4 ms·

IPFS was also one of the ways LLAMA was being well distributed. This is really a sad day for “people”

1attice 4y ago
I thought of this immediately too. I know that Meta is doing its level best to wipe LLaMA's weights from the face of the Internet, but I also suspect (and cannot prove) that various organs of the US state (CIA, NSA, FBI) are also trying. If they're not, I'm even more afraid, because one of their most important jobs is being alarmed so I don't have to be. I'd wager a few dollars that Cloudflare had a visit. And, as an aside, I'd say it's a matter of when, not if, a GPT-4-scale model is leaked, possibly GPT-4 itself. If not to the open Internet, then to China. OpenAI is (for now) staffed by humans, and humans are vulnerable to seduction, spearphishing and spycraft.
- Zuiii 4y ago
  lol what are you talking about? The weights are being linked to on Meta's repo right now: https://github.com/facebookresearch/llama/pull/73/files https://github.com/facebookresearch/llama/pull/73/files The idea that Meta is frantically running around trying to "wipe" these weights is ridiculous. They're either doing this to disclaim liability in case a new law imposes it retroactively, or they are trying to create uncertainty around its use to dissuade serious US competitors from using it for their own products. > but I also suspect (and cannot prove) that various organs of the US state (CIA, NSA, FBI) are also trying. I would be very surprised if they were given that it would only handicap the US while leaving other countries to advance unabated. Again, the only real thing facebook did here is show that we can have smaller models that are competitive to larger ones. The information to build and train these models is public. The code to do this is public. The capability to this so is public. The cat is out of the bag and humanity is better off for it. What humanity needs to do now is figure out how best to adapt.
  - 1attice 4y ago
    The tracker for that torrent was taken offline by Meta. There was a story about it on HN a while ago. Don't believe me? Try to torrent. Do homework. SMH.

I thought of this immediately too. I know that Meta is doing its level best to wipe LLaMA's weights from the face of the Internet, but I also suspect (and cannot prove) that various organs of the US state (CIA, NSA, FBI) are also trying. If they're not, I'm even more afraid, because one of their most important jobs is being alarmed so I don't have to be. I'd wager a few dollars that Cloudflare had a visit. And, as an aside, I'd say it's a matter of when, not if, a GPT-4-scale model is leaked, possibly GPT-4 itself. If not to the open Internet, then to China. OpenAI is (for now) staffed by humans, and humans are vulnerable to seduction, spearphishing and spycraft.

Zuiii 4y ago
lol what are you talking about? The weights are being linked to on Meta's repo right now: https://github.com/facebookresearch/llama/pull/73/files https://github.com/facebookresearch/llama/pull/73/files The idea that Meta is frantically running around trying to "wipe" these weights is ridiculous. They're either doing this to disclaim liability in case a new law imposes it retroactively, or they are trying to create uncertainty around its use to dissuade serious US competitors from using it for their own products. > but I also suspect (and cannot prove) that various organs of the US state (CIA, NSA, FBI) are also trying. I would be very surprised if they were given that it would only handicap the US while leaving other countries to advance unabated. Again, the only real thing facebook did here is show that we can have smaller models that are competitive to larger ones. The information to build and train these models is public. The code to do this is public. The capability to this so is public. The cat is out of the bag and humanity is better off for it. What humanity needs to do now is figure out how best to adapt.
- 1attice 4y ago
  The tracker for that torrent was taken offline by Meta. There was a story about it on HN a while ago. Don't believe me? Try to torrent. Do homework. SMH.