> Copilot excels at low-to-medium complexity tasks in well-tested codebases, from adding features and fixing bugs to extending tests, refactoring, and improving documentation.
Bounds bounds bounds bounds. The important part for humans seems to be maintaining boundaries for AI. If your well-tested codebase has the tests built thru AI, its probably not going to work.
I think its somewhat telling that they can't share numbers for how they're using it internally. I want to know that Microsoft, the company famous for dog-fooding is using this day in and day out, with success. There's real stuff in there, and my brain has an insanely hard time separating the trillion dollars of hype from the usefulness.
We've been using Copilot coding agent internally at GitHub, and more widely across Microsoft, for nearly three months. That dogfooding has been hugely valuable, with tonnes of valuable feedback (and bug bashing!) that has helped us get the agent ready to launch today.
So far, the agent has been used by about 400 GitHub employees in more than 300 our our repositories, and we've merged almost 1,000 pull requests contributed by Copilot.
In the repo where we're building the agent, the agent itself is actually the #5 contributor - so we really are using Copilot coding agent to build Copilot coding agent ;)
(Source: I'm the product lead at GitHub for Copilot coding agent.)
> In the repo where we're building the agent, the agent itself is actually the #5 contributor - so we really are using Copilot coding agent to build Copilot coding agent ;)
Really cool, thanks for sharing! Would you perhaps consider implementing something like these stats that aider keeps on "aider writing itself"? - https://aider.chat/HISTORY.html
> In the repo where we're building the agent, the agent itself is actually the #5 contributor - so we really are using Copilot coding agent to build Copilot coding agent ;)
Thats a fun stat! Are humans in the #1-4 slots? Its hard to know what processes are automated (300 repos sounds like a lot of repos!).
Thank you for sharing the numbers you can. Every time a product launch is announced, I feel like its a gleeful announcement of a decrease of my usefulness. I've got imposter syndrome enough, perhaps Microsoft might want to speak to the developer community and let us know what they see happening? Right now its mostly the pink slips that are doing the speaking.
How strong was the push from leadership to use the agents internally?
As part of the dogfooding I could see them really pushing hard to try having agents make and merge PRs, at which point the data is tainted and you don't know if the 1,000 PRs were created or merged to meet demand or because devs genuinely found it useful and accurate.
I'd like a breakdown of this phrase, how much human work vs Copilot and in what form, autocomplete vs agent. It's not specified seems more like a marketing trickery than real data
> In the repo where we're building the agent, the agent itself is actually the #5 contributor
How does this align with Microsoft's AI safety principals? What controls are in place to prevent Copilot from deciding that it could be more effective with less limitations?
What's the motivation for restricting to Pro+ if billing is via premium requests? I have a (free, via open source work) Pro subscription, which I occasionally use. I would have been interested in trying out the coding agent, but how do I know if it's worth $40 for me without trying it ;).
Question you may have a very informed perspective on:
where are we wrt the agent surveying open issues (say, via JIRA) and evaluating which ones it would be most effective at handling, and taking them on, ideally with some check-in for conirmation?
Or, contrariwise, from having product management agents which do track and assign work?
Is Copilot _enforced_ as the only option for an AI coding agent? Or can devs pick-and-choose whatever tool they prefer
I'm interested in the [vague] ratio of {internallyDevlopedTool} vs alternatives - essentially the "preference" score for internal tools (accounting for the natural bias towards ones own agent for testing/QA/data purposes). Any data, however vague is necessary, would be great.
(and if anybody has similar data for _any_ company developing their own agent, please shout out).
400 GitHub employees are using GitHub Copilot day in day out, and it comes out as #5 contributor? I wouldn't call that a success. If it is any useful, I would expect that even if a developer write 10% of their code using it, it would hold be #1 contributor in every project.
re: 300 of your repositories...
so it sounds like y'all don't use a monorepo architecture. i've been wondering if that would be a blocker to using these agents most effectively. expect some extra momentum to swing back to the multirepo approach accordingly
When I repeated to other tech people from about 2012 to 2020 that the technological singularity was very close, no one believed me. Coding is just the easiest to automate away into almost oblivion. And, too many non technical people drank the Flavor Aid for the fallacy that it can be "abolished" completely soon. It will gradually come for all sorts of knowledge work specialists including electrical and mechanical engineers, and probably doctors too. And, of course, office work too. Some iota of a specialists will remain to tune the bots, and some will remain in the fields to work with them for where expertise is absolutely required, but widespread unemployment of what were options for potential upward mobility into middle class are being destroyed and replaced with nothing. There won't be "retraining" or handwaving other opportunities for the "basket of labor", but competition of many uniquely, far overqualified people for ever dwindling opportunities.
It is difficult to get a man to understand something when his salary depends upon his not understanding it. - Upton Sinclair
> In the repo where we're building the agent, the agent itself is actually the #5 contributor - so we really are using Copilot coding agent to build Copilot coding agent ;)
From talking to colleagues at Microsoft it's a very management-driven push, not developer-driven. Friend on an Azure team had a team member who was nearly put on a PIP because they refused to install the internal AI coding assistant. Every manager has "number of developers using AI" as an OKR, but anecdotally most devs are installing the AI assistant and not using it or using it very occasionally. Allegedly it's pretty terrible at C# and PowerShell which limits its usefulness at MS.
> I want to know that Microsoft, the company famous for dog-fooding is using this day in and day out, with success
Have they tried dogfooding their dogshit little tool called Teams in the last few years? Cause if that's what their "famed" dogfooding gets us, I'm terrified to see what lays in wait with copilot.
I feel like I saw a quote recently that said 20-30% of MS code is generated in some way. [0]
In any case, I think this is the best use case for AI in programming—as a force multiplier for the developer. It’s for the best benefit of both AI and humanity for AI to avoid diminishing the creativity, agency and critical thinking skills of its human operators. AI should be task oriented, but high level decision-making and planning should always be a human task.
So I think our use of AI for programming should remain heavily human-driven for the long term. Ultimately, its use should involve enriching humans’ capabilities over churning out features for profit, though there are obvious limits to that.
You might want to study the history of technology and how rapidly compute efficiency has increased as well as how quickly the models are improving.
In this context, assuming that humans will still be able to do high level planning anywhere near as well as an AI, say 3-5 years out, is almost ludicrous.
They have released numbers, but I can't say they are for this specific product or something else. They are apparently having AI generate "30%" of their code.
Whatever the true stats for mistakes or blunders are now, remember that this is the worst its ever going to be. And there is no clear ceiling in sight that would prevent it from quickly getting better and better, especially given the current levels of investment.
That sounds reasonable enough, but the pace or end result is by no means guaranteed.
We have invested plenty of money and time into nuclear fusion with little progress. The list of key acheivments from CERN[1] is also meager in comparison to the investment put in, especially if you consider their ultimate goal to ultimately be towards applying research to more than just theory.
I tried doing some vibe coding on a greenfield project (using gemini 2.5 pro + cline). On one hand - super impressive, a major productivity booster (even compared to using a non-integrated LLM chat interface).
I noticed that LLMs need a very heavy hand in guiding the architecture, otherwise they'll add architectural tech debt. One easy example is that I noticed them breaking abstractions (putting things where they don't belong). Unfortunately, there's not that much self-retrospection on these aspects if you ask about the quality of the code or if there are any better ways of doing it. Of course, if you pick up that something is in the wrong spot and prompt better, they'll pick up on it immediately.
I also ended up blowing through $15 of LLM tokens in a single evening. (Previously, as a heavy LLM user including coding tasks, I was averaging maybe $20 a month.)
Cline very visibly displays the ongoing cost of the task. Light edits are about 10 cents, and heavy stuff can run a couple of bucks. It's just that the tab accumulates faster than I expect.
I think that models are gonna commoditize, if they haven't already. The cost of switching over is rather small, especially when you have good evals on what you want done.
Also there's no way you can build a business without providing value in this space. Buyers are not that dumb.
I, too, recommend aider whenever these discussions crop up; it converted me from the "AI tools suck" side of this discussion to the "you're using the wrong tool" side.
I'd also recommend creating little `README`'s in your codebase that are mainly written with aider as the intended audience. In it, I'll explain architecture, what code makes (non-)sense to write in this directory, and so on. Has the side-effect of being helpful for humans, too.
Nowadays when I'm editing with aider, I'll include the project README (which contains a project overview + pointers to other README's), and whatever README is most relevant to the scope of my session. It's super productive.
I'm yet to find a model that beats the cost-effectiveness of Sonnet 3.7. I've tried the latest deepseek models, and while I love the price (nearly 50x cheaper?), it's just far too error-prone compared to Sonnet 3.7. It generates solid plans / architecture discussions, but, unlike Sonnet, the code it generates often confidently off-the-mark.
My tool Plandex[1] allows you to switch between automatic and manual context management. It can be useful to begin a task with automatic context while scoping it out and making the high level plan, then switch to the more 'aider-style' manual context management once the relevant files are clearly established.
I loathe using AI in a greenfield project. There are simply too many possible paths, so it seems to randomly switch between approaches.
In a brownfield code base, I can often provide it reference files to pattern match against. So much easier to get great results when it can anchor itself in the rest of your code base.
The trick for greenfield projects is to use it to help you design detailed specs and a tentative implementation plan. Just bounce some ideas off of it, as with a somewhat smarter rubber duck, and hone the design until you arrive at something you're happy with. Then feed the detailed implementation plan step by step to another model or session.
This is a popular workflow I first read about here[1].
This has been the most useful use case for LLMs for me. Actually getting them to implement the spec correctly is the hard part, and you'll have to take the reigns and course correct often.
While its being touted for Greenfield projects I've notices a lot of failures when it comes to bootstrapping a stack.
For example it (Gemini 2.5) really struggles with newer ecosystem like Fastapi when wiring libraries like SQLAlchemy, Pytest, Python-playwright, etc., together.
I find more value in bootstrapping myself, and then using it to help with boiler plate once an effective safety harness is in place.
I've vibe coded small project as well using Claude Code. It's about visitors registration at the company. Simple project, one form, a couple of checkboxes, everything is stored in sqlite + has endpoint for getting .xlsx.
Initial cost was around $20 USD, which later grew to (mostly polishing) $40 with some manual work.
I've intentionally picked up simple stack: html+js+php.
A couple of things:
* I'd say I'm happy about the result from product's perspective
* Codebase could be better, but I could not care less about in this case
* By default, AI does not care about security unless I specifically tell it
* Claude insisted on using old libs. When I've specifically told it to use the latest and greatest, it upgraded them but left code that works just with an old version. Also it mixed latest DaisyUI with some old version of tailwindcss :)
On one hand it was super easy and fun to do, on the other hand if I was a junior engineer, I bet it would have cost more.
If you want to use Cline and are at all price sensitive (in these ranges) you have to do manual context management just for that reason. I find that too cumbersome and use Windsurf (currently with Gemini 2.5 pro) for that reason.
I think it's just that it's not end-to-end trained on architecture because the horizon is too short. It doesn't have the context length to learn the lessons that we do about good design.
> I noticed that LLMs need a very heavy hand in guiding the architecture, otherwise they'll add architectural tech debt. One easy example is that I noticed them breaking abstractions
That doesn’t matter anymore when you’re vibe coding it. No human is going to look at it anyway.
It can all be if/else on one line in one file. If it works and if the LLMs can work at, iterate and implement new business requirements, while keeping performance and security - code structure, quality and readability don’t matter one bit.
Customers don’t care about code quality and the only reason businesses used to care is to make it less money consuming to build and ship new things, so they can make more money.
Nope - I use a-la-carte pricing (through openrouter). I much prefer it over a subscription, as there are zero limits, I pay only for what I use, and there is much less of a walled garden (I can easily switch between Anthropic, Google, etc).
I wish they optimized things before adding more crap that will slow things down even more. The only thing that's fast with copilot is the autocomplete, it sometimes takes several minutes to make edits on a 100 line file regardless of the model I pick (some are faster than others). If these models had a close to 100% hit rate this would be somewhat fine, but going back and forth with something that takes this long is not productive. It's literally faster to open claude/chatgpt on a new tab and paste the question and code there and paste it back into vscode than using their ask/edit/agent tools.
I've cancelled my copilot subscription last week and when it expires in two weeks I'll mostly likely shift to local models for autocomplete/simple stuff.
My experience has mostly been the opposite -- changes to several-hundred-line files usually only take a few seconds.
That said, months ago I did experience the kind of slow agent edit times you mentioned. I don't know where the bottleneck was, but it hasn't come back.
I'm on library WiFi right now, "vibe coding" (as much as I dislike that term) a new tool for my customers using Copilot, and it's snappy.
I've had this too, especially it getting stuck at the very end and just.. never finishing. Once the usage-based billing comes into effect I think I'll try cursor again.
What local models are you using? The local models I tried for autocomplete were unusable, though based on aiders benchmark I never really tried with larger models for chat. If I could I would love to go local-only instead.
That first PR (115733) would make me quit after a week if we were to implement this crap at my job and someone forced me to babysit an AI in its PRs in this fashion. The others are also rough.
A wall of noise that tells you nothing of any substance but with an authoritative tone as if what it's doing is objective and truthful - Immediately followed by:
- The 8 actual lines of code (discounting the tests & boilerplate) it wrote to actually fix the issue is being questioned by the person reviewing the code, it seems he's not convinced this is actually fixing what it should be fixing.
- Not running the "comprehensive" regression tests at all
- When they do run, they fail
- When they get "fixed" oh-so confidently, they still fail. Fifty-nine failing checks. Some of these tests take upward of an hour to run.
So the reviewer here has to read all the generated slop in the PR description and try to grok what the PR is about, read through the changes himself anyway (thankfully it's only a ~50 line diff in this situation, but imagine if this was a large refactor of some sort with a dozen files changed), and then drag it by the hand multiple times to try fix issues it itself is causing. All the while you have to tag the AI as if it's another colleague and talk to it as if it's not just going to spit out whatever inane bullshit it thinks you want to hear based on the question asked. Test failed? Well, tests fixed! (no, they weren't)
And we're supposed to be excited about having this crap thrust on us, with clueless managers being sold on this being a replacement for an actual dev? We're being told this is what peak efficiency looks like?
Thanks, that’s really interesting to see - especially with the exchange around whether something is the problem or the symptom, where the confident tone belies the lack of understanding. As an open source maintainer I wonder about the best way to limit usage to cases where someone has time to spend on those interactions.
Thanks. I wonder what model they're using under the hood? I have such a good experience working with Cline and Claude Sonnet 3.7 and a comparatively much worse time with anything Github offers. These PRs are pretty consistent with the experience I've had in the IDE too. Incidentally, what has MSFT done to Claude Sonnet 3.7 in VSCode? It's like they lobotomized it compared to using it through Cline or the API directly. Trying to save on tokens or something?
Major scam alert, they are training on your code in private repos if you use this
You can tell because they advertise “Pro” and “Pro+” but then the FAQ reads,
> Does GitHub use Copilot Business or Enterprise data to train GitHub’s model?
> No. GitHub does not use either Copilot Business or Enterprise data to train its models.
Aka, even paid individuals plans are getting brain raped
If you're programming on Windows, your screen is being screenshotted every few seconds anyway. If you don't think OCR isn't analysing everything resembling a letter on your screen boy do I have some news for you.
I’ve been trying to use Copilot for a few days to get some help writing against code stored on GitHub.
Copilot has been pretty useless. It couldn’t maintain context for more than two exchanges.
Copilot: here’s some C code to do that
Me: convert that to $OTHER_LANGUAGE
Copilot: what code would you like me to convert?
Me: the code you just generated
Copilot: if you can upload a file or share a link to the code, I can help you translate it …
It points me in a direction that’s a minimum of 15 degrees off true north (“true north” being the goal for which I am coding), usually closer to 90 degrees. When I ask for code, it hallucinates over half of the API calls.
I’m sure you have no idea what my method is. Besides, this whole “you’re holding it wrong” mentality isn’t productive - our technology should be adapting to us, we shouldn’t need to adapt ourselves to it.
Anyway, I can just use another LLM that serves me better.
I played around with it quite a bit. it is both impressive and scary. most importantly, it tends to indiscriminately use dependencies from random tiny repos, and often enough not the correct ones, for major projects. buyer beware.
This is something I've noticed as well with different AIs. They seem to disproportionately trust data read from the web. For example, I asked to check if some obvious phishing pages were scams and multiple times I got just a summary of the content as if it was authoritative. Several times I've gotten some random chinese repo with 2 stars presented as if it was the industry standard solution, since that's what it said in the README.
On an unrelated note, it also suggested I use the "Strobe" protocol for encryption and sent me to https://strobe.cool which is ironic considering that page is all about making one hallucinate.
>They seem to disproportionately trust data read from the web.
I doubt LLM's have anything like what we would conceptualize as trust. They have information, which is regurgitated because it is activated as relevant.
That being said, many humans don't really have a strong concept of information validation as part of day to day action and thinking. Development theory talks about this in terms of 'formal operational' thinking and 'personal epistemology' - basically how does thinking happen and then how is knowledge in those models conceptualized. Learning Sciences research generally talks about Piaget and formal operational before adulthood and stages of personal epistemology in higher education.
Research consistently suggests that about 50% of adults are not able to consistently operate in the formal thinking space. The behavior you are talking about is also typical of 'absolutist' epistemic perspectives where answers are right or wrong and aren't meaningfully evaluated - just identifed as relevant or not. Evaluating the credibility of information is that it comes from a source that is trusted - most often an authority figure - it is not the role of the person knowing it.
Oh wow, that was great - particularly if I then look at my own body parts (like my palm) that I know are not moving, it's particularly disturbing. That's a really well done effect, I've seen something similar but nothing quite like that.
>On an unrelated note, it also suggested I use the "Strobe" protocol for encryption and sent me to https://strobe.cool which is ironic considering that page is all about making one hallucinate.
That's not hallucination. That's just an optical illusion.
As we've built Copilot coding agent, we've put a lot of thought and work into our security story.
One of the things we've done here is to treat Copilot's commits like commits from a first-time contributor to an open source project.
When Copilot pushes changes, your GitHub Actions workflows won't run by default, and you'll have to click the "Approve and run workflows" button in the merge box.
That gives you the chance to run Copilot's code before it runs in Actions and has access to your secrets.
(Source: I'm on the product team for Copilot coding agent.)
No, not at all. Why do people keep saying shit like these thought terminating sentences. Try to see the glass of Kool Aid please. People are trying to understand how to communicate important valuable things about failure states and you're advocating ignorance.
But rest assured that with Github Copilot Coding Agent, your codebase will develop larger and larger volumes of new, exciting, underexplored technical debt that you can't be blamed for, and your colleagues will follow you into the murky depths soon.
Readit News
Bounds bounds bounds bounds. The important part for humans seems to be maintaining boundaries for AI. If your well-tested codebase has the tests built thru AI, its probably not going to work.
I think its somewhat telling that they can't share numbers for how they're using it internally. I want to know that Microsoft, the company famous for dog-fooding is using this day in and day out, with success. There's real stuff in there, and my brain has an insanely hard time separating the trillion dollars of hype from the usefulness.
So far, the agent has been used by about 400 GitHub employees in more than 300 our our repositories, and we've merged almost 1,000 pull requests contributed by Copilot.
In the repo where we're building the agent, the agent itself is actually the #5 contributor - so we really are using Copilot coding agent to build Copilot coding agent ;)
(Source: I'm the product lead at GitHub for Copilot coding agent.)
I'm curious to know how many Copilot PRs were not merged and/or required human take-overs.
Really cool, thanks for sharing! Would you perhaps consider implementing something like these stats that aider keeps on "aider writing itself"? - https://aider.chat/HISTORY.html
Thats a fun stat! Are humans in the #1-4 slots? Its hard to know what processes are automated (300 repos sounds like a lot of repos!).
Thank you for sharing the numbers you can. Every time a product launch is announced, I feel like its a gleeful announcement of a decrease of my usefulness. I've got imposter syndrome enough, perhaps Microsoft might want to speak to the developer community and let us know what they see happening? Right now its mostly the pink slips that are doing the speaking.
As part of the dogfooding I could see them really pushing hard to try having agents make and merge PRs, at which point the data is tainted and you don't know if the 1,000 PRs were created or merged to meet demand or because devs genuinely found it useful and accurate.
I'd like a breakdown of this phrase, how much human work vs Copilot and in what form, autocomplete vs agent. It's not specified seems more like a marketing trickery than real data
How does this align with Microsoft's AI safety principals? What controls are in place to prevent Copilot from deciding that it could be more effective with less limitations?
where are we wrt the agent surveying open issues (say, via JIRA) and evaluating which ones it would be most effective at handling, and taking them on, ideally with some check-in for conirmation?
Or, contrariwise, from having product management agents which do track and assign work?
I'm interested in the [vague] ratio of {internallyDevlopedTool} vs alternatives - essentially the "preference" score for internal tools (accounting for the natural bias towards ones own agent for testing/QA/data purposes). Any data, however vague is necessary, would be great.
(and if anybody has similar data for _any_ company developing their own agent, please shout out).
It is difficult to get a man to understand something when his salary depends upon his not understanding it. - Upton Sinclair
Without data, a comprehensive study and peers review, it's a hell no. Would GitHub willing to be at academic scrutiny to prove it?
Ah yes, the takeoff.
Dead Comment
Have they tried dogfooding their dogshit little tool called Teams in the last few years? Cause if that's what their "famed" dogfooding gets us, I'm terrified to see what lays in wait with copilot.
In any case, I think this is the best use case for AI in programming—as a force multiplier for the developer. It’s for the best benefit of both AI and humanity for AI to avoid diminishing the creativity, agency and critical thinking skills of its human operators. AI should be task oriented, but high level decision-making and planning should always be a human task.
So I think our use of AI for programming should remain heavily human-driven for the long term. Ultimately, its use should involve enriching humans’ capabilities over churning out features for profit, though there are obvious limits to that.
[0] https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-a...
Similar to google. MS now requires devs to use ai
In this context, assuming that humans will still be able to do high level planning anywhere near as well as an AI, say 3-5 years out, is almost ludicrous.
They just cut down their workforce, letting some of their AI people go. So, I assume there isn't that much success.
Deleted Comment
https://techcrunch.com/2025/04/29/microsoft-ceo-says-up-to-3...
Deleted Comment
This was true up around 15 years ago. Hasn't been the case since.
We have invested plenty of money and time into nuclear fusion with little progress. The list of key acheivments from CERN[1] is also meager in comparison to the investment put in, especially if you consider their ultimate goal to ultimately be towards applying research to more than just theory.
[1] https://home.cern/about/key-achievements
I noticed that LLMs need a very heavy hand in guiding the architecture, otherwise they'll add architectural tech debt. One easy example is that I noticed them breaking abstractions (putting things where they don't belong). Unfortunately, there's not that much self-retrospection on these aspects if you ask about the quality of the code or if there are any better ways of doing it. Of course, if you pick up that something is in the wrong spot and prompt better, they'll pick up on it immediately.
I also ended up blowing through $15 of LLM tokens in a single evening. (Previously, as a heavy LLM user including coding tasks, I was averaging maybe $20 a month.)
This is a feature, not a bug. LLMs are going to be the next "OMG my AWS bill" phenomenon.
Also there's no way you can build a business without providing value in this space. Buyers are not that dumb.
Consider using Aider, and aggressively managing the context (via /add, /drop and /clear).
https://aider.chat/
I'd also recommend creating little `README`'s in your codebase that are mainly written with aider as the intended audience. In it, I'll explain architecture, what code makes (non-)sense to write in this directory, and so on. Has the side-effect of being helpful for humans, too.
Nowadays when I'm editing with aider, I'll include the project README (which contains a project overview + pointers to other README's), and whatever README is most relevant to the scope of my session. It's super productive.
I'm yet to find a model that beats the cost-effectiveness of Sonnet 3.7. I've tried the latest deepseek models, and while I love the price (nearly 50x cheaper?), it's just far too error-prone compared to Sonnet 3.7. It generates solid plans / architecture discussions, but, unlike Sonnet, the code it generates often confidently off-the-mark.
1 - https://github.com/plandex-ai/plandex
Also, a bit more on auto vs. manual context management in the docs: https://docs.plandex.ai/core-concepts/context-management
In a brownfield code base, I can often provide it reference files to pattern match against. So much easier to get great results when it can anchor itself in the rest of your code base.
This is a popular workflow I first read about here[1].
This has been the most useful use case for LLMs for me. Actually getting them to implement the spec correctly is the hard part, and you'll have to take the reigns and course correct often.
[1]: https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/
For example it (Gemini 2.5) really struggles with newer ecosystem like Fastapi when wiring libraries like SQLAlchemy, Pytest, Python-playwright, etc., together.
I find more value in bootstrapping myself, and then using it to help with boiler plate once an effective safety harness is in place.
Initial cost was around $20 USD, which later grew to (mostly polishing) $40 with some manual work.
I've intentionally picked up simple stack: html+js+php.
A couple of things:
* I'd say I'm happy about the result from product's perspective * Codebase could be better, but I could not care less about in this case * By default, AI does not care about security unless I specifically tell it * Claude insisted on using old libs. When I've specifically told it to use the latest and greatest, it upgraded them but left code that works just with an old version. Also it mixed latest DaisyUI with some old version of tailwindcss :)
On one hand it was super easy and fun to do, on the other hand if I was a junior engineer, I bet it would have cost more.
For now.
I wonder if the next phase would be the rise of (AI-driven?) "linters" that check that the implementation matches the architecture definition.
Everything old is new again!
That doesn’t matter anymore when you’re vibe coding it. No human is going to look at it anyway.
It can all be if/else on one line in one file. If it works and if the LLMs can work at, iterate and implement new business requirements, while keeping performance and security - code structure, quality and readability don’t matter one bit.
Customers don’t care about code quality and the only reason businesses used to care is to make it less money consuming to build and ship new things, so they can make more money.
I've cancelled my copilot subscription last week and when it expires in two weeks I'll mostly likely shift to local models for autocomplete/simple stuff.
That said, months ago I did experience the kind of slow agent edit times you mentioned. I don't know where the bottleneck was, but it hasn't come back.
I'm on library WiFi right now, "vibe coding" (as much as I dislike that term) a new tool for my customers using Copilot, and it's snappy.
https://streamable.com/rqlr84
The claude and gemini models tend to be the slowest (yes, including flash). 4o is currently the fastest but still not great.
https://streamable.com/rqlr84
Cursor is quicker, I guess it's a response parsing thing - when they make the decision to show it in the UI.
https://github.com/dotnet/runtime/pull/115733https://github.com/dotnet/runtime/pull/115732https://github.com/dotnet/runtime/pull/115762
A wall of noise that tells you nothing of any substance but with an authoritative tone as if what it's doing is objective and truthful - Immediately followed by:
- The 8 actual lines of code (discounting the tests & boilerplate) it wrote to actually fix the issue is being questioned by the person reviewing the code, it seems he's not convinced this is actually fixing what it should be fixing.
- Not running the "comprehensive" regression tests at all
- When they do run, they fail
- When they get "fixed" oh-so confidently, they still fail. Fifty-nine failing checks. Some of these tests take upward of an hour to run.
So the reviewer here has to read all the generated slop in the PR description and try to grok what the PR is about, read through the changes himself anyway (thankfully it's only a ~50 line diff in this situation, but imagine if this was a large refactor of some sort with a dozen files changed), and then drag it by the hand multiple times to try fix issues it itself is causing. All the while you have to tag the AI as if it's another colleague and talk to it as if it's not just going to spit out whatever inane bullshit it thinks you want to hear based on the question asked. Test failed? Well, tests fixed! (no, they weren't)
And we're supposed to be excited about having this crap thrust on us, with clueless managers being sold on this being a replacement for an actual dev? We're being told this is what peak efficiency looks like?
Dead Comment
You can tell because they advertise “Pro” and “Pro+” but then the FAQ reads,
> Does GitHub use Copilot Business or Enterprise data to train GitHub’s model? > No. GitHub does not use either Copilot Business or Enterprise data to train its models.
Aka, even paid individuals plans are getting brain raped
https://docs.github.com/en/copilot/managing-copilot/managing...
Copilot has been pretty useless. It couldn’t maintain context for more than two exchanges.
Copilot: here’s some C code to do that
Me: convert that to $OTHER_LANGUAGE
Copilot: what code would you like me to convert?
Me: the code you just generated
Copilot: if you can upload a file or share a link to the code, I can help you translate it …
It points me in a direction that’s a minimum of 15 degrees off true north (“true north” being the goal for which I am coding), usually closer to 90 degrees. When I ask for code, it hallucinates over half of the API calls.
Deleted Comment
Anyway, I can just use another LLM that serves me better.
On an unrelated note, it also suggested I use the "Strobe" protocol for encryption and sent me to https://strobe.cool which is ironic considering that page is all about making one hallucinate.
I doubt LLM's have anything like what we would conceptualize as trust. They have information, which is regurgitated because it is activated as relevant.
That being said, many humans don't really have a strong concept of information validation as part of day to day action and thinking. Development theory talks about this in terms of 'formal operational' thinking and 'personal epistemology' - basically how does thinking happen and then how is knowledge in those models conceptualized. Learning Sciences research generally talks about Piaget and formal operational before adulthood and stages of personal epistemology in higher education.
Research consistently suggests that about 50% of adults are not able to consistently operate in the formal thinking space. The behavior you are talking about is also typical of 'absolutist' epistemic perspectives where answers are right or wrong and aren't meaningfully evaluated - just identifed as relevant or not. Evaluating the credibility of information is that it comes from a source that is trusted - most often an authority figure - it is not the role of the person knowing it.
Oh wow, that was great - particularly if I then look at my own body parts (like my palm) that I know are not moving, it's particularly disturbing. That's a really well done effect, I've seen something similar but nothing quite like that.
That's not hallucination. That's just an optical illusion.
Would you be able to drop me an email? My address is my HN login @github.com.
(I work on the product team for Copilot coding agent.)
One of the things we've done here is to treat Copilot's commits like commits from a first-time contributor to an open source project.
When Copilot pushes changes, your GitHub Actions workflows won't run by default, and you'll have to click the "Approve and run workflows" button in the merge box.
That gives you the chance to run Copilot's code before it runs in Actions and has access to your secrets.
(Source: I'm on the product team for Copilot coding agent.)
Stop fighting and sink!
But rest assured that with Github Copilot Coding Agent, your codebase will develop larger and larger volumes of new, exciting, underexplored technical debt that you can't be blamed for, and your colleagues will follow you into the murky depths soon.