2x, not 10x: coding with LLMs in 2026 | Hacker News Reader

2x, not 10x: coding with LLMs in 2026

obryant.dev

280 pointstnisonoff243 comments

While I agree with the premise, I think this angle only applies on work one was going to do no matter what. The real power of these tools is that there are so many ideas people would like to try, but never have the time or motivation to pursue.

So the comparison is not only "built with and without LLM" but "would you even build this if you didn't have the LLM?". The gap in productivity in this case is much more wide.

> The real power of these tools is that there are so many ideas people would like to try, but never have the time or motivation to pursue.

This can be a negative multiplier: code I thought I wanted that gets immediately abandoned is a net-negative if no one else wants it (lets face it, this is the safest default posture for software of unknown providence).

In isolation, instant-abandonware takes up hdd space, burns dependabot's CPU-cycles, and wastes human attention when appearing in search results. In aggregate, it floods the zone with a deluge of forks with imperceptible differences between them, based on nit-picks, legitimate stand-out products will have a much harder time going forward.

Not every engineering effort is a product in search of market fit or a community.

Some things are already very useful as just a one shot. I just made a quick app to help me pack for a trip, it updated forecasts every day, let me know when rain entered the forecast at one of my stops and gave me a checklist that helped me quell my travel anxiety.

The greatest thing that LLMs have done is allow many to achieve things that they couldn't have before. I'm incredibly disinterested in "I can do the same thing I was already doing x% faster"

> gave me a checklist that helped me quell my travel anxiety.

it's truly fascinating how many positive descriptions of AI gesture at emotional management. I think that's the killer feature of this technology -- it makes people feel good, capable, reassured -- without the risk and vulnerability of interacting with another human.

I use Google Weather for my forecasts btw, no need to vibecode an app for that

Gathering forecasts for each leg of a trip only on the particular days for the particular location is tedious. Anyone can do it with google weather but why spend the time when you don't have to.

I can keep track of my expenses on a napkin but i'd much rather use a spreadsheet or dedicated app especially when that app is effectively free.

I asked Google Gemini to generate me a basic expense estimator to paste in ... Google Sheets. Every single formula needed editing. At least it generated the basic layout for me.

did you know that Google Weather has saved locations?

> Not every engineering effort is a product in search of market fit or a community

I think you misunderstood my point: by "legitimate stand-out products", I meant exactly that, with no connotation of commercialization. Maybe you can agree that having a high signal-to-noise ratio for (open source) projects is a desirable goal?

> Some things are already very useful as just a one shot.

I agree. I too have made or forked about a dozen apps and tools in the past few months. It would be dishonest not to consider the flipside, that this software is overfitted to the needs of a single person. Further, this hyper-bespoke software typically feature-complete within moments of the final prompt, and I have,on occasion, completely forgot about the tool/app I spent a weekend created, it clearly wasn't worth the effort I put in.

> The greatest thing that LLMs have done is allow many to achieve things that they couldn't have before.

Let's not pretend there isn't a cost to this.

> by "legitimate stand-out products", I meant exactly that, with no connotation of commercialization

The word "product" implies commercial.

> The word "product" implies commercial.

I disagree; but see where you're coming from. I'm chuckling at the irony of my word-choice: I initially had used "project" but nixed it because of its frequent association with Open Source. I instead opted for "product" as a broader term. For the sake of clarity, my original comment is referring to commercial and non-commercial software projects/products.

Spending 1 day vibecoding or modding an existing app to fit you better and not using it is much better than spending, 2 weeks and then using it from sunk cost falacy.

Also it teaches you that what you think you need and want is not what you need and want. This is why you are not using it

> having a high signal-to-noise ratio for (open source) projects is a desirable goal?

It’s not obviously true. A higher number of attempts, a larger talent pool, typically doesn’t change the average much (or it might even make the average go down), but tends to produce higher peak outcomes.

We see this everywhere (science, startups, sports, chess, etc).

If you want the best spreadsheet, game, or whatever app you want, you’re only interested in the few highest peaks.

So you do actually get better signal to noise with a larger wasteland of discarded attempts. The higher peaks make it easier to filter out the noise.

The goal you’re intrinsically motivated by seems different than this. That seems to be the whole disagreement.

This is wrong. Someone else wanting it is irrelevant. If it solves a need you have, that is a win. Not every piece of software needs to be a market fit. Not everyone wants to grind out an existence marketing software. And rapidly that type of business is going to go byebye

> This can be a negative multiplier: code I thought I wanted that gets immediately abandoned is a net-negative if no one else wants it (lets face it, this is the safest default posture for software of unknown providence).

obviously not ? negative result is still a result, just like in science. It adds new information ("approach X does not work" / "is useless") which is the only thing that matters

You made me do a double take with "providence". I think the word you wanted is "provenance", which comes from latin "provenire" (pro -> forth, venire -> come). English has too many similar words.

Unless you only run code that's protected by God. That's probably not a bad policy if you can verify it.

edit: Oh right, there's an OS for that https://en.wikipedia.org/wiki/TempleOS

We’re at the point where people who talk too fancy get accused of being LLMs, and the people who make human mistakes get scolded for grammar.

It seems like constructive criticism rather than scolding.

If I were running around saying "providence" instead of "provenance" I would want someone to tell me.

Yes, it was meant constructively and for exactly this reason. I interpreted what they wrote as what they meant, then realized it was a mistake that I could have easily made too.

Me too, but we're the minority. Most people are rolling their eyes every time this happens. They don't want constructive criticism any more than they want normal criticism.

I think the ideas worth pursuing still require substantial time and motivation, even with LLMs. Otherwise you just end up with a half-baked sloppy artifact that nobody will ever use or study, for a net negative benefit to humanity.

We have entered the era of small software, and no one knows what it will grow to.

Here's an example: my office has a few cars for employees to use as rentals. The number is small enough that it would never be worth any serious software dev to build a tool to manage, but large enough that its a moderate amount of work for someone to manage the requests/getting supervisor approvals/schedule changes due to breakdowns.

AI one-shot that guy a tool. Its now dead simple, he's got a calendar, automatic emails going to people's supervisors with click-here-to-approve links, rescheduling options, fleet management. It doesn't even look bad.

Who cares if its using some un-backed-up sqlite database in the backend, has some placeholder tab for a feature he changed his mind about, or violates the DRY principles a bunch or uses some inferior authentication mechanism. Its an in house tool, isn't mission critical, and it makes his life significantly easier.

Basically everyone is now a few prompts away from their own bespoke tools, and only they will be able to judge the benefit thereof.

Edit: to tie this more directly to the article, I would argue that this is an example of "infinity-x" coding, because the user was in fact not capable of coding a solution on their own without AI.

I agree.

In my business, I haven't found much area to use code. It's a pub, and we've long been low-tech. Cash register, no POS. I have a little code surrounding my own processes, but mostly it's manual. Hand-entering numbers in my spreadsheet, etc.

But what's interesting to me is that now I can probably program an esp32, or create a small mobile app for a mounted android tablet. I was a web dev in the past, and programming hardware was outside my skillset without dedicating some serious time to learning. Mobile I just always avoided--mostly the same reason.

Anyway, I've got some CYDs on my desk, we'll see what I can make with 'em. I want a kitchen ticketing system instead of the old hand-written ticket stubs, for starters.

Yes this is true to some extent. I've been using LLMs to run some computer vision tests, and I've certainly noticed myself running into the trap of "just one more AI experiment", or "just one more change" while neglecting to actually properly integrate the learnings into my mental model.

However, without an agent running its own experiments on a cloud GPU, would I realistically have invested my limited work hours and tried evaluating 10 different models, each with 10 different tuned parameters, to solve my specific use case?

Or would I have tried 1-2 models and spent my time trying to optimize those models?

I think there is some merit to the spray and pray approach when one is in the exploration phase of the solution space.

Also, on more than one occasion now, I have had fable halve the inference latency of a model simply because the original implementation from an academic included unnecessary GPU-to-CPU-to-GPU transfers or similarly inefficient operations. Those optimizations came at essentially 0 time cost to me and I can verify that the outputs are byte-identical. Pretty sweet!

People need to abandon the notion of every tool you build needing to fit the market. This is not the direction of software anymore when developers can all build tools that are bespoke to their own workflows and needs.

Well, I’m happy just to make half baked slop for myself. Highly personalized software that is poorly implemented and yet yields the result I desire.

E.g. software that generates these models that I can print

https://wiki.roshangeorge.dev/w/Blog/2026-06-30/Modeling_a_W...

https://wiki.roshangeorge.dev/w/Blog/2025-12-01/Grounding_Yo...

Or blog post authoring software

https://wiki.roshangeorge.dev/w/Blog/2026-04-25/The_rise_of_...

There were so many things that no one will ever study and won’t give humanity any benefit but I use everyday to make my life better. That’s enough. The value far exceeds $200/mo. I’m getting it for cheap and now that I have my GPUs and my models they can’t even take it from me in the future if they wanted, haha!

LLMs allow for human flourishing on a massive scale. One of the best inventions to occur in my life. Up there with the Internet/Web. Truly a marvelous time.

> There were so many things that no one will ever study and won’t give humanity any benefit but I use everyday to make my life better

Agreed. I haven’t been this excited by computers since I got broadband DSL in 1998.

i consider it as a very specialized subset (subgenre) of video games for a very narrow audience

You know how non-coders basically build crazy business apps in Excel? LLMs, even if they don't get any better, seem at least 10x more useful than Excel. Why do you say it's for a narrow audience?

i meant people who enjoy the idea of software even if they don’t necessarily need or enjoy the resulting software itself. it’s possible that some of the spreadsheet hacking has a similar flavor but lacks the natural language hook.

In this last week, I've performed experiments in an almost fully automated way - parallel to my main work, and easy to orchestrate.

This has led to 3 parallel pieces of adjacent work that each speed up our build by quite a drastic margin. When combined, this is a massive improvement. None of this would have happened in the old days, as the research itself takes a long time to babysit and a lot of options to check.

So I very much agree - the activation energy can be a lot lower on some kinds of tasks, and some of those get big returns for small inputs. It's not all like that, but part of the game is identifying when you can spot those high return efforts.

Figured out a better way of explaining it:

dev A knows exactly what the program should do and how to verify the AI output

dev B thinks they know what they are doing but are actually misguided by bad psycophantic AI output they have incorrectly verified.

both work on product C

or the AI could lead you to a completely unusable experiment and lead you to believe it was successful wasting weeks of work. Not saying it will happen to you, but there is a huge "survivorship" bias already in software for what tools are useful or not. Maybe it's the person and not the tools.

Yes, but only in the programming sense of productivity, not in the economic sense. When prices stop being subsidised and the novelty wears off, this will boil down to the question of how much money is that worth? We're already seeing an inflation in such projects, which are becoming hard to even show off. It could be similar to 3D printers. Fun and valuable, with a market size of $25B, but not quite world-changing (I'm not saying AI won't be world-changing, only that this tinkering aspect may not be).

Those sort of ideas are unlikely to yield major benefit, though.

This is basically replicating the plight of the solo open source dev, writ large. Individual programmers have long built the thing they've cared about on their own time (essentially "for free" because, despite kindergarten economics theory, a programmer cannot usually monetize a marginal hour). And it usually goes that the project never gets adopted anywhere. It might acrue more features and total man-hour effort than most of what FAANG does in open source to drown out the solo devs. But the market will decide that "no organizational buy-in" is a signal the project doesn't matter. Other developers will decide, "if he could do it, so could I" and also not adopt.

Same exact thing is happening and will continue to happen with all these generated "but we wouldn't have done it otherwise" projects. It's just very, very unlikely to go anywhere.

That which took very little effort to create will receive very little effort to promote.

I don't think this is the right framing, because I think it confuses the means and the ends.

Its like making a jig in woodworking. The measure of the jig's success is not whether it gets re-used or widespread adoption, its whether it made it easier to achieve some actual objective. Because the jig is a means to some other end.

Lots of these "we wouldn't have done it otherwise" applications are means, not ends.

That is true, but that suggests that the value of software is reverting to the value of a jig. Software that gets deployed to millions of people gets a large valuation because it provides small convenience times a million. In the era of bespoke software, you’re basically 100x-ing the production of a millionth of the value of commercial software. That’s why it isn’t all that impactful in the end and I think it’s worth highlighting.

It's the underpants gnomes again.

  - Vibe code a bunch of small projects that we couldn't justify ROI before
  - ???
  - Profit

LLMs open up huge possibilities to do work that you otherwise wouldn't do. I vibe coded a PDF viewer yesterday in a couple of hours using Claude and Grok. I've been itching for a native Mac app that meets some very basic requirements: one window with file tree on the left and tabbed view on the right, with the ability to call out to pdfgrep and jump to results. Like Zotero, but with better multi-file search and a viewer that doesn't choke on big files. Claude built it for me in an hour, and Grok iterated a couple of times to make it better. I didn't look at a single line of code. It's now my default PDF viewer, and I used it today to do a bunch of my real work.

It's a braindead simple program that mostly hooks together pre-existing functionality, it just so happened that none of the widely available apps had the specific mix of features I wanted. I could probably have done it myself in a week if I took time off my non-coding day job to figure out Swift and AppKit. But I wasn't going to do that. I’m psyched. I hate web apps and now I can just write my own for all the little things I use every day.

Cool story bro. I used a llm to help me name a variable.

Yes but I’ve seen some devs waste a lot of time using AI to build something that was a bad idea to begin with.

Without AI they might have first spent more time validating the idea was worth it.

For me it is exactly the opposite. AI makes it so easy to create test fixtures and run experiments that I now spend much more time validating ideas than I could ever afford to do before. When I write code for production, it's not "this ought to work", it's "here are the figures showing how well this works, on this dataset", where the dataset is also much larger than anything I would have used before, because I used AI to generate the tools which collected and organized it.

> it's not "this ought to work", it's "here are the figures showing how well this works, on this dataset"

The thing with constraints is that you focus of the thing with high value first. So you focus on the most promising ideas first or choose experiments that can get rid of most ideas. Instead of trying to validate each ideas and generate what is most likely noise to the decision process.

Like if I ever hire an assistant, I want like one to three options that are closely aligned to my needs, not a bible size report on 42 choices.

Isn’t it better to be able to quickly test ideas and see what works and what doesn’t?

> Without AI they might have first spent more time validating the idea was worth it.

Seems optimistic

Not even just building stuff with LLMS.

I'm doing analysis on stuff that we previously simply couldn't do in my company, it would take way too much time or effort, and we didn't have the manpower.

Yes, I work in hardware design ans the cost to code up the RTL for an idea and compare it to the existing code in terms of PPA, its very expensive in terms of engineering time. This week I had 5 different architecites coded, pushed through a functional test bench, bit accurate models created and then run through ppa analysis all done via llms. So its a task tbis project could not afford to do without llms, but can save us 10% in ppa

Im not so sure if this is a pro or a con. Take two extremes at product building: brute force vs deep research first. I guess llms allow you to experiment quickly with many approaches and see what sticks. Without llms, you need to spend more time researching and thinking what’s the most economical approach. Im biased towards the non-llm approach

I often use LLMs to build locally running web apps and console apps that I otherwise never would have built, because the value-add wouldn’t be worth, say, 40 hours of my free time.

Agree so much. So many small bugs, nits, tweaks I just send off to an LLM agent to figure it out.

the real power is making the bot search a large code base for the relevant code you want to change

that and 2x is "i'm 5'10" and i round up to 6'" low.

I'm probably a weird outlier. Coming from academia, it's ranging from 1x to infinity-x (as in, certain tasks wouldn't even be touched if not for AI).

For stuff that I'm used to (R) I can write nice and compact spaghetti (long %>% pipes). I'm not comfortable when a working script doesn't fit the screen (plus a few scrolls max). My style is probably easy only to me. When I teach, I don't teach it in particular.

AI gives me thousands of lines of codes for those. 10k once. It's cool if it suffices to source it all, but working with that is not pleasant.

But if I can reproduce a paper in a one-shot (it used to be an hour, but recently it got so much better), that's a task that would not have been even attempted years ago. And I'm talking about a methods paper with no available Github (or, as often happens, when the existing Github is useless)

I would have said 0x to infinity-x. Some stuff I think it would be a good idea and waste a bunch of time and then give up and just write it myself. And like you say, there’s lots of stuff I would just never have done, basically any kind of front end for example, if I didn’t have an LLM to do it.

I also do mostly research or one-of script development, for which I find LLMs less useful, as I want simplicity and I want to understand exactly what is happening. The only time I find LLMs help really there is if there’s some part I can abstract away like writing and interface to an api or something.

For me, I find the biggest benefit is that I actually do all the things I know I’m supposed to do, but usually don’t bother if I’m the only audience.The first time a model asked me, “do you want me to wrap this in an easy to use CLI?” I was like, “Yes. Yes I do.”

Interesting. Framing it as 1x to infinity-x matches my experience too. I've have good success with it reproducing papers with existing code, but not such much with one-shotting new code. Do you have any particular setup for this i.e. special validation prompts or multi-agent stuff or do you just ask something like reproduce paper X until it gets results Y?

I used to chat with "the paper" until both me and the AI were roughly on the same page at what's happening. It's important, as a "full read" of the "full paper" is not always used efficiently if not discussed, I've noticed. I then ask it for a detailed step by step realization (of the algorithm, or the paper/section/whatever), plus some context. Then I pass this as a prompt to a fresh new instance, and task it to implement in full based on the description (Gemini Deep Think was quite good with that).

Now I feel like 5.6 Sol Ultra is capable of doing roughly the same with its Agents, so it's getting easier in my experience. With the Codex, it can adjust or correct until the output is suitable.

I'm sure it depends on the field and the method.

Thanks a lot for sharing! I'll try the discussion to unpack context, that sounds interesting. I end up having such discussions in separate chats anyways to understand what's happening so it makes sense to do it upfront.

I have many problems dealing with statements that use Infinity, especially if I think about the math or physics implication of that :)

Infinity is a ridicilous statement. Even the work that wouldn't have been done before is now time away from work that needs to be done. At worst it's negative x.

Pretty sure this was one of those so-called “jokes” I have read about in a research paper once. American humor tends to rely on overstatement. This is contrast with British humor that more often builds on understatement.

It's a division by zero. We can change infinity to 'undefined'

I have a weird issue with using AI for coding. I can code something entirely by myself at my baseline speed; call it 1x. Or I can use Claude to do it, and it does it in 1/10 - 1/4 of the time. The problem, however, is that to review Claude’s code properly takes 2-3x the amount of time it would have taken me to write it all by hand.

So my two choices are basically “YOLO, LGTM” and hope I can revert if it breaks something, or to just write all the code by hand from the start. With the increased pressure for output, I’ve noticed both myself and coworkers tending more toward “commit and hope it works” over time. It’s sort of perverse incentives in a way...

This is what I've seen. There has been a middle-ground though, that I've seen. I was using Jetbrains' AI Assistant, or whatever it's called, during an interview for the (nearly) first time, and I only used it as auto-complete, but it was consistently, surprisingly, very accurate. I literally did write it faster. I scanned the literal next line, in the context of what I was writing, with all that context in my brain, and I saw what it wrote and went "yeah, that looks good". I'm a very fast typist (~> 100 wpm) and I still found it beneficial for writing out work fast. Further, since everything is still in my mental context and I can evaluate it at the same time, I feel like I can trust that type of code more.

But maybe that's just me.

How long ago was this?

I tried their "local ai autocomplete" thing a few years ago for a little while and it was hot garbage. It guessed the right/acceptable completion about 30% of the time at best.

I haven't looked at it since.

Within the last, like, ... 3 months, I think.

The models are good enough these days that I see it like managing a team of juniors. The LLMs need guidance and oversight, and about the same amount of time I'd spend reviewing code from a junior I spend on the LLM output. I think generally it's best practice to work in a way that "commit and hope it works (because it passes all the CI/CD automated testing and verification, and it's a change gated behind feature flags and has an identified and minimal blast radius, etc)" is possible

I feel like these days it's more like managing a team of seniors who are technically very competent, but may not always have the full domain context of the problem they're solving. And/or they forget parts, if that context is too large. And/or they have a poor temporal dimension (deprecated badly maintained information polluting the context).

They also tend to overengineer solutions if not guided well.

If you think about it, in that respect it's not that different from managing actual human senior engineers.

I think this is why focussing one's limited human attention more the input (defining clear requirements) as well as on validating the output (good CI/CD including end-to-end automated tests) is far more important than manual code reviews and micro-managing the development process.

Reviewing LLM code is quite different from junior code.

With juniors you tread carefully and give feedback only on important points to encourage growth.

With LLMs you channel the inner sailor and nitpick so much that even a senior would start to cry.

Those aren’t the only choices though. What you can do now is have a separate agent review the diff before you create the PR, or to leave feedback on the PR for the agent that wrote the code to fix.

You write the feature with one agent, and have different agents do the review, each with their own rules and context. Each review agent can even focus on different things, such as security, or adherence to your stylistic or architectural preferences.

This can find all kinds bugs and edge cases the first agent missed, and that you might never have thought of yourself, even with careful human review.

Economics drives behaviour more than people think. Your employer cares about results. It's 'YOLO, LGTM' all the way down now.

not to forget the cognitive strain of the constant back and forth between LLM and you.

One challenge is when knowing how the code should look like, the LLM solution always looks weird, and one tries to maunally steer against it, so accepting a bit of "good enough" is unavoidable to gain some productivity

This reminds me of themes I recently saw in [Harness Engineering is not Enough: Why Software Factories Fail](https://www.youtube.com/watch?v=Ib5GBkD555M) (Warning: the last 3 slides seem like an advertisement). One thing I liked is how Dex has a little graphic he glossed over showing software development is

- 25% planning & aligning with other teams

- 25% coding

- 25% testing/verifying

- 25% code review/rework

One argument was that agentic coding speeds up that coding part a bunch. So maybe there's 2x speedup in coding. But that's only a small speedup in the totality of everything software engineers do.

AI is great because my job is a lot easier, but it produces some extremely crappy code that would've NEVER passed a code review 1 year ago.

You could argue that it speeds up development by 5X or more, but then it slows down testing/verifying, code review, and in many cases it makes it impossible to review/rework by hand.

I just do what I'm told at work but even though I'm sure I'm in the minority I'm extremely skeptical that LLMs can produce any good-quality code.

Looks to me like we've just lowered the bar--by a lot--and stopped looking at the code that goes out. Every time I look into how Claude implemented something it's completely insane, with no way to refactor it or maintain it in the future.

> it produces some extremely crappy code that would've NEVER passed a code review 1 year ago

The 10x speed up comes from the fact that we have all given up on properly reviewing each others code and we just say “meh, it will be fine, lgtm”

This. The bar has never been lower.

Code review has always been pointless

I realize this like a decade ago. Security is basically the only thing I care about in code review.

I’ve come to realize that I don’t actually care what your code looks like. I’m going to need to learn it from scratch every time I use it. It’s going to look weird and foreign to me no matter what. It’s not worth my time nit-picking things that simply do not matter.

I’m glad somebody said it!

Yes, exactly my point.

BUT, to me if we hadn't lowered the bar MASSIVELY, it would NOT be fine.

I've seen people at my work that will push or even deploy features that don't do what their main purpose is and they just dismiss the issue saying it's a bug. As an example, a button that's supposed to open a modal window if you click on it. If you go test it and the modal doesn't open, they'll say "oh, there's a bug".

You developed or even deployed something that is wrong and completely useless. Just one year ago, this would be unheard of and you would've gotten a really hard time for it. Now it's just a bug and part of how we're doing things.

Like I said, I just do what I'm told at work but to me this is just wild. There's no way it won't come back to bite us.

I read this a lot, do you have examples of this? Last year the code produced by LLMs was pretty sloppy, but recent SOTA models seems pretty good to me. I'm curious if I'm missing something or it's folks using different models or difficulty levels.

From my experience, a snippet of code will look good.

But LLMs seem to have a hard time getting the big picture and reusing code that is already implemented and ALMOST does what you want vs. rewriting everything from scratch.

AI code is extremely difficult to read and follow. It's littered with hundreds of lines of comments and notes, often referring to other notes in other parts of the codebase, and often extremely out-of-date.

Claude seems to me to still focus on getting things done vs. doing it properly, so from my experience it would pretty often mess up your whole codebase just to be able to finish its task, vs. stopping and rethinking the approach.

After a few passes of that, with duplicate code with no underlying coherent vision, hundred of thousands of lines of documentation written as walls of text in markdown and weird coding decisions, your codebase is impossible to work out for a human.

I use Claude Opus 5 at high to plan, Claude Opus 5 at low to execute the plan.

I make it build things in small changes, I give very specific instructions about the architecture, and I make sure to point out existing functionality that can be used instead of writing something from scratch. But it's a lost battle because Claude can't learn and it's basically a black box.

It pretty shitty, but that's what they pay me for so I just do it.

> It's littered with hundreds of lines of comments and notes, often referring to other notes in other parts of the codebase, and often extremely out-of-date.

Adjustable by tooling and prompts - my favorite is a ban on comments >2 lines, a tagged index system for effectively in-repo RAG with short documents on each tagged topic like INFRA-DATABASE-SETUP.md - enforce with tooling that the tag matches the doc and vice versa. Enforce that any PR that has a tag change in it has the relevant doc changes.

> Claude seems to me to still focus on getting things done vs. doing it properly, so from my experience it would pretty often mess up your whole codebase just to be able to finish its task, vs. stopping and rethinking the approach.

Tell it your values - "I value correctness over getting things done, and getting things done properly over speed" solves many of those problems. I've actually spent a portion of today rejiggering my values documents because the models have improved enough I no longer need to be quite so careful about scoping.

I was having those same problems, but particularly using Fable / Sol to do the meta-work has largely eliminated them over the last ~3-4 months. Codebase is now a lean mean token-conserving machine.

Thanks! Many of your issues sound more agent / LLM specific than LLMs in general. At least compared to my experiences.

It also aligns with what I saw from Claude & Claude Code when I used it last year for a while. Now I use Codex and don't see (nearly) as much of that sort of behavior.

> But LLMs seem to have a hard time getting the big picture and reusing code that is already implemented and ALMOST does what you want vs. rewriting everything from scratch.

Yeah that's probably the weakest point of LLMs still. GPT-5.6 Sol got much better at this for me. I still usually end up doing 2-3 iterations of prompts and exploration to clean up various things but usually it's pretty light work now.

> hundred of thousands of lines of documentation written as walls of text in markdown and weird coding decisions, your codebase is impossible to work out for a human.

Okay Claude definitely seems to have an issue there. From what I've seen from coworkers using Claude it generates reams of endless docs. I resist the urge to `rm docs/planning*.md`! I don't think they get how bad Claude is at that.

A month back I tried the latest DeekSeek and it made reams of text back and forth with itself, but the code it output was reasonable and it didn't make pages of markdown files either.

> I make it build things in small changes, I give very specific instructions about the architecture, and I make sure to point out existing functionality that can be used instead of writing something from scratch.

That's a bummer. I found that once the models start having problems that it cascades.

I've also been able to keep steady progress on a 70k+ LOC GUI side project without the endless whack-a-mole of bugs using Codex and Sol. Squash a bug, review architecture, move on, etc.

> So maybe there's 2x speedup in coding. But that's only a small speedup in the totality of everything software engineers do.

Amdahl's Law should be familiar to anyone with a 4y computer science/engineering degree. Why aren't they applying it to their own throughput?

Well, there’s probably several reasons:

  1) because they’re not doing those others spheres of work
  2) because they don’t think about the work
     nor how tired they are afterwards
  3) they’re proselytizing AI work as the future and
     that conflicts with that vision
  4) the other shoe hasn’t dropped
  5) they really don’t see it
  6) some people genuinely hate programming and this
     helps them skip that.

I’m sure others as well.

Let's apply it:

- 25% planning & aligning with other teams

- 25% coding

- 25% testing/verifying

- 25% code review/rework

I'd say that thanks to LLM assistance I'm 10x faster at coding, code review/rework, and testing/verifying. (LLMs can partially automate testing/verifying too, and the code is higher quality now as well so less testing/verifying is necessary).

So that leaves us with:

- 92.5% planning & aligning with other teams

- 2.5% coding

- 2.5% testing/verifying

- 2.5% code review/rework

Obviously, that makes zero sense as a split. If you saw any organization doing that, you'd suggest having fewer teams, more silos, etc. Maybe you have designers produce code, instead of showing the designs to coders and having the coders implement it. Maybe you force all your engineers to dogfood the product that way they can identify issues themselves rather than needing QA teams to do it. And so on. so the last category, "planning & aligning with other teams", falls too.

The other hard to measure part is that I see/hear a lot of LLM usage going towards dev work that was never prioritized before.

Suddenly devs who were cranking out features with no interest in infrastructure are attacking giant refactors to make the code more understandable to the LLM.

Other devs are using LLMs to build themselves quality of life SDLC tools completely separate from the core code base.

Plenty of other examples of this.

The quality of my bash scripts is 100x what they used to be. Proper help messages, flag parsing, functions, internal variables, using sed all the time (I could never wrap my head around that).

Of course, the main issue is that they’re completely undebbugable now. My bash scripts used to be a sequential list of commands, now they’re 500 lines of variable laden functions.

Is my life any better? Dunno. But it’s satisfying (until there’s a bug)

> the main issue is that they’re completely undebbugable now

Well, that means the quality actually dropped then :). Looking impressive isn't equal to quality, understandability and reliability is

100x UX quality, not DX quality.

Not necessarily…. As with all engineering choices there are tradeoffs involved. There can be both advantages and disadvantages to using AI to build scripts. And more likely than not you’ll need AI to help debug the scripts when they have issues. Real tradeoffs to consider.

I switched to js for dev related scripting after seeing how much python scripting is used for portable c++ projects like chromium.

A node glob() or a regexp string.replace call is probably easier to read than spaghetti shell.

And your llm might do a much better job of creating clean, readable and testable code.

llms are 100X for me, but only 10X goes into actual shipping code. the rest is all on doing the dev process the way it actually should be done: incremental POCs, refining specifications, modularization, exhaustive test suites with 95%+ logical unit test coverage, fuzzing unit test coverage, exhaustive e2e testing, with a fuzzer harness over e2e tests for long term simulation and scale testing, bechmark gated self improvement loops for performance optimization, etc. i guarantee you i can ship higher quality 100% llm generated code than any human.

It scares me how much of this is true - people truly weren't doing all of this before; but like ... some of us (like myself) were, and we were trying to push other people to write good code like that - code that helped everyone else's future velocity be faster - including our own on the next MR and the next one after that.

I'm glad you're doing all those right things; but, really, some of us were already doing most of, if not all of that. And some of us (like myself) were doing it very quickly, too, since we'd been doing it that way over a decade and constantly getting faster.

Just one more data point but this is what I’m seeing too. The extreme ‘ruthless’ prioritization has stopped and devs (who want it) have breathing room to make the enhancements they’ve been wanting all along. Turns out what is good for AI is also good for regular old school SDLC. All you have to do is say it will make AI more effective.

You missed documentation. There was never time for that until now.

IME LLM docs are worse than useless the majority of the time. It's like most LLM writing, an inscrutable wall of text that simultaneously says nothing and too much at the same time.

I agree, but there's been a push at my company for humans to write docs to help the LLMs.

I use CC to do the testing and verifying in multiple deployment environments. It’s faster at things like coming up a plausible test payload or ten, querying a bunch of metrics from datadog, querying and verifying state from a db before and after testing, scanning all the deployed logs for relevant errors. For me that has been more of an uplift than the coding. I still feel like communication it’s worst at. So writing tickets, getting a decision out of a group, documenting only the important things, etc. it’s still not that good at

  > - 25% planning & aligning with other teams
  > - 25% coding
  > - 25% testing/verifying
  > - 25% code review/rework

legit question: has anyone got ai to do all the above? i've heard people trying ai-dlc [0] but it sounds like a slog...

[0] https://aws.amazon.com/jp/blogs/devops/ai-driven-development...

Agentic tools changed the workflow in our org, we were very tense before with a simple yet careful team process.. Since last winter we're now pushing a lot more but the teamwork (which was brittle before) is now mostly gone, everybody can roll on its own, but the review process didn't scale and now monthly deliveries are full of "seems to work". nobody checks properly, nobody reviews, nobody tighten any bolts.. and nobody cares much anyway, the system allows it.

The way I see it you should calibrate the way you work with LLMs based on how confident you are on that specific area, and if it's your responsibility to own/understand it. Here's how it feels for me:

* Learning stage: 0.5x - 1x. I change my system prompt to teacher mode, taking the productivity hit for actually learning the system/tool pays off dividends later. I change my system prompt to "teacher mode" and slowly loosen it as I get more confident.

* Working-knowledge: 2x - 3x. Once I am ramped up enough I feel like I can get a decent productivity boost. Most of the time is spent at the planning stage. This is my mode for areas I don't really own or care about, just need to get work done.

* Mastered: 10x+ I have been doing web front end for 12+ years, I can quickly review plan/implementations and for my initial prompt I already know most of what I want built.

1x == my speed before AI

> * Mastered: 10x+

This really needs to be calibrated to the type of work and complexity.

I can actually believe that LLMs would speed up basic web dev work in small, simple codebases 10X for simple requests.

These conversations usually turn into people talking past each other because they’re working on different things. For other less routine and more complex work, expecting a 10X productivity boost is not realistic at all. It doesn’t matter how good you get at writing prompts and reviewing plans. LLMs just don’t solve everything for you in a good way. Some times the true nature of the problem is revealed while implementing it and by deferring everything to an LLM you spend days throwing tokens at the wrong thing. There is a lot of work where the LLM speed up comes from helping you quickly search docs and codebases and double check your code, but handing the entire thing off to an LLM isn’t reasonable. These tasks aren’t going to reach this mythical 10X productivity boost that is genuinely achievable for much simpler work.

> Some times the true nature of the problem is revealed while implementing it and by deferring everything to an LLM you spend days throwing tokens at the wrong thing.

If you look at the code the LLM spat out (and you really, really should!) you will immediately notice that its shape is not what you thought it should be. You might not notice immediately if you're learning, but if you really "mastered" the domain, you will "just" see it. You will also recognize the problem with your assumptions, and immediately (or after some research) correct the prompt.

Looking at the diffs for everything the LLM does slows you down, of course. Not looking - or looking and not recognizing problems, for one reason or another - can be initially faster, but a single pathological case can eat both the time saving and tokens. For domains you truly "mastered", the current models can generate code as if they read your mind (because you can be that precise in the prompt, and quickly), so it's really glaring and very hard to miss when they somehow misread your mind.

It only works at the "mastered" / "unconscious competence" stages, and only in those narrow domains you truly mastered, but it does seem to work. Is it 10x? No idea, but there is a marked change in the speed boost when crossing from conscious to unconscious competence area, with everything else (harness, model) staying the same.

What they are saying is there are domains where there's no such thing as an expert who has the unconsious competence to immediately notice the LLM made glaring mistakes.

Some domains don't have the established regularity of architecture that it takes to achieve 10X efficiency. Two I can think of off the top of my head are embedded systems, with a variety of sensors, outputs, processing power, and memory, and novel protocols like ATProto, for which training data is thin on the ground.

Hey, could you share the prompt you're using for "teacher mode"?

Someone posted their AGENTS.md some time ago, that is used in practice to teach their students[1]. That might be useful to you in this context as well.

[1]: https://news.ycombinator.com/item?id=48359858

I'm probably in the 5x-10x range, but we are a very small team, and we've been accumulating a backlog of ideas over the years. I drive the architecture, and have almost complete autonomy over the work I do. I think someone in my shoes will see a greater result than someone trying to do the same thing in large team, with all the usual process and ceremony, only wanting to go faster.

Where as I'm stuck in a big team with a legacy codebase and lots of business rules, performing small incremental changes. So I spend my days banging the Esc key and shouting "Shut the f**k up!".

Luckily I work remote. I'd turn it off again, but my usage is monitored and I don't want to look like a Luddite.

So may be .75x to .8x range.

In my experience a legacy codebase is where LLM shines. What would take days to grok is understood and explained almost immediately. There was a bit of job security with being the one whose worked with IIFEs to manage state.

You must know the codebase well if the LLM is slowing you down.

Some of our biggest wins were in a legacy code base, where changes would require a lot of archaeology. I feel like Claude (and probably most coding LLMs) really shine when optimizing fat blocks of SQL, given the complex relational calculus.

There's someone I know that jumped head-first into AI... they have embraced everything agentic. While setting up mail on their new iphone they encountered a problem. They had claude go through and do a bunch of tests against the server... allegedly. It didn't do any actual testing, though it did generate a long-winded report of what it pretended to try, fully blaming the server the whole time.

Actual problem: he typed in the wrong password

Not sure what your point is, could you clarify? Are you saying that your friend is stupid? Or do you mean Claude is stupid? Both?

Claude lies, like all LLMs today.

Oh, right. Let's throw it all in the trash, totally useless

Why would you say that? It's incredibly useful technology that also lies. It's not binary. There are tradeoffs. Just be careful.

My point is that far too many people that offload everything to AI tend to just accept the text it generates without thought or verifying what it generates.

I recently removed about 50% of code in a feature submitted by a colleague, because it's a pile of over-engineered mess that either never gets used or caused trouble for us. We probably would not have added the code in the first place if we actually discussed the design. With the code removed, the feature is now much simpler and more maintenable.

And that's the second time this happened over the past few months.

One time before LLMs we outsourced this project. It was something to sign file headers, written in ruby by an overseas company. It was something like 4500 lines long, and I wound up reducing it to something like 550 lines.

My story, like yours, has nothing to do with the topic at hand.

It happened many times to me, as well. Humans write terrible code.

You didn't even mention if the code you're talking about was generated or not.

I think my last paragraph should have provided enough context.

I'm not sure about the x, but the first thing that arises from that is, I feel like in my case it's way higher than 2.

The 2nd thing is, how do I measure that.

---

In my case, the details of my work (Kinda DevOps, kinda Senior Dev) makes it that having an LLM to do the heavy lifting allows me to do things not only faster, but better, and across domains I do not hold expertise on.

An example of the effect of LLMs in my daily work is that I'm in the middle of a PHP upgrade for a rather large legacy application, and the "heavy lifting" is really out of the scale, letting me concentrate on what really matters, while at the same time if I do keep "the harness" tight I'm certain the results are the correct ones. Also correcting course is just as cheap.

Not having to worry on the tooling on exchange has the incredible desirable result of my velocity being incomparable to what it was before.

Then we have the side effect of how easy to do transfer knowledge: Rather than telling the QA guy how to do the work specific for this task, I defined a set of files (.md documents, skills, an off-the-shelf customised MCP server) that assist QA into doing the work in a way that helps me do my job better and faster.

There's also a clear possibility that what I'm doing will expand to the rest of the team I am in, completely altering the way in which we approach development.

If we take 'x' as 'mileage', yours might vary. Mine has, and I'm baffled at the positive net results I AM getting.

Also, this what I do (coding?) is extremely fun again.

Being able to work close to the speed of thought is the best high.

> allows me to do things not only faster, but better, and across domains I do not hold expertise on.

How would you know it's better when you have no expertise?

He didn't say that. He said [with my re-puncuation of it]:

"...allows me to do things not only faster, but better. And [allows me to do things] across domains I do not hold expertise on."

I have also done things in domains I don't hold expertise on. I'm a web dev, but I built a terminal TUI client yesterday. In a language I don't write.

He didn't say THAT was better, he said he's doing things faster and better. Presumably he's able to measure many of those things against how he did them before.

In the case of new domains, I know I can produce better output than my previous self, because the output compiles and does what I want. Previously I could not produce compiling output that did what I want. It's definitely better now.

Because it produces the desired output. The purpose of a program.

There are many domains where an intelligent human can act as a discriminator for output without knowing exactly in precise detail how the process itself works.

We had contractors come in to do a project for us once, they had a team of programmers in India as well as a team on the ground. The team in India would produce a bunch of code for them during the night, and they'd review it, and carry on. Things were being done very quickly. Management was happy. Finally after months of this they let us review their code. Turns out that the contractors were just reviewing the output, and they were nailing it. The code itself was absolutely unmaintainable. There was clearly one programmer that had never heard of a jsp, and wrote all of his java to produce html within the java code itself, and another who wrote all of their business logic in scriplets in the jsps themselves. Years later we finally got the okay from management to refactor the whole project. That refactoring project reduced the deployable artifact from over 700mb, to just under 40mb. It reduced a ton of security vulnerabilities and made the code vastly more maintainable, and significantly more performant.

The purpose of a program shouldn't be the only consideration unless it's 1 off scripts.

That is certainly true, but there are also many cases where someone without core expertise who believes themselves to be a capable discriminator actually is not.

Welcome to Earth. I see people parading as experts in all kinds of things all the time, without LLMs to back them up. Political opinions, economic theory, foreign policy, professional sports strategy, who is or isn't intelligent. People blast their very uninformed opinions about these constantly, without realizing they're uninformed.

This is a human thing, not an LLM thing.

I even have a pet joke about it. Like yesterday I built something in Rust with the LLM -- I'm a total novice in rust. I said "I'm so happy I became an expert in rust today." I built a shed, "so glad I'm an expert carpenter." Some -- maybe most -- people really attach that pin to their lapel with the same amount of experience.

Producing the desired output is the purpose of the program, but it's not the only consideration: there's performance, maintainability, security, and so on.

for me, almost all of the work is specs

I am no longer:

- reading docs for hours and hours

- typing (barely at all)

- writing code

- manually doing tight debug loops

- using an IDE

to do this I had to give up reading or even controlling the code and focusing on behavior/design-level control (not superficial, still dictating overall technical architecture)

i have agents doing everything from writing the code, verifying the code, hardening, increasing test coverage, analyzing behavior, algorithmic perf improvements, managing/deploying to cloud resources, etc... (pretty much everything)

and I am accomplishing projects that would take months or years in a fraction of the time.

that's way more than 10x.

somehow, this is harder and more cognitively demanding than writing code

> to do this I had to give up reading or even controlling the code and focusing on behavior/design-level control (not superficial, still dictating overall technical architecture)

how do you verify the behavior? are you still writing or at least reading tests or just doing manual testing?

I discuss the testing approach and coverage with the model before and after, sometimes in a fresh thread that does a static analysis.

interestingly my input is still pretty important

> interestingly my input is still pretty important

how do you know?

How much does it cost in GPU rental?

codex pro plan currently

I'm pretty sure that just a year or two ago, software engineers were saying that coding was only a small part of the job; the rest is meetings, dealing with requirements, design, testing, maintenance. So, how can we expect a 10x boost in a productivity from improving a small part of the job - especially now if we have to clean up any mistakes the LLM makes?

Yeah that's a good point, in order to see a 10x gain it would require at minimum coding be 90% of your day [7 hours, 20min) (or somehow for AI to replace meetings)

Now you actually have very long meetings with AI, if you consider all these long design / explorer / post-poc conversations "meetings".

You still have to know what you’re doing. That came from years of doing it by hand. Where is that necessary experience going to come from for the current and future generations of juniors?

In 10 years llms will handle it all from start to finish.

We have passed peak LLM already. In 10 years, something might be handling "it all" from start to finish, but it won't be LLMs.

Saying we're at the Peak of LLMS is like saying we never went further than everest, these are probabilistic state machines sure, but the real value is the effort put into tooling and training these models in RL situations.

Shipping. Breaking stuff. Fixing it. And discovering, repeatedly, that the tool sounded much more confident than it had any right to.

Like always. This is why the average carpenter you hire in 2026 still won’t do a perfect job, despite carpentry having existed for thousands of years.

So you don't think fundamentals built by actually coding will be necessary to produce well structured and maintainable code with an LLM?

> I have not yet taken up woodworking as a contingency plan for "Software Engineer" becoming an extinct profession

Between "stable" and "extinct" are many stages of software engineering labor market change, continually more difficult (for labor) as we move toward the latter state.

Software engineering is likely to become a skill/tool attached to another harder to automate ability going forward rather than a vocation/craft on its own.

To the extent that software engineers are employable, they will likely need some other x-factor adjacent to software skills, whether in the science, business, creative/artistic, or social domain. I've already seen this trend emerging in the startup world.

Totally depends. The real question is what's your average productivity increase?

I had a task to completely gut out a codebase to share with a vendor. I gave them my estimate - 2 weeks. Asked Claude to do it, and was done in an hour. Reviewed the changes, and it was perfect. This is an outlier of course, and it was a pretty basic codebase. But it's real world stuff.

Overall though, if I had the mental fortitude to work for 8 hours straight, I could easily average 5x my "normal" performance. But most days I can't perform at that level. Also I admit I am not a fast developer, I do a lot of testing and verifying as I'm paranoid.

Even this is overstating it IMO.

I already convert from multipliers to percentage of increase, so when someone claims 10x they very likely mean +100% productivity, and here 2x means +20% productivity, which seems about right. Nobody that was normally productive before LLMs has suddenly 10x'ed their output now.

The problem is that 20% productivity when it comes to generating code, really doesn't translate in 20% productivity increase overall, when you take into account the fact that the code quality is worse, the fact that writing code is actually not the majority of your time spent, and that people get burnt out from the usage.

2x is 100% increase.

I know, and I'm saying that 2x claims should be interpreted as an actual 20% increase, and 10x claims as an actual 100% increase. People are just bad at estimating their own productivity.

Ya, I feel 10% to 50% increase seems more realistic, when I see 100% to 900% increase claims it seems a bit extrapolated, at least I don't feel I'm seeing it at my work. And then it depends if we imply a given feature/fix/enhancement takes half the time as before, or we imply that we can get twice as many features/fix/enhancements done in the same amount of time as before. Those are not equivalent. I think the latter might be more accurate as to how it helps.

Yes, people (myself included) have a strong tendency to use hyperbolic factors, "ten times", "100x" because true estimates are very hard to do.

I am not a programmer but for me the return is 100x as I am automating stuff that I knew could be automated but did not know how to but at the same time it was not worth the effort to write it down and pay someone to code it for me. But now I have created so many python and bash scripts that save me hours and hours a month. In my opinion the people that can most take advantage of the programing capabilities are people that least use it for such or know the how they could automate a lot of the repetitive tasks they have to do at work or on their own time.

Now that you only work 3 days a year, what do you do with all the spare time?

Browse here and reddit and read a lot of fiction and lose a lot of money day trading futres.

yeah i can't say if it's 2x or 10x or 100x i don't know. i just know i barely open vim anymore. I still find my rate of development is limited by how much my single mind can hold at any given time. I am however accomplishing with a much higher rate of confidence many many more daily tasks usually by 10am... the limit is usually my willingness to imagine and think critically about an idea with enough clarity to result in something that make sense. It's a different energy then normal development but similar enough.

I agree with the general premise with respect to current SOTA, but for making nice plots that I would've never otherwise made or learned how to make (to the same standard), easily 100x.

> Never write READMEs, docstrings, or comments.

I found that latest codes don't write comments in code by default. And when they do, they write stupid shit like "This was code that did X, it was now removed".

You have to explcitly prompt them to write comments in code. They are still useful for you, the user. But are arguably useful for the model, too, given how many of them (especially Claude) only reads small chunks of files. So I'd rather have code comments than it reproducing a picture from incomplete data.

A small mention in your user-wide Claude.md (or Agents.md or whatever you use) about the desired types of comments goes a long way. The typical "why, not what", clarifying larger-picture stuff, and everything else you find in every book on code quality

I recently, with the help of Claude Code, made a tool which involved a database, two separate apps surfacing and interacting with various aspects of the data, and a user base of dozens, with some semi-specific domain knowledge on my part and an intention of re-using and building upon the minimum viable result.

OP's description very much mirrors my experience. I was able to do many things much more quickly with the help of the LLM, but there is no substitute for actual users interacting with the tool, saying, "I like this," or "no, this is wrong or needs work," or even, "here's something none of us thought of before, but now this tool makes me think XYZ would help us and might be achievable." That whole interaction takes real time and I don't know how you replace it with coding agents.

Secondly, on the matter of code structure, just on a qualitative level, I can see that claude will do things very efficiently on the way to a goal I give it, but it can't read my mind and know that I may want to repeat a specific pattern across two client apps. Or that its shortest-distance solution makes extensibility or broad applicability difficult. That I might want to share code and structure things in a certain way. Not without me saying so or, in many cases after it has built something workable, duplicating refactors I make with an eye for reusability or maintenance. Making those changes in time is important if you don't want to burn tokens later as the LLM tries to unravel its own spaghetti. And again, whether I am coding those intentions directly, or writing out detailed instructions in english, all of that takes time.

I’ve been vibe coding apps professionally for the last year and 100% agree with your first take. I really underestimated how much work there is in what I just call the “coordination” of developing software.

Users pointing out things that are missing, or me forgetting to tell them I’ve added something, or both of us having a different idea in our minds of what something should do. It’s still extremely time consuming.

> As for documentation, I've found this simple instruction to vastly improve LLMs' output:

> Never write READMEs, docstrings, or comments. I will write those myself later. And yes, I really mean this.

This is quite validating as I came to the exact same conclusion myself. We’re required to use an LLM for every task at work that touches code†, and I was really struggling to get Claude to stop with the long waffly comments that reiterate the next few lines of code in 3x as many characters, making contextless references to subtasks in whatever harness du jour we’re using this week.

No amount of examples or explanation of what I wanted would make it stop. And then I realised of course, I’m asking something which has no concept of meaning (or, indeed, anything) to only add meaningful comments. More fool me I guess.

Of course, it’s ultimately pointless given all of my colleagues are regularly opening PRs with more comments than code anyway. 80% of my code review responses these days are just increasingly exasperated “pointless comment, please remove”.

† This is just as infantilising as it sounds, by the way

+10, I can't stand AI-generated comments. Half the time they are paragraph-length translations of every 5 lines of code into English prose, or just regurgitating whatever prompt they were given. When they actually try making high level comments they are worse than useless. It's really hard to put into words exactly why. It's something about how they are very verbose, but at the same time vague and stated in a slightly-off manner. Or they will introduce new terminology not reflected in the code's nomenclature, sometimes terms that are just synonyms for terms that the code already uses. Something about it makes me instantly distrustful of any information the comment may actually convey.

locally correct and globally incoherent is how they seem to me. I, and a number of folks I have talked to, find some of them incomprehensible. Like each sentence makes sense, but the overall thing does not.

It's because they don't really have a mental model. RL with Verifiable Rewards can reward working code, but it can't tell you whether the prose make sense or not. In theory RLHF should be good with this. In practice, it appears to be a measure difficult to quantify and steer. I am not an insider though. So it's just a theory.

I have the same impression. It is an interesting phenomenon. Is it one of the ways we will gradually learn that AI in the form of LLMs is fundamentally lacking?

I use AI all the time, it gives a tremendous boost in speed for many tasks and it is most certainly here to stay, but there is something off about certain things, like code comments. Like a weird person.

I kind of wander if the comments are intentionally bad, providing programmers something to bike shed about. A sneakier approach to the flattery the models used to so use so heavily.

You can create a PostToolUse hook [0] which will automatically chastise Claude to shrink comments to one line. It's not perfect but better than the sea of prose it tends to generate by default. Bonus is that it would apply to your colleagues as well!

[0] https://github.com/chrisvariety/branch-fiction/blob/deb37f2b...

Good thinking! The multiline paragraph comments drive me up the wall.

No amount of agents.md / commit to memory updates has stopped agents from littering the code base with useless comments.

Think I'll add it to pre-commit hooks getting called by posttooluse

> We’re required to use an LLM for every task at work that touches code

Why not just write the code in the prompt so LLM can paste it.

I have the opposite problem with Code, it barely writes any comments even for fairly complex C++ code blocks (systems programming stuff), have to prompt it to add extra details about why it did something. Request in agents.md doesn't seem to help much.

I'm in the same boat. We have very little documentation where I work, and one of my goals was to improve that as I got onboard. We have a newer dev that started ripping +3K, +5K, +10K line PRs against our documentation repository.

I have tried to shut it down, it immediately nullifies the docs and make it genuine slop.

I truly believe that documentation should be human written and human readable. It's all of the stuff that AI does NOT pick up on. Small inconsistencies, only the necessary details. IE: Don't just rewrite the code that we have into readable english (We are software developers after all. We can read code.) - comments / documentation to me has always been the "in between the lines" stuff that can't be put into the code.

Even if that 2x were a representative number (although another HN story today says 1.1x - 10% improvement), what really matters is whether companies are overall seeing any AI spending return on the bottom line.

There is a lot of reason to suspect that at most companies, especially more established non-startup ones, there will close to zero bottom line benefit, because these companies already have free paid-for developer capacity (developer down time between project cycles) that presumably they would be utilizing if there were reason do to so. In a startup environment where there is zero down time, then productivity may at least show up in reduced time to market.

I think Uber's current experiment of limiting AI spend to 10% of salary per developer is interesting since it implies that spending more will not even recoup the extra token cost, although it does remain an experiment. Maybe they will see a revenue increase related to AI spending and choose to tweak that limit up, but it's also entirely possible that all they are doing is reducing developer workload by giving them a productivity tool, and there will be no financial benefit.

I agree the bottom line ROI largely isn't there and won't be in the future either. The upside for companies is not the bottom line but the potential savings from hiring fewer developers or cutting jobs.

As AI based development processes continue to improve, the confidence level employers will have in safely cutting jobs will increase. The longer term impact of AI is HR savings rather than introducing new capabilities into business.

> The longer term impact of AI is HR savings rather than introducing new capabilities into business.

I tend to agree, although this goes against the Dwarkesh narrative (apparently matching current SV zeitgeist) that there is some insatiable demand for "warehouses full of genius coders". If there really was demand for more coders (especially at the high prices the AI companies are hoping for), then companies would be hiring the unemployed developers available right now, not laying more off.

I wonder how many CEOs or CTOs appreciate the massive functional gap between an AI coder like Fable and a human developer - full general intelligence, with continual learning, theory of mind (so they understand what the boss wants, not just what s/he says), etc... all available now, not some 5-10 year AGI/ASI stretch goal (that may in fact take much longer - artificial brain, not just "AGI") ...

> Never write READMEs, docstrings, or comments. I will write those myself later. And yes, I really mean this.

The precise and rigorous practice of “engineering” in 2026.

When you look outside of coding, the productivity boost and time saving also hits when you're using the Ai as a thinking partner. eg there is a significant different between Opus and Fable when it comes to thinking, data analysis, strategy. I also think Grok 4.5 is extremely underrated, faster than any other model, less verbose, reliable. The human to agent trust factor is also under-considered - how much does the agent "get" you when you give it less.

If you're getting 2x you are likely at the limit of your own ability to coordinate or plan intelligent work effort, and you have plateaued such that even when models are twice as powerful as they are today you will still be getting 2x of your own human throughput.

We all have a limit after which we lack the attention (or attention to detail, or time, or energy, etc.) to meaningfully manage it.

Similarly, a manager may be able to handle a team of two very well but end up poorly managing a team of 20.

It used to be that knowing how to write code was a big factor in productivity. Now it is less of a factor compared to the many other cognitive and metacognitive faculties that working with agentic teams demand.

maybe i’m an edge case but for 99% of the day to day work i do (embedded, not x86), llms will confidently give me wrong code so many prompts in a row it’s a 0.5x factor. when i venture out to something more normie, like webdev on either side, it does better but still takes more time than if i sat and did it myself once over. i’m not exceptional, but certainly understand everything that has been generated at any rate.

i think LLMs are only a useful multiplier if you’re working on the plumbing parts, it’s never useful if you’re doing anything novel really.

if this bothers you, i am sorry and i wish you the best with your reddit trained auto complete. i write new code

I do a lot of novel work, that is not setting up a webapp - and totally concur on the novelty point with LLM's failing hard.

I can accept the author's premise that LLMs double coding productivity today.

Semiconductors following Moore's Law increased their density, which is a reasonable metric for effectiveness, by 40% a year for several decades. Leaving aside projections of AGI in two years, it's still a substantial impact if we can figure out how to increase coding productivity by 40% a year for two or three decades. I don't think that gets us to AGI or the Singularity, but it's similar to the impact of the steam engine, steel, or electricity.

These are "normal technologies" that were transformational, and that may be the path we are on with LLMs.

I won't argue this title isn't the author's experience, but it's misleading to rest a quantitative claim on a qualitative argument, while also refusing to fully exploit the technology.

From a firm that has fully embraced agentic engineering, they offered topline stats on their measurements of a 4x boost from November to February, and 8x on top from February to May. That's 32x since November 2025, coding with LLMs.

And like whatever, this is second hand and I'm not going to disclose the source. Actual hard, published, and peer-reviewed research is needed here to backup quantitative claims. 2x, 10x, 32x, whatever.

> LLMs' increased rate of adoption in 2026 is largely due to them becoming reliable enough to run effectively in automated feedback loops

Is that really from improvements in LLMs, or from improvements in the feedback loops?

I think about 80/20, harness vs llm progress

Does the number × depend on how big your thinking is? If you are only 2× then you need to think bigger, aim higher, etc. I think the limit on the 2× is not the ai? Where is the limit now? I think it's unknown.

Definitely feels like a lot more than 2x to me. Any time it feels like the agent is taking a long time I have to look in the mirror and say to myself "What could you have done in 5 minutes?"

> As such, I use LLMs mainly to produce a rough draft of the code which I then iterate on heavily, at least until I like the general structure.

Weird… I would have said this was how it was about 2 years ago, but no way in 2026.

My observation is that AI tools generally make people /less/ productive. They are more productive in that they produce more stuff, but they are less productive in that the stuff they are producing has lower inherent value delivery. The primary place where I see this in software teams as a Product person is in document creation. If you are using an AI tool to produce internal documents, this is likely a net-negative value activity that actually reduces the overall productivity of the team. Why? Because writing is thinking. By handing off document creation to an LLM for a document that's supposed to communicate important details between teams (e.g. technical design docs, requirements docs, strategy docs) you are actually handing off thinking, or rather handing off not-thinking as AI tools cannot "think". The outcome of is poorly "thought out" slop that generates more work for everyone involved to resolve /or/ everyone simply ignores the document and the previous processes stop being useful.

If anything, I'd say AI tools in /most/ enterprises as people are trying to use them now are at least a -30% of productivity. If used correctly for taking a human-written/thought requirements doc, converting it into an interactive prototype that can be critiqued and ultimately included as part of the solution alignment within the requirements doc, and then is handed to an engineering team that is effective at using humans and AI to produce code, then it's probably a benefit. But most businesses lack the internal rigor, quality culture, and data governance to support properly applying AI tools in a high context manner internal to their business.

Have you worked with codebases where people produced slop even before LLMs? In such codebases LLMs are a godsend. They can churn around huge amounts of code and find needles in the haystack.

Since I'm on the Product side these days, I don't spend much time delving into the codebase at this point in my career. I can believe it though, as I've done quite a few personal side projects using langchain and various harnesses. The problem I see is not necessarily that LLMs build things badly (although I know that can also be true), its that because people are using LLMs to generate documents, we are then building the wrong thing entirely (whether with an LLM or not, whether badly or not).

Removing human thought from the process of development is generally a net negative, IMO.

I spent 30min with Opus 5 generating some CSS and components for my Phoenix LiveView template project (which I use for starting new projects). Then I spent around 5 hours parsing and cleaning up the output. I could of course not have done that, but since it's my spare time project I can care about the code and quality.

I did get wins for sure (and I did save time), and I'm sure most people wouldn't even spend five minutes cleaning up, but this is definitely one place where people talk about different things regarding whatever x speedup they get.

> but since it's my spare time project I can care about the code and quality.

I suppose it's yet to be seen, but I'm seeing a lot of spaghetti code being dump into the codebase I'm working on at the moment by other devs. I don't think it's wrong per se, but human code would have thought more about the right abstractions and trade-offs for future maintainability.

There is a gap right now between what an AI can and can't do, and I don't think it's just a time and cost limitation either. It seems like current AIs are very good at writing good code at a surface level, but very bad code when you zoom out a little.

We have a lot of non-technical people committing code these days to constrained microservices and from time to time things break and I'll take a look and I'm always just like wtf am I reading? You see code so bad that you'd immediately fire a human engineer had the wrote it. Thing like explicit hacks to bypass errors that it should not be bypassing or mock data to fake some API that it wasn't able to access. Stuff which literally no human coder I've ever worked with would be contemplate doing...

I think it remains to be seen if the 2-10x speed up some are claiming to have today will persist if the junk code continues to grow.

Everyone can drive really fast if you give them a fast car, but very few can stay on the road when driving fast. LLMs are like free supercars for everyone, except people know nothing about how supercars work or how to drive them. Same thing with software, going fast is incredibly hard if you do it for any period of time, and I think this is something we will read more about in a couple of years, not that human written code never turned out to be complete disasters, but I believe with LLMs we will see at a completely different scale. And yes, car analogies suck.

I think for small projects you need to scale the coders mental model, which afaik works in the fastest time, when you let ai do the exploration and planning, but coder needs to write the code itself and then ai verifies. That generates the best of world codes and the coder at least has written the code. It’s scalable.

In projects that are so big that no single coder or no small group of coders is sufficient to grasp I don’t have personal experience, but my guess would be that they are just a cluster of other small projects.

I think it matters a lot to understand the kernel of AI in the past 10 years. It's the only way you can truly appreciate the exponential. It looks flat unless you see the dramatic leap in transformer architecture and scaling laws and reinforcement learning. I know it's been hyped to death but you have to see how those things are truly exponential at their core to appreciate how it's not just walking up one step or two steps faster and how it's walking up buildings etc.

When I approach a new area or task I first explore it, by hand, and implement, by hand, the task. This way, I get a deeper understanding and it gives me some momentum to llm the rest of the related tasks. Till I loose enough grip to "re-dive in" that area and to again, do stuff my hand. Some tasks, I only do by hand, and some tasks, I only use llm for.

Really depends on the task, my the project and on, and honesty, my instrest and availability at the moment.

Sometimes small tools are all you need, even if they're slop.

About a month ago I pirated an Argentinian movie and the only subtitles available in my language were out of sync and at a different speed/framerate so adjusting for delays wasn't enough. I was unsuccessful at fixing it with VLC and every other "online tool" I could find.

Knowing a srt file is just text with timestamps I vibe coded a python script to take in sample times throughout the movies so it could recalculate the rate and shifts and replace them in the file. It worked on the first try using only the deepseek web chat interface and my terminal.

Without AI I theoretically could have sat down with a pen and paper to figure out the math adjustment, then looked up python input handling syntax which I already forgot, typed something out and then hammered it into shape through trial and error over a few hours. But the friction and time investment of doing that would have been so great I would have just given up on watching the movie instead.

This is a dated take to me. I think the next stage in unlocking productivity is so called "loop engineering"; figuring out how to effectively not read all the code while ensuring quality. To me that means implementing statistical quality control and formal methods.

Before I get there I have to figure out how to reliably audit plan adherence. The problem is that when the specs are in natural language, as they are, you need a fallible LLM to verify it.

I agree that removing human from the loop is the next stage, but we're not there yet beyond small programs.

Agents left alone tend to create so much tech debt, that once the program becomes so messy that they can't fix one bug without creating two new bugs, it's too late to even clean that up.

The program will be super tidy in superficial aspects that linters catch (everything neatly formatted and verbosely commented), and roughly appear to do what it's supposed to do, but everything in between will be "I can't even".

We need something else than formal methods, because the problem is usually in lack of simplicity - you get four versions of the problem solved in four times in four in different ways, each uniquely flawed and just incompatible enough with the others that unifying them is too big and hairy for the agent, and will result in eight different glue adapters written in the process.

We built a conversational terraform platform at work, you describe what you need for your infra and it produces the terraform code following company standards and rules. What is important for us is the LLM never writes any code. It just extracts a structured spec from the conversation, and a deterministic engine (plain python, no LLM) renders the code from approved modules only. So there is no "code quality" discussion at all, and since the spec is structured and not natural language, you don't need another LLM to verify it, the risk moves to the extraction step which is small and easy to eval. But probably this only works in narrow domains where the output is composable from approved blocks (IaC, config etc), and its not a new idea, just old discipline behind a chat interface. But in that slice the speedup is way beyond 2x.

tf is yaml anyway, and llms can output it natively in structured output mode.

Sure, structured output gives you valid syntax, but thats not the problem we are solving. Our engine simply cannot produce anything outside of the approved module registry (pinned versions, mandatory tags etc). LLM with structured output still can produce plausible but wrong configs. Also same spec always renders the same code, thats what reviewers and auditors actually care

This will be a deeply unpopular opinion, but I have an 80k line Go project I've created where I haven't read the source, but would absolutely stake its quality against most hand-written projects of a similar size[0].

Recent LLMs struggle with the author's stated issues only inline: they're entirely capable of going back and evaluating codebases to find architectural issues and LLM slop signatures, especially when you use other models to check one model's output. This wasn't true until Fable-class models, but it's true now.

0: https://github.com/pjlsergeant/byre

I think it'd take some time for people to figure out what kind of harness or workflow would work best with each individual codebase or team culture. AI certainly can help speed things up a lot, but for now the humans driving the AI need discipline to follow good software-engineering practice, or at least have multiple AI agents critique each other's work, instead of just pushing out whatever one AI agent writes.

I was saying it's a 3x multiplier for me during the Opus 4.8/GPT 5.5 era. Now, I don't think I'm notably faster than that, but I'm taking on projects I couldn't have done without LLM assistance, and not just because of time, but because of knowledge.

It's become a lever that allows me to learn as I go while tackling problems that are far beyond what I already know. I know a lot, I've programmed for decades in a lot of languages in a lot of domains. But, with the current crop of the best LLMs, I can reach for bigger problems...and actually make progress. I've read books about DSP for audio, and have done little toy projects in the past, but with LLMs, I'm building complex and working synthesis engines in a couple of weekends. That would have required a few months of study and experimentation before. So, that's a huge multiple. Like 100x.

For things in my area of expertise? Probably still just 3x, maybe 4x, because it makes fewer mistakes I have to fix in code review. It still writes terrible docs, as the post mentions...they don't understand user desire, so they simply can't write documentation for a user to actually use. You can't prompt them to make really good docs, but I can usually prod their docs into coherence a bit faster than writing it myself.

But, it's the lever for doing things I've never done that is such an addictive thing. Which, I imagine is how non-technical people feel shipping their first web app or whatever with these things. For basic work, I think we're at a point where almost anyone can use an LLM to make working software (not necessarily secure or stable software, but working). But, I think we're also at a point where an expert can make that lever really do something, and I hope that means we'll begin to see extremely ambitious new software in addition to all the throwaway junk that's been proliferating at a frightening pace.

In short: Maybe don't make another fucking "memory" system for your chatbot so they can be your friend who remembers your birthday, and instead work on something meaningful.

I wholeheartedly agree on readme. AI generated readme irks me off almost immediately.

I am able to coax much more of the kind of thing you suggest that LLMS can not do by having it generate deterministic code and then narrate over the results. Hopefully that is not too negative and gives some helpful guidance.

I think LLMS have reached to a point where now its just about money. you have to decide whether you really want to spend your tokens to do X or you can just do it yourself and use the tokens for something else.

My measurements are Claude is 6x - 10x on prototyping and MVC then it starts to be 2x at most. You can also witness 0.5x (asking simple tasks might require more time than performing them by oneself)

Well, it's interesting. I've been using LLM's for coding for over two years. My findings are somewhat consistent: Feed it jobs in small chunks and it can be a great auto-complete, beyond that and it turns into a mess very quickly.

I'm two weeks into a project that has been interesting and also has confirmed the above yet again. I am porting a Windows application written in Python to C# using Avalonia UI.

I have never used C# or Avalonia UI. I've written Windows applications in other languages, never C#. I don't know the language, libraries, etc.

At first I told Codex: Here's the source code, port it. I just had to run that test.

Well, it didn't end well. I'll describe it as a frustrating set of prompts that seemed to result in the implementation going in circles with constant problems being introduced, breaking-fixing-breaking, etc.

I then started again with a clean repository and played the role of the architect with full documentation in the form of code. File-by-file, I had it port modules to effectively develop an operational foundation for the classes, methods, properties, abstractions, hardware interfaces, etc. in the original Python program. That went well, was very fast and a good experience. I am running Codex in JetBrains Rider and the integration is excellent. The native OpenAI Codex application is an absolute dog...it pegs all my cores at 100% while doing nothing.

Once all the underlying infrastructure was ported and, to the extent possible, individually tested, I threw UI integration at it. This happened quickly and OK from it's-ugly-but-I-can-use-it perspective. Codex seems to be way out of its element when it comes to UI/UX understanding. Funny examples like placing a button on top of an image with the image covering the button because it had a higher z order. Or finally placing the button at the correct z order but making it transparent on-hover. Funny stuff when you are moving slowly and you see it happen. It just proves that there is no understanding whatsoever.

With all of that and the experimentation, I'll estimate that a six month project will be cut down to three to four weeks. Another month and it will probably be a much better program with new features and more advanced capabilities.

And I have not touched a single line of code. Developing solid prompts is the absolute key, something that you can only really learn by doing and through lots of experimentation.

Yes, of course, I fed it working code. I think it could be very different if I were to start something from a blank slate.

> placing the button at the correct z order but making it transparent on-hover. Funny stuff when you are moving slowly and you see it happen. It just proves that there is no understanding whatsoever.

I don't think it proves that. It proves it doesn't run and visually inspect the program. That's because the harness is lacking, not the llm.

When you do web dev with the thing, you can have it screenshot the browser and it turns out it has no problem understanding things like "buttons should be visible." You just haven't allowed it to look. You yourself admit that a human makes the same mistake, but sees it.

Anyway, have your model build itself a way to take and review screenshots of the app.

> the dark ages of 2025

This made me chuckle. I do think we are seeing exponential improvements over the years and that we don't always notice it happening because we're right in it.

I feel LLM can only build what's known and popular.

Sailing into the relms of lesser explored area are painful.

It's like an intern but never learns, only getting replaced by better interns.

I mean the "x amount" is relative to your existing skill level, right?

It's worth asking ourselves "why does the x amount matter?" I get the desire to define KPIs to estimate productivity gains. But this all seems to be—rather rapidly—leading to an increasingly dehumanized reality (both figuratively and literally) so we can...produce more software? I love building software, and I enjoy using LLMs to help me do it, but there's just this weird vibe I can't quite shake about how we're trying to quantify all of this.

This has been my experience on a recent small project for a client. I took effort to understand and prepare the spec before feeding it to Claude Opus. Yet, the result was slop full of defects, and one that I did not understand well at that. I just did not have a mental model of it. I just thought, there's no way I could ever feel confident about this code, something is missing. It did not take me long to write the important parts myself, from scratch and develop that mental model.

Then I would use Claude to make a small fix, write a unit test (write a unit test, not suggest what they unit test should be) or write some less important UI code. It just felt like old school development, with some assistance. I think we may see a shift back to a "copilot" mode and by the way, I still use GitHub Copilot. The $19 subscription includes $30 worth of credits and the autocomplete in VS code that uses some cheap model is completely free. This autocomplete is very good and I would like the market to focus more on IDE integrations. The agentic thing is either ahead of its time or it will never have time. Time will tell. I think the LLM can only be as good as the training data, so it will always excel at small chunks, but to implement entire projects of which there's such variety, I would question that. And also, how using it to implement entire projects means the developer does not hold the program model in the mind, which leads to more issues long term.

By the way, in this way of working, I see no difference between Sonnet and Opus. Sonnet is good enough to use as an aid.

200x coding with LLM

I've found a variant of the Gell-Mann Amnesia effect with regards to LLM coding. They're really magical if you're not familiar with the language, domain, or frameworks you are trying to use. But if you're familiar with those things then you'll often catch them lying to your face and producing a lot of plausible nonsense.

It's dropping off hard, though. I think web-search fixed a lot of it.

It used to be terrible for car repair advice, now it's mostly right.

This ignores that non-developers have gone from 0x to 0.5x.

All these hot takes that LLMs are -10x, 100x, 2x are kind of silly. It really depends on:

What you are using it for? What language, what framework, how many LOC, what domain, how much docs can the LLM read, etc etc.

What are you optimizing for? Cost? Human knowing how your code works? Getting something out the door?

How Good You Are at Prompting the LLM for THAT specific set of parameters?

I find AI can save me like 5x time on something, and in other cases it's wasted 5x time. Overtime I'm hopefully learning how to use it better.

... the official return of pair programming! The best part: LLM's save you from having to argue with a human. Added bonus: now you get to master two tools instead of one.

Here’s another one claiming 10% https://news.ycombinator.com/item?id=49113774

>As such, I use LLMs mainly to produce a rough draft of the code which I then iterate on heavily, at least until I like the general structure

Tell me you just started with agentic programming without telling me you just started using agentic programming.

Look, don't get me wrong: new folks learning tech should absolutely write articles about it! But their claims might very well change once they learn more .. and I strongly suspect that will be the case here.

Tell me you're bad at software engineering without telling me you're bad at software engineering.

Tell me you're bad at people skills without telling me you're bad at people skills

Such is life.

Just wait until inference becomes 10x faster/cheaper.

Personally i'm only using local LLMs that i can run on my 7 year old PC (that also has a GPU with 24GB VRAM because reasons :-P) so i'm not sure how much that experience matches what others are doing (though roughly speaking what i see people complain about Claude doing doesn't feel that different from what i see the local stuff doing, so i guess the drawbacks aren't scaled down as model sizes increase).

I'm not sure about 1x, 2x or 10x increase as these metrics are about code written but that isn't a productivity metric (something pretty much every half-decent programmer would agree with before LLMs - remember stories about Bill Gates saying that more LoCs being good for software is like more weight is good for airplanes or Bill Atkinson's story about adding -2000 LoCs to improve QuickDraw?).

But they can certainly help "get you going" faster in that if you're stuck on something (for whatever reason - including "that feels too much drudgery") or have issues starting something, you can have an LLM take a stab at it and it'll produce "something". Sometimes it is enough by itself, but more often than not it'll need tweaks (either directly or having the LLM do it). I got to make a bunch of things i couldn't convince myself to do - e.g. an image viewer that doesn't suck (based on my arbitrary judgement), a game database, a script to convert a git repository into static html pages that kinda look like GitHub, etc.

They can also help find (and sometimes fix) bugs or other "code smell" issues. They're not that great for exact results (without tool calling -and knowledge on how to use them effectively- at least) but when it comes to fuzzy / vague stuff like "check out this code <code dump here> can you spot any issues?" they always tend to find some stuff (even if it is hallucinations :-P but sometimes they find actual issues too or whatever hallucination they come up with reveals some actual issues with the code that you didn't spot by yourself). I've been dabbling with Rust recently and asked Qwen 3.6 35B-A3B to judge my code and it wrote "6.5/10, will compile but looks like C in Rust" :-P.

The article says:

> Never write READMEs, docstrings, or comments. I will write those myself later. And yes, I really mean this.

And sure, LLMs aren't that great about those (i do let them leave whatever comments they want though and remove them later myself - i think those comments help during the generation/prediction - basically how they "think", kinda like the reasoning phase), but they can be very good at things like "here is the code, here is the documentation for it, spot discrepancies" (i had Devstral Small 2 do this and it hallucinated a few discrepancies but also found real stuff i missed in the docs).

I've tried to use Devstral Small 2 for some API docs but found it'd sometimes make assumptions about what function do or how. One approach that might work, but i haven't tried yet, is to write the "guide" myself, then have the LLM write the function docs using both the guide and the code as reference. The reason i think this will work is because it got things 95% correct just having access to the function code alone (and without the rest of the codebase), so the "guide" would help it reach 99%. I do not expect it to get to 100% so a manual edit pass will need to be done anyway (and i have an idea for a tool to assist in the manual edit pass for that - a tool that i'll probably get an LLM to write - BTW good luck coming up if all that stuff would increase or decrease productivity for an actual product and not some random stuff i'm toying with :-P).

One other thing i've also found LLMs useful recently is to have them use the stuff you make and see how they try to use it. I have an old project, a GUI toolkit i've been hacking on every now and then since 2011 or so, though it was never a priority. Yesterday i decided to dump all the header files to Qwen 35B-A3B (i use 4bit quantization that gives me a 256k context - it isn't particularly smart but it is neat to not have to micromanage context size like i have to do with Devstral Small 2 or Qwen 27B where both of them aren't very usable speedwise at anything above 32k context sizes).

Then i asked it to just make a few small programs and it did[0] (the shot shows a paint app, a calendar, a todo list and a unit converter). Pretty much every program found bugs in the library :-P and gave me ideas on how to improve things.

In general i get the impression that LLMs aren't great at architecting things but if you do the architecture yourself and write them a framework to use, they should do a fine job at it.

[0] http://runtimeterror.com/pages/iv/images/d50c436990db203a07f...

More than 2x for me. The models are accelerators, but the biggest gains came from building workflows around them. I spend most of my time documenting what I want, while implementation and much of the surrounding work become relatively cheap. I've found that AI amplifies engineering experience more than it replaces it. The better you are at framing problems and evaluating solutions, the more productive it becomes.

0.25x because I scope bigger now and then I may get a few "free" React components but largely doing all that work of putting it together.

Wouldn't trade it though. Feel like I can overall do more with less time and energy.

At the end of the day, AI is making me work more (good thing). If you count that as productivity, then sure.

> AI is making me work more (good thing)

How is that good?

It makes me feel fulfilled. It gives me things I want. Building is in my nature.

The work I'm doing is for myself (I'm not GP). But work isn't just "my job for the man." You can build things for your home, your shop, your family, your business if you have one.

Then call it your hobby, right?

I wrote a whole thing and realized the answer is a lot more philosophical.

I don't know why working more is good.

Maybe its dormant Protestant work ethic still finding a use case in my secular mind

You work more for your employer? At no additional cost to them?

Does a 60x speedup count? Are you still copying and pasting from chatgpt, because if so your definitely doing it wrong.

Wait, how about an infinite speedup? A friend of mine couldn't code, and now he can. So he is infinity-x better at coding thanks to LLMS!

Take that, you mere 10x-ers! Your days are toast!

Left behind, as the infinite-x rockstar developer roars off on their motorbike, trailing a plume of dust behind

If 60x was a thing, we would see many fantastic products appearing on the market, every product would have improved or goten major new interesting features. Yet it hasn't really happened.

10x or 60x of very little is still not much. "Working programmers" unable to write FizzBuzz exist.

10x or 60x of a negative value is a 10-60x bigger problem than that initial negative value. We all regularly commit things that are actually net negative.

Finally, 10x or 60x more productivity channeled into yet another static site generator, CMS, or programming language doesn't really help in improving products or developing new interesting features.

In short: LLMs offer real productivity gains, but they are still tools, and they are very new, so we misuse them a lot, in various ways (not limited to examples above). Unless they replace programmers and can consistently produce better than average code (even if they fail catastrophically in a few cases), they won't fundamentally change programming. They did change a lot, but not enough to get the result you're describing.

I believe I have seen 60x improvements on certain classes of bugs. Just not on development as a whole

And I've also seen the other side, where vibe coding on projects that had bad code quality to start with leads to bug riddled messes. Projects with beautiful and elaborate test harnesses, that break seconds after contact with a real human user

[flagged]

Skimmed, it's using the word llm a lot, no instances of the word Agent, codex, claude. The person writing this isn't using LLMs correctly.

Reply on news.ycombinator.com