Speaker A
Can you feel it? Do you feel that? Yeah. That, that is the end of programming as we know it.
Exploring the transformative impact of AI on programming, from Bun's rewrite to Eve Online's Python migration and future coding trends.
Key Takeaways
- AI can produce complex, reliable software when paired with strong verification systems.
- Human expertise remains crucial for guiding AI and ensuring quality in software development.
- The programming landscape is evolving, not ending, with AI augmenting rather than replacing developers.
- Large-scale code migrations like Eve Online's Python upgrade still face significant challenges.
- Future software development will increasingly rely on AI agents, but human direction is key.
What the video covers
- The video discusses the paradigm shift in programming driven by AI and advanced models.
- Focus on Bun's rewrite from Zig to Rust using AI, completed in 11 days with significant API costs.
- Comparison with Anthropic's AI-built C compiler, highlighting the importance of formal verification systems.
- Discussion on Eve Online's ongoing Python 2 to 3 migration and challenges involved.
- Emphasis on the necessity of human expertise and deep understanding despite AI advancements.
- Prediction that by 2025, AI-assisted coding will become mainstream in game development.
- Critique of the misconception that AI will fully replace human programmers.
- Insight into how verification systems and test suites enable reliable AI-generated code.
- Reflection on the evolving role of programmers as directors of AI agents rather than mere coders.
- Encouragement for programmers to deepen their knowledge and adapt to new AI-driven workflows.
Chapters
- 00:00Introduction: The End of Programming as We Know It
- 00:43New World of APIs and Frontier Models
- 01:34Human Review and the Rise of AI in Programming
- 02:16Bun's Zig to Rust Rewrite Using AI
- 02:59Verification Systems and AI's Capability
- 03:54Anthropic's AI-Built C Compiler Comparison
- 04:45Human Expertise vs AI: The Programmer's Role
- 06:15Eve Online's Python 2 to 3 Migration Challenges
- 07:01Budget, Verification, and Project Differences
- 07:51Future Predictions and Final Thoughts
Full Transcript — Download SRT & Markdown
Speaker A
This, of course, is motivated by the rewrite of Bun from Zig to Rust. But that's not the only software project making some big moves. Eve Online is moving from Python 2 to Python 3. I can't believe Eve Online was in Python.
Speaker A
[laughter] It just, it currently kind of blows my mind. We also have heavy hitters like Linus really adopting AI coding. And also a little quick prediction: by the end of 2025, everybody's going to be vibe coding video games with Gemini 3 Flash. So yes,
Speaker A
we are going to talk about it. Is this the end of programming as we know it?
Speaker A
Are we entering into a new world, a new and unseen world where access to APIs and frontier models is going to determine the quality of software you have? Or is it business as usual? Or is it actually something just a little bit
Speaker A
different? I suspect it might be the latter of the options. Before we answer the end of programming as we know it, we first have to answer who is today's sponsor and how should we thank them?
Speaker A
Did you know that I've done a lot of weird stuff that's difficult to explain to my father-in-law, such as building a tower defense game in a water tower, pretending to be Mark Zuckerberg in a social network parody in SF, by the way
Speaker A
stands for San Francisco, or even did a live Super Smash Brothers event at Open Sauce. [cheering] Well, if you didn't know those things, you should just go check out the Terminal channel. That's where I do all the fun group videos. Check it out.
Speaker A
Links in the description. All right. So, the end of programming as we know it and what, of course, is the end of programming. This is kind of a consistent argument that's been coming up lately, which is human review is out
Speaker A
the window. It's coming to an end. The days are numbered and that humans writing code, the days are numbered and that's that. And so, it's the end of an era, folks. We are packing it up. Pack your bags. Get the hell out of the
Speaker A
office. Agents are programming. Get used to it. Paul Dix, CEO of Influx Data/creator of InfluxDB, writes this kind of this paragraph right here, which I'd say is the central, like, argument that he's going to make over this, which
Speaker A
is the fact that AI wrote 1 million lines of code and then refined it over the course of the next couple months.
Speaker A
He's talking about the Bun's Zig to Rust rewrite, of which they've actually publicly released and it is actually going really, really well, which was accomplished over the course of 11 days and $165,000 of API spend using models that we may or
Speaker A
may not have access to over the course of the next couple months to produce reliable piece of software that is currently running on millions of developers' machines and is absolutely mind-blowing. I, I actually do agree. I do think that that is quite
Speaker A
mind-blowing. I am not terribly shocked that it worked, but it's still very impressive. It is definitely a step in a new direction, I would say. And you can say, well, it's not that impressive because they had an oracle to compare it
Speaker A
against. So, it was simple to go from one language to another. But I think that's selling the entire thing short.
Speaker A
If you can build a verification system and give proper direction, AI can produce highly complex, highly sophisticated piece of software and it can continue to refine it until it just works. You know what? I'm going to say something. I'm going to say something
Speaker A
and people aren't going to like it. I agree. Yeah. Totally, totally. Absolutely. The rewriting Bun in Rust isn't even Anthropic's first swing at this. You may not remember this, but Anthropic also built a C compiler using a team of parallel clouds and Opus 46.
Speaker A
This wasn't even that long ago. That was just earlier this year. Again, this had the same kind of condition that the Bun rewrite had, which is they really did have the perfect oracle. They had 30 years of tests, hard thought, hard one
Speaker A
kind of code and understanding of what is going on. Plus, they also had within the training weights the GCC compiler several times over. And so they were able to effectively regurgitate and recreate something through the direction and harness of test kind of redirecting
Speaker A
its flow to eventually where it built the highly complicated C compiler. And so I kind of look at these two things as rather similar. They both had exceptional build verification systems.
Speaker A
Bun, in particular, actually had not only an amazing test suite, it also had a formal specification which I would refer to as code that said this is how you should operate. So of course this can be done. And so I do think that he
Speaker A
is right in the sense that if you can provide a formal verification system and give direction, AIs can do incredible things. I will completely seed that ground. The only problem I've seen so far is to build something that's
Speaker A
actually that reliable, the verification system that you have is years of hard-fought knowledge. And I want to say this because I think there's a lot of people that when they see all this, they kind of get freaked out, right? You guys
Speaker A
are just panicking and you're like, gosh, now I'm going to be doing is just I'm just going to be only doing agent stuff. It's like, hold on, just chill out for a second. First off, if you're just a meat proxy, which by the way, I
Speaker A
don't like that term. I think that you're more than a meat proxy. Even if you are just a meat proxy, I think that that is a very devaluing of human, a human humanity statement and it kind of puts you on level with an LLM, which I
Speaker A
don't think you are. But that aside, I would like to just say that if all you do is just direct an AI to figure everything out, well, guess what? You're going to be the first one out the door.
Speaker A
Because you know what? Back in my little young years, okay, my little, little prime years when I was just my formative ones, okay? Those days, you know who got fired, and I've seen it multiple times?
Speaker A
Those that just simply copy-pasted from Stack Overflow and never really had a good understanding of the system. They never took the time to just simply learn the things. I see it as the same person.
Speaker A
Oh, you just copy from Faster Copy Machine. Yeah, sure. It can do a lot more. It's way more sophisticated, but what are you adding? Bro, I could just prompt an agent to do that job. I do agree that there is a change in how
Speaker A
software engineering is going to go on. I just don't know if it is the end of all programming. As far as I can tell, the more knowledge you have of code and the more you actually direct it into
Speaker A
good patterns, the dramatically further one can get. Also, the reason why I bring up Eve is that Eve Online has been in this process of doing a Python 2 to 3 transition since February 2023, even though this article just came out a
Speaker A
couple days ago. And they talk about how they're kind of dealing with some of the hard parts of translating a codebase from one language into another. And one of the hardest parts is how Python 2 and 3 can both compile the same code. And it
Speaker A
turns out they both operate differently. A classic example of this is division. In Python 2, 1 divided by 2 is zero integer division. In Python 3, it's 0.5.
Speaker A
It gets coerced into a floating point if there is a remainder. You can imagine this would, uh, cause chaos. Now, when I see this, I am shocked that this isn't just done by an LLM. Why aren't they doing it? Just like, like honestly, like
Speaker A
that seems like a pretty simple one to be like, "Yo, LLM, make sure all the operations are the same. Here's 20,000 lines. Do it." But for whatever reason, Eve has decided that this is going to be a slower process and they're going to
Speaker A
slowly go through it. Now, the stakes that they're operating at is fundamentally different, I think, than, uh, Bun. They also have a fundamentally different budget than, say, Bun. But there might even be something else. Maybe the verification system is actually
Speaker A
extremely hard for, uh, Eve to be able to have. Maybe it's just not even the same order.
Speaker A
their other output, they've actually been working quite a bit with LLM agents. So it's not like the team is just like, "Oh, we don't use generative AI." know they use generative AI quite a bit because they're in a 30-year-old
Speaker A
code base and a lot of it is just completely unknown because people have moved on or they've forgotten whatever they were doing at the time. Some of it is poorly documented. And so it's not a lack of them using AI. It's that there's
Speaker A
some other thing we don't understand. And of course, I made this joke at the beginning because I wanted you to think about something. This was tweeted in 2025. This is probably the most surefire one of all my predictions because we
Speaker A
already know the outcome of this one. No, not everybody is vibe coding games. No, it's actually still kind of hard.
Speaker A
All the demos you see are actually really crappy. And if you've tried it yourself, it's still not that easy. It turns out me with no knowledge of how 3D games work, I can't just magically make 3D game appear. I'm trying not to crash
Speaker A
out right now, just like I was trying not to crash out a year ago when I read this thing. And of course, I think my favorite is actually the Linus Torvolts one, which is Lionus has been really going hard on the old AI coding, it
Speaker A
turns out. But he just got done doing a graphics driver in which was a debug session from hell as he quoted uh from his change set right in here. But what made it insane is that the LLM said,
Speaker A
"Hey, this is actually an impossible and unsolvable bug. We can't do it. You just need to write a report and move on." But he said that he kept going. He did 24 debug patches and 18 kernel boots and he
Speaker A
found the problem, which is hilarious because the problem is simply a single line of code. It's just one. So again, I'm saying all these things because I wanted you to understand that yes, I I actually am on team like I am now. Okay,
Speaker A
I know it's been a while, but I am I'm I'm on it now that we are entering in a new era of programming. Agents have really officially gotten very very good and you are a bit foolish if you just
Speaker A
think they're idiots. Okay, it's time. It's time to go, okay, they're pretty good. Yes. Okay, fine. Awesome. But expertise clearly still plays a large role. If Lionus can't convince the AI agent that it can do something and he
Speaker A
you know he has access to that like super unguard raid fa Fable right because he's he's he's doing things in which Fable's always like bro I can't do that bro you're asking me to hack the planet I'm not going to do it by the way
Speaker A
fun side note I actually asked Fable to do a little bit of debugging on a game I was working on and it said hey I found a memory leak in your program and I was like oh bro what's what's the what's the
Speaker A
memory leak and it said I can't tell you that would be hacking thanks for the bug report. Super cool, Fable. Really appreciate that one. But nonetheless, it's the exact same thing that I said before. Just simply being a proxy for an
Speaker A
AI means that you at best can be as good as the worst programmer because the worst programmer, they have the same access to the same agents. So, what's going to actually change in the year of the new agent, the new programming? I
Speaker A
really do think technical expertise is going to be a gigantic differentiator. I don't feel like my tune has changed at all on that one. I think reading the friendly manual and learning and learning how to be curious and asking a
Speaker A
lot of questions will now just not pay 3% dividends. It's going to pay like 10% dividends. If anything, I'm more bullish than ever on learning. Okay, that's all I have to say for you now. Okay. Hey, the name the name is I don't think you
Speaker A
should even use the term me proxy. Okay, buddy. I think you're more valuable than that. Even if all you do is copy and pasta from the LLM and I and I would I really do I strongly encourage you to
Speaker A
stop doing that. Take your time, become good at it, read the friendly manual, ask questions, be curious, just do it. A gen.
Topics:AI programmingBun Zig Rust rewriteAnthropic AI compilerEve Online Python migrationsoftware verificationAI-assisted codingprogramming futuremachine learning code generationsoftware development trendsprogrammer role evolution
