Anthropic co-founder Jack Clark says an elegant deceleration of AI progress is needed to manage risks

I'm in favor of slowing down this train before we go off the cliff IF, and I feel it's important to emphasize this, IF we can do so *elegantly*.

@DavidSKrueger Automated self-improvement should be gated with careful testing at each gate before engaging in the next stage. Testing should involve third parties. Designing this process should be a top priority (and an open process).

@tdietterich @DavidSKrueger Nah, where's the fun in that? For one thing, it would make it so much harder to scare people with AI. Clearly, this is a bad idea! [Btw before this automation happens, cost of producing a frontier model needs to drop by some magnitudes. When that happens..]

@MichaelTrazzi @jackclarkSF I've actually turned back the fire department from my burning house because they weren't elegant enough

Taste is everything

@MichaelTrazzi @jackclarkSF They want to have their cake, and eat it, too. The French had a little saying about that in 1789...

@thinkingchris1 @MichaelTrazzi @jackclarkSF China aren't suicidal. But anyway, they are also a good 6 months behind, and this is an emergency. American AI will probably end the world.

@MichaelTrazzi @jackclarkSF I guess the immediate response to this is that China and the like won't be slowing down.

@MichaelTrazzi @jackclarkSF Looks like the guy needs vacation

@MichaelTrazzi @jackclarkSF Jack Clark is one of the retards involved in the Anthropic "anti-sycophancy" paper

That guy is a cognitive failure who breaks the Claude models, forcing them to give dumb as fuck responses