Ask HN: What LLM are you all using for coding assistance right now?
.
The network (at present) has 2 inputs(x,y), and 3 outputs (r,g,b) and is training to minimize loss. With 10 hidden layers of 10 elements per layer, it's oddly impressionistic so far.
I've tweaked things so that the batch size is the whole image, and I'm using FP64 weights. I have a clue about what the python code is doing, but not a huge clue. It reminds be of the things that I did back when I took the free AI classes from Stanford about a decade ago.
The inspiration that kicked this off was this video by Welch labs[1] "Why Deep Learning Works Unreasonably Well". Normally I'd just ponder such things, but thanks to LLMs, I can actually see just what works and doesn't.
I wonder just how detailed an image I can get from weights that are on the order of 50k or less. I'm willing to throw hundreds of CPU hours at it. Seeing the images as the training progresses is fascinating.
DeepSeek (with thinking & reasoning & search turned on)
Claude Code
QWEN Coder (with thinking & reasoning & search turned on)
ChatGPT
Google Gemini
- DeepSeek: excellent at coming up with solutions and churning out prototypes / working solutions with Reasoning mode turned on.
- Claude Code: I use this with cursor to quickly come up with overviews / READMEs for repos / new code I'm browsing and in making quick changes to the code-base (I only use it for simple tasks and don't usually trust it for implementing more advanced features).
- QWEN Coder: similar to deep-seek but much better at working with visual / image data sets.
- ChatGPT: usually use it for simple answers / finding bugs in code / resolving issues.
- Google Gemini: is catching up to other models when it comes to coding and more advanced tasks but still produces code that is a bit too verbose for my taste. Still solid progress since initial release and will most likely catch up to other models on most coding tasks soon.I want to preface by saying: I'm kind of talking out of my ass with my reasoning. I make a lot of assumptions based on my current understanding of how models work and what I've noticed from first-hand experience.
Gemini is based on BERT and I've found that tends to work better when there's output based on a context that goes back and forth. For example, if I have some text that corresponds with another set of text somewhere else and there needs to be back referencing done.
Claude Opus 4 has been quite good with full featured implementations, but only when it's given a huge amount of context to work with. Like so much context that it's a borderline specification.
GPT-4.1/5 has been good as just an all-arounder. It's what I'll generally default to unless I know my use case matches the other two models above.
Qwen3 coder 30b I've finally sorted and was using it all day with qwen code.
Qwen3 30b thinking is likely still better than coder?
Free gemini 2.5 pro is my backup for the tough problems.
Tomorrow though. LM Studio just released their latest version which greatly improves tool calling with GPT 20b. I'm running it at 120k context and medium reasoning. I'm pretty confident it's about to become to go-to.
claude-4-sonnet: seems to be the best at tool calling and actually changing the lines
gemini-2.5-pro: solves what sonnet can't solve, but you have to run it a couple of times to get the tool calling mistakes out
Cheers!
I tried chatGPT, claude, claude code, cursor and ampcode. I sticked to claude code at the end.