Excerpt:
“Even within the coding, it’s not working well,” said Smiley. “I’ll give you an example. Code can look right and pass the unit tests and still be wrong. The way you measure that is typically in benchmark tests. So a lot of these companies haven’t engaged in a proper feedback loop to see what the impact of AI coding is on the outcomes they care about. Lines of code, number of [pull requests], these are liabilities. These are not measures of engineering excellence.”
Measures of engineering excellence, said Smiley, include metrics like deployment frequency, lead time to production, change failure rate, mean time to restore, and incident severity. And we need a new set of metrics, he insists, to measure how AI affects engineering performance.
“We don’t know what those are yet,” he said.
One metric that might be helpful, he said, is measuring tokens burned to get to an approved pull request – a formally accepted change in software. That’s the kind of thing that needs to be assessed to determine whether AI helps an organization’s engineering practice.
To underscore the consequences of not having that kind of data, Smiley pointed to a recent attempt to rewrite SQLite in Rust using AI.
“It passed all the unit tests, the shape of the code looks right,” he said. It’s 3.7x more lines of code that performs 2,000 times worse than the actual SQLite. Two thousand times worse for a database is a non-viable product. It’s a dumpster fire. Throw it away. All that money you spent on it is worthless."
All the optimism about using AI for coding, Smiley argues, comes from measuring the wrong things.
“Coding works if you measure lines of code and pull requests,” he said. “Coding does not work if you measure quality and team performance. There’s no evidence to suggest that that’s moving in a positive direction.”
Yeah these newer systems are crazy. The agent spawns a dozen subagents that all do some figuring out on the code base and the user request. Then those results get collated, then passed along to a new set of subagents that make the actual changes. Then there are agents that check stuff and tell the subagents to redo stuff or make changes. And then it gets a final check like unit tests, compilation etc. And then it’s marked as done for the user. The amount of tokens this burns is crazy, but it gets them better results in the benchmarks, so it gets marketed as an improvement. In reality it’s still fucking up all the damned time.
Coding with AI is like coding with a junior dev, who didn’t pay attention in school, is high right now, doesn’t learn and only listens half of the time. It fools people into thinking it’s better, because it shits out code super fast. But the cognitive load is actually higher, because checking the code is much harder than coming up with it yourself. It’s slower by far. If you are actually going faster, the quality is lacking.
I code with AI a good bit for a side project since I need to use my work AI and get my stats up to show management that I’m using it. The “impressive” thing is learning new softwares and how to use them quickly in your environment. When setting up my homelab with automatic git pull, it quickly gave me some commands and showed me what to add in my docker container.
Correcting issues is exactly like coding with a high junior dev though. The code bloat is real and I’m going to attempt to use agentic AI to consolidate it in the future. I don’t believe you can really “vibe code” unless you already know how to code though. Stating the exact structures and organization and whatnot is vital for agentic AI programming semi-complex systems.