Mike LaBarbera (@ANSR42) header
Mike LaBarbera (@ANSR42) profile picture

Model Y + comma four. 14k+ open source miles. No longer renting FSD for $99/mo and training my own AI models at home.

Get $50 off a comma four ➡
Joined November 2021
Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)Photo by Mike LaBarbera (@ANSR42)
My FSD subscription ended last January, and this isn’t Tesla Autopilot. I installed a comma four in February, eight months of Level 2 ADAS I own outright in my Model Y. This is sunnypilot, an open source fork of @comma_ai's openpilot. No longer renting a closed system.
201
61
77
1,937
1,815,391
The openpilot community like @BarhamIsaac is always working to improve upon the already brilliant work of @comma_ai.
Nevermind. I fixed it. A little fiddling and some work with a canker and I've got turns as crispy as FSD with no navigation. There's absolutely 0 reason for me to type in an address for FSD. Just follow my blinkers. This is the magic of FOSS. @comma_ai
11
509
Mike LaBarbera retweeted
Nevermind. I fixed it. A little fiddling and some work with a canker and I've got turns as crispy as FSD with no navigation. There's absolutely 0 reason for me to type in an address for FSD. Just follow my blinkers. This is the magic of FOSS. @comma_ai
Hey @comma_ai Love everything about the product but I’d love it more if the model didn’t try to curb my wheels 😭 How do I help you guys get models with turning love?
4
3
1
32
4,751
A road-line AI model that only works on its own training drives doesn't tell you much, so I hid one drive from mine. The footage is from one road camera on my @comma_ai comma four. Then I ran the model on it. It holds lanes and road edges on the straight, and after the turn even in sun glare. Mid-intersection it predicts almost nothing, and at the end the right edge blobs onto a parked car.
1
1
17
1,326
Same drive from my last update.
I got ahead of myself on my Tesla Autopilot visualization. Cars drifted, details smeared, and the road underneath wasn't solid. So I went back to basics and started with the one thing it needs first: teaching a model where the road actually is. The first training run of my road-line and road-edge model took about 1 h 17 min on my RTX 4060. The labels are my @comma_ai comma four's own model output from about 20 hours of my drive logs (1,460 segments), locked to the road as the car moves. It's rough. Cross streets mostly stay blank unless I turned onto them, and the edges run right through parked cars. My model copies those mistakes for now, and I haven't tested it on the drives I held back yet. Next I clean up the labels and train it on more footage.
1
81
NVIDIA's BirdDreams generates a top down view of the road. Mine is measured from one @comma_ai comma four camera on a real drive, and it's rougher. Rebuilt it with GPS + OpenStreetMap for the roads, openpilot odometry for my path, YOLO11m for the other cars and my own depth model to place them. Theirs draws both the camera view and the top down map, trained on driving data where the cars were labeled with 3D boxes and headings. Mine has real flaws. My depth model still reads about 40% far, there's no heading model so some boxes face the wrong way, and the map is GPS snapped to OpenStreetMap, which is why it spins through the turn. Just a cool experiment I figured I'd try. Map data © OpenStreetMap contributors
3
28
1,881
Paying $104.45 a month for FSD is getting hard to defend. So far, over 14,000 miles in my Tesla Model Y with the @comma_ai Comma Four. It's all open-source driver assistance, hands-free on hardware you own. Pick between 350+ supported models. FSD's moat won't last long.
216
44
11
702
253,655
I got ahead of myself on my Tesla Autopilot visualization. Cars drifted, details smeared, and the road underneath wasn't solid. So I went back to basics and started with the one thing it needs first: teaching a model where the road actually is. The first training run of my road-line and road-edge model took about 1 h 17 min on my RTX 4060. The labels are my @comma_ai comma four's own model output from about 20 hours of my drive logs (1,460 segments), locked to the road as the car moves. It's rough. Cross streets mostly stay blank unless I turned onto them, and the edges run right through parked cars. My model copies those mistakes for now, and I haven't tested it on the drives I held back yet. Next I clean up the labels and train it on more footage.
3
1
1
27
2,165
This was my last update on my Tesla Autopilot visualization.
I rebuilt Tesla's Autopilot's visualization using only what my @comma_ai comma four logged on one drive. The bird's eye 3D view comes from those real logs. The cyan path is what the model planned, and it tracked the lead car. The other cars come from my own object detector run on the road video. The catch is cars still drift off the road and some face the wrong way. Next I'm working training the model to lock them in place. The real road camera sits in the bottom right hand corner, synced frame for frame. This was all rendered afterward on my RTX4060, so I think running it live on the Chestnut eGPU would be a good test.
3
381
Qwen VL can see my road just fine. Ask it where to drive on that road and it goes somewhere else. I gave @Alibaba_Qwen Qwen2.5 VL 7B 4 seconds of my comma four drive footage and asked where I'd go in the next 5 seconds. My blinker was on and my lane was ending. It planned to stay in it. Green is where I went. Red is where the AI guessed. After 5 seconds it was 10 ft off and on the road shoulder. That's not a knock on Qwen. This model was never built to plan a drive. Ran locally on my RTX 4060, on a recorded drive, not in the car.
2
31
2,322
Qwen does have one that is called Qwen-Drive, and it's opensource. Same clip, same question. That's my next test. Inspired by a commenter talking about throwing Qwen on a comma four and the KITScenes LongTail dataset: arxiv.org/abs/2603.23607
2
242
270 days ago with no plan beyond having lane assistance on unmarked roads in my Tesla Model Y. I stuck a @comma_ai comma four to my windshield. Nearly 600 posts, just under 2,000 replies, and 4.7 million impressions later, I'm training my own AI models and still learning more each day. Thank you to everyone who reads, replies, and pushes back. You've made this way more fun than I ever expected. I'm now 112 verified followers away from the 500 needed for the 𝕏 Original Content Rewards program. If you've enjoyed the content and I show up in your feed, a follow would mean a lot. Either way, thanks for being here.
I joined the @comma_ai fleet today. Installing in my Tesla Model Y 2020, updates to follow soon.
9
2
1
75
2,796
I told my small driving model to turn left, then right. It planned the same path both times. I'm distilling my own model from @comma_ai's big openpilot driving model. Progress has been slow, so I added a blinker input to see how it would react, hoping it would force the turns. Then I realized I never trained the model on a blinker. 🤦😅 Blue is forced left, orange is forced right. The model wasn't affected by the blinker I added. Still learning as I go.
1
23
1,960
Mike LaBarbera retweeted
It's super easy to install #DashKit in your #Tesla. It takes about 2 minutes. Watch our install guide! ⏱️
3
1
7
836
I rebuilt Tesla's Autopilot's visualization using only what my @comma_ai comma four logged on one drive. The bird's eye 3D view comes from those real logs. The cyan path is what the model planned, and it tracked the lead car. The other cars come from my own object detector run on the road video. The catch is cars still drift off the road and some face the wrong way. Next I'm working training the model to lock them in place. The real road camera sits in the bottom right hand corner, synced frame for frame. This was all rendered afterward on my RTX4060, so I think running it live on the Chestnut eGPU would be a good test.
8
2
2
76
4,033
My last update on the Tesla autopilot visualization:
Two dumb rules got the car directions right. My trained model, three runs in, still points some parked cars backwards. I'm rebuilding Tesla's Autopilot screen from my own drives recorded on my @comma_ai comma four and part of that is showing which way every car around me is pointing. I tried three ways to guess it. A simple lane rule and a motion rule both match the street, so the oncoming pickup and the red car face me. Run 3 of the model still flips a few cars around. It's a work in progress.
1
4
367
Tesla won't retrofit my HW3 Model Y to HW4, so I pulled a side panel and installed an RX 9060 eGPU myself. This is my @comma_ai Chestnut eGPU under the center console, used for running large open-source openpilot AI models, giving HW4-level compute. Plenty of room down there.
87
62
14
1,524
294,903
Sun glare is one of the best reasons to add lidar. Mainly because a camera only driving assistance system can become blinded and wobble. I've driven into a low sun many times with my @comma_ai comma four and I haven't seen it struggle yet.
9
24
4,823
Mike LaBarbera retweeted
CommaView is a small project. If it's useful to you, Support CommaView, in Settings › Support & diagnostics, opens our Buy Me a Coffee page: it helps cover hosting, testing and the iOS app. It unlocks nothing. buymeacoffee.com/commaview
1
6
221
Mike LaBarbera retweeted
No one is sending their 3 year old to daycare in a Cybercab.
96
22
11
551
27,536