Yes, this has been his project for many years now - he threw a fit and fled to China in 2016 when the NHTSA sent him a Special Order requesting test data for the Comma One product, which he then cancelled and released as pseudo open-source (the ML model is still a closed black box).
Yours is the very essence of a HackerNews comment. Comma is everything a startup should be. They are not wrapping a lame business model in CRUD and living off of malinvestment. They are solving ridiculously hard problem with a small team of very smart people. Their competition has burnt billions. Meanwhile, Comma is profitable, has a better safety model than anyone.
This is a straw-man. I did not compare Comma to any other startup, or some bizarre ideal of "what a startup should be."
I strongly disagree that Comma "have a better safety model than anyone" - their entire safety model revolves around two things, an assertion that the driver is fully responsible, and a poorly documented set of callback functions in their CAN-intercept board which, depending on the vehicle, do different things, but generally try to limit angular input and introduce an emergency override. This is smart, but the code is not documented well enough to be considered a "safety model" in any commercial sense, and the system, while it has a MISRA C linter plugged into it somewhere, is not built using safety-approved hardware or to any known safety standard. Plus, the fundamental approach on most cars involving CAN message-stuffing simply cannot be as reliable as an integrated system - in that sense, the entire model is flawed and cannot "have a better safety model than anyone."
And, Hotz explicitly and openly left the United States and cancelled his first Comma product launch, rather than respond to a letter NHTSA sent him asking questions about Comma's testing protocols and safety process. Call that what you will.
> pseudo open-source (the ML model is still a closed black box).
?
I am confused with this statement. Sure the ML model is a black box, but it's better than closed source completely with no model. It's more realistic to build the software yourself than training your own self driving ML system.
The most fundamental part of the system, the one which makes driving decisions, is not open. I did not say anything about whether or not this was "better" than the product being fully closed source, only that it is not truly open, and I fully believe this. "Open source autopilot" implies to me that the autopilot is open - that an end-user can inspect, audit, and attempt to understand the decisions their vehicle is making. This is not the case for Comma - rather, it is an open-source CANbus translation layer attached to a closed source autopilot.
When you say not everything is open source. I assume you mean the training code is not open source? I'm curious what you would want to learn from that? You wouldn't be able to actually train a model since you wouldn't have access to the data.
The end-user can inspect, audit and understand the decisions their vehicle is making. All you have to do is see how the neural network behaves for different inputs. That's the correct approach, whether you have access to the training code or not.
Comma don't even say _how_ the model works! What layers are there? What learning strategies are they using? What do they do? It's literally a black box! "All you have to do is see how it behaves for different inputs" is just black box reverse engineering! Machine Learning is NOT a magic black box.
Comma have constructed a "stack" of models, just as you would connect a series of functions to make a kernel in the mathematics sense, or a series of algorithms or instructions to make a program. And that stack is entirely closed.
https://medium.com/@chengyao.shen/decoding-comma-ai-openpilo... here is an example of reverse-engineering the driving model. If Comma released this exact sort of documentation, including what ML modeling strategies they were using, what each input and output parameter affected, and how the model was trained, I could maybe consider the system open.
The models are now saved in ONNX format. Which is the most readable format available. You can view the architecture of the model with a basic neural network viewer.
Again, I'm curious what you want to learn from the training code?
Chengyao's medium post is great, but it is only possible because the models, the code that runs them and the code that parses the outputs is fully open source.
My binary is saved in a PE format. Which is the most readable format available. You can view the architecture of the software by opening it in the basic Ghidra pseudocode decompiler. All Windows software is now "fully open source."
Chengyao's Medium post is advanced reverse-engineering work requiring a detailed knowledge of the appearance of specific ML algorithms saved in a binary format. And even with this knowledge, Chengyao was only able to _speculate_ about the behavior of the model and the desired response to certain inputs.
What would satisfy me from Comma, if they were aspiring to some kind of "open" label, would be a detailed document explaining each layer of the ML system and what its goals are - like Chengyao's Medium post, but without the need to reverse-engineer the system and attempt to infer its behavior!
Now, maybe Comma don't aspire to be truly open, in which case, that's fine - In that case, Comma is a closed model with an open-source CAN interceptor on top. So essentially, crowd-sourcing the tedious and high-liability parts (vehicle integration, driving video) while owning the valuable parts (training data and model architecture). Very cool!
What format would you rather the model be saved in? ONNX is the most cross platform and standard as far as I know, and it's also what we use internally.
It's not like a PE format which is compiled from something else higher level.
Isn't any ML model "just a bunch of weights," if you look at it right?
So, where does "modelV2" come from here, in the part that plans the lateral steering action? https://github.com/commaai/openpilot/blob/4ace476f14bb73c354... . It's a model. A video frame goes into the model, and somehow the desired path comes back out. That's the core of the driving system!
