Tesla‘s cars will stay on its current AI4 self-driving computer, with the next-generation AI5 chip going to the Optimus humanoid robot instead, Chief Executive Officer Elon Musk said on Sunday.
“It’s going to Optimus. Cars will remain at AI4. That’s enough to solve superhuman driving,” Musk wrote on X, in reply to an owner who asked when AI5 would be in new cars.
Musk made the remarks as he explained how the new Reality View display in Full Self-Driving (FSD) V15 is generated live, by a separate diffusion model that uses half of one AI accelerator on the car’s computer.
“We fork off from these UVE features, then run a separate ~1B diffusion model,” Musk wrote in reply to an owner’s review of the software.
Tesla began rolling out V15 to a first wave of customers on Saturday, and the update replaces the car’s on-screen visualization with what the company calls Tesla Reality View.
Over the following hours, Musk and members of Tesla’s AI team explained how the view is made and how much of the car’s computing power it uses.
A Smaller World Model
Musk started by describing a larger model.
“We’ve had a ~pixel-perfect world model since last year, which is critical for synthetic data generation,” he wrote. Synthetic data are computer-generated driving scenes used to train and test driving software.
The model in the car, he said, is a “(much) smaller version” of that system: “a conditional mean-flow model, based on the recent sequence of Tesla Universal Vision Encoder features.”
Those features “are the same ones that are plugged into the 3B encoder model that produces the driving actions,” Musk wrote. He did not spell out the units, but AI developers commonly use “B” for billions of parameters, a measure of a model’s size.
More on Tesla
- Tesla’s FSD V15 Goes to Wide Release Next Month, Musk Says23h ago
- Tesla Targets Cybercab Customers by Year-End, Says Most Stay Tesla-OwnedOctober 22
- Musk Says Optimus Will Take at Least a Year to Reach 20,000 Units a WeekOctober 21
- Tesla Starts FSD V15 Rollout, Adding Reality View and Speed-Camera Slowdown20h ago
Denoising Reality
A diffusion model, the technology behind many AI image and video generators, produces a picture by starting from random noise and removing it.
Tesla’s version starts from its own last frame. “This takes in the previous reality-view generation with added noise and then denoises it to the current visualization,” Musk wrote.
“It then produces a small video chunk (8 images at 36 fps, at the resolution of the UI screen),” he added. “This work is done by 1/2 of a TRIP engine every UVE frame at 18 hz and is spread over 4 cycles.”
Eight frames at 36 frames per second make up about 0.22 seconds of video, the same span as four cycles at 18 per second.
TRIP is the neural-network accelerator inside Tesla’s own self-driving chips. Musk’s wording indicates that each chip in the AI4 computer, also called Hardware 4, has three of them. Tesla has not published a specification sheet for AI4.
A vision encoder is the part of a neural network that turns camera images into features the rest of the system can use. Tesla’s V15 release notes do not use the Universal Vision Encoder name, but describe an “upgraded vision encoder shared between FSD, Robotaxi, and Optimus, which is trained on millions of internet videos.”
Squeezed for Compute
In a follow-up post, Musk said Reality View had to fit into what was left on the computer’s second chip.
“After we moved our tiny decoder into SRAM, we only have half a TRIP out of three on SoC B, for this reality visualizer, so spent the compute as best we could,” he wrote. SRAM is fast memory built into the chip itself.
“TRIP basically only handles 8-bit int and has to run softmax in emulation mode in 28 steps,” Musk added. Softmax is a calculation used throughout the transformer models behind modern AI systems.
He said the display still lags. “Currently, the latency is a bit high due to compute constraints, but we have ways to interleave newer UVE features along the diffusion transformer layers to cut roughly 35% of the latency,” he wrote.
Musk did not give a latency figure or a date for the fix.
On Saturday, Musk had called Reality View “real-time AI video generation (no gaming engine!) using just half of one AI4 TRIP accelerator.” Hours later, Ashok Elluswamy, Tesla’s Vice President of AI Software, wrote that the view was “Produced end-to-end, pixels-in-pixels-out.”
AI5 Goes to Optimus
Musk ended his follow-up with Tesla’s next chip. “AI5 is a gigantic improvement. Roughly 12X more compute,” he wrote.
The owner who asked about AI5 had written: “When will AI5 be in new cars? I feel like I should wait to upgrade from my AI3.”
In January, Musk wrote on X that “AI5 will make the cars almost perfect and greatly enhance Optimus.”
After Tesla taped out AI5 in April, Musk wrote: “A single AI5 has ~5 times the useful compute of a dual SoC AI4.” He did not say on Sunday what his 12X figure was measured against.
On the third-quarter earnings call on October 21, Musk said Tesla aimed to put AI5 into the Optimus humanoid robot “in approximately April of next year,” calling it “an order of magnitude more compute, AI5 versus AI4.”
Separately on Sunday, Musk wrote that “AI4 costs twice as much now due to memory and storage price hikes.”
Musk did not say whether the Cybercab, Tesla’s two-seat robotaxi, is covered by “cars will remain at AI4.” The Cybercab could run a pre-production AI4+ computer, two AI4 boards or an early build of AI5.
‘A Few Months Ago’
Tesla’s self-driving team said the display came together quickly.
Phil Duan, whom Tesla promoted to Director of Engineering for Autopilot in February and whose X profile says he leads self-driving at Tesla, posted a 51-second clip of Reality View on Saturday, U.S. time.
“Overheard in Palo Alto: GTA 6’s real-world mode just dropped on Teslas,” he wrote, referring to the Grand Theft Auto video game series.
Lawson Fulton, whose profile says he leads Reality View at Tesla, quoted the post. “Fun fact, when I was interviewing with @pduan 4 years ago he said: ‘One day, I want looking at the viz to feel just like you’re looking outside,'” Fulton wrote. “It didn’t feel possible at the time. But here we are!”
“And we only started working on it a few months ago,” Duan replied.
Hours earlier, he had written: “We’ve always wanted the UI to show the world through the eyes of the FSD model. We were just too busy solving the driving part first.”
On Sunday, Fulton quoted Musk’s explanation and wrote: “We are hiring world-class engineers at every level of the stack!”
Simulator Roots
The world model Musk referred to traces back to work Tesla presented in 2025.
In an October 2025 article on X summarizing his talk at the International Conference on Computer Vision, Elluswamy wrote that “at Tesla, we have developed a neural world simulator.”
“The world simulator is trained entirely by Tesla to produce all of the vehicle’s cameras and other sensor data,” he wrote. Engineers can connect it to the driving model to test it “in a closed-loop fashion,” the article said.
25 Times Safer
Tesla’s managers also expanded on the safety claim in the release.
“v15 is 25x safer than v14 across the fleet,” Duan wrote on Saturday. “That’s hard to notice on any individual drive because most drives are already very safe. The difference only becomes clear when you aggregate millions of miles.”
“Some people thought we meant 25% improvement, no we mean 25 times!” Musk wrote on Sunday.
Tesla’s release notes tie the figure to “the critical intervention rate.”






