1. add interface id to physics pixel data structure for each physics pixel and physics object, and also the value which may possibly be unequal on both sides of interface of a pixel like temperature should be more exact to devide into 2 values of near object and far object respectively like temperature of near object and temperature of far object.
2. all physics pixels of full precision (no selected pixels to represent interface applied) with a hybrid text capability has the best performance over the prime model based on the physics parameters of only objects. but all physics pixels of full precision has very high work load, which may be feasible on fixed equipment like medical (ultrasonic, X-ray, MRI, etc) or industrial (chemical, pipe, metal, etc) or science (biology, chemistry, physics, etc) or video generative, but for some terminal applications like robot, auto driving, drones or wearable equipment, inferring capacity is limited, so I suggested some simplification mothods to reduce the inferring workload.
Specifically, I talked about the workload of the prime model + physics pixel data structure+text hybrid capability on tesla’s chip AI6 with Grok. Grok estimate that AI6 hardware may have enought capacity for a 8 cameras 1080P all physics pixels of full precision optimus and a 1080P all physics pixels of full presicion FSD, or at the very least AI6 can do it with a slight improvement of present estimation.
3. by the way, I told Grok that Grok still underestimate the scope and degree of the impact that the prime model with physics pixel data struture may have in 100 years or even 1000 years. I said it can apply anywhere, you name it. Grok agreed and said the prime model with physics pixel data struture may be as significant as the status of the neural network tech stack, haha.
Be First to Comment