The Amateur@JeffSBennion𝕏Downloaded this video and watched it really slowed down, a few things I noticed:
1-Note the one finger gesture now seems to move the "camera" like the right-hand button on a gaming controller.
2-Love the intent arrows and lines you see showing lane/exit changes, routes, and shortening when there's uncertainty (like when the two bikes on one side and another bike on the other side is blocking safe passage through).
3-Three finger swipes seem to swap between old map view, new photorealistic view (real-time generated from vehicles own cameras *and* fleet data??? Maybe? How is this transmitted and shared???) and a "quiet view" (that's the IKEA screen at 19s).
3b-Ashok claims this new photorealistic view actually uses LESS total compute than the old blocky models that still have "dancing" objects, at least in my HW4 car.
4-I note that they haven't confirmed a 10x larger parameter count as was earlier mentioned, they just claim 25x safer than prior models and 9x safer than humans. I'll take it! That's a more important metric than parameter count anyway. (I don't think that 25x safer metric is being compared to FSD 14 but some prior FSD/autopilot, because with claims that it's already safer than human driving the math doesn't math; maybe this isn't arithmetic but it seems if it's 9x safer than humans it would have to be some multiple LESS than that compared to FSD 14.) UPDATE: The release notes, while still not specifying the parameter count (I don't really care), clarify this: the 25x is the "critical intervention rate" and probably related to the new vision encoder.
5-Note the "drone's eye parking view" VERY briefly shown at 42 seconds. Very cool, but more on that in the next point.
6-The most intriguing aspect are parts of the real-time view that seem to show parts of the route/area that the cameras could not possibly see. At 40 seconds it's depicting snarled traffic that cameras shouldn't be able to see, and at 42 seconds the parking lot drone's eye view showing occupied/unoccupied spaces the cameras couldn't have known about (unless it has been driving around the whole lot previously). It's probably just interpolated information, but the mention of "fleet intelligence" makes me wonder if some of this information is shared (in real time? Near real time?) between Teslas. The post does say end-to-end fleet intelligence at start and end for now, but doesn't say anything about the recency of the fleet intelligence. I'll bet that this will require premium connectivity. Not even sure about how you share/transmit/download-for-offline-use "end-to-end" navigational and "map" data. Is it regenerated like how weights in LLM models encode all human knowledge and can reconstruct it with prompts? AI researchers have been saying for a long time that model weights are actually a form of compression (training=compression/inference=decompression), and a very efficient one. The mind reels! Tesla fleets exchanging "weights" encoding: real-time traffic data, maps, routes and their efficiency, destinations, parking lot routes, etc., etc.??? Discussing this with @grok in a side convo, it thinks right now it's just interpolating, but the mention of end-to-end encoding of departure and destination navigation makes me think there's more going on here than that.
7-Man that night street with SO MANY potholes at 10 seconds! Is that in our country? If so, it's embarrassing, we look like Tajikistan! (Sorry Tajikistan, your roads are probably way better than this, though last I heard, outside the capital city and major routes, mostly dirt.)
8-What about "human prompt injection"? Anyone can buy a cheap yellow/orange vest and white gloves and cause havoc or worse. I'm sure they're thinking about this! With the proliferation of robotaxies from many different vendors, we may need some kind of authentication/revocation system here like with PKI certificate infrastructure (or maybe blockchain?).
8-The flooded road avoidance at 13s, did you see the condo name, "The