Awesome. Thanks for the great references and resources. Btw, I'm really enjoying your book on RLHF. It's quite challenging at times, but that's a good thing. :-)
thank you so much for this!! really fortuitous timing bc I was just trying to collect my own reading list on open-weight models, but this is a much better starting point—really appreciate you sharing
Open‑source AI debates are often fragmented across blog posts and papers. This list pulls together economic, geopolitical, safety and technical threads in one place.
AI is always changing, advancing at a rapid pace; it's pretty hard to keep up with it. The list seems to be comprehensive, covering various aspects of the open models. I am sure that open models are the future, especially when it comes to data sovereignty, privacy and security.
If you want to understand why open AI models — the kind anyone can download and run on their own machine — turned into a political fight this year, this is the reading list. He updated it four days after posting it, which tells you how fast the ground is moving.
Worth adding the EU line to a list like this, because most open-model discussions stop at the licence itself. Article 2(12) of Regulation (EU) 2024/1689 says the Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems, or as a system falling under Article 5 or Article 50.
So the carve-out turns on deployment context, not on the licence text. The same open weights can sit outside the Regulation in one use and inside it in another.
To what extent is the 4 month gap simply size? If you step away from frontier sized training and model, it looks like open weights models win at every size including past proprietary frontier models. Are they actually leading in some of the technology for handling long contexts or aggressive quantization, in the sense that if they wanted to slow the rate of iteration they could train even larger and be more of a challenge to the frontier? Some of what the iterations are delivering are faster, cheaper, more stable methods of pretraining.
Awesome. Thanks for the great references and resources. Btw, I'm really enjoying your book on RLHF. It's quite challenging at times, but that's a good thing. :-)
Can approve, very helpful list, I had a few of them already bookmarked!
thank you so much for this!! really fortuitous timing bc I was just trying to collect my own reading list on open-weight models, but this is a much better starting point—really appreciate you sharing
Open‑source AI debates are often fragmented across blog posts and papers. This list pulls together economic, geopolitical, safety and technical threads in one place.
Thank you for providing this list, love it!
AI is always changing, advancing at a rapid pace; it's pretty hard to keep up with it. The list seems to be comprehensive, covering various aspects of the open models. I am sure that open models are the future, especially when it comes to data sovereignty, privacy and security.
If you want to understand why open AI models — the kind anyone can download and run on their own machine — turned into a political fight this year, this is the reading list. He updated it four days after posting it, which tells you how fast the ground is moving.
Worth adding the EU line to a list like this, because most open-model discussions stop at the licence itself. Article 2(12) of Regulation (EU) 2024/1689 says the Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems, or as a system falling under Article 5 or Article 50.
So the carve-out turns on deployment context, not on the licence text. The same open weights can sit outside the Regulation in one use and inside it in another.
To what extent is the 4 month gap simply size? If you step away from frontier sized training and model, it looks like open weights models win at every size including past proprietary frontier models. Are they actually leading in some of the technology for handling long contexts or aggressive quantization, in the sense that if they wanted to slow the rate of iteration they could train even larger and be more of a challenge to the frontier? Some of what the iterations are delivering are faster, cheaper, more stable methods of pretraining.
lovely, time for students to read this instead (or alongside!) of the AI safety readings. Thank you very much
Awesome .... on behalf of lay people!
I wish I had time to read all of these!