"We keep using the term 'open source' in the context of large language models like Llama 2, but it’s not clear what we mean", or, rather, what "open source" means in world of floating point numbers as "code" https://
infoworld.com/article/370609
1/rethinking-open-source-for-ai.html
… by @mjasay for @InfoWorld with @Mashah08
@mjasay
-
Rethinking Open Source Definition for AI Language Models
By
–
-
Open Source AI: Legal Questions and Compliance Challenges
By
–
Well, I *am* a lawyer, but not a very good one. I don't have good answers right now. Just lots of questions. I'm confident that if we grapple with the questions, we'll sort it out (just as we did when applying open source to software distributed as services over a network)
-
Can AI Model Weights Be Licensed if Unmodifiable?
By
–
Talking with a friend, he noted that you can't modify the weights directly. You can't go in and change a floating point number and mess with it. You have to recompile those from somewhere else. Does a license on the weights (assuming he's right) make sense if true?
-
Rethinking Open Source Licensing for AI Model Weights
By
–
Also, do you think the current licensing of, say, Llama 2 does that? Meaning, applies the license to the weights? It seems like the license is more broadly applied than that, and that we may need some tweaks to the OSD to reimagine open source for AI.
-
Defining Open Source for Large Language Models Precisely
By
–
Yes, this is what I'm thinking (I'm in the middle of a post on the topic and really wanted your feedback). I think we're a bit too casual in how we apply "open source" to LLMs, & need to get crisp in how we think about it. No one is as far on the path as you are, so I pinged you
-
What Does Open Source Really Mean for LLM Weights?
By
–
But more importantly (for me), I'm grappling with what "open source" means in LLMs/AI. The source code isn't like what we'd find in, say, Datasette. It seems ppl think the license applies to the weights in an LLM, but they're just floating point numbers.
-
Redefining Open Source for Large Language Models
By
–
We casually apply the Open Source Definition as if it's dispositive here but, and I may be missing something here, it feels like LLMs are very different and the term "open source", or rather the definition, needs to be carefully recast for AI.
-
AI Model Licensing: Architecture, Weights, and Delta Distribution
By
–
Should the license be on the architecture? But, the same architecture with different weights can give you a completely different AI. Should the license be on the weights + architecture? Sometimes we ship deltas. Does the license apply to that?
-
Meta Llama 2 licensing: poison pill or innovation accelerator?
By
–
Hey, @simonw
, serious question(s). You call Meta's open, but not open source licensing of Llama 2 "a serious poison pill" but also that it "drove the pace of open innovation into hyperdrive." This seems like two opposing arguments (or potentially so)? https://
simonwillison.net/2023/Aug/27/wo
rdcamp-llms/#openly-licensed-models
… -
DevRel Strategy: Prioritizing Developer Needs Over Events
By
–
The message from developers was essentially: give me the information I need to get my job done with minimal fuss/friction. But too often DevRel was spending big on in-person events. Events have a place, but they're not The Place™. This reshaped how I thought of my mission.