Great write up! Thanks for sharing. I’ve been thinking a lot about the double-edged sword of open source models. You make a great point with model providers being able to monitor usage of their products over their API.
A safety layer with open source models does not exist. Once those weights are downloaded, they are out there forever—and people are able to customize them as they see fit. No change in law or policy can retroactively remove those capable LLMs from existence.
Anthropic was fortunate to catch this, but it certainly makes you wonder what’s going on undetected from locally hosted models.
Yes for sure. These openly accessible and downloadable models come with a freedom of usage that can go both ways. And it doesn’t seem like we’re very prepared for the malevolent case.
Great write up! Thanks for sharing. I’ve been thinking a lot about the double-edged sword of open source models. You make a great point with model providers being able to monitor usage of their products over their API.
A safety layer with open source models does not exist. Once those weights are downloaded, they are out there forever—and people are able to customize them as they see fit. No change in law or policy can retroactively remove those capable LLMs from existence.
Anthropic was fortunate to catch this, but it certainly makes you wonder what’s going on undetected from locally hosted models.
Yes for sure. These openly accessible and downloadable models come with a freedom of usage that can go both ways. And it doesn’t seem like we’re very prepared for the malevolent case.