Brilliant post Mark. I laughed out loud when I saw the Gaussian meme & the “I do ML to share cool stuff on Twitter”, basically described my past 3 years online.
Not only that you’ve put into words something I couldn’t put my finger on... why HuggingFace + Weights & Biases have quickly become two of my favourite ML companies.
"I’ve previously called Open AI a media company" this became true and permanent to me when they no longer were a non-profit, and they always have paid about the same ludicrous ML salaries.
A thought in the back of my mind (maybe too optimistic as an RL researcher) is that something for RL could work like hugging face... a open-source set of tools and infrastructure for making RL not impossible to use. Most SOTA algorithms don't even get reproduced on their simulated tasks...
1. Algorithms which I have seen supported in Stable Baselines which for some reason didn't reach mainstream success perhaps because RL isn't as widely useful or the community aspect never really kicked in.
2. Environments or Data loaders which was supposed to be gym but then not much happened in the project for a long time until distributed stuff was supported by Ray and then creating new environments had a better story with Unity. Unity is closed source but even then I love ML agents but didn't reach mainstream success perhaps because the cross section of ML devs with game devs is very small
In all cases, I have a lot to share on this so happy to brainstorm more 1:1 so lmk
Interesting post. V enjoyable read.
Brilliant post Mark. I laughed out loud when I saw the Gaussian meme & the “I do ML to share cool stuff on Twitter”, basically described my past 3 years online.
Not only that you’ve put into words something I couldn’t put my finger on... why HuggingFace + Weights & Biases have quickly become two of my favourite ML companies.
Taking notes and sharing this.
Interesting
"I’ve previously called Open AI a media company" this became true and permanent to me when they no longer were a non-profit, and they always have paid about the same ludicrous ML salaries.
A thought in the back of my mind (maybe too optimistic as an RL researcher) is that something for RL could work like hugging face... a open-source set of tools and infrastructure for making RL not impossible to use. Most SOTA algorithms don't even get reproduced on their simulated tasks...
So there's 2 pieces here the
1. Algorithms which I have seen supported in Stable Baselines which for some reason didn't reach mainstream success perhaps because RL isn't as widely useful or the community aspect never really kicked in.
2. Environments or Data loaders which was supposed to be gym but then not much happened in the project for a long time until distributed stuff was supported by Ray and then creating new environments had a better story with Unity. Unity is closed source but even then I love ML agents but didn't reach mainstream success perhaps because the cross section of ML devs with game devs is very small
In all cases, I have a lot to share on this so happy to brainstorm more 1:1 so lmk