>>106300283
There are new architectures being produced, LLM's are just where the money is right now no?. Over the past few months we already saw a Hierarchical Reasoning Model with only 27 million parameters outperform LLM's with tens of billions of parameters on the ARC AGI test (narrow, but it is a highly abstract task)
>>106299635
The current AI's abilities are not evenly distributed and despite not matching a 50th percentile human in all areas, they are useful. I have an engineer friend who is considered irreplacable and is a high performer using it for tasks.
Of course it is not autonomous and lacks abilities in cognitive areas, otherwise it would be able by definition to "specialise" and would by definition be capable of everything a 50th percentile human is (barring physical constraints such as actuator and battery tech)
>experiments
An ASI would have practically unlimited ability to run experiments in parallel at least. It would certainly have superhuman organisational capability and presumably would be capable of gathering resources at an accelerated rate.
Although yes there is some element that cannot be parallelised and would need to run in serial.
