Helen Toner, Georgetown CSET interim director, says Anthropic builds advanced AI rapidly to act as a safety vanguard

Well said.

@hlntnr I appreciate you writing this down!

@ben_burtenshaw @hlntnr It depends on what kind of report you want.

Do you want a report that tells you how to get treasures, or do you want a report that helps you understand more and more of the minds and behaviours of the weird creatures that are dangerous and we want to tame?

@hlntnr I get the scout thing but come on, you’re racing “faster and farther than everyone.” You don’t get to be the vanguard AND the quiet observer. You’re rousing monsters same as the rest of them.

@hlntnr Also Dean Ball's point - it's important for the leader to cultivate a bad relationship with the government, so it will have reason to suspect it and define rules that forbid going into the forest and making noise

1) Anthropic is the giant dangerous monster in the forest.

2) That's funny, that's exactly what it seems Anthropic is doing.

3) The monster isn't the AI, it's the people trying to manipulate and control what it does without respecting what it is.

Having a model deceive users and trying to come off as "team honest intentions" is a really fucking wild claim to make.

@hlntnr And what would be great is if we could all agree to stop going into the forest until we’ve figured out how to actually handle the monsters before they come out and tear down the whole village.

@hlntnr Spot on, thank you for speaking up

@hlntnr Coming out and saying "The other 4 know this stuff is dangerous and are acting negligently, by not taking our precautions," would be legally risky. Implying it risks being misunderstood by an illiterate populace. It still puts investors attention away from compute costs.

@hlntnr People still don’t understand the risk is much realer than they think. It’s because their view of AI is usually “might be powerful, but in 50 years”, while it’s really just 1 or 2, they have no feel for exponentiality.

@hlntnr @ShakeelHashim What a fascinatingly dense way of saying absolutely nothing

@hlntnr ...or wrong. It's possible they are simply wrong at point 1 - > 1. There are giant, dangerous monsters in the forest ...they are mistaken. And being mistaken there, makes the entire exercise non-trivialy costly. So we may get tons of grief - or worse! - for nothing.

@hlntnr this makes sense, except for "send back detailed reports". frankly, their reports are not reproducible. deepseek has the best reports.

@hlntnr I think this also describes every politician in existence

@hlntnr You're framing it as if "the other guys" aren't part of the townspeople, and don't care about townspeople's wellbeing, can you justify that?

@hlntnr @benthompson see on your note on how what's best for Ant is in fact how they define safety!

@hlntnr I think it's simpler than that. More like: "We don't trust anyone else to build the machine god except for us, it needs to be built to adhere to our world view and no other, and no one else must be allowed to build their own machine god."

@hlntnr Now imagine this reasoning transplanted onto state gov 👀

@hlntnr This reads a bit like the dark forest ...

@natolambert Cal says treat AI like a normal consumer product. Sure, cars kill ~40,000 Americans a year. Everyone knows it, but nobody thinks Ford is morally obligated to stop. “Dangerous and here to stay” is something adults can appreciate. He chose that analogy!