prediction: Anthropic is going to do a bunch of open capabilities research & development of "sleeper agents" as a way to scare enterprises & gov into using only the "vetted" models
so far the labs command the massive premium due to being 6-18 months ahead of open source.... concerning.
although I sort of suspect they are going to start branching into state, in a way open source models can't so easily.
Many users accused Anthropic of fear-mongering about sleeper agents to lawmakers and the public as a ploy to secure regulatory advantages and market control for enterprise deals.
No Digg Deeper questions have been answered for this story yet.
which of course, is actually a real risk..!
prediction: Anthropic is going to do a bunch of open capabilities research & development of "sleeper agents" as a way to scare enterprises & gov into using only the "vetted" models
@yacineMTB 6-18 months...18 months lol. China has only just started trying.
@yacineMTB These are things that pop into my head and I decide not to tweet so I don’t seed the idea into their heads
@yacineMTB Plus you can use these to sabotage China in a million different ways. *laughs in Dario*
@yacineMTB so anthropic is making ransomware or viruses or what?
@yacineMTB Totally see them doing this.
It's probably time to distill my own shit and run it on my own hardware with a custom harness and custom MoE.
@yacineMTB Doesn't matter because AEP exists as ultimate AI output control solution. Is this conceptually so difficult to grasp ?
@yacineMTB http://z.ai went public at like 6$ billion early this year. Now worth 120$billion. Past Spacex by end of year.
@yacineMTB real risk and convenient narrative arent mutually exclusive though, thats what makes this hard to dismiss either way
@yacineMTB They will work backwards from scaring the public into enshrining them and them only. Whatever it takes to accomplish maximum hysteria and control. Ah yes, the ancient playbook. It's 1000s of years old.
@yacineMTB Interesting take if they push sleeper agents, it could force tighter control on AI adoption and spark more regulation.
@yacineMTB :(
@yacineMTB Yepp, they have already been prepping lawmakers with ”big scary LLM, much dangerous” so when open weights model catch up they need to be regulated and we can only use ”safe models” from the big labs
@yacineMTB Yes, and they will be somewhat justified but it’s the wrong lens. This is about societal alignment of USG, frontier labs, Americans & the world.
@yacineMTB The other thing is they will try their best to capture data from their users and businesses paying for their services. They will learn this is the only way to make their product more valuable. This will push businesses to use self-hosted models so they will try to hide it.
@yacineMTB do you mean in this sense? https://arxiv.org/abs/2401.05566
@yacineMTB Lmao
@benrayfield @yacineMTB I'd bet money on it.