example.
(its choices were functools.singledispatch, functools.total_ordering, functools.cached_property)
Not sure about local code links.
This is one area where MCPs might actually be useful, https://context7.com/ being one of them. I haven't given it enough of a shot yet, though.
Are you new to it? There's a pretty standard arc that starts with how great it is and ends with all the "giving up on AI" blog posts you see.
I went through it to. I still use a chatbot as a better stack overflow, but I've stopped actually having AI write any code I use - it's not just the quality, it's the impact on my thinking and understanding that ultimately doesn't improve outcomes over just doing it myself.
I’ve usually been skeptical about people who get unreasonably good results and not surprised when they wake up a few weeks later and are disappointed. One area where I am consistently disappointed is when there are significant changes across versions: I had one argue over where I could write a switch in Java that catches null (you can in JDK 21) and lots of trouble with SQLAlchemy in Python which changed a lot between versions. I shudder to think what would happen if you asked questions about react-router but actually I shudder to think about react-router at all.
I've been using AI regularly since GPT 4 first came out a couple years ago. Over that time, various models from Sonnet to Gemini to 4o have generally been good rubber ducks. Good to talk to and discuss approaches and tradeoffs, and better in general than google + stack overflow + pouring over verbose documentation.
But I couldn't really "hand the models the wheel." They weren't trustworthy enough, easily lost the plot, failed to leverage important context right in front of them in the codebase, etc. You could see that there was potential there, but it felt pretty far away.
Something changed this spring. Gemini 2.5 Pro, Claude 4 models, o3 and o4-mini -- I'm starting to give the models the wheel now. They're good. They understand context. They understand the style of the codebase. And they of course bring the immense knowledge they've always had.
It's eerie to see, and to think about what comes with the next wave of models coming very soon. And if the last time you really gave model-driven programming a go was 6 months or more ago, you probably have no idea what's about to happen.
My most recent stab at this was Claude code with 3.7, circa March this year.
To be fair though, a big part of the issue for me is that having not done the work or properly thought through how a project is structured and how the code works, it comes back to bite later. A better model doesn't change this.
However, at least in my experience, nothing beats o3 for raw intelligence. It's a little too slow to use as the daily driver though.
It's kind of fun seeing the models all have their various pros and cons.
> To be fair though, a big part of the issue for me is that having not done the work or properly thought through how a project is structured and how the code works, it comes back to bite later. A better model doesn't change this.
Yes, even as I start to leverage the tools more, I try to double down on my own understanding of the problem being solved, at least at a high level. Need to make sure you don't lose the plot yourself.
It’s still slower going as the codebase increases in size, and this is my hypothesis for the huge variance; I was getting giddy at how fast I blew through the first 5 hours of a small project (perhaps in 30 mins with Claude) but quickly lost velocity when I started implementing tests and editing existing code.
To me, programming is a solo activity. "Chatting" with someone or something as I do it is just a distraction.
By the time, I'm about to prompt, I usually have enough information to just code it away. Coding is like riding a bicycle downhill. You just pay enough focus to ride it. It's like how you don't think about the characters and the words when you're typing. You're mostly thinking about what you want to say.
When there's an issue, I switch from coding to reading and thinking. And while the latter is mentally taxing, it is fast as I don't have to spell it out. And a good helper to that is a repository of information. Bookmarks to docs, documentation browser, code samples,.. By the times the LLM replies with a good enough paragraph, I'm already at the Array page on MDN.
Also writing test cases is how you experience the pain of having things that should not be coupled together. So you can go refactor stuff instead of having to initialize the majority of your software.
When they're all done I review the output and either clean it up a little and open a PR, or throw it away and tune my initial prompt and the memory bank and start over. They're just code-generating machines, not real programmers that it's worth iterating with – for one thing, they won't learn anything that way.
Fundamentally none of the issues inherent in LLMs will be fixed by increasing parameter count or better weights.
Shilling for the newest model is a dead end, better to figure out how to best put LLMs to use despite their limitations.
Um, ok just disregard my entire post where I talk about how my issues with using LLMs for programming were literally solved by better models.
Sorry but it is what it is.
I’m curious if/how you’ve built a career this way. Or is this account just a stress relief “shit-posting” persona?
Wouldn't be shocked if this is related to getting ramped up.
Switch to a language you don't know. How will that affect your AI usage? Will it go up or down over time?
Had similar experiences during the SO early days or when diving into open source projects. Suddenly you go from being stuck in your bubble to learning all kinds of things. Your code gets good. The returns diminish and you no longer curiously read unless the library is doing something you can't imagine.
But getting extreme concurrency outcomes or exotic hardware outcomes or low latency stuff or heavy numerics or a million other things? It's harder, the bots aren't as good. This is the divide: it can one shot a web page, even the older ones can just smoke a tailwind theme, headshot. A kernel patch to speed up a path on my bare metal box in a latency contest? Nah, not really.
But I see a lot of promise in the technology even if the sort of hype narrative around it seems pretty much false at this point: I still use the LLMs a lot. Part of that is that search just doesn't work anymore (although... Yandex got a lot better recently out of the blue of all things), and part of it is that I see enough exciting glimpses of like, if I got it hooked up to the right programming language and loaded the context right, wow, once in a while it just slams and it's kinda rare but frequent enough that I'm really interested in figuring out how to reproduce it reliably. And I think I'm getting better outcomes a little bit at a time, getting it dialed in. Two or three months ago an even with claude code would have me yelling at the monitor, now it's like, haha, i see you.
LLMs lower the technical debt spread by the clueless, to a lower baseline.
You were part of the clueless, so an LLM improved your code and lowered the technical debt you would have spread.
Have you compared that to your normal debugging thought processes? I get that you might be given another way to think about the problem but another human might be best for that, rather than a next token guesser.
I have a devil of a time with my team and wider, the younger ones mainly, getting them to pick up a phone instead of sending emails or chats or whatever. A voice chat can solve a problem within minutes or even seconds instead of the rather childish game of email ping pong. I do it myself too (email etc) and I even encourage it, despite what I said earlier - effective use of comms is a skill but you do need to understand when to use each variety.
This was a really fascinating project to work on because of the breadth of experiences and perspectives people have on LLMs, even when those people all otherwise have a lot in common (in this case, experienced programmers, all Recurse Center alums, all professional programmers in some capacity, almost all in the US, etc). I can't think of another area in programming where opinions differ this much.
As a recurse alum (s14 batch 2) I loved reading this. I loved my time at recurse and learned lots. This highlight from the post really resonates:
“ Real growth happens at the boundary of what you can do and what you can almost do. Used well, LLMs can help you more quickly find or even expand your edge, but they risk creating a gap between the edge of what you can produce and what you can understand.
RC is a place for rigor. You should strive to be more rigorous, not less, when using AI-powered tools to learn, though exactly what you need to be rigorous about is likely different when using them.”
This brings about an important point for a LOT of tools, which many people don't talk about: namely, with a tool as powerful as AI, there will always be minority of people with healthy and thoughtful attitude towards its use, but a majority who use it improperly because its power is too seductive and human beings on average are lazy.
Therefore, even if you "strive to be more rigorous", you WILL be a minority helping to drive a technology that is just too powerful to make any positive impact on the majority. The majority will suffer because they need to have an environment where they are forced not to cheat in order to learn and have basic competence, which I'd argue is far more crucial to a society that the top few having a lot of competence.
The individualistic will say that this is an inevitable price for freedom, but in practice, I think it's misguided. Universities, for example, NEED to monitor the exam room, because otherwise cheating would be rampant, even if there is a decent minority of students who would NOT cheat, simply because they want to maximize their learning.
With such powerful tools as AI, we need to think beyond our individualistic tendencies. The disciplined will often tout their balanced philosophy as justification for that tool use, such as this Recurse post is doing here, but what they are forgetting is that by promoting such a philosophy, it brings more legitimacy into the use of AI, for which the general world is not capable of handling.
In a fragile world, we must take responsibility beyond ourselves, and not promote dangerous tools even if a minority can use them properly.
This is why I am 100% against AI – no compromise.
"The disciplined minority can use AI well, but the lazy majority can't, so nobody gets to use it" I feel like I read this somewhere. Maybe a short story?
Should we ban calculators because some students become dependent on them? Ban the internet because people use it to watch cat videos instead of learning?
You've dressed up "hold everyone back to protect the incompetent" as social responsibility.
I never actually thought I would find someone who read Harrison Bergeron and said "you know what? let's do that!" But the Internet truly is a vast and terrifying place.
Regardless, I only advocate for restricting technologies that are too dangerous, much in the same way as atomic weapons are highly restricted by people can still own knives and even use guns in some circumstances.
I have nothing against the most intelligent using their intelligence wisely and doing more than the less intelligent, if only wise use is even possible. In the case of AI, I submit that it is not.
AI isn't nukes - anyone can train a model at home. There's no centralized thing to restrict. So what's your actual ask? That nobody ever trains a model? That we collectively pretend transformers don't exist?
You're dressing up bog-standard tech panic as social responsibility. Same reaction to every new technology: "This tool might be misused so nobody should have it."
If you can't see the connection between that and Harrison Bergeron's "some people excel so we must handicap everyone," then you've missed Vonnegut's entire point. You're not protecting the weak - you're enforcing mediocrity and calling it virtue.
Again, a rather knee-jerk reply. I am opening up the discussion, and putting out my opinion. I never said I should be God and arbiter, but I do think people in general should have a discussion about it, and general discussion starts with opinion.
> AI isn't nukes - anyone can train a model at home. There's no centralized thing to restrict. So what's your actual ask? That nobody ever trains a model? That we collectively pretend transformers don't exist?
It should be something to consider. We could stop it by spreading a social taboo about it, denigrate the use of it, etc. It's possible. Many non techies already hate AI, and mob force is not out of the question.
> You're dressing up bog-standard tech panic as social responsibility. Same reaction to every new technology: "This tool might be misused so nobody should have it."
I don't have that reaction to every new technology personally. But I think we should ask the question of every new technology, and especially onces that are already disrupting the labor market.
> If you can't see the connection between that and Harrison Bergeron's "some people excel so we must handicap everyone," then you've missed Vonnegut's entire point. You're not protecting the weak - you're enforcing mediocrity and calling it virtue.
What people call excellent and mediocre these days is often just the capacity to be economically over-ruthless, rather than contribute any good to society. We already have a wealth of ways that people can excel, even if we eradicated AI. So there's definitely no limitation on intelligent individuals to be excellent, even if we destroyed AI. So your argument really doesn't hold.
Edit: my goal isn't to protect the weak. I'd rather have everyone protected, including the very intelligent who still want to have a place to use their intelligence on their own and not be forced to use AI to keep up.
I see takes like this from time to time about everything.
They didn't say that.
As with all similar cases, they're allowed to advocate for whatever being dangerous, and you're allowed to say it isn't, the people who decide is all of us collectively and when we're at our best we do so on the basis of the actual arguments.
> AI isn't nukes - anyone can train a model at home.
(1) They were using an extreme to illustrate the point.
(2) Anyone can make a lot of things at home. I know two distinct ways to make a chemical weapon using only things I can find in a normal kitchen. That people can do a thing at home doesn't make the thing "not prohibited".
Most students today are AI fluent. Most teachers aren't. Students treat AI like Google Search, StackOverflow, GitHub, and every other dev tool.
I don't think the existence of the latter should mean we restrict access to AI for everyone, but I also don't think it's helpful to pretend AI is just this generation's TI-83.
"While some may use the tool irresponsibly, others will not, and therefore there's no need to restrict the tool. Society shouldn't handicap the majority to accommodate the minority."
You can choose to not engage with this critique but calling it a "false dichotomy" is in poor form. If anything, it makes me feel like you're not willing to entertain disagreement. You state that you want to start a discussion by expressing your opinion but I don't see a discussion here. I observe you expressing your opinion and dismissing criticism of that opinion as false.
In my mind a direct, good faith rebuttal would address the actual points - either disagree that the worst usage would lead to harm of the public or make a point (like the op tees up) that risking the public is one of worthy tradeoffs of freedom.
Again, even your expanded reply shrieks with false dichotomy. I never said ban every possible technology, only ones that are sufficiently dangerous.
No, they're arguing on the grounds that the tools are detrimental to the overwhelming majority in a way that also ends up being detrimental to the disciplined minority!
I'm not sure I agree, but either way you aren't properly engaging their actual argument.
The same line of thought can be used for any (new) tool, say a calculator, a computer or the internet. Shouldn't we try to find responsible ways of adopting LLMs, that empower the majority?
Yes, the same line of thought can. But we must also take power into account. The severity of the negative effects of a technology is proportional to its power and a calculator is relatively week.
> Shouldn't we try to find responsible ways of adopting LLMs, that empower the majority?
Not if there is no responsible way to adopt them because they are fundamentally against a happy existence by their very nature. Not all technology empowers, even when used completely fairly. Some technology approaches a pure arms race scenario, especially when the proportion of its effect is mainly economic efficiency without true life improvement, at least for the majority.
Of course, one can point to some benefits of LLMs, but my thesis is that the benefit/cost quantity approaches zero and thus crosses the point of diminishing returns to give us only a net negative in all possible worlds where the basic assumptions of human nature hold.
I could point out many more differences between the two situations but I won't because your lack of any intellectual effort doesn't even deserve a reply.
Crazy that I agreed with the first half of the sentence and was totally thrown off by the end. To me, “learning mode” is when I want the LLM. I’m in a new domain and I might not even know what to google yet, what libraries exist, what key words or concepts are relevant. That’s where an LLM shines. I can see basic generic code that’s well explained and quickly get the gist of something new. Then there’s “shipping mode” where quality is my priority, and subtle sneaky bugs really ought to be avoided—the kind I encounter so often with ai written code.
If the e-bike is an alternative to a road bike then yes. I'd argue that is almost never the case. The people I've spoken to are using them as an alternative to driving which is clearly beneficial to their fitness.
A ridiculous sentimental idea based on limited observation and bias against change that won't age well.
I ride about twice as much distance (mountain biking) after I got an ebike (per Strava). It's still a great workout.
Sample size one disclaimer...
A better biking analogy that I've used in the past is that it I wanted to go ride slick rock and have never ridden before, an ebike is not going to prevent me from endoing.
I’m a guy who likes to DO to validate assumptions: if there’s some task about how something should be written concurrently to be efficient and then we need some post processing to combine the results, etc, etc, well, before Claude Code, I’d write a scrappy prototype (think like a single MVC “slice” of all the distinct layers but all in a single Java file) to experiment, validate assumptions and uncover the unknown unknowns.
It’s how I approach programming and always will. I think writing a spec as an issue or ticket about something without getting your hands dirty will always be incomplete and at odds with reality. So I write, prototype and build.
With a “validated experiment” I’d still need a lot of cleaning up and post processing in a way to make it production ready. Now it’s a prompt! The learning is still the process of figuring things out and validating assumptions. But the “translation to formal code” part is basically solved.
Obviously, it’s also a great unblocking mechanism when I’m stuck on something be it a complex query or me FEELING an abstraction is wrong but not seeing a good one etc.
I really enjoyed this article. The numerous anecdotes from folks at RC was great. In particular thanks for sharing this video of voice coding [1].
This line in particular stood out to me that I use to think about LLMs myself:
"One particularly enthusiastic user of LLMs described having two modes: “shipping mode” and “learning mode,” with the former relying heavily on models and the latter involving no LLMs, at least for code generation."
Sometimes when I use Claude Code I either put it in Plan Mode or tell it to not write any code and just rubber duck with it until I come up with an approach I like and then just write the code myself. It's not as fast as writing the plan with Claude and asking it to write the code, but offers me more learning.
Is anyone John Henry-ing this question and having parallel teams build the same product at the same time?
--
"Technology Review," the magazine of MIT, ran a short article in January called "Housebreaking the Software" by Robert Cowen, science editor of the "Christian Science Monitor," in which he very sensibly said: "The general-purpose home computer for the average user has not yet arrived.
Neither the software nor the information services accessible via telephone are yet good enough to justify such a purchase unless there is a specialized need. Thus, if you have the cash for a home computer but no clear need for one yet, you would be better advised to put it in liquid investment for two or three more years." But in the next paragraph he says "Those who would stand aside from this revolution will, by this decade's end, find themselves as much of an anachronism as those who yearn for the good old one-horse shay." This is mostly just hot air.
What does it mean to be an anachronism? Am I one because I don't own a car or a TV? Is something bad supposed to happen to me because of that? What about the horse and buggy Amish? They are, as a group, the most successful farmers in the country, everywhere buying up farms that up-to-date high-tech farmers have had to sell because they couldn't pay the interest on the money they had to borrow to buy the fancy equipment.
Perhaps what Mr. Cowen is trying to say is that if I don't learn how to run the computers of 1982, I won't be able later, even if I want to, to learn to run the computers of 1990. Nonsense! Knowing how to run a 1982 computer will have little or nothing to do with knowing how to run a 1990 computer. And what about the children now being born and yet to be born? When they get old enough, they will, if they feel like it, learn to run the computers of the 1990s.
Well, if they can, then if I want to, I can. From being mostly meaningless, or, where meaningful, mostly wrong, these very typical words by Mr. Cowen are in method and intent exactly like all those ads that tell us that if we don't buy this deodorant or detergent or gadget or whatever, everyone else, even our friends, will despise, mock, and shun us the advertising industry's attack on the fragile self-esteem of millions of people. This using of people's fear to sell them things is destructive and morally disgusting.
The fact that the computer industry and its salesmen and prophets have taken this approach is the best reason in the world for being very skeptical of anything they say. Clever they may be, but they are mostly not to be trusted. What they want above all is not to make a better world, but to join the big list of computer millionaires.
A computer is, after all, not a revolution or a way of life but a tool, like a pen or wrench or typewriter or car. A good reason for buying and using a tool is that with it we can do something that we want or need to do better than we used to do it. A bad reason for buying a tool is just to have it, in which case it becomes, not a tool, but a toy.
On Computers Growing Without Schooling #29 September 1982
by John Holt.
You say this like it should give him more credibility. He created a homeschooling methodology that scores well below structured homeschooling in academic evaluations. And that's generously assuming it's being practiced in earnest rather than my experience with people doing it (effectively just child neglect with high minded justification)
I have absolutely no doubt that a quack like John Holt would love AI as a virtual babysitter for children.
He also both points out the silliness of the pitch for the hype-y tech of the day and calls that same tech a tool some might use with good reason.
That doesn't read to me like "tech hyped as inevitable is never okay" but more like "ignore sales pitches and evaluate the hyped tech for yourself." Your own judgment of today's hyped tech might be "kill it with fire." Still, it seems fully in sync with the spirit of that passage for Recurse to ask for a bunch of smart people's varied opinions on it and share them with the world.