manoftheforest

@LimitingThe What about us people, who like driving? You lost me when you detached the steering wheel, from wheels. Love the instant speed through 😁

This is the clearest 'convergence' roadmap Elon has provided thus far for Tesla's AI chips. I'll try to provide more color here with my own thoughts/speculation/interpretation: 1) AI4 will provide self-driving safety beyond human levels. In either AI4 or AI5 there will also need to be enough compute/reasoning to allow you to converse with the car in tricky situations. I'll be interested to see which version that shows up in. 2) AI5 will push the long tail of 9's to the point where human fatalities are almost eliminated. It will also allow an initial Optimus product for limited deployment for specific use cases or early adopters. Optimus will require significantly more compute than vehicles. 3) AI6 will open enough functionality in Optimus that it can be mass manufactured and deployed for most use cases. Elon's companies will start building their data center inference compute around this chip. +It will still be used for vehicles. 4) AI7 will allow for higher operating temperatures and also potentially better radiation hardness. The higher operating temperatures will allow for lower radiator mass for space-based compute, reducing the launch costs. The radiation hardness isn't necessary for the first low orbit space-based data centers because off-the-shelf chips are suprisingly radiation hard, but could improve capital efficiency. This chip will also be capabale of large scale training compute, offering Elon's companies an alternative to Nvidia. DOJO never died, it just transmuted. Notably, Elon formerly said AI8 was going to be "out of this world" so he's potentially advancing the timeline or at least which version that capability will show up in. This would put mass deployment of orbital compute at around 2029-2030? +It will still be used for vehicles and Optimus