I don't see how more advanced models won't get gated to specific known KYC'd entities. Classification-style guardrails will never be sufficient. Distillation attacks too are really hard to prevent. Open-source models can have their guardrails easily stripped away so it'll be incredibly dangerous to continue to release more and more capable OSS models that can and will be used to give bad actors 100x leverage.
I feel like this headline is a bit over-stated. There is not a ton of evidence it was about a jailbreak, and neither was there evidence that is was about retribution.
Ultimately, I bet Anthropic is fine with this because they needed to take Fable down to improve the guardrails (that were getting a ton of pushback) and they consider treating Fable as "too dangerous" to just be good PR hype for them. And they just get a little more anti-Trump "cred".
So the article calls it "knowledge gaps". Has technical expertise ever mattered when the law wants to ban or restrict something it doesn't like? The DMCA comes to mind.
This is mostly a restatement of https://news.ycombinator.com/item?id=48552687
Ok, we'e moved the comments thither, except the ones that are only relevant to current article.
you missed the chance to also say hither
To expand on this:
Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak (theregister.com) 398 points | 6 hours ago | 223 comments
This is a frustrating article - it provides no new information at all to support the claim that it was "never about a jailbreak".
I suspect there's more to the story than has been reported too, but I'd like information to help turn those suspicions into something more concrete.
Yes, this is just an even shorter rehash of what has been said several times now.
https://archive.is/3rY1H
This article tries to make it sound like Anthropic is playing 4-D chess and sacrificed a pawn to force a better future outcome.
This seems too simple, and too complex at the same time.
I don't see how more advanced models won't get gated to specific known KYC'd entities. Classification-style guardrails will never be sufficient. Distillation attacks too are really hard to prevent. Open-source models can have their guardrails easily stripped away so it'll be incredibly dangerous to continue to release more and more capable OSS models that can and will be used to give bad actors 100x leverage.
Should be pointed out this is an opinion article
This is an opinion piece.
I feel like this headline is a bit over-stated. There is not a ton of evidence it was about a jailbreak, and neither was there evidence that is was about retribution.
Ultimately, I bet Anthropic is fine with this because they needed to take Fable down to improve the guardrails (that were getting a ton of pushback) and they consider treating Fable as "too dangerous" to just be good PR hype for them. And they just get a little more anti-Trump "cred".
TechCrunch articles should be ignored into oblivion.
I think this is pretty low quality content for HN.
So the article calls it "knowledge gaps". Has technical expertise ever mattered when the law wants to ban or restrict something it doesn't like? The DMCA comes to mind.
[dead]
[flagged]
Look at how the Trump administration treats Canada, it's the same thing. They lie and make up reasons to punish countries that hurts their feelings.