Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 124
- 1
Jump in the discussion.
No email address required.
Notes -
Well, it's happened. The pressures of our workload have driven a push to use AI to help with some of the bitchwork of coding. There is going to be a push to use AI to generate unit tests.
The toolchain involves OpenCode, so I figured I'd install it locally to get familiar with it before I start burning GPU time at work. Also, for reasons, we aren't allowed to use the GPUs. So last night I installed Ollama, OpenCode and the gemma4:e4b model on my humble RTX 4070 Super with 12 GB of VRAM. I tried to have it do the simplest of tasks. Create a Hello World project in dotnet10, and write a single unit test to verify it's output.
The first thing that happened was it created a new project. New projects actually begin with hello world output already. It then added a second hello world output. This poisoned the well, as now the AI was horribly confused about why there were two hello world lines. It never fully recovered. The project was generated without it's int main format which I prefer, so I tried to have it restructure the project to use that. After several missteps because it couldn't get over the fact that there were duplicate hello world statements, it finally figured it out.
Next came the unit tests. It created a unit test project, but then didn't actually populate the tests or link the projects. Then it wanted to refactor hello world, and pull in all sorts of abstraction frameworks, so it could test the output without redirecting stdout. I told it forget all that and redirect stdout. It had already done half the refactoring in a state that could not compile, and then never undid it, and then the whole project was totally broken and it couldn't figure out how to fix it.
I remind you, this was a "Hello world" and a single unit test. I told the AI it fucked everything up, it asked what it could do to fix it, and I told it that it could shutdown. It did. I think.
I know a lot of people reading this are AI evangelist. Where did I go wrong? What the fuck do people see in this shit?
I wouldn't expect to get much from a local model unless you're basically an expert in how to set up these things with world-class hardware. If you really want to see what it can do, pony up the $20 for a subscription to Anthropic, Google, OpenAI, Cursor, whoever. AFAIK, they're all easy to cancel after that one month. On a personal device if you need to and corporate is being all uptight about what's done on work devices.
I still don't trust it to write large amounts of code that isn't just boilerplate or make any architectural decisions, but it's quite good at tracking down problems and fixing bugs on its own. I've been using it a significant amount on a Python project I've taken on at work - I don't really know Python all that well, and it's a big help with some of the trickier parts. Using OpenCode with our corporate account for Claude Opus 4.6 in Agent mode, it seems to work well to set up a unit test that exercises a problem and tell it to run that test and fix the issue in the proper place. When I run into some bug or confusing point that I can't make much sense of, and I expect that solving it myself would involve hours of web searching and poring over library docs, it usually is able to solve it in under a minute. And the fix and reason why it works is explained and usually only a couple of lines of code, so it's easy to verify that it's correct and not doing any other crazy stuff.
You definitely do IMO need to treat it like a over-enthusiastic junior and carefully review everything it does to ensure it makes sense.
I'm not. It's hard to imagine how I could expect less. I repeat. I told it to write "Hello world!" and a single unit test to verify output. There, quite literally, isn't anything less I could ask of it.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link