Hacker News Reader

Comment by nextaccountic | Hacker News Reader

Training is still very expensive.. and if it becomes cheaper, then a next-gen model with better performance will be expensive again.

It may be very hard to train cutting edge models if you only have consumer hardware

View on HN

Either way, it’ll get to a point where a few generations removed from cutting edge will be “good enough” for existing problem spaces to a degree models are further commoditized. Like any other tech (e.g. iPhones) Next-gen will enable new emergent use cases or continue to eke out diminishing advantages over competition.

Why do you need cutting edge models if you can solve your problems with less?

The gap between the value added by GPT 3.5 and GPT 4 to my every day tasks is immense. Most people, outside of these circles, don't realize that such a large gap exists.

Could I ask what exactly you get from 4 that you don't get from 3.5?

Personally it’s not one specific function but the answer I’m looking for [e.g, code that just works] is significantly better from GPT-4 imo. The better the output is, the less time I spend and the more likely I am to default to using the model in the first place.

I see. So even though 3.5 would attempt to answer your question, the results from 4 are significantly better?

They are required if there is an arms-race between actors wielding models. For example, to detect increasingly authentic deepfakes.

Sure but the clandestine actors most of us worry about don't need that level of sophistication.

Reply on news.ycombinator.com