Yeah, I have to agree with you. For example, I would have no problem using a decently tested LLMs for engineering simply because Engineering usually accounts for errors and uses appropriate factors to accommodate them. Sure LLMs could be get more accurate in future, but I believe the error will reduce asymptotically. Essentially, more accurate LLMs get, it will get that much harder to increase the accuracy. There is always a price to pay, IMO.