An Honest Review of AI Programming: Useful for Search, Useless for Code
After being given a Claude subscription and told to 'tokenmaxxing,' a game developer tried AI tools with skepticism. Over three months, they found LLMs surprisingly useful for searching internal company knowledge bases and summarizing information, but warns against using them to write code. The article highlights the risk of hallucinations, the importance of checking primary sources, and the tendency of AI to generate plausible-sounding but incorrect answers, especially for niche technical questions.
Please don’t ask it to write code.
- tangotaylor
> In a past life I had to argue every year to renew a license for a profiling tool that cost about 20 EUR a month. I’ve heard since that everyone at the company is now getting a Claude subscription, even non-programmers.
Underrated quote. I also found this frustrating. At one company I was at (a very old company which was trying to pivot to software engineering), we had hellish bureaucratic fights with the IT department to get access to Pycharm, Obsidian, and even GitHub. But then the AI craze dropped and management just gave us all GitHub Copilot access without us even asking.
Cory Doctorow's book The Reverse Centaur's Guide to Life After AI talks more about this. It's just a modern symptom of an age-old power struggle. The workers want more control over their craft, including quality standards and tools, but their bosses want more control over the workers.
What's happening now is bosses are feeling pressure from investors to show productivity gains from using AI, so bosses panic-push AI within their companies. Which leads to misaligned incentives like tokenmaxxing.
- adamddev1
Here's a guy who does more than make simple, bloated end-user apps. He tries to dig into the code and make novel performance optimizations etc. He sees that AI can't code everything.
- dr0idattack
I liked the two original experiments in the article, which in a microcosm gels with my 1+ year of deep agentic development. But I would like a whole article about his experience, and with which model (Opus? Fable? erg, Sonnet?) and effort he used.
I've progressed in using latest Claude-kins and the GPTs as usually competent teammate/buddies, and generally know to sort out the fluff confidence with the realz (shoot, that was how I was when I was but a wee little coder lad: overconfident because of an error-free compile and one non-segfault run.)
You have to put in the time, the skill creation, the system prompt/personalization, the (sometimes adversarial) automation, the testing, verification, kicking down the loop castles (as usually caused by being cheeky with highest effort levels.)
- spprashant
> While I have found LLMs useful for researching and planning code changes, my attempts at actually making them write code have been quite lackluster. I found them to be slow and expensive to generate, for a mediocre result.
I think this observation is generally true for the kind of problems the author is working on.
But I would not make the leap to avoid using LLMs for any kind of code writing. LLMs do fantastically well in the 95%+ of the code that engineers spend time on. And for those we should leverage the technology.
It is upto us as engineers to figure out when to stop using LLMs. We are smarter than just dumping logs and half dozen specialized markdown files to a LLM and have it figure out solutions.
- btbuildem
It reads like something written by a time traveler from the past.