zlacker

> There is no way to stop self hosted models.

edit: Current models- sure, but they will soon be outdated. I think the idea is to strangle the development of comparable, SoTA models in the future that individuals can self-host; OpenAI certainly won't release their weights, and they'd want the act of releasing weights without a license to be criminalized. If such a law is signed, it would remove the threat of smaller AI companies from disintermediating OpenAI, and individuals from collaborating to engage in any activity that results in publicly available model weights (or even making the recipe itself illegal to distribute)

>>sangno+(OP)
I thought we got away from knowledge distribution embargos via 1A during the encryption era.

Even if it passed, I find it hard to believe a bunch of individuals couldn't collaborate via distributed training, which would be almost impossible to prohibit. Anyone could mask their traffic or connect to anon US VPN to circumvent it. The demand will be there to outweigh the risk.

>>sangno+(OP)
You can't strangle the development of such models because the data comes from anywhere and everywhere. Short of shutting off the entire Internet, there's nothing a government can do to prevent some guy on the opposite side of the world from hoovering up publicly accessible human text into a corpus befitting an LLM training set.

>>10000t+gB
It costs a lot of money to train foundation models, that is a big hurdle to open source models which can strangle further development.

Open source AI needs people with low stakes (Meta AI) who continue to open source foundation models for the community to tinker with

>>sangno+(OP)
I have a question, AI is not exclusively to use with data from the internet right?, eg you can throw a bunch of text and ask to order it on a table with x columns, this will need data from the internet? I guess not, you can self host and use it exclusively with your data

>>helloj+1c
> distributed training

Unfortunately this isn't a thing. Eg too much batch norm latency leaves your GPUs idle. Unless all your hardware is in the same building, training a single model would be so inefficient that it's not worth it.