Tesla is training a new FSD model with ~10X the parameters and a big improvement to video compression loss.
Probably ready for public release end of next month if testing goes well.
Video compression improvements (reducing data loss during training) have contributed ~20-30% additional uplift in past updates by enabling higher-fidelity inputs.
From Phil Beisel [x-Apple / x-Rivian (tech team founder)] who is making an educated guess about the video compression.
Raw frame data from 8 cameras is extremely large, and transmitting it over cellular or Wi-Fi is both time- and cost-prohibitive. Tesla likely uses a lossy compression model to reduce this burden, but then decompresses it back into a raw photon representation during training.
Inference (i.e., real-time processing on the car) always uses raw photon data directly.
This still bypasses the traditional ISP (see article). And to be clear, this isn’t like JPEG-style compression, which discards a lot of fidelity. Tesla’s method compresses the data for transport while preserving much of the original signal quality.
A 10x parameter increase could yield ~3-5x improvement in miles per intervention (MPI). The baseline version in Austin is about one intervention per 5000-15000 miles per critical intervention. With video compression adding ~1.5-2x uplift, overall ~4-10x better reliability is plausible in optimistic scenarios. This might push to 20000 to 150000 Miles per critical intervention.
In a recent post, Elon mentioned a new FSD model—assumed to be version 14—that has 10x the parameters of version 13.
He also referred to a “big improvement to video compression loss.” This might sound confusing, given prior comments about using raw photon counts.
However, I… https://t.co/xKJf9Ac9nS pic.twitter.com/Xin02L5eL0
— phil beisel (@pbeisel) August 6, 2025
Qualitative Improvements in Perception and Decision-Making
10x parameters + better compression should enhance handling of edge cases (e.g., construction zones, pedestrians, weather), reducing hesitations and false positives by 50-70%.
Smoothness and Speed: Faster processing (from efficient memory use) could cut latency by 20-30%, making drives feel more human-like.
Autopilot and FSD have about 6.7 million miles per crash and if this is improved by 2-3x better then the miles per crash could be one in 13-20 million miles.
Tesla’s robotaxi pilot in Austin (launched mid-2025) has driven geofence expansions and software tweaks flowing to consumer FSD.
There should be better pickup/dropoff logic.
Urban-specific enhancements (e.g., 2x fewer interventions in dense traffic after one month of testing).
Tesla is working as quickly as possible to get 100+ Teslas operating for autonomous ride-hailing (can’t use the word “taxi” or “cab” in California) in the Bay Area and allow anyone to request a ride. Allowing anyone to request a ride for the entire Bay Area service area would likely need 1000+ Tesla autonomous vehicles.
Tesla is training a new FSD model with ~10X params and a big improvement to video compression loss.
Probably ready for public release end of next month if testing goes well.
— Elon Musk (@elonmusk) August 6, 2025
We are working as quickly as possible to get 100+ Teslas operating for autonomous ride-hailing (can’t use the word “taxi” or “cab” in California) in the Bay Area and allow anyone to request a ride
— Elon Musk (@elonmusk) August 6, 2025
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Which is it? FSD and autonomy are great and ready for the big time. Or, FSD will benefit from 10X more parameters.
Does FSD need AI5 and AI6, why? What happens to the AI3 and AI4 cars out there?
It seems there is an infinite pool of cases that need to be handled…before the $T money pump starts flowing.
FSD and Tesla autonomy are already making billions but FSD and robotaxi going scaled without safety monitor is a month to a few months away.
The 10X parameters is for the AI4 cars. There is some distillation type shrinking of the models to fit what they can onto AI3.