AI coding tools can reduce productivity
secondthoughts.ai
secondthoughts.ai
That seems obvious, but a consequence of that is that people who are sceptical of ai (like me) only use it when they've exhausted other resources (like google). You ask very specific questions where not a lot of documentation is available and inevetably even o3 ends up being pretty useless.
Conversely there's people who love ai and use it for everything, and since the majority of the stuff they ask about is fairly simple and well documented (eg "Write me some typescript"), they rarely have a negative experience.
- Some people simply ask a lot more questions than others (this ignores whether they like or dislike AI), i.e. some people simply prefer to find things out more by themselves, and thus also use other resources like Google or Stack Overflow as a last resort. So their questions to an AI will likely be more complicated, because they already found out the easy parts by themselves.
- If I have to make the effort to explain to the AI in a sufficiently exhaustive way what I need (which I often have to do), I expect the answers of the AI to be really good. If it isn't, having explained my problem to the AI was simply a waste of time.
I find the worst part to be when it doesn't correct flaws in my assumptions.
For example, yesterday I asked it "what is the difference between these two Datadog queries"? And it replied something that was semi-correct, but it didn't discover the fundamental flaw - that the first one wasn't a valid query because of unbalanced parens. In fact, it turns out that the two strings (+ another one) would get concatenated and only then would it be a valid query.
A simple "the first string is not a valid query because of a missing closing paren" would have saved a lot of time in trying to understand this, and I suspect that's what I would have received if I had prompted it with "what's the problem with this query" but LLMs are just too sycophantic to help with these things.
But most other models don't.
I do have a custom instruction in place to ask if I'm aware of concepts related to my question - perhaps in coming up with these, it notices when something relevant hasn't been mentioned.
Many folks I know are skeptical of the hype, or maybe full on anti/distrustful, due to reasons I think are valid. But many of those same people have tried llm tools, maybe chatgpt or copilot or cursor, and recognize the value even w/ huge misgivings. Some of have gone further with tools like claude code and seen the real potential there, quite a step beyond fancy auto-complete or just-in-time agents...but even there you can end up in rabbit-holes and drowning in horrible design.
In your incredibly reductive scale, I'm closer to 'love' than 'skeptical', but I'm often much of both sides. But I'd never write a prompt like 'write me some typescript' for any real work, or honestly anything close to that, unless its just for memes or demonstrations.
But no-one who programs for a living uses prompts like that, at least not for real work. That is just silly talk.
The tone of your comment suggests that my comment upset you, which wasn't my intent. But you have to try to be a little generous when you read other peoples stuff, or these discussion will get very tedious quickly.
If we accept that AI is a tool, then then problem is the nature of the tool as it will vary heavily from individual to individual. This partially accounts for the ridiculous differences from self reported accounts of people, who use it on a regular basis.
And then, there is a possibility that my questions are not that unusual and/or are well documented ( quite possible ) so my perception of the usefulness of those answers is skewed.
My recent interaction with o4 was pretty decent on a very new ( by industry standards ) development and while documentation for it exists, it is a swirling vortex of insanity from where I sit. I was actually amazed to see how easily 4o saw some of those discrepancies and listed those to me along with likely pitfalls that may come with it. We will be able to find if that prediction holds v.soon.
What I am saying is that it has its uses.
A tool that constantly adapts to how it is used will frequently be misaligned with user intent. Language models constantly change their own behavior based on the specific phrasing you gave it, the context you deployed it in, and the inherent randomness in token generation. Its capacity to be used as a tool will be inherently limited by this unpredictability.
That might seem like a distinction without a point but people are not tools and people are useful. If someone argues people are tools then they don't fit your definition any more than an LLM does.
LLMs are as useful as people I would say, and as unpredictable. I run a small consulting company and we develop custom software for businesses, government, whomever. I often have to interface with other teams and other people. To get the most out of working with people you have to know how to talk with them.
Also, I noticed you didn't say anything negative about LLMs :) You just compared it w/ tooling. So I'm curious, do you think outside of the unsuitability of consideration as a tool, do you find LLMs useful?
You have any example questions where o3 failed to be helpful?
I use it pretty similarly to you, only resorting to it to unblock myself basically, otherwise I'm mostly the one doing the actual work, LLMs help with specific functions or specific blockers, or exploring new "spaces". But almost all the times I've gotten stuck, o3 (and o3-pro mode) managed to unstuck me, once I've figured out the right way to ask the question, even when my own searching and reading didn't help.
For research I'm enjoying asking ChatGPT to annotate its responses with sources and reading those; in some cases I've found SIGGRAPH papers that I wouldn't have stumbled upon otherwise, and it's nice to get them all in a response.
ChatGPT (4o, if it's of any interest) is very knowledgeable about DirectX12 (which we switched to just this week) and I've gained tons of peripheral knowledge with regards to the things I've been battling with, but only one out of four times has it been able to actually diagnose directly what the issue was; three separate times it's been something it didn't really bring up or note in any meaningful regard. What helped was really just me writing about it, thinking about everything around it and for that it's been very helpful.
Realistically, if someone let an agent running on this stuff loose on our code base it would likely end up wasting days of time and still not fix the issue. Even worse, the results would have to be tested on a specific GPU to even trigger the issue to begin with.
It seems to me that fancy auto-complete is likely the best this would be able to do still, and I actually like it for that. I don't use LLM-assisted auto-complete anymore, but I used to use GitHub Copilot back in 2022 and it was more productive than my brief tests of agents.
If I were to regularly use LLMs for actual programmit it would most likely be just for tab-completion of "rest of expressions" or 1 line at a time, but probably with local LLMs.
For hard questions, I prefer to use my own skills, because AI often regurgitates what I'm already aware. I still ask AI in the off-chance it comes up with something cool, but most often, I have to do it myself.
The asymptote approached by software engineering is GDP=$0 because all problems are solved by maximally efficient automation. Never gonna happen, but progress on that path is a decent proxy for efficiency.
Too often the job is about introducing problems that are good for some company's bottom line, but that's the opposite of efficiency.
Some of the most productive devs don't get paid by the big corps who make use of their open source projects, hence the constant urging of corps and people to sponsor projects they make money via.
What about countries? In my Poland $25k would be an amazing salary for a senior while in USA fresh grads can earn $80k. Are they more productive?
... at the same time, given same seniority, job and location - I'd be willing to say it wouldn't be a bad heuristic.
10k PLN monthly would be a very good salary - and also perhaps a norm, as far as I know from experience and talking to peers.
I never understood how there can be a significant wage gap between these countries and Western Europe for jobs that are anyway done online.
I made great money running my own businesses, but the vast majority of the programming was by people I hired. I’m a decent talent, but that gave me the ability to hire better ones than me.
Changing jobs typically brings a higher salary than your previous job. Are you saying that I'm significantly more productive right after changing jobs than right before?
I recently moved from being employed by a company to do software development, to running my own software development company and doing consulting work for others. I can now put in significantly fewer hours, doing the same kind of work (sometimes even on the same projects that I worked on before), and make more money. Am I now significantly more productive? I don't feel more productive, I just learned to charge more for my time.
IMO, your suggestion falls on its own ridiculousness.
ON the other hand it makes no sense from some points of view. For example, if you get a pay rise that does not mean you are more productive.
i have to bill my clients and have documented around 3 weeks of development time saved by using LLMs to port other client systems to our system since December. on one hand this means we should probably update our cost estimates, but im not management so for the time ive decided to use the saved time to overdeliver on quality
eventually clients might get wise and not want to overdeliver on quality and we would charge less according to time saved by LLMs. despite a measured increase in "productivity" i would be generating less $ because my overall billable hour % decreases
hopefully overdelivering now reduces tech debt to reduce overhead and introduces new features which can increase our client pipeline to offset the eventual shift in what we charge our clients. thats about all the agency i can find in this situation
Like can we determine the productivity of doctors, lawyers, journalists, or pastry chefs?
What job out there is so simple that we can meaningfully measure all the positive and negative effects of the worker as well as account for different conditions between workers.
I could probably get behind the idea that you could measure productivity for professional poker players (given a long enough evaluation period). Hard to think of much else.
The British government (probably not any worse than anyone else, just what I am most familiar with) does measure the productivity of the NHS: https://www.england.nhs.uk/long-read/nhs-productivity/ (including doctors, obviously).
They also try to measure the performance of teachers and schools and introduced performance league tables and special exams (SATS - exams sat at various ages school children in the state system, nothing like the American exams with the same name) to do this more pervasively. They made it better by creating multi-academy trusts which adds a layer of management running multi-schools so even more people want even more metrics.
The same for police, and pretty much everything else.
And to be fair, some crud work is repetitive enough so it should be possible to get a fair measure of at least the difference in speed between developers.
But that building simple crud services with rest interfaces takes as much time as it does is a failure of the tools we use.
Yes, yes we can.
Programmers really need to stop this cope about us being such special snowflakes that we can't be assessed and that our maangers just need to take that we're worth keeping around on good faith.
Of course we can. But can we do it in a meaningful way, such that the metric itself doesn't become a subject to optimization?
"When a measure becomes a target, it ceases to be a good measure"
By making the metrics part of a sustaintable company-wide goal. If there's a company-wide goal to increase X kind of revenue by Y% making actionable targets on how a team can contribute (not lazy shit like "our changes should contribute Z% of that Y%"), and within that create for a person another smaller metric based on that.
Also, medical facilities… you certainly could define it as profit, but that bothers me and many other people.
You could define it as patients seen, or “cured” but that incentivizes very quick but probably poor care.
You could define it as intensity of treatment or amount of care given, but you’d probably end up in a situation where 1 incredibly sick person has every doctor treating them.
You could define it as…
In real world, most things don't work out that way. What metrics do you use to measure surgeons' success? If you use fatality rate, then as a result surgeons will refuse to do more risky surgeries which will put their ratings at risk, which makes the healthcare worse, instead of better.
Could you make an effort to explain how, or at the very least link to some reasoning? Otherwise your comment is basically the equivalent of “nuh-uh”, which doesn’t meaningfully contribute to the discussion.
> Programmers really need to stop this cope about us being such special snowflakes
Which is not at all what is happening in your parent comment. On the contrary, they’re putting developers on even footing with other professions.
You can look at the kind of work they're doing, how effective their solutions are, and how long it takes them to do it. That's the basics of it across a wide range of professions. Now, there's no one-size-fits-all metric or formula you can just calculate based on objective facts for most of this, because the work is more varied than e.g. factory work, but it's also not impossible to make the comparison, if you actually understand the work reasonably and you use judgement.
In the case of this study, because the assignment of the comparison they were doing was random, then just measuring time to completion across a range of tasks is a perfectly reasonable metric, because there's nothing to really bias the outcome, just a lot of factors that add noise instead. But it is worth noting that the result is a very broad average, and there is likely a very complicated distribution of details underneath, which is much harder to measure.
AKA, be subjective! Which people are wary of, because what it brings is politics and tribalism.
Like I get that in SWE (like all other fields), managers have to make judgement calls and try to evaluate which reports contribute the most, but the GP post seemed surprised that this wasn't a solved problem by now, which just seems incomprehensible to me.
At the end of the road. Patient outcome and contentedness compared to others with similar indications. Patients seen and all that is that sort of short-term BS that you see everywhere that's giving metrics a bad name. It'd be like determining a mechanic's productivity by how many times he twisted a wrench.
Which would incentivise doctors to refuse to treat patients who are more ill, lest they risk their ratings go down.
Well I would first of all remark that this doesn’t seem and to be how it’s normally done as I’ve never been asked to rate my “contentedness” or similar with my medical care.
And where is the “end of the road?” Most medical interventions could be plausibly evaluated at all manner of different intervals.
Also, “similar indications” is doing a lot of work here. Patient outcomes are often influenced more by the individual than the doctor. By the time you bucket all the patients by age, diet, activity level, smoking status, alcohol intake, metabolic health, bmi, family history, etc…buckets are going to be pretty tiny. Clinics and hospitals aren’t that big, there won’t be anything to compare. If you only bucket the most obvious categories like age, you’ll have comparisons, but it will just be noise.
and how you would achieve it? "similar indications" would be coming from doctor that you are trying to rate
rating "contentedness" gets you doctors prescribing useless medications to keep patients happy
expert surgeons have often bad survival rates as they get complicated cases, and trying to rate how complicated cases are to compare two experts would be nightmare as bad as rating doctors - so you only replace one hard problem with another as hard problem
We have that shit like that too in SWE. Lines of code, github issues closed, features shipped, etc…
The hard thing is occupations where the quantity of effort is unrelated to the result due to the vast number of confounding factors.
https://secondthoughts.ai/p/ai-coding-slowdown
HN discussion: https://news.ycombinator.com/item?id=44526912
Management is forced to rely on various metrics which are gamed or inaccurate.
I suppose it would be simpler to compare productivity for people working on standard, "normalized" tasks, but often every other task a programmer is assigned is something different to the previous one, and different developers get different tasks.
It's difficult to measure productivity based on real-world work, but we can create an artificial experiment: give N programmers the same M "normal", everyday tasks and observe whether those using AI tools complete them more quickly.
This is somewhat similar to athletic competitions — artificial in nature, yet widely accepted as a way to compare runners’ performance.
Focusing on the speed of outputting code or closing tickets is shortsighted from a software engineering standpoint.
My entire life, I have written “ship” software. It’s been pretty easy to say what my “product” is.
But I have also worked at a fairly small scale, in very small teams (often, only me). I was paid to manage a team, but it was a fairly small team, with highly measurable output. Personally, I have been writing software as free, open-source stuff, and it was easy to measure.
Some time ago, someone posted a story about how most software engineers have hardly ever actually shipped anything. I can’t even imagine that. I would find that incredibly depressing.
It would also make productivity pretty hard to measure. If I spent six months, working on something that never made it out of the cręche, would that mean all my work was for nothing?
Also, really experienced engineers write a lot less code (that does a lot more). They may spend four hours, writing a highly efficient 20-line method, while a less-experienced engineer might write a passable 100-line method in a couple of hours. The experienced engineers’ work might be “one and done,” never needing revision, while the less-experienced engineer’s work is a slow bug farm (loaded with million-dollar security vulnerability tech debt), which means that the productivity is actually deferred, for the more experienced engineer. Their manager may like the less-experienced engineer's work, because they make a lot more noise, doing it, are "faster," and give MOAR LINES. The "down-the-road" tech debt is of no concern to the manager.
I worked for a company that held the engineer Accountable, even if the issue appears, two years after shipping. It encouraged engineers to do their homework, and each team had a dedicated testing section, to ensure that they didn't ship bugs.
When I ask ChatGPT (for example) for a code solution, I find that it’s usually quite “naive” (pretty prolix). I usually end up rewriting it. That doesn’t mean that’s a bad thing, though. It gives me a useful “starting point,” and can save me several hours of experimenting.
The usual counter-point is that if you (commonly) write code by experimenting, you are doing it wrong. Better think the problem through, and then write decent code (that you finally turn into great code). If the code that you start with is that as "naive" as you describe, in my experience it is nearly always better to throw it away (you can't make gold out of shit) and completely start over, i.e. think the problem through and then write decent code.
I find they often cause more trouble than they are worth, because they are completely wrong, and need to be “unlearned.”
"Experimenting" is a vital part of my process. I call it "Evolutionary Design,"[0] and it involves a lot of iteration. I have found that it's vital to UI[1], because I can almost never predict how UI will act, when actually presented to the user. The same goes for a lot of communication workflows. I have to "run it up the flagpole, and see who salutes." I almost always find that my theorized approach has issues, and I need to make changes. The old "Measure twice; cut once" approach to software development has caused me great trouble, over the years, and I have found that I need to adjust to new tools, and new contexts.
For example, right now, I am revamping one of my UI widgets[2]. It started as a minor tweak for iOS26, but I realized that it's a bit "long in the tooth," and that I can make it more robust, simple, and usable. I have been running the test harness all morning, seeing issues, and going back to the code, and tweaking.
[0] https://littlegreenviper.com/evolutionary-design-specificati...
How I measure performance is how many features I can implement in a given period of time.
It's nice that people have done studies and have opinions, but for me, it's 10x to 20x better.
Someone already operating at the very limit of their abilities doing stuff that is for them high complexity, high cognitive load, detail intense, and tactically non-obvious? Even a machine that just handed you the perfect code can't 20x your real output, even if it gave you the source file at 20x your native sophistication you wouldn't be able to build and deploy it, let alone make changes to it.
But even if it's the last 5-20% after you're already operating at your very limit and trying to hit your limit every single day is massive, it makes a bunch of stuff on the bubble go from "not realistic" to "we did that".
A key skill is to sense when the AI is starting to guess for solutions (no different to human devs) and then either lean into another AI or reset context and start over.
I'm finding the code quality increase greatly with the addition of the text 'and please follow best practices because will be pen tested on this!' and wow.. it takes it much more seriously.
It would be interesting to set up a MCP style interface, but even me copy/pasting between windows was constructive.
The time this worked best was when I was building a security model for an API that had to be flexible and follow best practices. It was interesting seeing ChatGPT compare and contrast against major API vendors, and Claude Code asking the detailed implementation questions.
The final output was a pragmatic middle-ground between simplistic and way too complex.
Most of the coding needed to give people CRUD interfaces to resources is all about copy / pasting and integrating tools together.
Sort of like the old days when we were patching all those copy/paste's from StackOverflow.
Too little of full stack application writing is truly unique.
Also I disagree. For web dev atleast, most people are just rewriting the same stuff in a different order. Even though the entire project might be complex from a high level perspective, when you dive into the components or even just a single route it ain't "high complexity" at all and since I believe most jobs are in web / app dev which just recycles the same code over and over again that's why there's a lot of people claiming huge boosts to productivity.
The difficult part is reading thousand lines of unfamiliar code to measure the impact of a fix, finding the fix by reasoning about the whole moduke, designing a feature for long term maintainability,…
Note that all of them requires thinking and not much coding. Coding is easy, especially when you’ve done all the (correct?) thinking beforehand.
When you zoom in, even this kind of work isn't uniform - a lot of it is still shaving yaks, boring chores, and tasks that are hard dependencies for the work that is truly cognitively demanding, but themselves are easy(ish) annoyances. It's those subtasks - and the extra burden of mentally keeping track of them - that sets the limit of what even the most skilled, productive engineer can do. Offloading some of that to AI lets one free some mental capacity for work that actually benefits from that.
> Even a machine that just handed you the perfect code can't 20x your real output, even if it gave you the source file at 20x your native sophistication you wouldn't be able to build and deploy it, let alone make changes to it.
Not true if you use it right.
You're probably following the "grug developer" philosophy, as it's popular these days (as well as "but think of the juniors!", which is the perceived ideal in the current zeitgeist). By design, this turns coding into boring, low-cognitive-load work. Reviewing such code is, thus, easier (and less demoralizing) than writing it.
20x is probably a bit much across the board, but for the technical part, I can believe it - there's too much unavoidable but trivial bullshit involved in software these days (build scripts, Dockerfies, IaaS). Preventing deep context switching on those is a big time saver.
Tons of dev work is not exciting, I have already launched a solo dev startup that was acquired, and the 'fun' part of that coding was minimal. Too much was the scaffolding, CRUD endpoints, web forms, build scripts, endpoint documentation, and the true innovative stuff was such a small part of the whole project. Of the 14 months of work, only 1 month was truly innovative.
Yeah, I'm not a dev but I can see why this is true, because it's also the argument I use in my job as an academic. Some people say "but your work is intellectually complex, how can you trust LLMs to do research, etc.?", which of course, I don't. But 80% of the job is not actually incrementally complex, it's routine stuff. These days I'm writing the final report of a project and half of the text is being generated by Gemini, when I write the data management plan (which is even more useless) probably 90% will be generated by Gemini. This frees a lot of time that I can devote to the actual research. And the same when I use it to polish a grant proposal, generate me some code for a chart in a paper, reformat a LaTeX table, brainstorm some initial ideas, come up with an exercise for an exam, etc.
When I said that after you've done all the other stuff, I was including cutting all the ridiculous bullshit that's been foisted on an entire generation of hackers to buy yachts for Bezos and shit.
I build clean libraries from source with correct `pkg-info` and then anything will build against it. I have well-maintained Debian and NixOS configurations that run on non-virtualized hardware. I use an `emacs` configuration that is built-to-specifications, and best-in-class open builds for other important editors.
I don't even know why someone would want a model spewing more of that garbage onto the road in front of them until you're running a tight, optimized stack to begin with, then the model emulates to some degree the things it sees, and they're also good.
Lagged-ass electron apps are a choice: run neovim or emacs or zed, I have Cursor installed, once in a while I need vscode for something, but how often is someone dictating my editor?
I have to target OCI container platforms for work sometimes, that's what Arion and nix2container are for. Ditto package managers: uv and bun exist and can interact with legacy requirements.txt and package.json in most cases.
Anything from a Helm chart to the configuration for ddagent can be written from nixlang and into a .deb.
My current job has a ton of Docker on GCE running TypeScript, I have to emit compatible code and configuration, but no one stands over my shoulders to make sure I'm doing the Cloud Approved jank path or having a bash script or Haskell program print it. I have a Jank Stack Compatibility Layer that builds all that nonsense.
Job after job there's a little setup cost and people look at me funny, 6 months in my desk is an island of high-velocity sanity people are starting to use because I carry a "glory days FAANG" toolkit around and compile reasonable plain text into whatever ripoff cloud garbage is getting pimped this week.
It's a pretty extreme workplace where you can't run reasonable Unix on your own machine and submit compiler output instead of typing for the truly mandatory jank integration points.
Maybe, but I don't feel (of course, I could be wrong) that doing boring tasks take away any mental capacity; they feel more like fidgeting while I think. If a tool could do the boring things it may free my time to do other boring work that allows me to think - like doing the dishes - provided I don't have to carefully review the code.
Another issue (that I asked about yesterday [1]) is that seemingly boring tasks may end up being more subtle once you start coding them, and while I don't care too much about the quality of the code in the early iterations of the project, I have to be able to trust that whatever does the coding for me will come back and report any difficulties I hadn't anticipated.
> Reviewing such code is, thus, easier (and less demoralizing) than writing it.
That might well be true, but since writing it doesn't cost me much to begin with, the benefit might not be large. Don't get me wrong, I would still take it, but only if I could fully trust the agent to tell me what subtleties it encountered.
> there's too much unavoidable but trivial bullshit involved in software these days (build scripts, Dockerfies, IaaS). Preventing deep context switching on those is a big time saver.
If work is truly trivial, I'd like it to be automated by something that I can trust to do trivial work well and/or tell me when things aren't as trivial and I should pay attention to some detail I overlooked.
We can generally trust machines to either work reliably or fail with some clear indication. People might not be fully reliable, but we can generally trust them to report back with important questions they have or information they've learnt while doing the job. From the reports I've seen about using coding agents, they work like neither. You can neither trust them to succeed or fail reliably, nor can you trust them to come back with pertinent questions or information. Without either kind of trust, I don't think that "offloading" work to them would truly feel like offloading. I'm sure some people can work with that, but I think I'll wait until I can trust the agents.
How much of the code you write is actually like this? I work in the domain of data modeling, for me once the math is worked out majority of the code is "trivial". The kind of code you are talking about is maybe 20% of my time. Honestly, also the most enjoyable 20%. I will be very happy if that is all I would work on while rest of it done by AI.
Like: Why isn’t this working? Here Claude read this like 90 page PDF and tell me where I went wrong interfacing with this SDK.
Ohh I accidentally passed async_context_background_threading_safe instead of async_context_thread_safe_poll and it’s so now it’s panicking. Wow that would have taken me forever.
I'm leaning into the future growth of AI capabilities to help me here, otherwise I'll have to do it myself.
That is a tomorrow problem, too much project structure/functionality to get right first.
With most projects where innovation is a key requirement, the goal isn't to write textbook quality code, it's to prove your ideas work and quickly evolve the project.
Once you have an idea of how it's going to work, you can then choose to start over from scratch or continue on and clean up all the bits you skipped over.
Right now I'm in the innovation cycle, and having AI able to pick up whole API path strategies and pivot them, is incredibly amazing.
How many times have you used large API's and seen clear hands of different developers and URI strategies, with an AI, you just pivot.
Code quality and pen tests are critical, but they can come later.
In my experience, no.
These kind of shortcuts taken at the beginning of the project is why velocity have a sharp descent after some times. Because you’re either spending time undoing all of it (unlikely to be allowed) or you’re fighting in the code jungle trying to get some feature out.
https://repo.autonoma.ca/notanexus.git
I don't know the PDF.js library. Writing both the client- and server-side for a PDF annotation editor would have taken 60 hours, maybe more. Instead, a combination Copilot, DeepSeek, Claude, and Gemini yielded a working prototype in under 6 hours:
https://repo.autonoma.ca/notanexus.git/tree/HEAD/src/js
I wrote maybe 3 lines of JavaScript, the rest was all prompted.
How do you know? Seems to me you’re making the exact same estimation mistake of the people in the study.
> Instead, a combination Copilot, DeepSeek, Claude, and Gemini yielded a working prototype in under 6 hours
Six hours for a prototype using four LLMs? That is not impressive, it sounds insane and a tremendous mess that will take so long to dig out of the prototype stage it’ll effectively require a rewrite.
And why are you comparing an LLM prototype to a finished product “by hand” (I surely hope you’re not suggesting such a prototype would take sixty hours)? That is disingenuous and skewing the numbers.
I have over 20 years of web development experience and 40 years of general experience writing software. I wrote the authors and they confirmed my thoughts:
"I totally believe it! Per the paper abstract, we find many factors driving results - and one of the factors is how experienced the developers are on the codebase, and how big/complex the codebases are.
"Given that this was a new and unfamiliar domain and new codebase, I would expect there to be much more speedup than the domain we studied!"
> Six hours for a prototype using four LLMs?
They have limits on the number of queries, so I used four different LLMs in tandem to circumvent query limits. I didn't write it four times using four different LLMs.
> it sounds insane and a tremendous mess
I posted the code. It's well organized, has few (if any) encapsulation violations, sticks to OOP quite well, works, and---if I knew the PDF.js API---would be fairly easy to maintain.
Yes, I stand by my claim that writing this annotation editor (PHP, HTML, CSS, and JS) would take me about 60 hours by hand and about 6 hours using the LLMs.
I am wondering, what sort of tasks are you seeing these x20 boost?
Words without actions are junk. You are asserting something you have no proof for. Proove it then. Amaze us all with your productivity, out in the open. Shred those pilled up open issues on open source projects and then give us a report of how fast-easy it.
If it is "easily true" you'll be done by next month
But I agree, you don't need to prove anything to me. You need to proove it to yourself first.
I scoped out a body of work and even with the AI assisting on building cards and feature documentation, it came to about 2 to 4 weeks to implement.
It was done in 2 days.
The key I've found with working as fast as possible is to have planning sessions with Claude Code and make it challenge you and ask tons of questions. Then get it to break the work into 'cards' (think Jira, but they are just .md files in your repo) and then maintain a todo.md and done.md file pair that sorts and organizes work flow.
Then start a new context, tell it to review todo.md and pick up next task, and burn through it, when done, commit and update todo.md and done.md, /compact and you're off on the next.
It's more than AI hinting at what to do, it's a whole new way of working with rigor and structure around it. Then you just focus fire on the next card, and the next, and if you ever think up new features, then card it up and put it in the work queue.
If one of these things isn’t true, you’re either a fool or those productivity increases aren’t real.
10x to 20x is in relation to time, so something that would have taken 2 weeks (80 hours) would be done in 8 hours to be 10x.
> Do you clock off at Monday lunchtime and spend the rest of the week playing video games? Did your boss fire nineteen developers and give their jobs to you?
In other words, how are you taking advantage of all that extra time you claim to have?
A simple example: if someone patents a machine that makes canned tuna 10 times faster than how they're currently being made, would tuna factories make 10 times more money? The answer is obviously no. Actually, they'd make the same money as before, or even less than that. Only the one who makes such a machine (and the consumers of tuna cans) would be benefited.
Claude code has made bootstrapping a new project, searching for API docs, troubleshooting, summarizing code, finding a GitHub project, building unit tests, refactoring, etc easily 20x faster.
It’s the context switching that is EXTREMELY expensive for a person, but costless for the LLM. I can focus on strategy (planning features) instead of being bogged down in lots of tactics (code warnings, syntax errors).
Claude Code is amazing, but the 20x gains aren’t evenly distributed. There are some projects that are too specialized (obscure languages, repos larger than the LLM’s context window, concepts that aren’t directly applicable to any codebase in their training corpus, etc). But for those of us using common languages and commodity projects, it’s a massive force multiplier.
I built my second iOS app (Swift) in about 3 days x 8 hours of vibe coding. A vocab practice app with adjustable learning profile, 3 different testing mechanisms, gamification (awards, badges), iOS notifications, text to speech, etc. My first iOS app was smaller, mostly a fork of another app, and took me 4 weeks of long days. 20x speed up with Claude Code is realistic.
And it saves even more time when researching + planning which features to add.
There should be a FOSS project explosion if those numbers were true by now. Commercial products too.
Jokes aside, if 20x was on the table for any kind of meaningful work we wouldn't need to wait for much of anything, entire parts of industry would be invented and technically reworked by now. It's most likely ~1.25x for what is mostly trivial work that approaches 95% boilerplate and zero actual design work.
If you read calrain's posts the 20x number is taken from the "fact" that sometimes (not consistently or most of the time) something that was estimated at 2 weeks or 80 hours (who knows what it was and how that number came to be?) took 2 hours instead. That's not just some minor detail; it's just not a sound way of thinking about productivity increases.
I ended up asking it how it wanted to work and would an 'AdminKit Template' work to get things moving.
It recommended AdminKit and that was a good move.
For me, custom UI's aren't a big part of the solution, I just need web pages to manage CRUD endpoints to manage the product.
AdminKit has been a good fit so far, but it was a fresh start, no migration.
It mentioned AdminKit and it worked out pretty well.
Recently, there was story about developer who was able to crush interview and got parallel full-time jobs in several start-ups. Initially he was able to deliver but then not so much.
Somehow your case is reminding this to me, where AI is this overemployed developer.
So I work 8 hours a day (to get money to eat) and code another 4 hours at home at night.
Weekends are both 10 hour days, and then rinse / repeat.
Unfortunately some projects are just hard to do and until now, they were too hard to attempt to solve solo. But with AI assistance, I am literally moving mountains.
The project may still be a failure but at least it will fail faster, no different to the pre-AI days.
It means you can replace a whole team of developers alone.
I can believe that some tasks are speed up by 10x or even 20x, but I find very hard to believe it's the average of your productivity (maintaining good code quality)
So me finishing a carded up block of work that is expected to take 2 weeks (80 hours) and I get it done in 1 day (8 hours) then that would be a 10x boost.
There are always tar pits of time where you are no better off with AI, but sometimes it's 20x.
I've setup development teams in the past, and have have been coding since the late 70's, so I am sort of aware of my capabilities.
It super depends on the type of work you're doing.
I mean, it's literally unbelievable.
This is absurd measuring. You can’t in good faith claim a 20x improvement if it only happens “sometimes” and other times it’s a time sink.
The more detail you keep providing in this thread, the clearer it becomes your assessment lands somewhere between the disingenuous and the delusional.
Does that mean you deliver the same amount of code in the same time with 20x less bugs?
Or the same quality code in 20x less time?
Or 10x less bugs in 2x less time?
If you had a hammer which could drive a nail through a plank 20x faster but took 60x longer to prepare before each strike, claiming 20x gains would be disingenuous.
Sometimes it doesn't help at all. Other times it spits out several hours of work in seconds.
It's like asking what is the weighted average of 1 and infinity? Even if you can quantify how many 1s and how many infinities there are, the answer is always going to be nonsensical.
(I don't think it's 20x, it's most likely hyperbole. People aren't that unique and it's not hard to see that people who use LLMs are often lulled into thinking they're more valuable to them than they actually are, especially when they "do more", i.e. they're a magic little person program that seems to do tasks on their own as opposed to glorified auto-complete that probably by raw numbers is actually more productive.)
Stage magicians say that the magic is done in the audiences memory after the trick is done. It's the effect of the activity.
AI coding tools makes developers happier and able to spend more brain power on actually difficult things. But overall perhaps the amount of work isn't in orders of magnitudes it just feels like it.
Waze the navigation app routes you in non standard routes so that you are not stuck in traffic, so it feels fast that you are making progress. But the time taken may be longer and the distance travelled may be further!
Being in stuck traffic and not moving even for a little bit makes you feel that time has stopped, it's boring and frustrating. Now developers need never be stuck. Their roads will be clear, but they may take longer routes.
We get little boosts of dopamine using AI tools to do stuff. Perhaps we used these signals as indicators of productivity "Ahh that days work felt good, I did a lot"
You're not "stuck in traffic", you are the traffic. If the app distributes users around and this makes it so they don't end up in traffic jams, it's effectively preventing traffic jams from forming
I liked your washing machine vs. sink example that I see you just edited out. The machine may do it slower and less efficiently than you'd do in the sink, but the machine runs in parallel, freeing you to do something else. So is with good use of LLMs.
For Waze, even if you are traffic and others go around you, you still may get there quicker and your car use less energy than taking the suggested route that feels faster. Others may feel happier and feel like they were faster though. Indeed they were faster but might have taken a longer journey.
Also, generally most people don't use the app around here to effect significant road use changes. But if they did im not sure (but I'm having fun trying to think) what metaphor we can apply to the current topic :)
Can't help but note that in 99% cases this "difficult things" trope makes little sense. In most jobs, the freed time is either spent on other stupid tasks or is lost due to org inefficiencies, or is just procrastinated.
Please don't speak for all developers when you say stuff like this
AI coding tools make me miserable to use
When a measure becomes a target, it ceases to be a good measure.
You can build a new product company with 20 people. Probably in the same domain as you are in right now.
So were the people taking the study. Which is why we do these, to understand where our understanding of ourselves is lacking.
Maybe you are special and do get extra gains. Or maybe you are as wrong about yourself as everyone else and are overestimating the gains you think you have.
Example: using LeafletJS — not hard, but I didn't want to have to search all over to figure out how to use it.
Example: other web page development requiring dropping image files, complicated scrolling, split-views, etc.
In short, there are projects I have put off in the past but eagerly begin now that LLMs are there to guide me. It's difficult to compare times and productivity in cases like that.
When I'm working with platforms/languages/frameworks I am already deeply familiar with I don't think they save me much time at all. When I've tried to use them in this context they seem to save me a bunch of time in some situations, but also cost me a bunch of time in others resulting in basically a wash as far as time saved goes.
And for me a wash isn't worth the long-term cost of losing touch with the code by not being the one to have crafted it.
But when it comes to environments I'm not intimately familiar with they can provide a very easy on-ramp that is a much more pleasant experience than trying to figure things out through often iffy technical documentation or code samples.
Leaflet doc is single page document with examples you can copy-paste. There is page navogation at the top. Also ctrl/cmd+f and keyword seems quicker than writing the prompt.
Still, when I simply told Claude that I wanted the pins to group together when zoomed out — it immediately knew I meant "clustering" and added the proper import to the top of the HTML file ... got it done.
> To directly measure the real-world impact of AI tools on software development, we recruited 16 experienced developers from large open-source repositories (averaging 22k+ stars and 1M+ lines of code) that they’ve contributed to for multiple years. Developers provide lists of real issues (246 total) that would be valuable to the repository—bug fixes, features, and refactors that would normally be part of their regular work. Then, we randomly assign each issue to either allow or disallow use of AI while working on the issue. When AI is allowed, developers can use any tools they choose (primarily Cursor Pro with Claude 3.5/3.7 Sonnet—frontier models at the time of the study); when disallowed, they work without generative AI assistance. Developers complete these tasks (which average two hours each) while recording their screens, then self-report the total implementation time they needed. We pay developers $150/hr as compensation for their participation in the study.
So it's a small sample size of 16 developers. And it sounds like different tasks were (randomly) assigned to the no-AI and with-AI groups - so the control group doesn't have the same tasks as the experimental group. I think this could lead to some pretty noisy data.
Interestingly - small sample size isn't in the list of objections that the auther includes under "Addressing Every Objection You Thought Of, And Some You Didn’t".
I do think it's an interesting study. But would want to see if the results could be reproduced before reading into it too much.
But in terms of pure statistical validity, I don't think it matters.
I think that's where you get 10-20x. When you're working on niche stuff it's either not gonna work or work poorly.
For example right now I need to figure out why an ffmpeg filter doesn't do X thing smoothly, even though the C code is tiny for the filter and it's self contained.. Gemini refuses to add comments to the code. It just apologizes for not being able to add comments to 150 lines of code lol.
However for building an ffmpeg pipeline in python I was dumbfounded how fast I was prototyping stuff and building fairly complex filter chains which if I had to do by hand just by reading the docs it would've taken me a whole lot more time, effort and frustration but was a joy to figure out with Gemini.
So going back to the study, IMO it's flawed because by definition working on new features for open source projects wouldn't be the bread and butter of LLMs however most people aren't working on stuff like this, they're rewriting the same code that 10000 other people have written but with their own tiny little twist or whatever.
My analogy to this is seeing people spend time trying to figure out how to change colors, draw shapes in powerpoint, rather than focus on the content and presentation. So here, we have developers now focusing their efforts on correcting the AI output, rather than doing the research and improving their ability to deliver code in the future.
Hmm...
> I think this for me is the most worrying: "You can see that for AI Allowed tasks, developers spent less time researching and writing code".
A Kindle is exactly the kind of device you would research and educate yourself via and the quantity of books has nothing to do with the reading of them or contents thereof. Terrible comparison.
When I’m in the “zone” I wouldn’t go near an LLM, but when I’ve fallen out of the “zone” they can be useful tools in getting me back into it, or just finishing that one extra thing before signing off for the day
I think the right answer to “does LLM use help or hinder developer productivity” is “it depends on how you use them”
Arguably, on a single coding task, I don't really move that much faster. However, I have much, much more brain capacity left both while coding and when I'm done coding.
This has two knock on effects:
1. Most simply, I'm productive for longer. Since LLMs are doing a lot of the heavy lifting, my brain doesn't have to constantly think. This is especially important in time periods where I'd previously have too little mental energy to think deeply about code.
2. I can do other things while coding. Well, right now, Cursor is cycling on a simple task while I type this. Most days, though, I'm responding to customers, working on documentation/planning, or doing some other non-coding task that's critical to my workflows. This is actually where I find my biggest productivity gains. Instead of coding THEN X, I can now do coding WITH X.
Context shifting while trying to code seems like a bad idea to me
Maybe you're some incredible multi -tasking genius able to change tasks rapidly without losing any of the details, but I suspect if most people tried this workflow they would produce worse code and also whatever their other output is would be low quality too
This type of stuff is great to have an LLM tackle while I work on something else.
I'm a frontend guy, been using Claude Code for a couple of weeks now. It's been able to speed up some boilerplate, it's sped up a lot of "naming is hard" conversations I like to have (but my coworkers probably don't, lol), it's enabled me to do a lot more stuff in my most recent project.
But for a task or two I suspect that it has slowed me down. If I'm unable to articulate the problem well enough and the problem is hard enough you can go in circles for awhile. And I think the nature of "the right answer is just around the corner" makes it hard to timebox or find a specific point where you say "yup, time to ditch this and do it the old-fashioned way". There is a bit of a slot-machine effect here.
Likely more, as it takes longer for you to activate your brain when your first thought is to ask an LLM rather than solve it yourself. Its like people reaching for a calculator to do 4+5, that doesn't make you faster or more accurate.
edit: should have mentioned the low-level stuff I work on is mature code and a lot of times novel.
Whether or not a general purpose foundation model for coding is trained on more backend or frontend code is largely irrelevant in this specific context.
I work at a shop where we do all custom frontend work and it's just not up to the task. And, while it has chipped in on some accessibility features for me, I wouldn't trust it to do that unsupervised. Even semantic HTML is a mixed bag: if you point out something is a figure/figcaption it'll probably do it right, but I haven't found that it'll intuit these things and get it right on the first try.
But I'd imagine if you don't care about the frontend looking original or even good, and you stick really closely to something like tailwind, it could output something good enough.
And critically, I think a lot of times the hardest part of frontend work is starting, getting that first iteration out. LLMs are good for that. Actually got me over the hump on a little personal page I made a month or so ago and it was a massive help. Put out something that looked terrible but gave me what I needed to move forward.
I ended shoehorned into backend dev in Ruby/Py/Java and don't find it improves my day to day a lot.
Specifically in C, it can bang out complicated but mostly common data-structures without fault where I would surely do one-off errors. I guess since I do C for hobby I tend to solve more interesting and complicated problems like generating a whole array of dynamic C-dispatchers from a UI-library spec in JSON that allows parsing and rendering a UI specified in YAML. Gemini pro even spat out a YAML-dialect parser after a few attempts/fixes.
Maybe it's a function of familiarity and problems you end using the AI for.
Yes.
>in domains where you have trouble judging the quality
Sure, possibly. Kind of like how you think the news is accurate until you read a story that's in your field.
But not necessarily. Might just be more "I don't know how do to <basic task> in <domain that I don't spend a lot of time in>", and LLMs are good at doing basic tasks.
Recently, my company has been investigating AI tools for coding. I know this sounds very late to the game, but we're a DoD consultancy and one not traditional associated with software development. So, for most of the people in the company, they are very impressed with the AI's output.
I, on the other hand, am a fairly recent addition to the company. I was specifically hired to be a "wildcard" in their usual operations. Which is too say, maybe 10 of us in a company of 3000 know what we're doing regarding software (but that's being generous because I don't really have visibility into half of the company). So, that means 99.7% of the company doesn't have the experience necessary to tell what good software development looks like.
The stuff the people using the AI are putting out is... better than what the MilOps analysts pressed into writing Python-scripts-with-delusions-of-grandeur were doing before, but by no means what I'd call quality software. I have pretty deep experience in both back end and front end. It's a step above "code written by smart people completely inexperienced in writing software that has to be maintained over a lifetime", but many steps below, "software that can successfully be maintained over a lifetime".
You can tweak the prompt a bit to skew the probability distribution with careful prompting (LLMs that are told to claim to be math PHDs are better at math problems, for instance), but in the end all of those weights in the model are spent to encode the most probable outputs.
So, it will be interesting to see how this plays out. If the average person using AI is able to produce above average code, then we could end up in a virtuous cycle where AI continuously improves with human help. On the other hand, if this just allows more low quality code to be written then the opposite happens and AI becomes more and more useless.
When it comes to software the entire reason maintainability is a goal is because writing and improving software is incredibly time consuming and requires a lot of skill. It requires so much skill and time that during my decades in industry I rarely found code I would consider quality. Furthermore the output from AI tools currently may have various drawbacks, but this technology is going to keep improving year over year for the foreseeable future.
In both of these cases, I found that just the smart auto-complete is a massive time-saver. In fact, it's more valuable to me than the interactive or agentic features.
Here's a snippet of some code that's in one of my recent buffers:
// The instruction should be skipped if all of its named
// outputs have been coalesced away.
if ! self.should_keep_instr(instr) {
return;
}
// Non-dropped should have a choice.
let instr_choice =
choices.maybe_instr_choice(instr_ref)
.expect("No choice for instruction");
self.pick_map.set_instr_choice(
instr_ref,
instr_choice.clone(),
);
// Incref all named def inputs to the PIR choice.
instr_choice.visit_input_defs(|input_def| {
self.def_incref(input_def);
});
// Decref all named def inputs to the SIR instr.
instr.visit_inputs(
|input_def| self.def_decref(input_def, sir_graph)
);
The other nice benefit is that I don't have to trust the LLM. I can evaluate each snippet right there and typically the machine does a good job of picking out syntactic style and semantics from the rest of the codebase and file and applying it to the completion.
The snippet, if it's not obvious, is from a bit of compiler backend code I'm working on. I would never have even _attempted_ to write a compiler backend in my spare time without this assistance.
For experienced devs, autocomplete is good enough for massive efficiency gains in dev speed.
I still haven't warmed to the agentic interfaces because I inherently don't trust the LLMs to produce correct code reliably, so I always end up reviewing it, and reviewing greenfield code is often more work than just writing it (esp now that autocomplete is so much more useful at making that writing faster).
For frontend though? The stuff I really don't specialize in (despite some of my first html beginning on FrontPage 1997 back in 1997), it's a lifesaver. Just gotta be careful with prompts since so many front end frameworks are basically backend code at this point.
Things like "apply this known algorithm to that project-specific data structure" work really well and save plenty of time. Things that require a gut feeling for how things are organized in memory don't work unless you are willing to babysit the model.
Just last week I had to review some monstrosity of a FE ticket written by one of our backenders, with the comment of "it's 90% there, should be good to takeover". I had to throw out pretty much everything and rewrite it from scratch. My solution was like 150 lines modified, whereas the monstrous output of the AI was non-functional, ugly, a performance nightmare and around 800 lines, with extremely unhelpful and generic commit messages to the tune of "Made things great!!1!1!!".
I can't even really blame them, the C-level craze and zeal for the AI shit is such that if you're not doing crap like this you get scrutinized and PIP'd.
At least frontenders usually have some humility and will tell you they have no clue if it's a good solution or not, while BEnders are always for some reason extremely dismissive of FE work (as can be seen in this very thread). It's truly baffling to me
AI isn't very good at being concise, in my experience. To the point of producing worse code. Which is a strange change from humans who might just have a habit of being too concise, but not by the same degree.
In my experience, review was inadequate back before we had AI spewing forth code of dubious quality. There's no reason to think it's any better now.
An actually-useful AI would be one that would make reviews better, do them itself, or at least help me get through reviews faster.
One: The work to get code to a reviewable point is significant. Skipping it, either with or without AI, is just going to elongate the review process.
Two: The whole point of using AI is to outsource the thought to a machine that can think much faster than you can in order to ship faster. If the normal dev process was 6 hours to write and 2 hours to review, and the AI dev process was 1 hour to write and 8 hours to review, the author will say "hey why is review taking so long; this defeats the purpose". You can't say "code review fixes these problems" and then bristle at the necessary extra review.
Wouldn't it be the opposite? I'd expect the code would be 47% longer because it's worse and heavier in tech debt (e.g. code repeated in multiple places instead of being factored out into a function).
All source code is technical debt. If you increase the amount of code, you increase the amount of debt. It's impossible to reduce debt with more code. The only way to reduce debt is by reducing code.
(and note that I'm not measuring code in bytes here; switching to single-character variable names would not reduce debt. I'm measuring it in statements, expressions, instructions; reducing those without reducing functionality decreases debt)
I think you instead meant to say more business logic implemented in code is more technical debt, not necessarily just more code.
These were maintainers of large open source projects. It's all relative. It's clearly providing massive gains for some and not as much for others. It should follow that it's benefit to you depends on who you are and what you are working on.
It isn't black and white.
There are some very good findings though, like how the devs thought they were sped up but they were actually slowed down.
It could happen that the impact of using AI depends of the task at hand, the capability of the SWE to pair programming with it, and of the LLM used, to such an extend that those factors were bigger that the average effect on a bag of tasks, in this case the large deviation from the mean makes any one parameter estimation void of useful information.
I thought it was the model, but then I realised, v0 is carried by the shadcn UI library, not the intelligence of the model
With the headline we can easily guess that the study should be flawed, with the sample not representative, or developers not expert enough with the AI or so, and then, they give a very well done list of all valable objections with arguments about why they don't think that contradict the study.
That replied to all the questions I could have had in the end.
Like what if by focusing on LLMs for productivity we just reinforce old-bad habits, and get into a local maxima... And even worse, what if being stuck with current so-so patterns, languages, etc means we don't innovate in language design, tooling, or other areas that might actually be productivity wins?
I expect it'll balance.
They're not great at business logic though, especially if you're doing anything remotely novel. Which is the difficult part of programming anyway.
But yeah, to the average corporate programmer who needs to recreate the same internal business tool that every other company has anyway, it probably saves a lot of time.
??? Like everyone else? This is why LLMs have large context windows.
> But the model can generalize well enough to handle the task.
Sounds like what you're doing is normal corporate business logic. If it's able to generalize it it's probably not particularly novel.
Also, you're contradicting yourself: you said LLMs aren't great at 'anything remotely novel,' then dismissed my novel inference as routine because 'everyone' does it with context windows. Which is it? Either novel input is challenging for LLMs or it's routine and expected: contradicting your claim that LLMs can't handle novel scenarios.
The bar to beat is time it takes for human intelligence + Cursor Tab to complete a task, but in practice the bar becomes "feeling" like its faster
I'm sorry, but it feels to me like this research has only proven that developers tend to underestimate how long a task is supposed to take, with or without AI.
In no way did they actually measure how much faster a specific task was when performed with and without AI?
You have two tasks:
- Task 1 - 3h
- Task 2 - 1h
Then you randomly tell the dev, ok do Task 1 without AI, and Task 2 with AI.
Then you measure the actual time it took.
Their estimate for AI task missed the mark by 19%, but those without AI were done 20% faster then estimated.
At the time of estimating they didn't know if the task would need to be done with AI or not.
Over IDK, 2-3 hours I got something that seemed on its face to work, but:
- it didn't use the pub/sub API correctly
- the 1 low-coverage test it generated didn't even compile (Go)
- there were a bunch of small errors it got confused by--particularly around closures
I got it to "90%" (again though it didn't at all work) with the first prompt, and then over something like a dozen more mostly got it to fix its own errors. But:
- I didn't know the pub/sub API--I was relying on Cursor to do this correctly--and it totally submarined me
- I had to do all the digging to get the test to compile
- I had to go line by line and tell it to rewrite... almost everything
I quit when I realized I was spending more time prompting it to fix things than it would take me to fully engage my brain and fix them myself. I also noticed that there was a strong pull to "just do one more prompt" rather than dig in and actually understand things. That's super problematic to me.
Worse, this wasn't actually faster. How do I know that? The next day I did what I normally do: read docs and wrote it myself. I spent less time (I'm a fast typist and a Vim user) overall, and my code works. My experience matches pretty well w/ the results of TFA.
---
Something I will say though is there is a lot of garbage stuff in tech. Like, I don't want to learn Terraform (again) just to figure out how to deploy things to production w/o paying a Heroku-like premium. Maybe I don't want to look up recursive CTEs again, or C function pointers, or spent 2 weeks researching a heisenbug I put into code for some silly reason AI would have caught immediately. I am _confident_ we can solve these things without boiling oceans to get AI to do it for us.
But all this shit about how "I'm 20x more productive" is totally absurd. The only evidence we have of this is people just saying it. I don't think a 20x productivity increase is even imaginable. Overall productivity since 1950 is up 3.6x [0]. These people are asking us to believe they've achieved over 400 years of productivity gains in "3 months". Extraordinary claims require extraordinary evidence. My guess is either you were extremely unproductive before, or (like others are saying in the threads) in very small ways you're 20x more productive but most things are unaffected or even slower.
Respectfully, this is user error.
First prompt:
``` Build a new package at <path>. Use the <blah> package at <path> as an example. The new package should work like the <blah> package, but instead of receiving events over HTTP, it should receive events as JSON over a Google Pub/Sub topic. This is what one such event would look like:
{ /* some JSON */ } ```
My assumptions when I gave it the following prompt were wrong, but it didn't correct me (it actually does sometimes, so this isn't an unreasonable expectation):
``` The <method> method will only process a single message from the subscription. Modify it to continuously process any messages received from the subscription. ```
These next 2 didn't work:
``` The context object has no method WithCancel. Simply use the ctx argument to the method above. ```
``` There's no need to attach this to the <object> object; there's also no need for this field. Remove them. ```
At this point, I fix it myself and move on.
``` There's no need to use a waitgroup in <method>, or to have that field on <object>. Modify <method> to not use a waitgroup. ```
``` There's no need to run the logic in <object> inside an anonymous function on a goroutine. Remove that; we only need the code inside the for loop. ```
``` Using the <package> package at <path> as an example, add metrics and logging ```
This didn't work for esoteric reasons:
``` On line 122 you're casting ctx to <context>, but that's already its type from this method's parameters. Remove this case and the error handling for when it fails. ```
...but this fixed it though:
``` Assume that ctx here is just like the ctx from <package>, for example it already has a logger. ```
There were some really basic errors in the test code. I thought I would just ask it to fix them:
``` Fix the errors in the test code. ```
That made things worse, so I just told it exactly what I wanted:
``` <field1> and <field2> are integers, just use integers ```
I wouldn't call it a "conversation" per se, but this is essentially what I see Kenton Varda, Simon Willison, et al doing.
In my opinion, it takes several weeks of active use to nail down your preferred workflow with these tools and to get a meaningful understanding of their abilities and limitations.
I.e., yes they hallucinate and don't have great understanding of truth/fact (however you choose to define those terms), but you need to develop an intuition for how to work around those issues and how to recognize the problems in your setup that increase the likelihood of the LLM heading down false paths. This intuition cannot come until you fight through the initial struggle period.
In some ways, it's similar to picking up emacs/vim and learning the shortcuts. It's a negative to your velocity until it's not, and once you overcome that initial hurdle, your productivity takes off. Admittedly, it's not for everyone (I never bothered to learn the ins and outs of vim bindings because my bottleneck isn't my speed of writing code), but it provides a huge productivity boost for those types of engineers.
Coming back to my main point: your LLM needs quite a bit of guidance in the early stages, especially as you're feeling out what types of tasks it's able to knock out the park and what types of tasks it'll struggle with. For instance, in the example you gave here, I wonder what would happen if you asked it to present you a detailed plan before it gets to writing any code and to provide a list of assumptions it is making? You will need to do a bit of review with it before you let it go execute the plan (siilar to how a junior engineer would come to you with questions before being able to handle certain tasks).
I also recommend writing up a thorough self-review checklist that it stored in your repo (e.g. in an AGENTS.md file) that provides the customized instructions you want your LLM to follow (it won't always do so but it helps a ton). Otherwise, each new session is essentially starting over without it learning, which is pretty frustrating.
I'm happy to talk more because I'm pretty optimistic about LLMs and enjoy using them in my day-to-day where appropriate.
And finally, I'm not sure how much you've thought about giving them more autonomy, but I do recommend doing so if you have a safe, sandboxed environment. The real magic and productivity boost of LLMs come when you give them some more autonomy and provide them with tools to figure out the problems they encounter, unlocking your time to be spent on higher-leverage tasks such as designing systems and processes. If it can run linters, unit tests, and grep your codebase during its development process and use this to iterate, you'll have a much more fun time.
Does this help?
I guess the tricky bit is, nobody knows what the future looks like. "The internet is a fad" in 1999 hasn't aged well, but a lot of people touted 1960s AI, XML and 3d telivisions as things that'd be the tools in only a few years.
We're all just guessing till then.