OpenPilot isn't Open(github.com)
github.com
OpenPilot isn't Open
https://github.com/commaai/openpilot/issues/21480
10 comments
Android, chrome, and vscode are all touted as open source by much of their functionality relies on closed sources binaries.
You are right, the only thing that seems not to be open in Open Pilot is that. But apart from the code to train the model, the project is almost entirely open source. I don't think that's as bad as, say for example, OpenAI
I think that I am fine with OpenAI, especially after spending the last week experimenting with Common Lisp examples (for the next edition of my book; look in src/openai in https://github.com/mark-watson/loving-common-lisp)
I use pretrained BERT models that I don’t have the resources to train myself. I don’t have problems paying a very small fee to run OpenAI’s models as an API that can do things like:
I just tried a random question:
cl-user> (openai:answer-question "What rivers flow in Arizona?" 60) "The Colorado, the Gila, the Little Colorado, the Salt, the Verde, the San Pedro, the Santa Cruz, the San Juan, the Little Colorado, the Rio Grande, the Gila, the San Pedro, the Santa Cruz, the San Juan, the Little Colorado"
Not bad, but I think it missed a few. The OpenAI model can do a lot more things besides just answering questions:
1. Given a list of food ingredients, write a complete recipe including directions
2. Summarize text
3. Grammer correction
4. Translate languages (limited to English to French for now)
5. Answer any general question (I have an example below)
6. Write an advertisement from a product description
7. It is used at GitHub to build a service that writes software, given an English description of the problem
8. Generate spreadsheet data from data descriptions in English text
9. Write short stories, given a starter paragraph.
10. etc.
I use pretrained BERT models that I don’t have the resources to train myself. I don’t have problems paying a very small fee to run OpenAI’s models as an API that can do things like:
I just tried a random question:
cl-user> (openai:answer-question "What rivers flow in Arizona?" 60) "The Colorado, the Gila, the Little Colorado, the Salt, the Verde, the San Pedro, the Santa Cruz, the San Juan, the Little Colorado, the Rio Grande, the Gila, the San Pedro, the Santa Cruz, the San Juan, the Little Colorado"
Not bad, but I think it missed a few. The OpenAI model can do a lot more things besides just answering questions:
1. Given a list of food ingredients, write a complete recipe including directions
2. Summarize text
3. Grammer correction
4. Translate languages (limited to English to French for now)
5. Answer any general question (I have an example below)
6. Write an advertisement from a product description
7. It is used at GitHub to build a service that writes software, given an English description of the problem
8. Generate spreadsheet data from data descriptions in English text
9. Write short stories, given a starter paragraph.
10. etc.
Hotz has mentioned that the training code is not public because it wouldn't benefit anyone apart from those that are trying to wholesale copy the entire project. You can see model architecture and gleam insight from the decisions that they made. You can run simulations with the model and replay scenarios. You can add support for currently unsupported vehicles. Dev tools let you see signals generated.
There's a lot of stuff you can do. But what they don't let you do is easily clone the report, throw a few million in compute and repackage what they had built. I think it's a reasonable restriction.
There's a lot of stuff you can do. But what they don't let you do is easily clone the report, throw a few million in compute and repackage what they had built. I think it's a reasonable restriction.
> ... it wouldn't benefit anyone apart from those that are trying to wholesale copy the entire project.
Is that not a valid thing for a project called "Open" Pilot?
Is that not a valid thing for a project called "Open" Pilot?
Love it ... “Open as in OpenAI” should be a standard phrase.
Just like “Free as in beer”
> "This repo doesn’t have any training code. The training code is not public."
Also on HN top page:
"Bullshit ability as an honest signal of intelligence."
https://journals.sagepub.com/doi/full/10.1177/14747049211000...
/lol
"Bullshit ability as an honest signal of intelligence."
https://journals.sagepub.com/doi/full/10.1177/14747049211000...
/lol
i dont believe that this complaint is valid unless the project itself provides no prototypical training data. if we produce a 3d gun design and someone asks for the 3d printing spools, are we gonna tell the design maker that its not trully open?
so unless someone tells me this repo provides no bootstrap training data, this is more a choosy beggar.
so unless someone tells me this repo provides no bootstrap training data, this is more a choosy beggar.
AKA the OpenAI definition of Open