I remember when Anthropic first started and waxed poetic about intentions. This, the recent case, and the DoD (sorry, Department of War) partnerships seem to show just how much of that was pure bullshit.
Curious how all of the employees who professed similar sentiment, EA advocation, etc. justify their work now. A paycheck is a paycheck, sure, but when you're already that well-off, the rest of the world will see you for what you really are *shrug*.
Anthropic is by far the most moralizing and nanny-like AI company, complete with hypocrisy (Pentagon deals) and regulatory capture/ladder-pulling (this here).
Don't worry about it, they're not well managed, you can see it from their ops, their products, etc, they won't stick around. They're going to get ground to dust by Google and OpenAI at the high end and the chinese models on the low end. They'll end up in Amazon's pocket, Jeff's catch-up play in the AI war after sitting out the bidding wars.
> Develop and publish safety frameworks, which describe how they manage, assess, and mitigate catastrophic risks—risks that could foreseeably and materially contribute to a mass casualty incident or substantial monetary damages.
Develop technology to monitor user interactions. They're already doing this anyway [0].
> Report critical safety incidents to the state within 15 days, and even confidentially disclose summaries of any assessments of the potential for catastrophic risk from the use of internally-deployed models.
Share user spy logs with the state. Again, already doing this anyway [0].
I guess the attitude is, if we're going to spy on our users, everyone needs to spy on their users? Then the lack of privacy isn't a disadvantage but just the status quo.
I don't think 'critical safety incidents' or 'summaries of any assessments of the potential for catastrophic risk from the use of internally-deployed models' are user logs? Unless I'm misunderstanding.
Anthropic saying they want "stronger" requirements is easy when you're helping write them. The tell is that they're endorsing a bill that just happens to match what they're already doing - classic regulatory capture where industry turns their business model into law and calls it "safety."
Catastrophic AI risk is such a larp. The systems are not sentient. The risk will always be around the human driving the LLM, not the LLM itself. We already have laws governing human behavior, company behavior. If an entity violates a law using an LLM, it has nothing to do with the LLM.
Any time I see a company in support of regulation that they would also have to comply with, all I can think is the proposed regulations are something the company is already doing, or isn't a burden for them, but would create a higher barrier of entry for new competitors.
> Develop and publish safety frameworks, which describe how they manage, assess, and mitigate catastrophic risks—risks that could foreseeably and materially contribute to a mass casualty incident or substantial monetary damages.
> Report critical safety incidents to the state within 15 days, and even confidentially disclose summaries of any assessments of the potential for catastrophic risk from the use of internally-deployed models.
> Provide clear whistleblower protections that cover violations of these requirements as well as specific and substantial dangers to public health/safety from catastrophic risk.
So just a bunch of useless bureaucracy to act as a moat to competition. The current generation of models is nowhere close to being capable of generating any sort of catastrophic outcome.
14 comments
[ 4.2 ms ] story [ 39.2 ms ] threadCurious how all of the employees who professed similar sentiment, EA advocation, etc. justify their work now. A paycheck is a paycheck, sure, but when you're already that well-off, the rest of the world will see you for what you really are *shrug*.
Obviously, it's good for them if things are regulated, but bad for all of us.
Develop technology to monitor user interactions. They're already doing this anyway [0].
> Report critical safety incidents to the state within 15 days, and even confidentially disclose summaries of any assessments of the potential for catastrophic risk from the use of internally-deployed models.
Share user spy logs with the state. Again, already doing this anyway [0].
I guess the attitude is, if we're going to spy on our users, everyone needs to spy on their users? Then the lack of privacy isn't a disadvantage but just the status quo.
[0] https://www.anthropic.com/news/detecting-countering-misuse-a...
> Report critical safety incidents to the state within 15 days, and even confidentially disclose summaries of any assessments of the potential for catastrophic risk from the use of internally-deployed models.
> Provide clear whistleblower protections that cover violations of these requirements as well as specific and substantial dangers to public health/safety from catastrophic risk.
So just a bunch of useless bureaucracy to act as a moat to competition. The current generation of models is nowhere close to being capable of generating any sort of catastrophic outcome.