It all started when I came to the realization that I haven’t been seeing any new FSD Beta test drive videos in the news any more, and I started wondering why. The answer is because Tesla hasn’t updated the FSD Beta since FSD Beta 8.2 came out on March 4th, 2021. For Tesla owners and non-Tesla owners alike, this article should be pretty interesting, as we have made a timeline of both Autopilot progress since 2014 as well as FSD updates, but before we get to that, there are some basics that you should be aware of.
As a non-Tesla owner, you might think that Tesla pushes out an update to owners once a month. The reality, however, is somewhat different. In fact, Tesla pushes out updates once or sometimes even twice in one week. These updates are usually small bug fixes, and in some cases are specific to certain regions or versions. Tesla’s updates usually come in the following format: 2021.4.18.2 — which is the latest version that was released on June 5th. Here is another example: 2020.48.37.2. Now, you might think that this second update is one from last year, and while, yes, the last 2020 updates did have a very similar number, the year 2020 for most Teslas ended with 2020.48.35.9. Instead, 2020.48.37.2 was an update released on April 29th this year at the exact same time as version 2021.4.15.5. The difference is that the 2020 one was intended for Tesla vehicles with a version 1 media control unit. Those vehicles still get updates, but they are a lot less frequent and contain fewer goodies than the main update branch (limited by hardware). Nonetheless, kudos to Tesla for still supplying Teslas built almost 10 years ago with the latest UI and software.
I believe I have said this once before, but I will say it again: there isn’t another company on earth that is as good as Tesla at keeping all hardware up to date with software, despite the fact that in the factory all kinds of component upgrades under the hood quietly take place on a weekly basis. If you have ever had to deal with hardware drivers on Windows, you might be in a better position to appreciate the magnitude of Tesla’s accomplishment.
Now, here is a timeline of Tesla Autopilot starting from when HW1 was first implemented in 2014:
So, yes, that timeline turned out to be more comprehensive than I had originally planned, but it should at least put things into perspective. For example, did you realize that from the time the FSD beta was first launched until now represents only 9.5% of Autopilot history? Or that it took less than 5 years to get to the FSD Beta from the moment Tesla parted ways with Mobileye? The Windows 10 operating system has been around longer than that. In fact, had Tesla parted ways with Mobileye just a few months later, the time period would have completely coincided with the US election, from when Trump won until the day Biden won.
Considering that Tesla is doing something that has never been done before, I think all the criticism we hear from earnings report to earnings report about Tesla being late is a bit hollow.
FSD Beta Update Timeline
The news of today has to do with an FSD milestone — though, not a very good one. As was mentioned in the introduction of this article, it has now been 100 days since the last update, so let me put that into perspective. Here is a list of Tesla updates made available to FSD Beta drivers, when they were released, and how long it had been since the previous update:
FSD Beta begins
- 2020.40.8.10 | October 21 | Tesla releases FSD Beta to early testers
- 2020.40.8.11 | October 24 | 3 Days since last update
- 2020.40.8.12 | October 31 | 7 Days since last update
- 2020.44.10.2 | November 14 | 14 Days since last update
FSD Beta 6
- 2020.44.15.3 | November 27 | 13 Days since last update
- 2020.44.15.4 | November 30 | 3 Days since last update
FSD Beta 7
- 2020.48.10.1 | December 14 | 14 Days since last update
FSD Beta 8
- 2020.48.12.5 | December 24 | 10 Days since last update
- 2020.48.26.1 | December 30 | 6 Days since last update
- 2020.48.35.1 | January 17 | 18 Days since last update
- 2020.48.35.6 | January 22 | 5 Days since last update
FSD Beta 8.1
- 2020.48.35.7 | January 26 | 4 Days since last update
FSD Beta 8.2
- 2021.4.11.1 | March 4 FSD | 37 Days since last update
Today — 100 Days since last update
Just to help paint a picture, this kind of update schedule is also very normal for regular Tesla update releases, and for the non–FSD Beta vehicles, updates have continued rolling out at this same uninterrupted pace.
Elon’s Update Forecasts
Just for fun, we used our internal database of Elon Musk tweets (available to paying subscribers) to also keep track of Elon’s promises regarding this latest FSD Beta update, which is 84–97 days late. I’m snipping or paraphrasing the important information here:
- March 6 — “Release is 2-3 weeks” (March 20-27)
- March 6 — “Release in 10 days” (March 16)
- March 9 — “End of next week” (March 20)
- April 9 — “Almost ready with FSD V9 Pure vision, no radar” (Almost?)
- April 15 — “Major improvements are being made to the vision stack every week. Beta button hopefully next month.” (May)
- April 15 — “Button timing of May is aspirational. Depends on how well limited beta of V9.0 goes, but I would be surprised if wide beta (aka button) is later than June.” (May, maybe June)
- April 29 — “Probably 2 weeks” (May 13)
- May 12 — “a week or two to polish pure vision FSD & v9 beta will release” (May 19-26)
Radar and 360 Video Combine Like Oil & Water
Now comes the most important part of this article, and it’s the reason for why this update is so late. On the surface, it may seem like it is the removal of radar that has caused the delay, but the root of the problem actually lies a lot deeper.
For quite some time, Elon has been talking about changing the way Autopilot works. Rather than process the video feeds of individual cameras, the computer will now stitch all that footage together into one massive 360 picture and both process and learn from that. In all these years, Elon was so optimistic about Autopilot progress partially because the vehicle should be able to learn from data and improve almost exponentially. However, the reason that didn’t happen was because learning from multiple separate cameras just didn’t work. In fact, if you watch the FSD Beta YouTubers, you will hear them say how quickly the car is learning and improving with every update that comes out once every couple of days. That is how Elon kept promising us it would work.
Radar is Like a Narrator Who Tells You What You are Looking At
Tesla has slowly been transition to 360 video — or 4D, if you will — for quite a long time. We know for a fact that it was working on this during all of 2020 and is still working on it today. Yes, some neural nets already run on 4D, but part of them still run on 2.5D, and the fact that both can run in parallel is truly a testament to the sheet might of the Tesla HW3 SoC. Once Tesla finishes this upgrade and no longer needs to run 2.5D, that will free up a lot of capacity in its SoC, which can then be dedicated to better processing or run more processes simultaneously. Basically, there will be more things the car can keep in mind and account for while driving.
However, when Tesla came around to switching radar telemetry from 2.5D to 4D, that is when they ran into problems. It makes sense, one or two forward facing camera, matching their telemetry to that of a radar, while complicated, is doable. However, doing this to a sphere of 360 video, it just does not work. If you have ever tried to video edit 360 video before, as I have, you might start to understand the difficulty. Rather than simplify the process, as 360 does, radar is a step in the opposite direction, complicating things by trying to tell the cameras what it thinks the cameras should be seeing, rather than just focusing on looking at it with your own 8 eyes. It’s like a real-life narrator, intended for the blind — you stand in front of a chair and it says “chair.” For a blind person, that might be nice, but for you, it would definitely get on your nerves.
Why It’s Taking So Long
So, Tesla has quite correctly made the decision to remove the radar. This eliminates the shortfalls of radar, like phantom braking caused by signs and overpasses, but also eliminates the strengths of radar, like identifying what is and is not stationary. Vision is able to do so, but it seems likely that Tesla had not previously focused on that because it was not a priority. Tesla likely did not foresee that 360 video and radar telemetry would be like oil and water. So, it is taking a painfully long amount of time to get Autopilot back to the same level of reliability as it was before.
This all fits in logically with how Tesla explained its FSD improvement process. It chooses the weakest chain in the link (like stop signs), collects data on it, processes it, strengthens it a few orders of magnitude, and then moves on to the next weakest link. However, the task of recognizing moving vs. stationary objects is a lot more difficult than recognizing stop signs. Like I said, it all makes a lot of sense given what we know.