Start your week with a bad take!

MMaximilian Schwarzmüller
Computing/SoftwareSmall Business/StartupsManagementInternet Technology

Transcript

00:00:00Today I woke up to an interesting post by Guillermo Rauch, the CEO of Wurzel, about
00:00:06which I want to talk and also about the implications and my thoughts on the role of code and how
00:00:11important the actual code is, today and in the future.
00:00:16Now in this post Guillermo says "Code is an output, nature is healing.
00:00:21For too long we treated code as input.
00:00:23We glorified it, hand formatted it, prettified it, obsessed over it.
00:00:28We built sophisticated GUIs to write it in, IDEs.
00:00:31We syntax-highlit, tree-set, mini-mapped the code, keyboard-triggers, inline-autocompletes,
00:00:37ghost-text.
00:00:38What color scheme is that?
00:00:40We stayed up debating the ideal length of APIs and function bodies.
00:00:44Is this API going to look nice enough for another human to read?
00:00:48We're now turning our attention to the true inputs, requirements, specs, feedback, design
00:00:54inspiration, crucially production inputs.
00:00:58Our coding agents need to understand how your users are experiencing your application, what
00:01:04errors you're running into, and turn that into code.
00:01:07We will inevitably glorify code less as well as coders.
00:01:12The best engineers I've worked with always saw code as a means to an end anyway, an output
00:01:17that's bound to soon be transformed again."
00:01:20Wow, there's a lot in this post!
00:01:23Now I have a lot of thoughts about this.
00:01:27Now, obviously, and I think we can all see this, the entire role or the entire world of
00:01:35programming is changing rapidly.
00:01:38There's lots of stuff going on there and especially I and I guess many people in the industry can
00:01:46feel that since December it accelerated again with Opus 4.5 and Cloud Code and now also
00:01:53Codex and we got new models and new tools coming up almost every week.
00:01:58Things are changing quickly.
00:02:01And on a personal note, of course, I'm trying to keep up with that also regarding the content
00:02:06I create, regarding the courses I create and of course how I work, which is the foundation
00:02:10of that all, and that's why I released new courses on Cloud Code and Codex and I'm by
00:02:16the way running a big promotion right now where you can get my annual membership which gives
00:02:21you access to all the courses for a super low price.
00:02:24But it is having an impact on all of us.
00:02:27That is my point here.
00:02:28It's changing what we do, which content I create and of course how we build software.
00:02:35However, I'm not so sure if we really move away from the code and to me this post reads
00:02:42like code doesn't matter anymore and I think I disagree with that quite a bit.
00:02:50Now I will of course admit and obviously it's true that in the past, yes, we did obsess over
00:02:58API design, over naming of variables and functions, of the aesthetics of the code.
00:03:07And I would actually argue parts of this are still important today and very likely also
00:03:15in the future.
00:03:17Because even if the role of humans shifts away from writing the code towards reviewing it,
00:03:24what is easier to review, a messy codebase with 10,000 lines of codes that are partially
00:03:33redundant or unnecessarily complex with weirdly named or shaped functions or a clean codebase?
00:03:41What will be easier to review?
00:03:44And even if parts or even big parts of that review process shift towards machines and AI
00:03:50in the future, which of course is possible, even then I'm not convinced that the code quality
00:03:59does not matter at all anymore in that world.
00:04:02And if it's just that you're paying for all those tokens that are being generated and reviewed,
00:04:09if something can be built cleaner, it will very likely consume less tokens for both generating
00:04:17that code as well as for reviewing it.
00:04:19And token cost looks like it will be very, very important in the future because right
00:04:25now we're all on those subsidized plans by Enthropic, by OpenAI.
00:04:30We're not paying the real token cost, the $20 or $200 subscriptions you might have and
00:04:37I have.
00:04:39These are not subscriptions where those vendors necessarily make any money off.
00:04:44The true token cost is higher than that.
00:04:47And it's quite likely to go up in the future or what we pay to go up in the future when
00:04:54those companies need to earn money at some point.
00:04:57Obviously, maybe due to all the technical progress that will only start at a point where the token
00:05:02cost as a whole got down a bit, but we don't know that yet with certainty.
00:05:07We don't know how that market will look like in the future.
00:05:10And even if it would go down theoretically, of course, if companies can charge you a certain
00:05:16price, which may be substantially above the true price, well, they will happily take that
00:05:22margin.
00:05:23But yeah, that's a totally different topic.
00:05:25But tokens and token efficiency matters today and very likely also matters in the future.
00:05:31And that, of course, is related to code quality because in general, good code, clean code,
00:05:39is code that is not allowed to grow endlessly and become endlessly complex and complicated.
00:05:47Now, of course, there is an argument to be made that programmers in the past also sometimes
00:05:55had the tendency to write a bit too much code, maybe in certain points, to come up with a
00:06:02clean API and maybe some unnecessary abstractions, which they didn't really need at the point
00:06:08of time where they were implemented.
00:06:09We've probably all been there, where we've worked on a side project with zero users, and
00:06:15we decided that we want to implement our database access such that we could easily swap the database,
00:06:22even though we never intended to do that.
00:06:24And yeah, therefore, we wrote some adapter that was, of course, way more complex than
00:06:29it needed to be.
00:06:30We've all been there.
00:06:31And was it great?
00:06:32No, it was not, and it will not be today or in the future, no matter if that's code written
00:06:37by you or by AI.
00:06:39And right now, and that can, of course, change, but right now, in my experience, these AI models
00:06:45and tools like codecs, cloud code, they have a tendency to overcomplicate things, to introduce
00:06:54unnecessary complexity, to suggest changes that, yeah, really don't make a lot of sense
00:07:02that are unnecessarily complex, codecs especially right now has a tendency to never remove any
00:07:10code, just add more and more and more code and add fallback code and legacy handling.
00:07:15And you have to explicitly almost force it to delete code and to let go of old APIs.
00:07:22So that, of course, can all change, but these are things you have to fight about for today.
00:07:30And that is, of course, kind of what this post says doesn't matter anymore.
00:07:35But I say, yes, it does matter.
00:07:37You, as a developer, you have to take those fights and you have to care about the actual
00:07:43code and the code quality today and very likely also in the future.
00:07:49Obviously, nobody knows what the future holds, but it will always be important that you have
00:07:56a code base that doesn't grow unnecessarily complex for many reasons, token efficiency
00:08:01being one of them.
00:08:02Another reason, of course, and also the review part, by the way, another reason, of course,
00:08:08is that code, of course, has an impact on the performance of an application.
00:08:18Now, I'm not sharing any groundbreaking truth here or news, but of course, this is something
00:08:26that's easy to overlook.
00:08:28You could say that with AI generating code, the actual code doesn't matter too much anymore
00:08:34because it's easy to refactor, regenerate, replace.
00:08:38And that is all true if you ignore the token cost part again.
00:08:43But, of course, the code that's being generated has an impact or could have an impact on the
00:08:51performance of your application.
00:08:53There are multiple ways of querying a database.
00:08:56You could run nested queries or unnecessarily many queries because, hey, the result is the
00:09:03same.
00:09:04You got the data you need.
00:09:05Yeah, but a poorly optimized database query, of course, can show its true cost as soon as
00:09:12you do have a significant amount of users.
00:09:15The same, of course, is true for nested loops or all kinds of stuff that can creep into
00:09:20your code base that leads to worse performance.
00:09:23And that is why, of course, still the code quality matters and optimizing the code matters.
00:09:30You can definitely say that future AI models and tools will be very good at that and will
00:09:37be able to do that, that future models will be able to produce better code right from the
00:09:43start and also evaluate and then improve code.
00:09:47And that is absolutely possible.
00:09:50That still doesn't defeat my point that code quality matters, though, no matter if a future
00:09:55model generates it or you as a developer.
00:09:58Now, of course, what is the big difference is that in the world where the AI is able to
00:10:04do all of that, we humans are taken out of the loop.
00:10:08And I guess that is kind of the point Guillermo is making here.
00:10:13But of course, we don't know how exactly that future will look like.
00:10:16Now what I will say about this post is, and I get it, that is just how the internet is
00:10:23these days.
00:10:24You have to be provocative.
00:10:25You have to kind of be edgy so that people like me create a video about it.
00:10:30I totally get it.
00:10:31But of course, in this post, it sounds like all these things we did in the past were pretty
00:10:37stupid.
00:10:38Right?
00:10:39Building dedicated tools, GUIs that make it easier to write code, discuss API design.
00:10:45That's pretty stupid.
00:10:46Well, no, that shows that you care.
00:10:51And as a human, you want to care about your craft, about your work.
00:10:56That's like saying, oh, pretty stupid that designers used to first sketch on a piece
00:11:02of paper and then use dedicated software like Photoshop to turn this into actual digital
00:11:07art.
00:11:08No, that is not stupid.
00:11:10That is exactly what sets a human that cares about their craft and their job apart from
00:11:17someone who doesn't.
00:11:19And framing this as kind of stupid or weird or something that is gone for good is just
00:11:29a bad take.
00:11:30It's just not cool in any way.
00:11:32And of course now we're building new software, by the way.
00:11:36We're building new software for the agentic engineering world, something like the Codex
00:11:43app and all these AI agents and agent apps and GUIs we have these days.
00:11:49That's the new kind of software we're building and we'll keep on doing stuff like that.
00:11:54And we'll of course invent new ways to discuss about like how a properly written skill for
00:12:01agents should look like or how to set up a good agents MD file or whatever.
00:12:07We'll keep on discussing until we're totally out of the loop and we as humans don't need
00:12:14to do anything anymore.
00:12:16Which is a point I don't think will arrive and I certainly don't hope will arrive.
00:12:22But until that point, it's good to have opinions about the stuff you work on and to care about
00:12:29the stuff you work on.
00:12:31And I think that is really important.
00:12:33So yeah, I think code, the structure of the code, that will stay important.
00:12:41The only question is if the AI can do it all on its own and produce perfect code.
00:12:47And I don't know that and nobody can know right now.
00:12:50You can't look into the future.
00:12:53But in general, as a human, you should care about what you're doing.
00:12:57That is what sets the people that can be super successful and have fun at their work apart
00:13:03but from the people that don't, I think.

Key Takeaway

While AI is rapidly transforming programming into an agentic process, maintaining high code quality remains essential for human reviewability, application performance, and managing the economic costs of AI tokens.

Highlights

Discussion of Guillermo Rauch's controversial post suggesting code is now an output rather than an input.

The speaker's disagreement with the idea that code quality and aesthetics no longer matter in the AI era.

Economic implications of code efficiency regarding 'token costs' for AI generation and review.

Observations on the tendency of current AI models like Codex to overcomplicate codebases and avoid deleting legacy code.

The importance of maintaining human pride in craftsmanship and the role of developer opinions in software engineering.

Concerns about application performance resulting from poorly optimized AI-generated database queries and nested loops.

Timeline

Introduction and the 'Code as Output' Philosophy

The speaker introduces a provocative post by Vercel CEO Guillermo Rauch, which argues that the industry is moving away from 'glorifying' code as a primary input. Rauch suggests that focus is shifting toward true inputs like requirements and design, with AI agents handling the messy output of code. The speaker acknowledges the rapid acceleration of this trend due to new models like Opus 4.5 and Codex. This shift is already impacting how content is created and how software is built globally. It sets the stage for a debate on whether the human relationship with code is fundamentally broken or merely evolving.

The Case for Continued Code Quality and Review

The speaker expresses strong skepticism toward the idea that code quality is becoming irrelevant. Even if humans move from writing to reviewing, a clean and well-named codebase is significantly easier to audit than a messy one. Redundant or unnecessarily complex code poses a burden for both human eyes and automated systems. The argument posits that aesthetics and structure are functional requirements for maintainability, not just vanity. This section emphasizes that the readability of an API remains a critical factor in software longevity.

The Hidden Economics of Token Costs

A major argument is presented regarding the financial cost of AI-generated code through the lens of 'token efficiency.' Longer, more complex code requires more tokens to generate and more tokens to process during an AI-driven review, leading to higher expenses. The speaker warns that current AI subscription prices are likely subsidized and will eventually rise to reflect true market costs. Therefore, writing concise, clean code becomes a matter of economic necessity rather than just a stylistic choice. Efficiency in code directly translates to lower operational costs in an AI-integrated development lifecycle.

AI's Tendency Toward Overcomplication

The speaker shares personal observations on current AI tools, noting their tendency to add unnecessary abstractions and 'fallback' code. Tools like Codex often struggle with deleting code, preferring to stack new layers of complexity onto existing structures. Developers must 'fight' these models to keep codebases lean and prevent them from growing uncontrollably. This creates a new role for the developer as a gatekeeper against 'software bloat' introduced by generative agents. The section highlights the practical reality that AI does not yet possess the human instinct for simplicity.

Performance Impact and the Future of Engineering

The discussion shifts to the technical performance of applications, where poorly optimized AI code can cause significant issues. For example, inefficient database queries or nested loops might work for a single user but fail under heavy production loads. The speaker argues that while future AI might get better at optimization, the quality of the underlying logic still dictates user experience. He also defends the 'craftsmanship' of developers, likening it to a designer's use of Photoshop or sketches. Caring about the fine details of one's work is what distinguishes a professional from a casual user.

Conclusion: Craftsmanship in an Agentic World

The final segment focuses on the new tools being built for 'agentic engineering,' such as the Codex app and specialized GUIs. Even in this new world, humans will find new things to debate and refine, such as the structure of 'agents.md' files. The speaker asserts that humans should remain in the loop and continue to have strong opinions about their work to stay successful. He concludes that while the future is uncertain, the value of caring about the output—code—is a timeless trait of a good engineer. Ultimately, he hopes that humans will never be totally removed from the creative process of problem-solving.

Community Posts

View all posts