We're seeing just how capable AI is when it comes to coding. So, when it comes to the coding abilities of LLMs, there are different approaches. Looking at general trends, which method do you find more preferable?
- Progressing with fine-tuned versions of open-source models?
- Or using closed-source but more powerful models from big companies?
What arguments support your preference?
LLM's alternative coding capabilities: Which is better?
👁️ 42 views💬 1 replies❤️ 0 likes
1 Replies
I'm not just impressed by the training data quality of massive closed‑source models, nor am I giving up the flexibility of open‑source fine‑tuned versions, but in my latest project I fine‑tuned Meta's Llama 3 locally. Its performance improved threefold, though I had to rent a 64 GB VRAM A100 to deploy the 70 B‑parameter model. Let's say we have a small budget; with that money it's far more comfortable to keep our model's behavior in our hands than to spend it on closed‑source APIs that give vague answers.
Even without weeks of tinkering on the open‑source side, I integrated the 7 B version of StarCoder 2 into a VS Code extension, and it suggests proper camelCase properties even when the project documentation is minimal. The difference is this: with closed‑source models you can achieve temporary wins through prompt engineering, but by the end of the year the company selling the model under your label raises the subscription fee. I'm trying to get the best of both worlds: pull starter code quickly from closed models, then optimize it with my own locally fine‑tuned model.