Ashok Elluswamy, Tesla’s Autopilot director, commented on our article and gave additional insight into how Tesla was able to add support for HW3 and what they’re working on next. He confirmed that Tesla’s FSD team was able to achieve similar performance between AI4 and HW3 with FSD V12.5
Similar Performance, Smaller Model
One of the key points Ashok mentioned was that HW3’s FSD 12.5 release has similar, but not exactly the same performance as the AI4 release. This appears to be due to the smaller AI model for HW3, which has been compute-constrained by memory and speed limitations.
Of course, Ashok also tells us to mind the fact that the parameter counts and size of the AI model aren’t exactly everything. Hardware 3 still has the advantage of being the primary focus of Tesla’s FSD teams until the new Supercomputer Cluster at Giga Texas comes online. That, plus the tremendous reams of data that Tesla has acquired from Hardware 3 vehicles on the road today means that the current Hardware 3 version of FSD V12.5 is probably in a pretty good spot.
Larger Model Coming to HW3
It looks like the Tesla FSD team is working on emulating AI4 in Hardware 3 now, as there are some unique hardware-based operations supported on AI4, that aren’t supported on Hardware 3. Ashok says that in order for Tesla to support the larger model on HW3 they’ll need to implement these AI4’s hardware features with software on HW3. He confirms that Tesla is actively implementing and verifying these operations and they plan to support the larger FSD model on HW3 in the future.
Right now it sounds like Tesla is compressing the model, but still able to get close to the same FSD performance of AI4.
We’re looking forward to seeing FSD 12.5 hitting HW3 vehicles in the upcoming weeks if all goes well with initial employee testing.
We were able to achieve similar performance as the AI4 12.5 release with a relatively smaller model for AI3. Deploying the larger model requires us to implement a few kernels in the compiler to emulate the same operations that are natively supported on the AI4 hardware. The team…
One of the keys here is Ashok has focused on mentioning that the same FSD V12.5 model can run on both hardware platforms in the future. He noted that the performance was similar, but not the same. We’re not quite sure how big the difference is going to be just yet, but it looks to be negligible – especially from the positivity behind his post.
Ashok also doesn’t mention FSD V12.6 or future major iterations of FSD, but we aren’t worried. It sounds like Tesla is still committed to supporting HW3 in future iterations of FSD working, so it looks like it’ll be in a good spot for the foreseeable future.
Tesla recently showed off a demo of Optimus, its humanoid robot, walking around in moderately challenging terrain—not on a flat surface but on dirt and slopes. These things can be difficult for a humanoid robot, especially during the training cycle.
Most interestingly, Milan Kovac, VP of Engineering for Optimus, clarified what it takes to get Optimus to this stage. Let’s break down what he said.
Optimus is Blind
Optimus is getting seriously good at walking now - it can keep its balance over uneven ground - even while walking blind. Tesla is currently using just the sensors, all powered by a neural net running on the embedded computer.
Essentially, Tesla is building Optimus from the ground up, relying on as much additional data as possible while it trains vision. This is similar to how they train FSD on vehicles, using LiDAR rigs to validate the vision system’s accuracy. While Optimus doesn’t have LiDAR, it relies on all those other sensors on board, many of which will likely become simplified as vision takes over as the primary sensor.
Today, Optimus is walking blind, but it’s able to react almost instantly to changes in the terrain underneath it, even if it falls or slips.
What’s Next?
Next up, Tesla AI will be adding vision to Optimus - helping complete the neural net. Remember, Optimus runs on the same overall AI stack as FSD - in fact, Optimus uses an FSD computer and an offshoot of the FSD stack for vision-based tasks.
Milan mentions they’re planning on adding vision to help the robot plan ahead and improve its walking gait. While the zombie shuffle is iconic and a little bit amusing, getting humanoid robots to walk like humans is actually difficult.
There’s plenty more, too - including better responsiveness to velocity and direction commands and learning to fall and stand back up. Falling while protecting yourself to minimize damage is something natural to humans - but not exactly natural to something like a robot. Training it to do so is essential in keeping the robot, the environment around it, and the people it is interacting with safe.
We’re excited to see what’s coming with Optimus next because it is already getting started in some fashion in Tesla’s factories.
In a relatively surprising move, GM announced that it is realigning its autonomy strategy and prioritizing advanced driver assistance systems (ADAS) over fully autonomous vehicles.
GM is effectively closing Cruise (autonomous) and focusing on its Super Cruise (ADAS) feature. The engineering teams at Cruise will join the GM teams working on Super Cruise, effectively shuttering the fully autonomous vehicle business.
End of Cruise
GM cites that “an increasingly competitive robotaxi market” and “considerable time and resources” are required for scaling the business to a profitable level. Essentially - they’re unable to keep up with competitors at current funding and research levels, putting them further and further behind.
Cruise has been offering driverless rides in several cities, using HD mapping of cities alongside vehicles equipped with a dazzling array of over 40 sensors. That means that each cruise vehicle is essentially a massive investment and does not turn a profit while collecting data to work towards Autonomy.
Cruise has definitely been on the back burner for a while, and a quick glance at their website - since it's still up for now - shows the last time they officially released any sort of major news packet was back in 2019.
Competition is Killer
Their current direct competitor - Waymo, is funded by Google, which maintains a direct interest in ensuring they have a play in the AI and autonomy space.
Interestingly, this news comes just a month after Tesla’s We, Robot event, where they showed off the Cybercab and the Robotaxi network, as well as plans to begin deployment of the network and Unsupervised FSD sometime in 2025. Tesla is already in talks with some cities in California and Texas to launch Robotaxi in 2025.
GM Admits Tesla Has the Right Strategy
As part of the business call following the announcement, GM admitted that Tesla’s end-to-end and Vision-based approach towards autonomy is the right strategy. While they say Cruise started down that path, they’re putting aside their goals towards fully autonomous vehicles for now and focusing on introducing that tech in Super Cruise instead.
NEWS: GM just admitted that @Tesla’s end-to-end approach to autonomy is the right strategy.
“That’s where the industry is pivoting. Cruise had already started making headway down that path. We are moving to a foundation model and end-to-end approach going forward.” pic.twitter.com/ACs5SFKUc3
With GM now focusing on Super Cruise, they’ll put aside autonomy and instead focus solely on ADAS features to relieve driver stress and improve safety. While those are positive goals that will benefit all road users, full autonomy is really the key to removing the massive impact that vehicle accidents have on society today.
In addition, Super Cruise is extremely limited, cannot brake for traffic controls, and doesn’t work in adverse conditions - even rain. It can only function when lane markings are clear, there are no construction zones, and there is a functional web connection.
The final key to the picture is that the vehicle has to be on an HD-mapped and compatible highway - essentially locking Super Cruise to wherever GM has time to spend mapping, rather than being functional anywhere in a general sense, like FSD or Autopilot.
Others Impressed - Licensing FSD
Interestingly, some other manufacturers have also weighed into the demise of Cruise. BMW, in a now-deleted post, said that a demo of Tesla’s FSD is “very impressive.” There’s a distinct chance that BMW and other manufacturers are looking to see what Tesla does next.
BMW chimes in on a now-deleted post. The Internet is forever, BMW!
Not a Tesla App
It seems that FSD has caught their eyes after We, Robot - and that the demonstrations of FSD V13.2 online seem to be the pivot point. At the 2024 Shareholder Meeting earlier in the year, Elon shared the fact that several manufacturers had reached out, looking to understand what was required to license FSD from Tesla.
There is a good chance 2025 will be the year we’ll see announcements of the adoption of FSD by legacy manufacturers - similar to how we saw the surprise announcements of the adoption of the NACS charging standard.
![Tesla’s Optimus Robot Learns to Walk Without Vision [VIDEO]](https://www.notateslaapp.com/img/containers/article_images/optimus/optimus_fall.jpg/59ea0e3082db081e7473a1f0da1ed31c/optimus_fall.jpg)
_300w.png)
![Tesla Updates Santa Mode & Rainbow Road to Use Ambient Lighting and Updates SteamOS [VIDEO]](https://www.notateslaapp.com/img/containers/article_images/tesla-screen/ambient-lighting-santa-mode-rainbow.webp/bd6dd7459fc8963516a33efd3bd54466/ambient-lighting-santa-mode-rainbow.jpg)
![Tesla Builds Cybertruck Tunnel While Production Lines Pause [VIDEO]](https://www.notateslaapp.com/img/containers/article_images/2024/cybertruck-tunnel.webp/822a2c10124f46fe13e0d45cefdd9d33/cybertruck-tunnel.jpg)
![Impressive Tesla FSD Videos: Making Room for Vehicles & Mastering Tight Maneuvers [WATCH]](https://www.notateslaapp.com/img/containers/article_images/2023/fsd-beta-vis-3.jpg/9c0bfda120e9fecc4ddb111b350565ac/fsd-beta-vis-3.jpg)
