In the past splitting diffs was an important process to speed up and simplify code review.
But now I would rather make one PR per service to simplify CI and deploy. As a result PRs become much much larger.
Exactly, big PRs are where review turns into skimming. I thought that since agents raise most of the PRs now, they should attach the full change context with reference paths. Then a review agent with full project memory can flag anything off on it, also give a short summary to reviewer as well so reviewer only has to review the summary and decide the approval.
If you're still writing migrations you're still bot in the edge and are missing out on some good gains. We got rid of out database and now just query the LLM whenever we need to fetch customer account details etc. Our delivery speed has quadrupled this last week alone
these models with passable agent performance have been around for a year now. how many months of days is that? are you saying you're getting an entire year's worth of programming done every 12 days? more than two years' worth a month? over 12 months thats more than 24 years, practically an entire career.
Do you think you've achieved that output of your entire career before ai this year?
For me at the current moment in time its more like 2 days of work turned to one day, give or take. And I'm not that good of a programmer.
I think so! From my point of view I don't think that these models have been passable for around a year now. Seems to me that opus 5.5 and fable 5.2 are really ramping things up.
Especially on the vision side of things, and being able to be steered in a direction I'm finally happy with.
Give you a few examples:
- Couple of projects at work that where shelved due to being too long, have been merged in a few weeks.
- I generated a whole icon pack in a weekend using fable 5.2 (haven't been able to make this with any other model)
- On the personal side, I build a few smart devices (PCB included and housing printing) in a few days. Something that would have taken me who knows what long!
- I have agents roaming the app findings ~20 visual quirks at work per day. I used to take a whole day to patch between 5-10.
Obviously for some projects things don't change and it's still 1:1 speed, but other 1:100 even..!
I am shipping so much more, but I don't know that our customers would necessarily know it. Total rewrites that needed to be done but were getting pushed off due to the scope are now getting done. Security issues are getting patched. Support ticket open to resolution time is dropping. So yes, shipping a lot faster to create a much better project rather than just focusing on features.
> Total rewrites that needed to be done but were getting pushed off
I am doing this now with a massive internal app. Our internal customers see this as a drop-in performance because one of the senior team members, me, is not adding new features. But we are addressing a massive technical debt in about one-fifth the time it would've taken without the use of LLMs.
As we’ve seen happening in the open source projects with the piles of bugs being reported to them so much of what’s going on doesn’t seem to be new features, but just longstanding technical debt coming to the fore.
Maybe AI can’t write that new feature for us really fast. Maybe we asked. Maybe the code was complete garbage and it made it clear we really don’t want to keep piling on top of it and need to fix the issues we’ve known about for a long time. Adding new features at the old pace was a problem. We can’t handle the new.
Between that and the fact that some kinds of tech debt are much easier to pay off now that’s where a lot of our “new capacity” is going. Shoring things up.
do those rewrites really need to be done still? what are you gaining?
In the old days we rewrote for a gain in productivity by humans or a gain in performance which resulted in non-trivial dollars. If AI is writing and testing the code, what benefits are you getting from the rewrite?
If it's actually fixing performance metrics in a way the user can feel or that hit the bottom line it's worthwhile. I'm a lot more suspicious when it is said that the user wouldn't notice them, though.
The excitement of LLMs has been around the productivity of the engineers, not the quality of the projects. Traditionally, we have bottlenecks at review and QA stages for new work, and a lot of orgs turn on the firehose of new code without fixing those bottlenecks.
If you want to drive at high speed, you need to build a highway first.
I'm shipping faster, but not as fast if the output was the same as before. AI allows for deeper code, all the nice to haves you wouldn't normally do. Test coverage is now fully complete as well, and I mean full tilt mocks and every edge case, not just the happy path plus bugs found in testing.
Not everyone has the process and positioning needed to take advantage of AI, so it’s not surprising that some people swear it doesn’t help. But those people need to understand that other people do have the process and positioning and AI is dramatically changing things.
Personally, we’re just beginning to ship faster because we’re a big ship and turn slowly. We’re still getting up to speed but we’re also bogged down with bug fixes, security fixes, and the review bottleneck. Reviews are especially tough now that every commit is much larger because it is now much more correct than the crap we were submitting before AI.
Also, many people are simply lying and hate AI. There are people I’ve encountered who look for any evidence to justify not embracing it. It’s frankly tiring but I understand some people cannot handle change and are frankly terrified of what is happening to their careers and their concept of self-worth.
I’m not sure I’ve really seen it within the community of programmers. It was certainly more there early on when the models were much worse, but I’ve seen it fading.
I have seen it among “the general public“. I’ve seen normal people complain about things in open-source projects or other contexts where to whatever degree they accepted programming took time before now they think it’s all just a magic wand with Claude, no matter the level of complexity.
I haven’t heard a single person at my company in management at any layer complain that we’re not working fast enough now that we have AI. One of my friends has 100% told me their management is pulling that. The management over there is on the everyone should do four times as many tickets, and use at least X amount of tokens per day. otherwise you’re slacking off train.
That’s not the case at my company at all. It’s available to use but they’re not forcing it or firing people who aren’t using it. It’s just a tool we can use.
If you follow various data sources like the StackOverflow developer survey or DORA from 2023 onwards, by this theory about 77% - 90% have been fired or pushed out of the industry.
What do you mean software development was arduous? It was pretty chill. You will start working on something, start crafting code carefully with time to review, optimize, etc.
Now it feels like everyone is in a rush and there is no time for crafting anything. Just vibe code a pile of shit that looks good but it's unnecessarily complex.
You are going to see how arduous software development is in a few years when no one understands the systems they have built
This is something that Brooks had reflected on in "The Mythical Man Month":
"One occasionally reads newspaper accounts of how two programmers in a remodeled garage have built an important program that surpasses the best efforts of large teams. And every programmer is prepared to believe such tales, for he knows that he could build any program much faster than the 1000 statements/year reported for industrial teams. Why then have not all industrial programming teams been replaced by dedicated garage duos? One must look at what is being produced."
Brooks argues that the coding that happened in the garage is only a fraction of what you need for the final product:
"It costs nine times as much. But it is the truly useful object, the intended product of most system programming efforts."
What we're seeing right now does not feel much different. You can greatly accelerate a first draft, a greenfield project, an experiment, or at least the initial iteration of one. Changing a live production system or a widely used library is an entirely different proposition.
People are running away with it in hilarious ways using all sorts of storyboarding, 3d environments, gameplay, etc etc.
I work in a highly regulated environment. My AI spend is ~$300/week and it has been tremendously helpful in troubleshooting incidents(!), increasing test coverage, gaining understanding of legacy/inherited code, applying fixes, etc. But shipping new features is something that impacts us in many many many ways and the uptake there has been quite slow.
That argues for immutability of production throughput, and yet most professional SWEs this days would be ashamed to deliver only 1000 statements/year of production code.
Clearly, acceleration is possible. So maybe we should move on to the conversation of what will make further acceleration possible, instead of arguing it only works for iteration one.
Over the years many people built clones of things like Twitter, Dropbox, uuhh. Loads of things really. The core of Twitter is trivial, it's basically like my first PHP tools creating shoutboxes for websites.
Then the next challenge is scaling it to the world - and Twitter themselves struggled with that, the first years were full of fail whales, or stories of servers churning for ages whenever Justin Bieber made a tweet because (for some reason) they updated everyone's feeds even if they weren't actually online / looking at them.
But solving that also wasn't what made Twitter or Facebook popular, and no LLM will ever solve that because none of the successful software projects were successful because of code, code volume, development speed, number of people involved, or anything like that.
(That said I am convinced you can do the same amount of coding with less people at good enough quality, and said quality will improve over time as we've seen in the past ~2 years. The providers just need to add the software quality guardrails that we as software engineers learned over the years, and I'm sure there's dozens of companies working on just that)
The trillions-dollars question is how come software itself doesn't get improved by (at least) an order of magnitude, if not more.
More to the point, the price of RAM that has gone berserk. Easy solution, let's then make our OSes, browsers and almost all the relevant software that we use day to day an order a magnitude more nimble, for example if a Chrome instance now eats 1-2GB of memory then make it it at most 100-200MBs of said memory while doing the exact same things (if not more). The much lauded AI revolution in code-writing hasn't created that, not sure if it will ever do, it will just add slop after slop after slop, in fact it wouldn't surprise me if the memory requirements for something like Chrome won't increase as a result of the recent vibe-coding mania.
But by this point the AI zealots are already running the asylum, so nothing to be done.
> The trillions-dollars question is how come software itself doesn't get improved by (at least) an order of magnitude, if not more.
The sad thing is those using AI to code (and I certainly do) don't realize they're producing more but just more of the same turds they were already producing.
And in addition to that they're producing turds nobody gives a shit about (ah!).
Where, indeed, are the both less memory-hungry, less CPU-intensive and more secure rewrites of the pigs we work with on a daily basis?
Nowhere to be seen.
Because the AI zealots are way too busy adding crap features over crap features.
I am writing code 5-10x faster and I can feel it go, but I'm still waiting 2 days for the IT to install something on my useless locked down laptop, still waiting a week for a trivial corporate firewall change, still waiting 2-3 weeks for business to make up their minds about whatever the hell they want to sell now. So I am shipping faster, but actually not.
Yeah, our CI/CD pipeline is still a dumpster fire. Management refuses to use AI even though it would help them more than us. Our build team hasn’t really used it to improve our build times.
101 comments
[ 0.25 ms ] story [ 30.0 ms ] threadDo you think you've achieved that output of your entire career before ai this year?
For me at the current moment in time its more like 2 days of work turned to one day, give or take. And I'm not that good of a programmer.
Especially on the vision side of things, and being able to be steered in a direction I'm finally happy with.
Give you a few examples:
- Couple of projects at work that where shelved due to being too long, have been merged in a few weeks.
- I generated a whole icon pack in a weekend using fable 5.2 (haven't been able to make this with any other model)
- On the personal side, I build a few smart devices (PCB included and housing printing) in a few days. Something that would have taken me who knows what long!
- I have agents roaming the app findings ~20 visual quirks at work per day. I used to take a whole day to patch between 5-10.
Obviously for some projects things don't change and it's still 1:1 speed, but other 1:100 even..!
I am doing this now with a massive internal app. Our internal customers see this as a drop-in performance because one of the senior team members, me, is not adding new features. But we are addressing a massive technical debt in about one-fifth the time it would've taken without the use of LLMs.
As we’ve seen happening in the open source projects with the piles of bugs being reported to them so much of what’s going on doesn’t seem to be new features, but just longstanding technical debt coming to the fore.
Maybe AI can’t write that new feature for us really fast. Maybe we asked. Maybe the code was complete garbage and it made it clear we really don’t want to keep piling on top of it and need to fix the issues we’ve known about for a long time. Adding new features at the old pace was a problem. We can’t handle the new.
Between that and the fact that some kinds of tech debt are much easier to pay off now that’s where a lot of our “new capacity” is going. Shoring things up.
In the old days we rewrote for a gain in productivity by humans or a gain in performance which resulted in non-trivial dollars. If AI is writing and testing the code, what benefits are you getting from the rewrite?
Fixing performance bottlenecks caused by bad architecture. Fixing bad architecture so that we can ship even faster in the future.
If you want to drive at high speed, you need to build a highway first.
Claude Code TUI, probably the poster child for AI-driven coding, itself is still going to sh*t.
Google Search: 1998
Gmail: 2004
Google Map: 2005
Google Docs: 2006
Chrome: 2008
vs
Claude 3 (+ tool calling, web search): 2024
Claude Code: 2025
Claude Design, Sheet, Data: 2026
"We" (whatever it means) are shipping faster now, no?
I think people already have forgotten how long and arduous software development was just a few years ago.
The biggest bottleneck now is CI/CD/Experimentation with respect to engineering, and product alignment with respect to the business.
Personally, we’re just beginning to ship faster because we’re a big ship and turn slowly. We’re still getting up to speed but we’re also bogged down with bug fixes, security fixes, and the review bottleneck. Reviews are especially tough now that every commit is much larger because it is now much more correct than the crap we were submitting before AI.
Also, many people are simply lying and hate AI. There are people I’ve encountered who look for any evidence to justify not embracing it. It’s frankly tiring but I understand some people cannot handle change and are frankly terrified of what is happening to their careers and their concept of self-worth.
I have seen it among “the general public“. I’ve seen normal people complain about things in open-source projects or other contexts where to whatever degree they accepted programming took time before now they think it’s all just a magic wand with Claude, no matter the level of complexity.
I haven’t heard a single person at my company in management at any layer complain that we’re not working fast enough now that we have AI. One of my friends has 100% told me their management is pulling that. The management over there is on the everyone should do four times as many tickets, and use at least X amount of tokens per day. otherwise you’re slacking off train.
It’s all in who you talk to.
Engineers that have only started full time work post 2023 haven't experienced the counterfactual of not having AI.
They see all the mundane business/eng bottlenecks that are pervasive now and don't truly know how much better it is now than before
The programmers who are left is an artificial selection of AI boosters.
Now it feels like everyone is in a rush and there is no time for crafting anything. Just vibe code a pile of shit that looks good but it's unnecessarily complex.
You are going to see how arduous software development is in a few years when no one understands the systems they have built
"One occasionally reads newspaper accounts of how two programmers in a remodeled garage have built an important program that surpasses the best efforts of large teams. And every programmer is prepared to believe such tales, for he knows that he could build any program much faster than the 1000 statements/year reported for industrial teams. Why then have not all industrial programming teams been replaced by dedicated garage duos? One must look at what is being produced."
Brooks argues that the coding that happened in the garage is only a fraction of what you need for the final product:
"It costs nine times as much. But it is the truly useful object, the intended product of most system programming efforts."
What we're seeing right now does not feel much different. You can greatly accelerate a first draft, a greenfield project, an experiment, or at least the initial iteration of one. Changing a live production system or a widely used library is an entirely different proposition.
Low risk work is tremendously accelerated. Just look at the memes that people re building. One of my favorite right now are derivations of this thing: https://x.com/jamesonhodge/status/2118733265372860826
People are running away with it in hilarious ways using all sorts of storyboarding, 3d environments, gameplay, etc etc.
I work in a highly regulated environment. My AI spend is ~$300/week and it has been tremendously helpful in troubleshooting incidents(!), increasing test coverage, gaining understanding of legacy/inherited code, applying fixes, etc. But shipping new features is something that impacts us in many many many ways and the uptake there has been quite slow.
https://chimera-vocal-synth.vercel.app/
Thanks for the share, I love it.
Clearly, acceleration is possible. So maybe we should move on to the conversation of what will make further acceleration possible, instead of arguing it only works for iteration one.
Then the next challenge is scaling it to the world - and Twitter themselves struggled with that, the first years were full of fail whales, or stories of servers churning for ages whenever Justin Bieber made a tweet because (for some reason) they updated everyone's feeds even if they weren't actually online / looking at them.
But solving that also wasn't what made Twitter or Facebook popular, and no LLM will ever solve that because none of the successful software projects were successful because of code, code volume, development speed, number of people involved, or anything like that.
(That said I am convinced you can do the same amount of coding with less people at good enough quality, and said quality will improve over time as we've seen in the past ~2 years. The providers just need to add the software quality guardrails that we as software engineers learned over the years, and I'm sure there's dozens of companies working on just that)
More to the point, the price of RAM that has gone berserk. Easy solution, let's then make our OSes, browsers and almost all the relevant software that we use day to day an order a magnitude more nimble, for example if a Chrome instance now eats 1-2GB of memory then make it it at most 100-200MBs of said memory while doing the exact same things (if not more). The much lauded AI revolution in code-writing hasn't created that, not sure if it will ever do, it will just add slop after slop after slop, in fact it wouldn't surprise me if the memory requirements for something like Chrome won't increase as a result of the recent vibe-coding mania.
But by this point the AI zealots are already running the asylum, so nothing to be done.
The sad thing is those using AI to code (and I certainly do) don't realize they're producing more but just more of the same turds they were already producing.
And in addition to that they're producing turds nobody gives a shit about (ah!).
Where, indeed, are the both less memory-hungry, less CPU-intensive and more secure rewrites of the pigs we work with on a daily basis?
Nowhere to be seen.
Because the AI zealots are way too busy adding crap features over crap features.
Pre coding : Requirements, grooming, solutioning (cross product specially), design, LLD
Post coding: Cross product testing, E2E integration testing, CRQ approvals, Deployment, pilot, roll out.
SWEs don't just do coding. AI has only solved that and testing to an extent.