Ron Gilbert (GrumpyGamer)My Dinner With AI
Kasiandra Richmond> Which brings me to a point. [...]. But at some point I’ve spent so much time writing the spec that I could have just done it myself. A
This is what I was observing during our AI Workshop.
> I hear about new devs freaking out when they run out of tokens for AI. They are lost without it.
Same story I observed with the Cloud and new Ops.
> I’m going to enjoy programming and not being a baby sitter to an AI programmer.
The best lesson one can learn from this.
hogsy@grumpygamer good write-up which matches my experience experimenting with it. You're spot on about managers seeing it as magic... Been seeing this happen where I work; I was highlighting the potential shortcomings and time it can waste when used improperly, and the response was like dealing with a cult.
I get the impression, particularly with Claude, some people feel like they need to justify their expense for their new toy and will defend it at all costs.
Ein „Primat“@grumpygamer thanks ❤️
Stephen Illingworth@grumpygamer I keep trying this stuff every few weeks and have zero success every time. Even just initialising a project is a challenge. I just don't get it.
Stuart McHattie 👨🏼💻@JetSetIlly which tools are you trying? As someone who has never failed to use Claude Code for a programming task, I find a zero success rate really hard to believe.
I know how to write the software manually. I’ve done it for almost 40 years! But AI has changed what I get to focus on. I can worry about the finicky details of how something works without worrying quite so much about all the boilerplate that enables my choices.
@JetSetIlly ok fair dos. Though I see you’re using glm-4.7-flash for the model which Anthropic haven’t optimised Claude Code to work with so it could be an incompatibility. And being a local model, it could be a bit weak compared to real Claude models.
Have you tried your local model with OpenCode? https://opencode.ai
It’s more suited to non-proprietary models.
@sdjmchattie I guess this stuff just isn't for me. But that's okay because I like programming.
If I were a painter, I wouldn't stop painting just because somebody invented the camera obscura.
Timberwolf@grumpygamer This mirrors my experience so much, even down to the random appearances of `const` in C code. There's a point I reach in pretty much every extended session with a coding agent where I start thinking, "I'm using exactly the techniques I'd use to manage a lazy employee who doesn't realise how blatantly obvious their attempts to avoid doing the job they've been asked to do are".
muddle@grumpygamer Liking the Dinner with Andre shot.
alethenorio@grumpygamer good read. It's not the first time I hear similar experiences from people using C or C++. I program Go on a daily basis (Cloud systems) and I find that it gets things right more often than not. Maybe certain types of environments are more suited for the current iteration of AI?
Ask it to fix a flaky bug though and anything can happen 😂
Future SprogThis has been my experience as well:
> The clearer you are the better AI is. But at some point I’ve spent so much time writing the spec that I could have just done it myself. And it’s not just a high level spec, you have to get into the weeds.
PeevishDave@grumpygamer 100%. I'd add one great problem is AI tries always to tackle the problem head-on,
while 90% of the times it's better to work on the problem than on the solution.
That is, simplify/transform the problem to a point the solution becomes straighforward.
I guess that's why it failed everytime I tried to let it solve some vector math problem,
and in the end I had to do the work.
while 90% of the times it's better to work on the problem than on the solution.
That is, simplify/transform the problem to a point the solution becomes straighforward.
I guess that's why it failed everytime I tried to let it solve some vector math problem,
and in the end I had to do the work.
@grumpygamer About copilot, I know it's lame, but I'd like to have some different additional buttons on the keyboard
to accept suggestions or just pause them (it kills me when I just wanted to add a 'tab'). I guess such shortcuts are remappable
but for some reason I don't want to do it.
Also, funny thing. I made a dialogue tool in vscode before the copilot support, and now it butts in and suggests unsolicited lines of dialogue.
So I need to turn it on and off every time I switch from/to code
to accept suggestions or just pause them (it kills me when I just wanted to add a 'tab'). I guess such shortcuts are remappable
but for some reason I don't want to do it.
Also, funny thing. I made a dialogue tool in vscode before the copilot support, and now it butts in and suggests unsolicited lines of dialogue.
So I need to turn it on and off every time I switch from/to code
felix (grayscale) 🐺@grumpygamer the agentic coding sweet spot is when you tell an agent to make a change that passes the testsuite. it might not get it right first try, but it can retry on its own until it does pass tests, and in the meantime you can play Freecell or burn yourself out checking the output of the other 79 agents doing stuff for you.
and you can ask the agent to first write tests for you, which works as well as anything else.
(of course, testsuites are hard to do for videogames)
mirabilos@grumpygamer no. It does not need to change.
One simply does not use the planet-burning fashtech.
nikolasdi@grumpygamer Great article.