스크립트

00:00:00"Spotify ships 4,500 production deploys a day
00:00:05and 73% of all pull requests are now AI-assisted."
00:00:10Wow.
00:00:11Especially the 4,500 production deploys a day,
00:00:15that is really impressive.
00:00:16And I'm not sure if it's impressive in a good way.
00:00:19And there is a funny comment below that post on X,
00:00:23which I enjoyed quite a bit.
00:00:25Are they doing a PR, a pull request for every song?
00:00:28And yeah, it certainly seems like.
00:00:30I mean, this is from the Claude Devs X account.
00:00:35It's an advertisement for Claude Code,
00:00:37because as we learn in this thread,
00:00:39they are of course using Claude Code
00:00:42and anthropic models for shipping that much.
00:00:45And of course, if you watched some of my earlier videos,
00:00:48you know, I also use AI-Agents, of course.
00:00:50It is the future.
00:00:52It is how software development works these days.
00:00:54And I have a Claude Code course where I share all my insights,
00:00:57and tips and tricks, and how to use this tool effectively.
00:01:00But 4,500 production deploys a day, really?
00:01:05I mean, yeah, Spotify clearly is a large company.
00:01:08That's no secret, but you have to wonder what are they doing?
00:01:12I am an active Spotify user.
00:01:15And I feel like the app, well, has not changed that much,
00:01:21certainly not to the better over the last year or so.
00:01:26And it truly is too much for me to grasp what could be in those 4,500 production deploys a day.
00:01:37And please, and this is no joke, I absolutely mean it.
00:01:40Please share in the comments what you think they are shipping all day.
00:01:44Now, the 73% of all pull requests, yeah, sure.
00:01:48I mean, AI-assisted can be anything theoretically.
00:01:51Obviously, a simple completion, tab completion, old school style,
00:01:58whilst you are reviewing the PR and maybe making some changes could count as AI-assisted.
00:02:04And of course, having an AI agent analyzed that would also be AI-assisted.
00:02:08And frankly, I wonder that it's not 100% because AI, I would imagine,
00:02:13would be involved in pretty much all PR reviews or the creation of pretty much all PRs.
00:02:20But yeah, I'm sure this number will change.
00:02:22Wouldn't be strange to see that.
00:02:26But that first number is really kind of a trend we can see in general, right?
00:02:34We had the entire token maxing debate a couple of months ago.
00:02:38Then many companies realized that this is too expensive and just telling your software engineers
00:02:44to burn as many tokens as possible is probably not a good strategy
00:02:48because it's pretty easy, obviously, to burn tokens and doesn't guarantee good results.
00:02:53Now we seem to be shifting to an era where it's quite popular to be shipping a lot.
00:02:58Because, of course, with AI, you can, right?
00:03:01You can ship a lot.
00:03:02You can have like five agents running simultaneously using Gitwork trees,
00:03:06working on four different projects at a time.
00:03:08And yeah, you can parallelize to some degree.
00:03:12You can build lots of stuff with AI for sure.
00:03:15No doubt about that.
00:03:17And it can accelerate you as a software developer.
00:03:20It can also slow you down.
00:03:21We've all been there where you went down a road with AI,
00:03:25where you became more and more of a vibe coder
00:03:28because you became too lazy to actually evaluate what the AI generated
00:03:32because it just feels so fast to keep on going.
00:03:36And then at some point, you have no understanding of your code base at all,
00:03:39but something doesn't work or you're not happy with the direction it's taking
00:03:42and you abandon that project or you undo a lot.
00:03:48And well, we can argue if you're really being more productive then.
00:03:52But that aside, and we're all still trying to figure out how to best use AI,
00:03:58you can be more productive with it.
00:04:01I mean, even if you would just give it a pretty clear plan of what it should do
00:04:05and it would be a tiny feature,
00:04:06just having it spit out large chunks of code can speed you up.
00:04:10I'm not sure if it's like a 10x improvement or even more as some people claim.
00:04:14And I'm not sure if 4,500 deploys a day is the measure we should go for.
00:04:22I think we should really aim for figuring out how to use AI effectively,
00:04:28how to, sure, speed up our software delivery,
00:04:32but maybe, maybe also focus a bit on the quality, not just the quantity.
00:04:37Right now, we're all in that quantity era, right?
00:04:39Token maxing number of deploys.
00:04:41Quality also matters and AI can help with that.
00:04:45But good old testing, also human testing,
00:04:48not just automated tests generated by AI,
00:04:51where it also has a tendency or can have a tendency of writing tests that pass
00:04:56but don't actually test anything.
00:04:57But that aside, good old manual testing
00:05:00and really just injecting your taste and your decisions as a human into the workflow.
00:05:07These are all probably good ideas, I feel like.
00:05:09And yeah, will be interesting to see when this trend ends.
00:05:13I just stumbled across that number, wanted to share my amazement of that number.
00:05:18And as I mentioned before, I really want to hear,
00:05:20what do you think is Spotify shipping all day long?
00:05:25Because I really just can't grasp it.
00:05:29Sorry.

핵심 요약

While AI-driven development enables rapid deployment rates like Spotify's 4,500 daily updates, prioritizing quantity risks neglecting human oversight and code quality.

하이라이트

  • Spotify executes 4,500 production deployments per day.

  • AI-assisted tools now facilitate 73% of all pull requests at Spotify.

  • Increased deployment frequency results from the use of AI agents and automated coding tools like Claude Code.

  • Over-reliance on AI for code generation can lead to 'vibe coding,' where developers lose understanding of the codebase.

  • High-volume deployment strategies often prioritize quantity over code quality.

타임라인

High-Frequency Deployment and AI Integration

  • Spotify records 4,500 production deployments every day.
  • AI-assisted systems handle 73% of total pull requests.
  • Anthropic's Claude Code is used to facilitate this high volume of shipping.

Spotify maintains a high velocity of software delivery, utilizing AI agents to manage a large portion of code changes. This volume is attributed to the integration of AI tools within the development lifecycle. The classification of 'AI-assisted' includes everything from standard tab completion to sophisticated AI agents analyzing pull requests.

Trends in AI-Driven Development

  • Development practices have shifted from 'token maxing' to maximizing the number of deployments.
  • AI agents allow for parallel development by working on multiple projects simultaneously.
  • Excessive reliance on AI generation can reduce a developer's grasp of their own codebase.

Software development trends have moved past the initial phase of simply burning tokens for AI outputs. The focus now rests on high-frequency shipping enabled by concurrent AI agents. However, this speed can lead to laziness, where developers fail to properly evaluate generated code, potentially slowing progress in the long run.

Balancing Speed with Quality

  • Effective software delivery requires a balance between speed and code quality.
  • AI-generated automated tests may pass without actually validating core functionality.
  • Human intervention remains necessary for testing and strategic decision-making.

Relying solely on AI to maintain high deployment counts is insufficient if the output lacks quality. Human testers and critical decision-makers are essential to verify that code functions correctly beyond passing automated checks. Future success in AI-assisted development depends on integrating human intuition with automated acceleration.

커뮤니티 글

모든 글 보기