Starting to suspect that Anthropic's putative security and safety considerations are largely posturing and performative.

I think self-consistency is such a rare virtue nowadays that people can't conceive of an org that sticks to its mission

agree with them or not, all steps from Ant's leadership are very predictable after conditioning on their core values

the anthropic silent limits thing is not "dishonest", not "a terrible idea", not "evil", not "dangerous" etc etc. it is just very disingenuous (but also very on brand) for them to claim they do it for "safety concerns". that is all.

the silently failing strategy is a bit weird. I can speculate it is to prevent jailbreak, but not doing same for other safety risks is one rare case of non-consistent strategy.

@natolambert I understand the cynicism by default, but I don't understand what they could do differently, that would distinguish between legitimate safety restrictions and so-called safety washing. Their actions are compatible with both

@natolambert They are already working on their next super model. It's called Fairy Tale

@soldni How is hiding the fact that it is not helping in line with any of the core values they claim to uphold?

@natolambert But the public is not directly going to be using any of these models (outside of individual developers). Eventually things are going to be abstracted out as products and that's all the public is going to see. My two cents.

@natolambert tech bros on X are much dumber than genpop

@tunguz

@tunguz OG OpenAI marketing.

@tunguz I believe you are correct sir

@natolambert The most obviously evil LLM company by far, IMO.

@natolambert can't you always buy data from a company outside US that is not AI lab?

@natolambert I’m not touching this one

@Smit_Chaudhary3 I presume their argument would be that showing would make jailbreak easier

as long as the actually response does not violate constitution by not giving incorrect answers, i can see how this would be their preferred choice

(not my area of expertise, sitting debate out)

@natolambert Safety washing critique has been the constant narrative since the Claude launch — if the technical audits (e.g. the 2026 alignment scout report) don't shift this room's perception, public trust is a long shot.

@lev_ey @natolambert yea people are just salty, which is understandable

I think its probably really hard to both be the leading research lab and also slow down progress.

Plus, money is quite attractive and I think it gets harder and harder to get money if you're slowing down - both from consumers + investors.

They are trying to strike a very careful balance imo, but you can only slow down so much before you make yourself irrelevant or consumers lose interest and move on.

@natolambert regulators aren't on x anyway