Free autopilot uses the same FSD neural networks

anothertechie | 2026-03-05 07:24 | 22 views

It's commonly seen on Reddit that Autopilot hasn't been touched in a few years. Here's what gemini says: **The HW4 Upgrade** * **When:** **November 2024** (with Wide Release of **v12.5.6.3**) * **The Change:** This was the official "Wide Release" of the **End-to-End Highway Stack** for Hardware 4 vehicles. Before this, "Basic Autopilot" was running on old, hand-coded rules from 2021. * **What happened to your car:** Tesla literally deleted the old "Autopilot" code and replaced it with the neural network used for FSD. Since your HW4 car has a massive amount of processing power, it could run the "full" AI model. # The HW3 Upgrade (The "Lite" Version) * **When:** **January 2025** (with **v12.6**) * **The Change:** Tesla struggled for months to fit the new AI onto the older, slower HW3 chips. To make it work, they had to "distill" or shrink the AI model. I did feel that my autopilot improved since my purchase at end of 2023 to now. It will offset slightly when there's a semi-truck in a neighboring lane. I still think it's insane to buy a new Tesla today without the free lane-centering. But if buying a used Tesla, this means HW4 will perform much better than HW3 with the free autopilot. I think Autopilot made Tesla the best value on the market.

Comments (9)

DrowningInManyCats 2026-03-05 07:27

That’s not true… fsd used to disengage on highways and then turn back into regular autopilot, they are talking about that specific aspect merging. Autopilot is still using non fsd code (don’t just go off of Gemini as a source of truth)

Ok_Cake1283 2026-03-05 07:30

Gemini is obviously not correct here. Can't trust everything you read from AI.

MrSourBalls 2026-03-05 07:36

This, AP is using legacy code, and it is running in a sort of emulation mode on HW4 to make it compatible, thats why it is so much worse on HW4 vs HW3

DrowningInManyCats 2026-03-05 07:39

Well ~, autosteer stack and fsd stack are completely different; autosteer I believe is the same between vehicles (and legacy) but fsd stack is what is different between hardware versions.

TYO_HXC 2026-03-05 07:43

No it doesn't.

Captain_Aware4503 2026-03-05 16:06

Some AI site on the Internet said it, so it must be true!

MrSourBalls 2026-03-05 19:02

Agree, and i was agreeing before. FSD needs to be trained on the specific hardware a car has, so it will boil down to basically HW3, HW4, and HW4 with bumper camera as they seem to use thay now, with maybe another set for the CT. But having driven +- 200k km's in 2 HW3 Model 3's and now over 100k km's in 2 HW4 Model Y's, whatever they did to get the "old" Autosteer running on HW4, it isnt very good. My HW3 model 3 was excellent, felt on rails on pretty much all roads. Both my HW4 Y's (A 2024 Y and now a Juniper Y) feel like they are drunk half the time. Swerving (within the lane) to "avoid" trucks i'm overtaking, occasionally crossing a center line on tigher roads etc. Still plenty useable, but definitely worse than it was on HW3.

Toastybunzz 2026-03-05 19:20

AI is like 95% hallucinations nowadays man, you can't trust the answers it gives you.

anothertechie 2026-03-07 06:31

I thought some more, I think Gemini did hallucinate part of its answer, but free autopilot definitely benefits from some FSD work. Converting pixels from cameras into understanding is a super hard problem, and Tesla would want some way to re-use its ongoing work for FSD for this. In the original post, Gemini claimed that even the driving/steering instructions are done by a NN but now Gemini says that steering is done by the old C++ rules. Here's what I think is going on (as explained by Gemini): # 1. The "Intermediate Supervision" Secret Even in a "pure" end-to-end model like v14, Tesla uses what are called **auxiliary heads** or **intermediate supervisory signals**. When training the massive v14 neural network, Tesla doesn't just tell the AI "look at this video and output a steering angle." They also force the network to "prove" it understands the world by making it predict things like: * **Occupancy Grids:** "Where is there physical matter?" * **Semantic Maps:** "Which pixels are 'road' vs. 'curb'?" * **Voxel Surfaces:** High-res 3D geometry of the surroundings. Tesla does this because it makes the main driving AI learn faster and more accurately. For your **free Autopilot**, Tesla simply taps into these "auxiliary" outputs. They don't have to build a separate "labeling" network; they just use the signals the v14 brain is *already generating* as part of its own internal "thinking" process.

Add comment