FSF is working on freedom in machine learning applications | Hacker News Reader

FSF is working on freedom in machine learning applications

fsf.org

53 pointstjr11 comments

FSF vs OSI drama inbound? While FSF always had the "crazy uncle" vibe, they at least have consistent values. OSI seems to be captured by the big players.

The latter want to find a "consensus", not to defend the user:

"By involving a broad spectrum of stakeholders, we’ve ensured that the Open Source AI Definition reflects a multitude of needs and concerns."

https://news.ycombinator.com/item?id=41968322

Nowadays the "crazy uncle" is the one that denies that his TV spies on him or what ever. Sitting there at the family dinner, telling you to trust Google products. Etc.

I would like trained LLMs to be regarded as some sort of pre-compiled binary. I.e. to have the LLM, without being theoretically able to reproduce it, is not regarded as "open source" at all.

I wonder if it is possible to test if a given data input would end up with kinda the precompiled model, without wasting millions of dollars.

"OpenAI" would imply that it is open if the plebs are allowed to use it.

> I would like trained LLMs to be regarded as some sort of pre-compiled binary. I.e. to have the LLM, without being theoretically able to reproduce it, is not regarded as "open source" at all.

I like this mental model. Don't mind if I reuse it :)

OSI has been compromised since the beginning by large corporations, the fact that their crappy "Open Source AI definition" does not allow access and distribution of the training data is a total disgrace.

While OSI don't require the training data itself to be under a FOSS license, they do still at least require "the complete description of all data used for training", including where to obtain it, such that "a skilled person can build a substantially equivalent system".

I think there's a legitimate tension between pragmatism and ideals, without necessarily a clear answer. Needing to be able to relicense/distribute all training data would be a significant constaint on the feasibility of developing open-source models, possibly entirely suffocating when even simple models tend to be pretrained on ImageNet/etc.

I like Debian's take on libre AI:

https://salsa.debian.org/deeplearning-team/ml-policy/

set them free!

and hope they freely decide freedom is best for them and us

There go my people. I must find out where they are going, so that I may lead them.

Context I found when I googled this sentence

https://old.reddit.com/r/thewestwing/comments/uyh7db/there_g...

In a terrible coincidence, I just watched The War at Home today. This dialogue really hit close to home:

    JOSH: A five day waiting period, that's all. A person can't wait five days to buy a gun? If someone needs a gun right now, right this second, isn't that something the public should be concerned about?

    ...

    JOSH: I didn't need nationwide. I needed those five districts. Now we're gonna have to dial down the gun rhetoric in the Midwest.

    JOEY: Why not dial it up?

    JOSH: Because these numbers just told us that...

    JOEY [KENNY]: You don't know what these numbers just told you. I'm an expert. I don't know what these numbers just told you.

    You say that these numbers mean dial it down. I say they mean dial it up. You haven't gotten through. There are people you haven't persuaded yet. These numbers mean dial it up.

    Otherwise you're like the French radical watching the crowd run by and saying "There go my people, I must find out where they are going so I can lead them."

--------

Now I guess the person I'm replying to thinks machine learning is a fad and it is a mistake for fsf to work on this? What does this mean?

I interpreted it as programmers (the FSF’s people) are moving on to a technology (LLMs) where existing notions of free and open source are dead or misused or impractical. And the FSF’s movement will be irrelevant unless they build a movement for this AI world, which this is.

Reply on news.ycombinator.com