When they are trained for a specific task, they do noticably better. Hallucinations aren’t guaranteed to never happen but it is a lot closer to pure accuracy with niche AI. LLMs that you see are made for pleasurable conversation. AI/ML other than LLMs can become really to the T. I’ve seen YouTube videos testing such.
“noticably better” than what? You need to ask yourself that question.
Machine learning solutions should never be in control of any irreversible actions. Slowing down a car due to perceived danger? Okay, worst case driver has to accelerate again (emergency braking decisions must not be made by such a system either to avoid being rear-ended). Swerving to avoid an obstacle? Absolutely not!
As I said in another comment: There are driving assistants that work deterministically. Anything else can give visual or auditory clues but shouldn’t be allowed to control much if anything else.
The question of how good an ML system is, is irrelevant as they are non-deterministic by nature, and their legitimate use cases are elsewhere.
Than humans and even expert humans. For specific, narrow tasks super human behavior has long been achieved in many classification tasks. Like tumor detection in XRays for example.
ATC jobs are extremely stressful and complex and they have surprisingly little digital support systems. A digital system could for instance easily suggest an optimal route through the airspace based on the aircraft’s capabilities and data and other airplanes in the air space before the pilots even checked in with the ATC. Today ATCs still have to request and suggest course changes via radio even though both them and pilots are sitting in front of computers.
When they are trained for a specific task, they do noticably better. Hallucinations aren’t guaranteed to never happen but it is a lot closer to pure accuracy with niche AI. LLMs that you see are made for pleasurable conversation. AI/ML other than LLMs can become really to the T. I’ve seen YouTube videos testing such.
“noticably better” than what? You need to ask yourself that question. Machine learning solutions should never be in control of any irreversible actions. Slowing down a car due to perceived danger? Okay, worst case driver has to accelerate again (emergency braking decisions must not be made by such a system either to avoid being rear-ended). Swerving to avoid an obstacle? Absolutely not!
As I said in another comment: There are driving assistants that work deterministically. Anything else can give visual or auditory clues but shouldn’t be allowed to control much if anything else.
The question of how good an ML system is, is irrelevant as they are non-deterministic by nature, and their legitimate use cases are elsewhere.
Than humans and even expert humans. For specific, narrow tasks super human behavior has long been achieved in many classification tasks. Like tumor detection in XRays for example.
ATC jobs are extremely stressful and complex and they have surprisingly little digital support systems. A digital system could for instance easily suggest an optimal route through the airspace based on the aircraft’s capabilities and data and other airplanes in the air space before the pilots even checked in with the ATC. Today ATCs still have to request and suggest course changes via radio even though both them and pilots are sitting in front of computers.
It’s like Ubuntu vs firmware. One is made to be able to do a ton of things, and one is made for a very specific purpose.