Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
1–10 of 17 posts
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#2So it's okay when big American corps raid the internet ignoring any terms of service or licenses they see in order to train models they rent back to us, but when a foreign entity trains off of Anthropic it's illegal?
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#3So it's okay when big American corps raid the internet ignoring any terms of service or licenses they see in order to train models they rent back to us, but when a foreign entity trains off of Anthropic it's illegal?
From the tweet, Anthropic's point is that distillation is Ok, unless new model has safeguards removed or used for military or surveillance purposes.
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#4So it's okay when big American corps raid the internet ignoring any terms of service or licenses they see in order to train models they rent back to us, but when a foreign entity trains off of Anthropic it's illegal?
From the tweet, Anthropic's point is that distillation is Ok, unless new model has safeguards removed or used for military or surveillance purposes.
I don’t think so. It reads much more like “distillation is okay when you do it to your own models.”
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#5So it's okay when big American corps raid the internet ignoring any terms of service or licenses they see in order to train models they rent back to us, but when a foreign entity trains off of Anthropic it's illegal?
From the tweet, Anthropic's point is that distillation is Ok, unless new model has safeguards removed or used for military or surveillance purposes.
The fact that they're calling it an "attack" implies otherwise.
I find the entire premise of this announcement absurd. Fraudulent accounts? They're just accounts. They paid for the access the same as any other. They're accessing Claude just like a human (or *claw) would.
There's no argument against their strategy that doesn't make them complete hypocrites in respect to how they got the model training data in the first place.
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#6Pot, meet kettle!
I don’t think I’m the only one feeling some schadenfreude at this news. I suppose it’s ok when you’re a hot Silicon Valley scale-up to slurp up the rest of the worlds data for free and then hire hot shot lawyers to defend you against all the creatives you ripped off, but when it’s the “evil” Chinese doing the same to you it’s a dastardly “attack”?
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#7Earlier quoted context omitted.
From the tweet, Anthropic's point is that distillation is Ok, unless new model has safeguards removed or used for military or surveillance purposes.
The fact that they're calling it an "attack" implies otherwise. I find the entire premise of this announcement absurd. Fraudulent accounts? They're just accounts. They paid for the access the same as any other. They're accessing Claude just like a human (or *claw) would. There's no argument against their strategy that doesn't make them complete hypocrites in respect to how they got the model training data in the firs…
> them complete hypocrites in respect to how they got the model training data in the first place.
sure, hypocrisies is part of rules for big games: politics and business.
> Fraudulent accounts? They're just accounts.
they tell the story in blog post, that they don't allow claude in China, but those labs use some proxy services to access claude and mix traffic with regular users to hids its activities
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#8I don't think this counts as distillation. Distillation is when you use a teacher model to train a student model, but crucially, you have access to the entire probability distribution of the generated tokens, not just to the tokens themselves. That probability distribution increases tremendously the strength of the signal, so the training converges much faster. Claude does not provide these probabilities. So, Claude was used for synthetic training data generation, but not really for distillation.
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#9I don't think this counts as distillation. Distillation is when you use a teacher model to train a student model, but crucially, you have access to the entire probability distribution of the generated tokens, not just to the tokens themselves. That probability distribution increases tremendously the strength of the signal, so the training converges much faster. Claude does not provide these probabilities. So, Claude…
Sampling repeatedly gives them an estimate of the probability distribution in any case though.
Re: Alleged Distillation Attacks by DeepSeek, Moonshot AI, and MiniMax
#10I don't think this counts as distillation. Distillation is when you use a teacher model to train a student model, but crucially, you have access to the entire probability distribution of the generated tokens, not just to the tokens themselves. That probability distribution increases tremendously the strength of the signal, so the training converges much faster. Claude does not provide these probabilities. So, Claude…
Sampling repeatedly gives them an estimate of the probability distribution in any case though.
That would be an interesting paper actually; what is the optimal sampling technique given you only have access to the token outputs. Surely someone has already done it.