| ▲ | unknownfuture 6 hours ago | |||||||||||||||||||||||||
That's still not the kind of efficiency the article is talking about. They're referring to raw training and inference scaling and it's relationship (or lack thereof) to traditional economies of scale that we've seen with past technologies where they get cheaper as adoption increases, not more expensive. Its true that the models requiring fewer turns and tokens due to increasing sophistication improves cost efficiency for users but that doesn't address the fundamental computational scaling problems of LLM architectures. | ||||||||||||||||||||||||||
| ▲ | simianwords 5 hours ago | parent [-] | |||||||||||||||||||||||||
This doesn't mean anything. Here are the facts 1. models themselves are getting cheaper - around 5000x in the past 1.5 years 2. ironically, if it were fully cheap, the same crowd would say that this would make AI bubble pop because where would they get moat? 3. training is also getting cheaper, the whole lifecycle to train and do inference on models from 1 year ago is on the whole 50x cheaper or so I also think the author's point is a bit nebulous > Chatbot companies are aware that their products are inefficient. Some have found techniques for improving performance, but they have not yielded significant gains Performance has increased. To you specifically: what metric would falsify the claim that "they have not yielded gains" and "their products are inefficient"? The same metric should apply to - industries like steel - pharma - internet/cloud computing - automobile - consumer electronics All of them have had high impact. So any criticism on inefficiency should be unique to LLM's and shouldn't apply to all of them. Please answer. | ||||||||||||||||||||||||||
| ||||||||||||||||||||||||||