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.”
So is this just early adaptation problems? Or are we starting to find the ceiling for Ai?
Its early adoption problems in the same way as putting radium in toothpaste was. There are legitimate, already growing uses for various AI systems but as the technology is still new there’s a bunch of people just trying to put it in everything, which is innevitably a lot of places where it will never be good (At least not until it gets much better in a way that LLMs fundementally never can be due to the underlying method by which they work)
bright white teeth are highly overrated, glow in the dark teeth, well…wouldn’t a cheap little night light work even better than a radioactive mouth?
“Work” at what purpose, selling product and making investors money? Presumably, no.
My job has me working on AI stuff and it reminds me a lot of Internet technology back in the 90s.
For instance: I’m creating a local model to integrate with our MCP server. It took a lot of fiddling with a Modelfile for it to use the tools the MCP has installed. And it needs 20GB of VRAM to give reasonably accurate responses.
The amount of fiddling and checking and rough edges feel like writing JavaScript 1.0, or the switchover to HTML4.
Companies get a lot of praise for having AI products, but the reality isn’t nearly as flashy as they make it out to be. I’m seeing some usefulness in it as I learn more, but it’s not nearly what the hype machine says.
I also remember the Internet being fiddly as fuck and questionably useful during the dialup days.
AI is improving a lot faster than Internet did. It was like a decade before we got broadband and another before we had wifi.
By that logic, people shitting on AI will look very quaint in a decade or so.
The Internet is and always will be fiddly. We just keep making it so easy that it looks like magic.
“Why do I have to take 5 extra steps to just quickly save a file onto my computer, without needing literally everything on the cloud, especially if I am on a laptop on a device currently in airplane mode, most likely in a literal airplane in an area without reliable Internet connectivity?”
Also consider that there are places - third world nations, and so very MANY areas within supposedly “first-world” ones - that do not have reliable Internet, even today. The KISS principle still applies now, as it did back then too. Your argument screams privileged access, without acknowledging those basic precepts, including perpetual access to subscription services, which must always be maintained, e.g. even after someone retires.
And I disagree in that arguments of the form “LLMs currently do not perform better than my own human effort, in my inexperienced hands at least” will be outdated a decade from now. If LLMs get better, then they will become the musings of people who struggled with early tech before it was fully ready, which does not somehow invalidate their veracity especially in the historical sense.
Those of us with eyes have already seen the ceiling of currently available GenAI “solutions,” which is synonymous with early adoption problems.
The technology will evolve, and the same basic problems will exist. The article has good points about how structured acceptance criteria will need to be more strictly enforced.
Early adaptation and rushed implementation. There may be a bubble bursting for the businesses who tried to “roll out something fast that is good enough to get subscribers for a few months so we can cash in.” However, this is just the very beginning of AI.
Removed by mod