AI coding assistance discussion

[igor_kavinski]

I'm in Heaven. Did I die and end up in Heaven?

ARC A770 16GB thought for SIX SECONDS and solved the problem at a blistering speed of 150.53 tokens per seconds.

 

[igor_kavinski]

With AI this good, who wants to code??? https://forums.anandtech.com/threads/the-ai-discussion-thread.2611025/post-41443464

 

[igor_kavinski]

Qwen3 30B A3B Q8 is broken, at least for Win32 C++ programming prompts. It repeated itself four times in about an hour, thinking from scratch about the problem. Had to end it because it seemed to be stuck in a thinking loop.

 

[igor_kavinski]

AI debugging: https://forums.anandtech.com/thread...s-to-have-promise-software-debugging.2629706/

 

[Sgraffite]

I have used a couple LLMs such as ChatGPT, Claude, etc... but the way I am using them I am asking a question and then making changes to code based on the response. I guess there are fancier AI versions that can act directly upon your code. Some can even create ,run, fail and keep iterating on unit tests in the background until the unit tests pass. Maybe that is what having an AI agent means?

Does anyone have experience with the more advanced AI like I described and if so which one(s)?

 

[igor_kavinski]

Does anyone have experience with the more advanced AI like I described and if so which one(s)?

Haven't experienced but heard great things about: https://aider.chat/

 

[Sgraffite]

I was trying out Claude.AI and ended up paying for Pro. Now I have been messing around with Claude Code the last couple weeks and it seems pretty good. It has an MCP so it can take actions. It can write unit tests and iterate on them until they work. It can run curl commands, stop/start servers, etc... There's a ton of features I have not even used yet.

 

[igor_kavinski]

His prompts are on his github so that's convenient to test out your LLM hardware performance.

 

[igor_kavinski]

I was trying out Claude.AI and ended up paying for Pro.

I think this guy also used Claude to do something pretty complex: https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md

What I'm curious about is the cost, though. How much did all this cost? Claude isn't free.

I do hate all this AI hype talk without anyone really mentioning the cost involved. Usage limits exist and it is not clear how much typical usage costs.

 

[Sgraffite]

I think this guy also used Claude to do something pretty complex: https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md

What I'm curious about is the cost, though. How much did all this cost? Claude isn't free.

I do hate all this AI hype talk without anyone really mentioning the cost involved. Usage limits exist and it is not clear how much typical usage costs.

I've been using it almost every night now on personal projects to prototype different things. The plan I am on is less than $20 USD a month.

I had a Commander Keen project from a couple years ago. I set up Pocketbase and had Claude make a tile mapper tool and a level editor. It saves the levels to the database so you can share them. I had it implement the tile palette, minimap, undo/redo an eyedropper to choose a tile and a bunch of other stuff, it is pretty neat. The tilemapper saves a lot of time as I can dynamically choose tiles and save them to the database as well. It saves the tile names and sprite sheet x & y coords. I can redefine any tile and save it, then test a level again and the change will be there because the game loads the tiles from the db each time you run it.

 

[igor_kavinski]

I've been using it almost every night now on personal projects to prototype different things. The plan I am on is less than $20 USD a month.

How many requests per day do you think it can manage on that plan before it tells you that you have run out of tokens?

 

[Sgraffite]

How many requests per day do you think it can manage on that plan before it tells you that you have run out of tokens?

It's not that straightforward in my experience. From what I can tell now this plan is broken into 5 hour sessions. Once you begin/resume a conversation the 5 hour period starts. You have some number of allowed usage during that 5 hour period, likely not just specific to tokens but it is hard to tell. If you slow play it you can go the whole 5 hours without getting timed out. If you ask it more complicated things you can burn through that usage in an hour or less and then have to wait 4 hours to resume it.

For example have it make some minor CSS changes you might not even get limited within that 5 hour session. Having it write frontend/backend unit tests and debugging deadlocks/async bugs on a page using web sockets, it would chew through it quite quickly.

I would say if I am actively using it every chance I get, I usually won't get timed out for about 2.5-4 hours depending on complexity. If you don't get timed out within the 5 hours I'm pretty sure you just silently start another 5 hour session.

 

[igor_kavinski]

I would say if I am actively using it every chance I get, I usually won't get timed out for about 2.5-4 hours depending on complexity. If you don't get timed out within the 5 hours I'm pretty sure you just silently start another 5 hour session.

Thanks for the great explanation!

Still not sure if I want to use it though. When I'm trying to make something work, I tend to want to finish it ASAP. Hitting a usage limit and having to wait hours will be insanely frustrating for me.

 

[Sgraffite]

Thanks for the great explanation!

Still not sure if I want to use it though. When I'm trying to make something work, I tend to want to finish it ASAP. Hitting a usage limit and having to wait hours will be insanely frustrating for me.

It can potentially get you there way faster so that seems like it would meet your goal, it just depends. Writing tools for side projects that I never really plan on finishing I would never do normally, but if AI can make it in 1-2 hours and add a bit of polish in another hour, that is also a lot of fun for me in general. Making a level editor and tile mapper manually would have taken me a couple weeks of after hours time normally and now I could do it while also doing some dishes or vacuuming or working out, etc...

It's probably one of the best uses of money I have spent in a while.

 

[void_pointer]

I have been doing a little hobby coding since late last year with various local models as well as ChatGPT, Claude, Gemini, etc. and I have found that local models almost never meet the performance of the major online providers. Finally I gave up and just went for the Github Copilot subscription.
I am tempted to say that Microsoft is pushing Copilot too hard with its integration into VS Code (my favorite IDE), but I have to admit that I never got around to trying to enable a local agent in Code because the local agents I was using weren't up to the task.
Just recently, I used Copilot to create a quick demo of a classic Snake browser game. I could probably cook all this up pretty quick myself since I have been using OpenGL since the 90s, but I wanted to see how quickly I could get Copilot to do the same thing. The result: in about three hours, if memory serves, I had a 3d snake game in my browser. Pretty awesome.
Would I be able to do the same thing if I hadn't been a c/c++ slave from the day? I don't think so. IMO, the AI is more likely to be helpful than harmful for people who can already code, but noobs beware.

 

[igor_kavinski]

The result: in about three hours, if memory serves, I had a 3d snake game in my browser. Pretty awesome.

WebGL based? Screenshot if you don't mind? A video would be even better!