Hacker News Reader

Hear hear. I know a 3 person startup that has a "lead AI researcher" who is trying to train and fine-tune models. That's not their startup's purpose though... they have an actual product. So wtf are they doing? The lead AI guy thinks he's going to compete with these big companies and it's total fantasy.

LLMs are a commodity

View on HN

That does indeed sound crazy. But finetuning is also a commodity these days. You can train a good Mistral LoRA in under 24 hours on a single consumer GPU. We’re talking about $10 of compute.

You can run a dozen of these LoRAs atop the same base model on the same infrastructure for a dozen specific use cases.

The inference quality, performance and cost can all be substantially better than GPT4 with prompting.

Doesn't it entirely depend on how specialized the training data for a given fine tuned model might be?

Reply on news.ycombinator.com