In late 90s, my thesis "Recognition of 3D objects using Artificial Neural Networks" was a model that were aimed to identify real world objects in robotic applications, and trained with RST 3D object database (rotation, transition and scale invariant ). It is later used in military applications like automatic target detection in radars . (Long after I left Airforce) This was in late 90s and I was not early in this area as there were thousands other applications like that.
The AI you are using today is another type of artificial neural network that is trained with blocs of text (LLM) that predicts most likely words given a sentence . Not much different than what we had for decades.
These are still like first computer from 1946 (ENIAC) , that weights 30 tons with 18000 vacuum tube transistors.
Current AI models use separate data and processing units which require enormous energy to transfer and process the data. And need to be retrained from the scratch for more advance learning unlike human brain that contains memory and processing unit in the same entity and retrains itself on the fly mostly during sleep, using very low energy .
There will be a gradual evolution in AI over time , converging to biological systems
Therefore AI will be a cyclical technology with many local bubbles and pops unlike crypto or internet that were one and done.
Many bottlenecks you know of right now will disappear in time as AI systems converge to biological systems and many semiconductor chip companies like NVDA will disappear like Kodak if they dont adapt and innovate.
Next big revolutionary advancement in AI will be the systems with completely different topology containing memory and data processing uniformly distributed in the same unit with self training capability . Whoever invents these chips and design these units will be the biggest richest tech company in the world and almost all other tech will either work for it or disappear.