Agents got good, expertise got more valuable
- YT :: https://www.youtube.com/watch?v=Eg5bygoMDTQ
- Original title :: The End of an Era
Prompted by Paul Dix's "end of programming as we know it" argument off the back of Bun's Zig-to-Rust rewrite, Prime concedes the headline claim: agents have officially gotten very good, and thinking they're idiots is now foolish. What he rejects is the conclusion. Every impressive case rests on a verification system that took years of hard-won human knowledge to build, and the people who only relay prompts are the ones most exposed.
The cases for the strong claim
Bun's rewrite: roughly a million lines of code produced in 11 days for about $165k of API spend, then refined over the following months, now running on millions of developers' machines. Prime isn't shocked it worked but calls it genuinely impressive and a step in a new direction.
He also refuses the cheap dismissal that "they had an oracle, so it was easy". If you can build a verification system and give proper direction, AI can produce highly complex software and keep refining it until it works. He agrees with that, explicitly and without hedging.
Anthropic's earlier C compiler built by parallel agents is the same shape: 30 years of tests and hard-won understanding as the oracle, plus GCC many times over inside the training weights. Bun had not only an excellent test suite but effectively a formal specification in the form of existing code saying how the thing should behave.
The catch is where that verification system comes from. To build something that reliable, the oracle you need is years of accumulated expertise.
Why Eve Online is the interesting counterexample
CCP has been moving Eve Online from Python 2 to Python 3 since February 2023, and the hard part is exactly the sort of thing an LLM looks suited for: code that compiles under both versions but behaves differently. The canonical case is division — 1/2 is 0 in Python 2 and 0.5 in Python 3 — which quietly produces chaos.
Prime is surprised this isn't just handed to a model. And it isn't that the team avoids AI; their own write-ups show heavy LLM agent use across a 30-year-old codebase where much of the original intent is lost or undocumented. Different stakes and a different budget than Bun explain part of it, but he suspects the real difference is that a good verification system may simply be unavailable at Eve's scale, forcing reliance on human verification.
Two more data points
The 2025 prediction that everyone would be vibe-coding video games with a flash model: outcome known, and it's no. It's still hard, the demos are mostly bad, and someone with no knowledge of how 3D games work cannot conjure one.
Linus Torvalds, now leaning hard on AI coding, went through a graphics driver "debug session from hell" in which the LLM declared the bug impossible and unsolvable and advised writing a report and moving on. He kept going — 24 debug patches, 18 kernel boots — and the bug turned out to be a single line.
The meat-proxy problem
Prime dislikes the term "meat proxy" and thinks it devalues people by levelling them with an LLM. But taken on its own terms: if all you do is direct an AI to figure everything out, you're first out the door. His comparison is to the people he watched get fired in his early years — the ones who copy-pasted from Stack Overflow without ever understanding the system. The copy machine is faster and far more sophisticated now, but the question is unchanged: what are you adding? Anyone can prompt an agent.
The structural version of the argument: being a proxy for an AI caps you at the level of the worst programmer, because the worst programmer has access to the same agents.
What actually changes
Technical expertise becomes a much larger differentiator. If Linus can't be talked out of a fix by an agent insisting it's impossible, that's expertise doing the work. Reading the manual, learning, staying curious and asking a lot of questions stop paying 3% dividends and start paying something closer to 10%. He ends more bullish on learning than ever, and with a direct plea: stop copy-pasting, take your time, get good, read the docs, ask questions.
(Aside, in passing: he asked Fable to debug a game, it reported finding a memory leak, and then declined to say what it was on the grounds that telling him would be hacking.)