>>13442715I'm not too well versed in AI myself, but I'll consider my knowledge to be above anyone in this thread. Or atleast 99% of this thread.
There's two modes for AI. Training and Inference.
1) Training machines require great deal of computing power. A high end GPU from nVidia/AMD can do training on consumer level AI tasks and it would take couple of hours/days worth of stuff to chew through a neutral net that learns to detect text to voice speech, image recognition, face recognition, etc. A cluster of consumer GPU can chew through these tasks in few hours what it may take few days. Further more, a single Google TPU can do these in few minutes what it takes a cluster of consumer GPU few hours. A cluster of Google TPU can do these in few seconds what a single TPU can do in few minutes and so on. Google has thousands of TPU clusters. Amazon has thousands of their own TPU type clusters, so does Microsoft, Oracle, Huawei, etc
We absolutely have the hardware for training. But what about inference? What is inference? Inference machine just does pattern matching in real life, or in pseudo-real time. For example, Apple/Google smartphones have tiny inference AI chips that can detect faces/cars/objects/etc. Tesla cars has a slightly larger 100 watt inference machine in them that do neural net driving. Comma AI uses an inference machine from another consumer level smartphone to do something similar to Tesla's AI but is much smaller. So we can already do inferences on a stupid low budget for decent results.
The government, google, apple, huawei, any large tech company can scale these tiny inference machines to extremely large inference machines that do not care about power usage or cost. So hardware absolutely isn't the issue for any large company/organization.
