Usage does not correlate with valued output or ROI. This article smells like a bad attempt to say hey guys our customers see ROI without any actual evidence that they do.
Also, somewhat amusingly, the “founder” consistently being at the top of the use curve might just have something to do with everyone else also using it, but that more implies people are using it because the chief at the top wants them to use it… not because it’s actually useful. A pattern that’s typical of bad AI deployments.
> Usage does not correlate with valued output or ROI
This sounds possible but how can we know? It could just as well correlate. Effort of all kinds correlates with success even when it's not obvious or not... linear.
Frankly, it seems obvious to me that AI has negative efficiency (0.25x as efficient as the old way? 0.125x as efficient as the old way? 0.0625x as efficient as the old way?)
If you make an effort to use the new tools effectively, the gains are wild. Don't fall into a grumpy luddite trap, the train is leaving the station and you'll struggle to catch up if you don't learn and grow.
Not using AI for software development in 2027 will be like only knowing how to program via punchcard in 2010.
It's new. It's different. It's hard. It's your job. Learn how to use it effectively, embrace the new abundance mindset.
This looks like measuring what is easy to do, rather than what really matters.
PR open counts, issues created , ceos/founders spending more time on linear don't automatically lead to better outcomes (in my experience they are often negatively correlated:-) )
This looks like measuring what is easy to do, rather than what really matters.
Even that's hard. There aren't enough signals to attribute changes directly to AI, so these apps seem to correlate the signal that the user was interacting with AI to the changes they made e.g "Bob used AI at that time, and they opened a PR at a similar time, so Bob probably used AI to make that PR."
Until the tooling for gathering data on AI usage improves the data will be fairly interesting because correlations often point to something related, but won't be a source of truth.
Yeah all of these along with raw token usage are metrics that were being used around December to February by people who just didn't have anything to go off yet.
Skill/Hook usage rates, budget spend, auto-approval rate, focus area heatmaps, MTTR, MTTD are all there now
I'd be interested to overlay, I don't know, customer satisfaction or anything that can show the follow-on effect of all this output. Linear won't have that information.
My guess is some will jump up (where the team has managed to make themselves move effective and responsive) and many will plummet (doesn't need explaining).
I'd be interested to overlay, I don't know, customer satisfaction or anything that can show the follow-on effect of all this output. Linear won't have that information.
Very few businesses can accurately attribute customer value to the work they do, especially once they're passed start-up scale. A mature company makes lots of small changes and they're rarely measurable.
This looks like a "We are so AI native and efficient!!" article. At least, they could delve deeper into how they define the metrics and how they collected the data.
this seems inappropriate. I think its a bad paradigm that just because you use a platform's service, they get intimate details about your usage. And for them to be so bold about publishing the statistics they've stolen from their customers data? Gives me a reason to never recommend my org use this platform.
I think their point is that there's business value in the usage data and linear is using that value in a way that benefits them but not the customers they got it from.
It reminds me of matt levine's reframing of insider trading where it's not about fairness it's about theft. You're supposed to get secret insights and use them to get an edge. What you can't do is get an edge for yourself with secret insights that your employer got.
So it roughly comes down to "that data is valuable and rightfully belongs to the originating company." Which then makes this a contract diligence type situation.
My company uses Linear. The data presented in this blog post is worthless to me and the company. It can be crudely summarized as “agents and agentic development processes are conducting more Linear operations.” Duh!
I’d rather it was published for free in a blog post than sold without my knowledge. As a linear user. This shit doesn’t matter, everyone in the industry know these kinds of stats are being tracked, I’m happy for a company to be transparent about it.
Why would a company want to leak its processes and workflows to another company in a capitalist system. Seems sloppy and a short sighted transfer of wealth to external stockowners.
Oh hey! My industry. Guess what? ADP doesnt just yoink your data. ADP conducts voluntary surveys on the scale of hundreds of thousands. Voluntary. ADP also pays for it for the most part. Do you think Linear conducted..voluntary surveys here?
This is pretty interesting but I wish it would have been refined in two ways:
- The prose before the data appears to be AI generated. Not a big deal but it makes the reader work harder to figure out what's actually being said.
- Linear didn't control for their platforms AI changes over the past year. The platform has become much more AI integrated, so a lot of numbers will move. I'm not sure how you do it but this is only useful signal with a control.
I program as a hobby, personal projects because I can.
I recently set up a local llm to see what the fuss is about and other than the few ringer solutions my experience is as you described. 2min promping, 5min waiting, 3hrs debugging or just doing it myself.
I am very likely doing it wrong, and it does speed up some aspects, but I wouldn't say I trust llm code any more than my own. Until it runs and throws an error, the llm is 100% confident that it has written perfect code.
Models that run on (average) consumer hardware are not even close to comparable to models like Fable or Sol. Its like comparing an ant to the largest dinosaur.
A normal agentic loop will have the agent using a type system and basic tests to do some basic validation of changes. A good agentic loop would give the agent a very easy way to verify if it’s on the right track. I think agents are better than many humans at writing error free code (runtime errors, not bugs. The code could still be buggy or incorrect.)
Fundamentally though, an agent cannot produce great code, because great code requires intent, which is the opposite of the statistical mean.
You will get a solution that works with a proper workflow, but you won't get one that scales or would be truly maintainable.
Which is also what you get with random midwit drive-by contributors, but faster. I'll give it that.
Agreed on the point of intent, and with agents in the loop, it often drifts from the original vision. Epiq allows you to time travel the board, and then track intent, and vision drift. You get to replay the evolution of the board over time and correlate it with the corresponding commits and stats. In practice, it's been a useful way to spot vision drift and understand why/when/how a feature changed direction.
If can put properly engineered intent in the prompt that is verifiable, it works wonders. Anything that can defaults to the llm doing its way, you're right, it just can't converge to good, not with proper constraints.
If you built a task management system, you'd have very different code bases depending on whether it's for internal use at a mid-size development org or as a SaaS.
So I wonder whether, in your experience, the results you've seen, could have improved by providing sufficient context? - or what context was given.
I.e. if you have the agent that same context, as one of your colleagues would have/require to solve a problem.
You've missed my point. I didn't dismiss agents. I did dismiss the industry.
I don't need to add more context to a statement that operates on a layer above where context injection would influence it. It is a conceptual impossibility. Not a technical roadblock.
Why not just write the code yourself? To me it seems like methodically constructing the pull request by hand is probably faster than lazily prompting and re-prompting.
What I've observed is that by prompting for longer I get to keep my brain focused on the architectural ideas (networking, protocols, data structures, etc) rather than worrying about the most performant/elegant implementations.
I always enjoyed writing code and I am very good at writing very performant and elegant code, but it would sometimes come at the detriment of focusing on the code and not the design.
Absolutely, have warned my juniors of this. Doesn't stop malicious compliance though, I can just prompt and burn tokens with source material from my assigned tickets for no reason perfectly fine.
I look at my colleagues' screens and they're prompting shit like 'restart this program' and 'is [service] running correctly'. I have below average prompt frequency because I know crazy shit like #!/bin/bash and ps aux. It's so goddamn insane.
And to top it all off, my token count is above the average, it's just the prompt count that is low. Got questioned about it earlier this week.
I tried hard to tokenmaxx one month and couldn’t do it as for most tasks, I’ve already figured out the solution while reading the ticket and vim/emacs-fu is more enjoyable than prompting and waiting. And for other tasks, I just knew the technology to do something with the least amount of effort.
It used to be that one person had one to three codebases they knew intensely at my company. If you needed a bug in codebase X fixed, person Y was the one to do it and if they aren't available, person Z can do it, just not as quickly.
Now every person on my team has to handle tickets for every single codebase. There are about two dozen different large codebases involved here.
It's a ludicrous antipattern because person Y still needs to review the PR that person A generated for codebase X, and it will take them about as much time to wrangle the 2000 line PR (oh boy do LLMs love their mocks for unit tests) as it would have been for them to do the 50 line code change.
On top of that, it has "allowed" us to add feature after feature onto codebases not designed for them without refactoring. Is it good that this Flask API went from a purpose built service that interacted with the data analytics for product A stored in database X, and now our sales guys can sell product B, C, and D stored in database X? Uh, I'm sure it's great for them. Oh, and now it all can be stored in database X, Y, or Z depending on what the customer wants or what sales promised them. Great. Now I'm looking at a Flask app.py that's 12,000 lines of repeated code.
It has allowed poor designs to still produce working code. For a while. We seem to be getting a lot of bugs lately that look really bad to customers because it's for really simple shit. And I can't help but notice that happening to all the various products and sites I use too...
> It's a ludicrous antipattern because person Y still needs to review the PR that person A generated for codebase X, and it will take them about as much time to wrangle the 2000 line PR (oh boy do LLMs love their mocks for unit tests) as it would have been for them to do the 50 line code change.
Do you have any users? Is the software important? I'd find a new job if I were you
> On top of that, it has "allowed" us to add feature after feature onto codebases not designed for them without refactoring. Is it good that this Flask API went from a purpose built service that interacted with the data analytics for product A stored in database X, and now our sales guys can sell product B, C, and D stored in database X? Uh, I'm sure it's great for them. Oh, and now it all can be stored in database X, Y, or Z depending on what the customer wants or what sales promised them. Great. Now I'm looking at a Flask app.py that's 12,000 lines of repeated code.
Companies are tracking token usage across the board. You are in trouble for too less or too many. Unfortunately the token usage is the only measurable thing for most folks so everyone is playing the game, otherwise how else would Anthropic and OpenAI make the money
Writing code directly takes longer to warm up. Usually, I'd keep tens of thousands of lines in my head. In the past, I spent a lot of time designing error propagation and execution contexts. (Talented people might figure it out right away, but unfortunately I don't have that kind of talent.) So I'd have to think about things like Result<T> and how far to propagate errors—and worry about whether my approach would conflict with the existing codebase.
But these days, AI just generates code following the existing patterns of the codebase. In the past, staring at a blank screen meant going through a checklist of things to design—starting from policies and writing everything down step by step. Now, I just ask AI and it gives me a template—which is great. Then if the AI makes a mistake, I fix it manually.
Of course, I still hand-code sometimes—but only in the areas I enjoy. Most of the time, I use AI coding. Both are fun, and they complement each other in interesting ways. Doing both together is actually enjoyable.
I have never found AI code generators to be capable of generating code following existing patterns (and really moreso have been disappointed about AI code generators' capabilities in this regard compared to Emacs/LSP tooling).
I could see it being true in a very regimented design, but I'm still skeptical, I'd like to see details of that design, and I question whether very strict adherence to design patterns limits the ability of the LLM to generate useful features.
For me, coding is either a flow state (when I have huge swaths of code to write) or deeply thinking about just a few lines (mostly to figure if I caught everything). The latter case is when I procrastinate the most, because I usually have an idea of a solution, but don’t want to do the work of actually verifying if it’s correct. Using AI then is skipping the enjoyable part (writing the code) to the unpleasant part (formulating hypothesis and verifying them).
I feel similarly, but at the same time, I think I am the exact opposite. I actually find formulating hypotheses more fun.
For hobby projects or things I start casually, I usually do not think about errors and such at all. When it is a tool I want to build or need for myself, I really do not care about that part.
In my case, I do not contribute to open source at all. Mostly, I deliver code for factory systems or specific companies, and usually, there are strict enterprise requirements. (To be precise, there is always that mandatory code the lead developer on their end dictates, right?) That kind of code is mostly no fun, but it has to meet their requirements and often clashes with my own style. Having AI write that code for me is a huge relief.
In that sense, I think it is just a difference in personality and preferences. I originally became a programmer because I wanted to make games. I started programming because I found it fascinating to see things drawn and displayed on the screen. Becoming a programmer was all because making Flash games was so much fun... So in that regard, for me, writing code is just 'drawing what I want on the screen', which is why I guess I do not mind if the code is written by AI.
When I contribute to other people's projects, I do not use AI for anything other than English translation, but for my own projects, I have no hesitation.
Is this really just a difference in inclination? It is not that I did not enjoy writing code, but rather that seeing what I want rendered on the screen brings me more joy.
When the concept of 'vibe coding' first came out, I really hated it (since my knowledge was earned over 4 to 5 years of getting scolded by lead developers as a subcontractor and factory software provider). But thinking about it, what I really wanted to do as a developer was just to build the worlds I envisioned, so I decided not to let it bother me too much.
We talk often here on HN, and I really enjoy debating with you. I learn a lot from you.Mr."skydhash", I actually remember you quite often, and I even steal a few keywords from your posts sometimes. Because we have different tendencies, we occasionally clash, but having these conversations is exactly what makes it enjoyable.
Thank you for always replying. Have a great day, and I hope this does not offend you in any way.
> Is this really just a difference in inclination? It is not that I did not enjoy writing code, but rather that seeing what I want rendered on the screen brings me more joy.
I've been tinkering with things since a very young age, started learning about computers in middle school and really started with programming in college (I had the basics since high school, but I was interested more in 3D modeling). So writing code is more like tinkering for me and I don't really particularly care about the result other than making it happen (correctly). Once it's done, it's no longer a subject of intellectual interest.
So the joy of creating a program is in the creation itself. Once it's done I merely use it (or maintain it if it's part of the work).
I don't condemn AI use, even when doing vibe coding. My main issue is with the hypers stating that it's ok to lower a codebase quality or encouraging recklessness (and the dubious anecdotes) in a collaborative settings. If you can ensure quality and collaborate easily with your colleagues, go ahead. If you can't, then you shouldn't send PRs around.
> Thank you for always replying. Have a great day, and I hope this does not offend you in any way.
Have a great day too. I always appreciate the different point of views on a subject. It's a big world and everyone has their own perspectives.
>If you can't, then you shouldn't send PRs around.
I think this is exactly why our differences emerge.
I rarely collaborate with colleagues. In contract delivery work, that is simply how things operate. Usually, after the architecture is divided into modules, I take on the role of implementing one entire area from start to finish. Because of this, I actually have almost no experience with direct code level collaboration.
While multiple developers typically share a single code base and constantly exchange PRs, I take full responsibility for the internal implementation within the designed I/O interfaces, which seems to be where our divergence stems from.
When the modules are finally integrated, it only becomes a matter of accountability. In that sense, aside from my own website, I might not actually be doing any sustainable development. To be honest, as you know if you try AI vibe coding, the AI's abstraction and my abstraction are different. Because I am not used to its structure, it is not easy for me to manually fix the code generated by AI. Even if I do fix it, I mostly just tweak the surface level. In that regard, I completely agree that there are valid concerns regarding long term maintenance. However, since meeting strict deadlines and ensuring the required behavior are more important to me than long term maintainability, I tend to be more lenient toward AI generation.
It seems we reached different conclusions because we operate in completely different domains. It is always fascinating to see how perspectives differ depending on the field when having these conversations. Have a nice day.
> I spend 45 minutes writing it. then i git commit and move on cuz i made something good.
I cannot express strongly enough how much I'd rather work with you than with someone writing a 1000-word essay on how AI code generation is actually good enough this time.
Well you got me. I’d agree about that workflow, but something more rapid like a few seconds generating code and then a couple minutes reading it can be more engaging to me than writing it out yourself. I think it’s near time we all stop having such strong opinions about the matter either way personally.
Yeah I kinda agree. The AI boosters have been super annoying for a while.
I’m not anywhere close to an AI booster but I find value in it. I think we should think of it less like some intelligent being or “agent” and more as a code generation tool. It would both be more productive and healthier.
> I think we should think of it less like some intelligent being or “agent” and more as a code generation tool.
I think the jury is still out on whether AI code generation is worthwhile at all. Maybe it's a done deal for management because it allows them to spy on their workers, but I think it's less efficient and significantly more dull than writing code by hand.
It’s worthwhile to me under specific conditions, mainly that the codebase has existing patterns and abstractions that the LLM utilizes, and they’re actually good. It’s not a common case but I’m lucky that one of the codebases I’m in is one I built myself mostly by hand, and LLMs are really effective in it at the moment. My main concern is making sure the generated code maintains the high standard, but it’s definitely saving me time.
I’ve kind of changed my mind on prompting, it’s definitely a skill. It’s a skill based on your own skills in the domain. I’m at the point where I can get the LLM to generate the same code (roughly speaking) that I would have written. So it’s basically generating the same thing I would write, just faster. So it’s like reading your own code. Using it as a crutch to do things you aren’t capable of is where people run into trouble. That’s where the massive amounts of code review come into play. For me, I’m only ever reviewing 100-300 loc changes at a time. Often less. Because I know what I’m doing and can break things down into manageable diffs.
Can’t see myself going back, but also can’t see myself doing it without the experience I have without LLMs. Which is a bit of an issue for new developers. Not sure what the solution is for that.
I’ve reached a similar conclusion, but there’s a part that worries me: the expertise that allows us to judge AI’s output was itself built by doing the work we’re now delegating. So there’s a risk that our judgement will decay over time. I’ve been thinking about the problem as choosing where we can afford to “borrow” comprehension, versus where we need to keep exercising it, and how to “claim back” the critical comprehension we lost.
A possible heuristic: Everywhere the LLM made a design decision is probably a place to revisit later. You can ship it now, but to stay in the game you need to stay in the loop.
If you need the LLM to explain to you the concept behind something - or even if you figure out yourself what it did - that's likely wrong and will lead to the undesirable outcomes you've mentioned.
As long as it just types and doesn't think otoh, you should be good. There is - I'd argue - little to no value in manually typing if (foo) {}. And it's also great for bouncing ideas around.
That plus if you’ve ever stared at your code and then searched StackOverflow to see if you could find a better way of doing it, it’s like having that running continuously.
> Yeah it’s great for prompting for improvements, granted you can actually judge the suggestions for yourself. They’re often slightly or majorly bad.
I agree with this comment.
That being said, the major errors in some of the suggestions make me terrified of the people who say things like "AI is great for researching and strategizing".
I don't think these people are on the whole engaged or knowledgeable enough to understand the dead ends that AI research is leading them to or the concepts that it doesn't (or can't) discuss.
Really I just feel like AI has turned subpar engineers into dangerous, supercharged Dunning-Kruger cases. Everyone backing AI code generation online acts like a "10x developer", and I think the point of the AI coding tools are to make you feel that way.
I think most of you would be better off slowing down and writing the code by hand in a focused, methodical manner.
"Pull requests are up 111% in two years".
Would be more honest to say that the number of pull requests "detected" by linear are up XXX%.
Because it only works if you setup git repo tracking and use it properly. And at that point it is not obvious if more teams are using linear and using it correctly, or if the number of PR really increased that much!
Doubling PRs isn't necessarily unbelievable. AI reduces ceremony around PRs, makes the small shitty changes that you never had time to do possible and the larger cleanups at least doable.
I might build a Chrome extension that looks for “AI” in HN submission titles and shows a fake comment at the start saying “AI is all slop garbage that’s total slop and I hate it. AI sucks and it’s slop!”
It will save people time reading the 5-10 other identical vacuous comments
Rebuttal to what? I’m complaining that it’s a totally pointless comment that could be used on any AI related post. The comment just says “AI is slop garbage that sucks”.
What am I going to say to that? Obviously it can’t suck that badly or there wouldn’t be a large majority of programmers using it, plenty of famous programmers and developers of languages praising it, millions of people paying for $200 plans. It doesn’t need to be Carmack-level to get a massive amount of use out of it when you can just say “implement this chart here, give me some ideas on speeding up this algorithm, write a function that converts this heightmap to a .fbx” and it writes it.
> Time spent on customer requests, docs, and projects held steady [..] AI has so far changed how teams execute far more than how they decide what to build
I think the measurement for this may be flawed. We do mostly use AI to decide how to build. But what we build is influenced by AI-driven research into a problem or task. That's largely done in coding and desktop AI tools, not Linear Asks/AI.
I'm working on accelerating my team's work by implementing AI-driven code pipelines with guardrails to eliminate as much unnecessary review time as possible. Also making a chatbot for turning repetitive tasks & PRs into buttons, and an "architectural guidance" chatbot that gives advice tailored to our business, software/system architecture, cloud, standards, etc. This puts AI and automated jobs in the center of both how (automated task) and what (architecture guidance).
But this has a not-so-great implication for Linear. With my tools, a human never has to touch a ticket, so we could use any ticketing system with an API or CLI. Linear is a great product because they made a great interface. What happens when I replace their interface with a chat bot?
Usage does not correlate with valued output or ROI. This article smells like a bad attempt to say hey guys our customers see ROI without any actual evidence that they do.
Also, somewhat amusingly, the “founder” consistently being at the top of the use curve might just have something to do with everyone else also using it, but that more implies people are using it because the chief at the top wants them to use it… not because it’s actually useful. A pattern that’s typical of bad AI deployments.
> Usage does not correlate with valued output or ROI
This sounds possible but how can we know? It could just as well correlate. Effort of all kinds correlates with success even when it's not obvious or not... linear.
Frankly, it seems obvious to me that AI has negative efficiency (0.25x as efficient as the old way? 0.125x as efficient as the old way? 0.0625x as efficient as the old way?)
AI has created a generation of 0.125x engineers.
Skill issue.
If you make an effort to use the new tools effectively, the gains are wild. Don't fall into a grumpy luddite trap, the train is leaving the station and you'll struggle to catch up if you don't learn and grow.
Not using AI for software development in 2027 will be like only knowing how to program via punchcard in 2010.
It's new. It's different. It's hard. It's your job. Learn how to use it effectively, embrace the new abundance mindset.
I hope you don't have children.
This looks like measuring what is easy to do, rather than what really matters.
PR open counts, issues created , ceos/founders spending more time on linear don't automatically lead to better outcomes (in my experience they are often negatively correlated:-) )
This looks like measuring what is easy to do, rather than what really matters.
Even that's hard. There aren't enough signals to attribute changes directly to AI, so these apps seem to correlate the signal that the user was interacting with AI to the changes they made e.g "Bob used AI at that time, and they opened a PR at a similar time, so Bob probably used AI to make that PR."
Until the tooling for gathering data on AI usage improves the data will be fairly interesting because correlations often point to something related, but won't be a source of truth.
Yeah all of these along with raw token usage are metrics that were being used around December to February by people who just didn't have anything to go off yet.
Skill/Hook usage rates, budget spend, auto-approval rate, focus area heatmaps, MTTR, MTTD are all there now
I'd be interested to overlay, I don't know, customer satisfaction or anything that can show the follow-on effect of all this output. Linear won't have that information.
My guess is some will jump up (where the team has managed to make themselves move effective and responsive) and many will plummet (doesn't need explaining).
Then there might be something to look at.
I'd be interested to overlay, I don't know, customer satisfaction or anything that can show the follow-on effect of all this output. Linear won't have that information.
Very few businesses can accurately attribute customer value to the work they do, especially once they're passed start-up scale. A mature company makes lots of small changes and they're rarely measurable.
This looks like a "We are so AI native and efficient!!" article. At least, they could delve deeper into how they define the metrics and how they collected the data.
this seems inappropriate. I think its a bad paradigm that just because you use a platform's service, they get intimate details about your usage. And for them to be so bold about publishing the statistics they've stolen from their customers data? Gives me a reason to never recommend my org use this platform.
The data is aggregated, and you cannot possibly identify a single user from what's been published. I see no issue here.
I think their point is that there's business value in the usage data and linear is using that value in a way that benefits them but not the customers they got it from.
It reminds me of matt levine's reframing of insider trading where it's not about fairness it's about theft. You're supposed to get secret insights and use them to get an edge. What you can't do is get an edge for yourself with secret insights that your employer got.
So it roughly comes down to "that data is valuable and rightfully belongs to the originating company." Which then makes this a contract diligence type situation.
My company uses Linear. The data presented in this blog post is worthless to me and the company. It can be crudely summarized as “agents and agentic development processes are conducting more Linear operations.” Duh!
lol isn't this what google said and yet..the NSA cometh
I’d rather it was published for free in a blog post than sold without my knowledge. As a linear user. This shit doesn’t matter, everyone in the industry know these kinds of stats are being tracked, I’m happy for a company to be transparent about it.
Why would a company want to leak its processes and workflows to another company in a capitalist system. Seems sloppy and a short sighted transfer of wealth to external stockowners.
Uhh that's how a lot of economic data works too. Guess how we get a lot of jobs data... ADP
Oh hey! My industry. Guess what? ADP doesnt just yoink your data. ADP conducts voluntary surveys on the scale of hundreds of thousands. Voluntary. ADP also pays for it for the most part. Do you think Linear conducted..voluntary surveys here?
This is pretty interesting but I wish it would have been refined in two ways:
- The prose before the data appears to be AI generated. Not a big deal but it makes the reader work harder to figure out what's actually being said.
- Linear didn't control for their platforms AI changes over the past year. The platform has become much more AI integrated, so a lot of numbers will move. I'm not sure how you do it but this is only useful signal with a control.
These days, my work has become: generate code for 20 minutes, then spend an hour reading it.
And then more hours cleaning it up and re-prompting.
I program as a hobby, personal projects because I can.
I recently set up a local llm to see what the fuss is about and other than the few ringer solutions my experience is as you described. 2min promping, 5min waiting, 3hrs debugging or just doing it myself.
I am very likely doing it wrong, and it does speed up some aspects, but I wouldn't say I trust llm code any more than my own. Until it runs and throws an error, the llm is 100% confident that it has written perfect code.
Models that run on (average) consumer hardware are not even close to comparable to models like Fable or Sol. Its like comparing an ant to the largest dinosaur.
You don’t even have to go that far, recent Opus models are quite impressive.
A normal agentic loop will have the agent using a type system and basic tests to do some basic validation of changes. A good agentic loop would give the agent a very easy way to verify if it’s on the right track. I think agents are better than many humans at writing error free code (runtime errors, not bugs. The code could still be buggy or incorrect.)
Fundamentally though, an agent cannot produce great code, because great code requires intent, which is the opposite of the statistical mean.
You will get a solution that works with a proper workflow, but you won't get one that scales or would be truly maintainable. Which is also what you get with random midwit drive-by contributors, but faster. I'll give it that.
Agreed on the point of intent, and with agents in the loop, it often drifts from the original vision. Epiq allows you to time travel the board, and then track intent, and vision drift. You get to replay the evolution of the board over time and correlate it with the corresponding commits and stats. In practice, it's been a useful way to spot vision drift and understand why/when/how a feature changed direction.
https://ljtn.github.io/epiq/
> because great code requires intent
If can put properly engineered intent in the prompt that is verifiable, it works wonders. Anything that can defaults to the llm doing its way, you're right, it just can't converge to good, not with proper constraints.
If you built a task management system, you'd have very different code bases depending on whether it's for internal use at a mid-size development org or as a SaaS.
So I wonder whether, in your experience, the results you've seen, could have improved by providing sufficient context? - or what context was given.
I.e. if you have the agent that same context, as one of your colleagues would have/require to solve a problem.
"You're holding it wrong"
You've missed my point. I didn't dismiss agents. I did dismiss the industry.
I don't need to add more context to a statement that operates on a layer above where context injection would influence it. It is a conceptual impossibility. Not a technical roadblock.
Try using Fable and report back. Local LLM is to Fable as Little Tike car is to a Porsche.
I don't think setting up a local LLM is a reasonable way to get a good idea of how enterprises are using this stuff.
Local llms aren’t super exciting unfortunately.
Why not just write the code yourself? To me it seems like methodically constructing the pull request by hand is probably faster than lazily prompting and re-prompting.
What I've observed is that by prompting for longer I get to keep my brain focused on the architectural ideas (networking, protocols, data structures, etc) rather than worrying about the most performant/elegant implementations.
I always enjoyed writing code and I am very good at writing very performant and elegant code, but it would sometimes come at the detriment of focusing on the code and not the design.
because my manager will ping me and say "anon you aren't prompting enough" like they never heard of Goodhart's law before.
Write a script to make random prompts and use tokens. Not like they look at what you’re actually prompting.
No, don’t be silly, they get ai to spy on you en masse instead. Nobody has to look at anything anymore for it to be actionable.
Absolutely, have warned my juniors of this. Doesn't stop malicious compliance though, I can just prompt and burn tokens with source material from my assigned tickets for no reason perfectly fine.
I look at my colleagues' screens and they're prompting shit like 'restart this program' and 'is [service] running correctly'. I have below average prompt frequency because I know crazy shit like #!/bin/bash and ps aux. It's so goddamn insane.
And to top it all off, my token count is above the average, it's just the prompt count that is low. Got questioned about it earlier this week.
I tried hard to tokenmaxx one month and couldn’t do it as for most tasks, I’ve already figured out the solution while reading the ticket and vim/emacs-fu is more enjoyable than prompting and waiting. And for other tasks, I just knew the technology to do something with the least amount of effort.
> I have below average prompt frequency because I know crazy shit like #!/bin/bash and ps aux. It's so goddamn insane.
I think AI made everyone so dumb that I'm a 10x engineer now.
In my experience?
It used to be that one person had one to three codebases they knew intensely at my company. If you needed a bug in codebase X fixed, person Y was the one to do it and if they aren't available, person Z can do it, just not as quickly.
Now every person on my team has to handle tickets for every single codebase. There are about two dozen different large codebases involved here.
It's a ludicrous antipattern because person Y still needs to review the PR that person A generated for codebase X, and it will take them about as much time to wrangle the 2000 line PR (oh boy do LLMs love their mocks for unit tests) as it would have been for them to do the 50 line code change.
On top of that, it has "allowed" us to add feature after feature onto codebases not designed for them without refactoring. Is it good that this Flask API went from a purpose built service that interacted with the data analytics for product A stored in database X, and now our sales guys can sell product B, C, and D stored in database X? Uh, I'm sure it's great for them. Oh, and now it all can be stored in database X, Y, or Z depending on what the customer wants or what sales promised them. Great. Now I'm looking at a Flask app.py that's 12,000 lines of repeated code.
It has allowed poor designs to still produce working code. For a while. We seem to be getting a lot of bugs lately that look really bad to customers because it's for really simple shit. And I can't help but notice that happening to all the various products and sites I use too...
> It's a ludicrous antipattern because person Y still needs to review the PR that person A generated for codebase X, and it will take them about as much time to wrangle the 2000 line PR (oh boy do LLMs love their mocks for unit tests) as it would have been for them to do the 50 line code change.
Do you have any users? Is the software important? I'd find a new job if I were you
> On top of that, it has "allowed" us to add feature after feature onto codebases not designed for them without refactoring. Is it good that this Flask API went from a purpose built service that interacted with the data analytics for product A stored in database X, and now our sales guys can sell product B, C, and D stored in database X? Uh, I'm sure it's great for them. Oh, and now it all can be stored in database X, Y, or Z depending on what the customer wants or what sales promised them. Great. Now I'm looking at a Flask app.py that's 12,000 lines of repeated code.
This is terrifying to me. I'm sorry.
Companies are tracking token usage across the board. You are in trouble for too less or too many. Unfortunately the token usage is the only measurable thing for most folks so everyone is playing the game, otherwise how else would Anthropic and OpenAI make the money
Writing code directly takes longer to warm up. Usually, I'd keep tens of thousands of lines in my head. In the past, I spent a lot of time designing error propagation and execution contexts. (Talented people might figure it out right away, but unfortunately I don't have that kind of talent.) So I'd have to think about things like Result<T> and how far to propagate errors—and worry about whether my approach would conflict with the existing codebase.
But these days, AI just generates code following the existing patterns of the codebase. In the past, staring at a blank screen meant going through a checklist of things to design—starting from policies and writing everything down step by step. Now, I just ask AI and it gives me a template—which is great. Then if the AI makes a mistake, I fix it manually.
Of course, I still hand-code sometimes—but only in the areas I enjoy. Most of the time, I use AI coding. Both are fun, and they complement each other in interesting ways. Doing both together is actually enjoyable.
> But these days, AI just generates code following the existing patterns of the codebase
Is this sarcasm?
I have never found AI code generators to be capable of generating code following existing patterns (and really moreso have been disappointed about AI code generators' capabilities in this regard compared to Emacs/LSP tooling).
I could see it being true in a very regimented design, but I'm still skeptical, I'd like to see details of that design, and I question whether very strict adherence to design patterns limits the ability of the LLM to generate useful features.
It does so better the more "standard" the "existing patterns" are. :^)
For me, coding is either a flow state (when I have huge swaths of code to write) or deeply thinking about just a few lines (mostly to figure if I caught everything). The latter case is when I procrastinate the most, because I usually have an idea of a solution, but don’t want to do the work of actually verifying if it’s correct. Using AI then is skipping the enjoyable part (writing the code) to the unpleasant part (formulating hypothesis and verifying them).
I feel similarly, but at the same time, I think I am the exact opposite. I actually find formulating hypotheses more fun.
For hobby projects or things I start casually, I usually do not think about errors and such at all. When it is a tool I want to build or need for myself, I really do not care about that part.
In my case, I do not contribute to open source at all. Mostly, I deliver code for factory systems or specific companies, and usually, there are strict enterprise requirements. (To be precise, there is always that mandatory code the lead developer on their end dictates, right?) That kind of code is mostly no fun, but it has to meet their requirements and often clashes with my own style. Having AI write that code for me is a huge relief.
In that sense, I think it is just a difference in personality and preferences. I originally became a programmer because I wanted to make games. I started programming because I found it fascinating to see things drawn and displayed on the screen. Becoming a programmer was all because making Flash games was so much fun... So in that regard, for me, writing code is just 'drawing what I want on the screen', which is why I guess I do not mind if the code is written by AI.
When I contribute to other people's projects, I do not use AI for anything other than English translation, but for my own projects, I have no hesitation.
Is this really just a difference in inclination? It is not that I did not enjoy writing code, but rather that seeing what I want rendered on the screen brings me more joy.
When the concept of 'vibe coding' first came out, I really hated it (since my knowledge was earned over 4 to 5 years of getting scolded by lead developers as a subcontractor and factory software provider). But thinking about it, what I really wanted to do as a developer was just to build the worlds I envisioned, so I decided not to let it bother me too much.
We talk often here on HN, and I really enjoy debating with you. I learn a lot from you.Mr."skydhash", I actually remember you quite often, and I even steal a few keywords from your posts sometimes. Because we have different tendencies, we occasionally clash, but having these conversations is exactly what makes it enjoyable.
Thank you for always replying. Have a great day, and I hope this does not offend you in any way.
Although it has nothing to do with me, what a pleasure it was to read your comment.
> Is this really just a difference in inclination? It is not that I did not enjoy writing code, but rather that seeing what I want rendered on the screen brings me more joy.
I've been tinkering with things since a very young age, started learning about computers in middle school and really started with programming in college (I had the basics since high school, but I was interested more in 3D modeling). So writing code is more like tinkering for me and I don't really particularly care about the result other than making it happen (correctly). Once it's done, it's no longer a subject of intellectual interest.
So the joy of creating a program is in the creation itself. Once it's done I merely use it (or maintain it if it's part of the work).
I don't condemn AI use, even when doing vibe coding. My main issue is with the hypers stating that it's ok to lower a codebase quality or encouraging recklessness (and the dubious anecdotes) in a collaborative settings. If you can ensure quality and collaborate easily with your colleagues, go ahead. If you can't, then you shouldn't send PRs around.
> Thank you for always replying. Have a great day, and I hope this does not offend you in any way.
Have a great day too. I always appreciate the different point of views on a subject. It's a big world and everyone has their own perspectives.
>If you can't, then you shouldn't send PRs around.
I think this is exactly why our differences emerge.
I rarely collaborate with colleagues. In contract delivery work, that is simply how things operate. Usually, after the architecture is divided into modules, I take on the role of implementing one entire area from start to finish. Because of this, I actually have almost no experience with direct code level collaboration.
While multiple developers typically share a single code base and constantly exchange PRs, I take full responsibility for the internal implementation within the designed I/O interfaces, which seems to be where our divergence stems from.
When the modules are finally integrated, it only becomes a matter of accountability. In that sense, aside from my own website, I might not actually be doing any sustainable development. To be honest, as you know if you try AI vibe coding, the AI's abstraction and my abstraction are different. Because I am not used to its structure, it is not easy for me to manually fix the code generated by AI. Even if I do fix it, I mostly just tweak the surface level. In that regard, I completely agree that there are valid concerns regarding long term maintenance. However, since meeting strict deadlines and ensuring the required behavior are more important to me than long term maintainability, I tend to be more lenient toward AI generation.
It seems we reached different conclusions because we operate in completely different domains. It is always fascinating to see how perspectives differ depending on the field when having these conversations. Have a nice day.
Felt that. I already gave DeepSeek the HTML and it still said the UI was good to go, then told me it has no vision.
Like the Titan submarine team's moto mine is - real men test in production.
The LLM produce so much code that the best I can is skim and look for obvious flaws, also pass it trough adversarial one.
I spend 45 minutes writing it. then i git commit and move on cuz i made something good.
who is winning here? lol
> I spend 45 minutes writing it. then i git commit and move on cuz i made something good.
I cannot express strongly enough how much I'd rather work with you than with someone writing a 1000-word essay on how AI code generation is actually good enough this time.
That's gotta be at least 10x or 20x more efficient than the old way of doing things.
It could be. Or it could be 1x, or 0.2x. You don’t have enough information to make that judgment.
I was joking. The workflow doesn't seem particularly fast or engaging in my opinion. AI code generation doesn't seem worthwhile to me.
Well you got me. I’d agree about that workflow, but something more rapid like a few seconds generating code and then a couple minutes reading it can be more engaging to me than writing it out yourself. I think it’s near time we all stop having such strong opinions about the matter either way personally.
> I think it’s near time we all stop having such strong opinions about the matter either way personally.
I don't think this is reasonable given how abusive the pro-AI rhetoric has been for years now.
I’m entitled to my opinion about your opinion just like you’re entitled to your opinion about my opinion about your opinion.
If you don't accept my opinion as your own you're going to lose your job and become a member of a permanent underclass.
You make it sound like it’s a motivated opinion.
Yeah I kinda agree. The AI boosters have been super annoying for a while.
I’m not anywhere close to an AI booster but I find value in it. I think we should think of it less like some intelligent being or “agent” and more as a code generation tool. It would both be more productive and healthier.
> I think we should think of it less like some intelligent being or “agent” and more as a code generation tool.
I think the jury is still out on whether AI code generation is worthwhile at all. Maybe it's a done deal for management because it allows them to spy on their workers, but I think it's less efficient and significantly more dull than writing code by hand.
It’s worthwhile to me under specific conditions, mainly that the codebase has existing patterns and abstractions that the LLM utilizes, and they’re actually good. It’s not a common case but I’m lucky that one of the codebases I’m in is one I built myself mostly by hand, and LLMs are really effective in it at the moment. My main concern is making sure the generated code maintains the high standard, but it’s definitely saving me time.
I’ve kind of changed my mind on prompting, it’s definitely a skill. It’s a skill based on your own skills in the domain. I’m at the point where I can get the LLM to generate the same code (roughly speaking) that I would have written. So it’s basically generating the same thing I would write, just faster. So it’s like reading your own code. Using it as a crutch to do things you aren’t capable of is where people run into trouble. That’s where the massive amounts of code review come into play. For me, I’m only ever reviewing 100-300 loc changes at a time. Often less. Because I know what I’m doing and can break things down into manageable diffs.
Can’t see myself going back, but also can’t see myself doing it without the experience I have without LLMs. Which is a bit of an issue for new developers. Not sure what the solution is for that.
I’ve reached a similar conclusion, but there’s a part that worries me: the expertise that allows us to judge AI’s output was itself built by doing the work we’re now delegating. So there’s a risk that our judgement will decay over time. I’ve been thinking about the problem as choosing where we can afford to “borrow” comprehension, versus where we need to keep exercising it, and how to “claim back” the critical comprehension we lost.
> So there’s a risk that our judgement will decay over time.
It's happening in real time and those of us who don't use AI code generators can tell.
A possible heuristic: Everywhere the LLM made a design decision is probably a place to revisit later. You can ship it now, but to stay in the game you need to stay in the loop.
If you need the LLM to explain to you the concept behind something - or even if you figure out yourself what it did - that's likely wrong and will lead to the undesirable outcomes you've mentioned.
As long as it just types and doesn't think otoh, you should be good. There is - I'd argue - little to no value in manually typing if (foo) {}. And it's also great for bouncing ideas around.
That plus if you’ve ever stared at your code and then searched StackOverflow to see if you could find a better way of doing it, it’s like having that running continuously.
Yeah it’s great for prompting for improvements, granted you can actually judge the suggestions for yourself. They’re often slightly or majorly bad.
> Yeah it’s great for prompting for improvements, granted you can actually judge the suggestions for yourself. They’re often slightly or majorly bad.
I agree with this comment.
That being said, the major errors in some of the suggestions make me terrified of the people who say things like "AI is great for researching and strategizing".
I don't think these people are on the whole engaged or knowledgeable enough to understand the dead ends that AI research is leading them to or the concepts that it doesn't (or can't) discuss.
Really I just feel like AI has turned subpar engineers into dangerous, supercharged Dunning-Kruger cases. Everyone backing AI code generation online acts like a "10x developer", and I think the point of the AI coding tools are to make you feel that way.
I think most of you would be better off slowing down and writing the code by hand in a focused, methodical manner.
I think it's worth noting that headlines like "AI adoption has spread to every function" is only limited to roles covered/tracked by Linear.
Recent studies by e.g. Google show a much broader range of adoption.
Nice data!
Tim (author) if you're there: it'd be amazing to see the split of which agents people are using, if you have that data.
I didn't know Linear has "AI features". Linear is boring, but that's actually fine by me.
I use LLMs to write my code, but this does not show up in this data.
"Pull requests are up 111% in two years". Would be more honest to say that the number of pull requests "detected" by linear are up XXX%. Because it only works if you setup git repo tracking and use it properly. And at that point it is not obvious if more teams are using linear and using it correctly, or if the number of PR really increased that much!
Doubling PRs isn't necessarily unbelievable. AI reduces ceremony around PRs, makes the small shitty changes that you never had time to do possible and the larger cleanups at least doable.
Double the PRs doesn't mean double the output.
The AI Slopologists strike again. More garbage by garbage people.
Is this just an opinion? If so, fair.
If it's an attempt at rebutting their claims etc, it'd be easier to interact if you provided some data, or concrete observations :)
Can't tell if you mean the people the article is talking about or the article itself
¿Por qué no los dos?
I might build a Chrome extension that looks for “AI” in HN submission titles and shows a fake comment at the start saying “AI is all slop garbage that’s total slop and I hate it. AI sucks and it’s slop!”
It will save people time reading the 5-10 other identical vacuous comments
Do you have a rebuttal or are you just mad that people don't like AI?
It sucks at generating code and AI true believers have made my job hell.
Rebuttal to what? I’m complaining that it’s a totally pointless comment that could be used on any AI related post. The comment just says “AI is slop garbage that sucks”.
What am I going to say to that? Obviously it can’t suck that badly or there wouldn’t be a large majority of programmers using it, plenty of famous programmers and developers of languages praising it, millions of people paying for $200 plans. It doesn’t need to be Carmack-level to get a massive amount of use out of it when you can just say “implement this chart here, give me some ideas on speeding up this algorithm, write a function that converts this heightmap to a .fbx” and it writes it.
> Time spent on customer requests, docs, and projects held steady [..] AI has so far changed how teams execute far more than how they decide what to build
I think the measurement for this may be flawed. We do mostly use AI to decide how to build. But what we build is influenced by AI-driven research into a problem or task. That's largely done in coding and desktop AI tools, not Linear Asks/AI.
I'm working on accelerating my team's work by implementing AI-driven code pipelines with guardrails to eliminate as much unnecessary review time as possible. Also making a chatbot for turning repetitive tasks & PRs into buttons, and an "architectural guidance" chatbot that gives advice tailored to our business, software/system architecture, cloud, standards, etc. This puts AI and automated jobs in the center of both how (automated task) and what (architecture guidance).
But this has a not-so-great implication for Linear. With my tools, a human never has to touch a ticket, so we could use any ticketing system with an API or CLI. Linear is a great product because they made a great interface. What happens when I replace their interface with a chat bot?