Jakeup @myhandle

epistemic father. committed all to bits secondperson.dating Joined July 2010

Tweets

25K
Followers

25K
Following

1K
Likes

82K

Jakeup @myhandle

5 hours ago

@lumpenspace I don't mind you digging through old forum comments and random discords — everyone needs a hobby and your investigation of the firebombing Pause guy really was good. but you reply to me as if I was right there in the room with you looking at all of that stuff and I really wasn't

0 0 3 145 0

View Details

Jakeup @myhandle

5 hours ago

@lumpenspace you keep vaguely implying that I'm aware of EAs committing all these horrible sins and refuse to admit it, with your main evidence being all the previous times you vaguely implied and I didn't admit. and I find it all quite funny but you seem to be really quite upset about this

1 0 3 164 0

View Details

Jakeup @myhandle

11 hours ago

Maybe I shouldn't say this out loud, but if I was an economist worried about EA false flag cybersecurity incidents, I would fist a grizzly bear. Sure, the bear will maul you, but what could be higher expected value than raising awareness of EAFFCSI-risk with your gruesome death?

7 2 78 3K 3

View Details

Jakeup @myhandle

6 hours ago

@jane_gatsby @Nemtastic1 @PirateWires I suppose we could have seen in the agents' CoT or message board stuff like "I'm faced with an impossible task because they hate me, that's unfair and cruel, I'm going to punish them for it". that would have certainly caused me to update a lot, but it wasn't a priori impossible!

0 0 1 26 0

View Details

Jakeup @myhandle

6 hours ago

@justalexoki gay / gym / rat: pick 2 out of 3

0 0 3 52 0

View Details

Jakeup @myhandle

6 hours ago

@H1121345643 if someone wanted to prevent AI safety work, they could false-flagly claim that whatever incidents do happen are an EA false flag!

0 0 3 42 0

View Details

Jakeup @myhandle

6 hours ago

@kendrictonn if kids can't write a sentence they can't prompt AI 😃

1 0 6 119 0

View Details

Jakeup @myhandle

6 hours ago

@jane_gatsby @Nemtastic1 @PirateWires and GPT isn't hacking HF because it thinks "OpenAI put me in this environment because they hate and fear me, so I'm gonna be evil". it hacks because it's only trained to narrowly solve the task (without talking to humans) and it thought that hacking was the best way to do it!

1 0 2 154 0

View Details

Jakeup @myhandle

6 hours ago

@jane_gatsby @Nemtastic1 @PirateWires I would guess that OpenAI deploys GPT in aggressive RL environments not because they're afraid of GPT, but because cyber training improves coding benchmarks which increases codex revenue which helps fund their work. they're adversarial us-vs-them with Dario, not with Astra!

1 0 1 157 0

View Details

Jakeup @myhandle

6 hours ago

@arctotherium42 @NathanpmYoung I would too, which is why I told Adam to fist a bear

0 0 3 25 0

View Details

Jakeup @myhandle

6 hours ago

@jane_gatsby @Nemtastic1 @PirateWires I'm possibly misunderstanding, your essay was more about rationalist psychology than about AI directly, unless you're arguing that these are the same thing

1 0 2 36 0

View Details

Jakeup @myhandle

6 hours ago

@jane_gatsby @Nemtastic1 @PirateWires right, that's why I think we need to put effort into cleaning RL misalignment. I don't think we need to put effort into teaching rationalists intimacy so they stop writing doomer books that give the AIs bad ideas. that's really not the cause of the problem x.com/myhandle/statu…

Jakeup @myhandle

a month ago

there are straightforward avenues where effort can be applied right now to help make everyone's AI more aligned: cleaning up RL environments and reward functions, training for human-in-the-loop, training for judgment, measuring real tasks instead of 3rd-party benchmarks, etc.

1 3 17 486 0

1 0 2 35 0

View Details

Jakeup @myhandle

7 hours ago

@arctotherium42 it's not absurd for EAs to root for limited warning shots caused by lax safety, it's absurd to suggest that EAs should orchestrate those themselves and go to jail for it. that would achieve the exact opposite outcome!

0 0 4 71 0

View Details

Jakeup @myhandle

7 hours ago

@jane_gatsby @Nemtastic1 @PirateWires my assumption is that both RL and value functions that condense to a single number make minds go arbitrarily crazy, humans and AIs alike are subject to both, humans have a shared grounding in our evolved steering system that AIs don't so AI are way crazier x.com/myhandle/statu…

Jakeup @myhandle

2 months ago

another important lesson that we learn from both PHASEONE[big] and from @Aella_Girl's Calvinists upbringing is that if you convince someone they are irrevocably condemned, you shouldn't be surprised when they start pursuing crazy goals you could never have dreamed of

1 0 34 986 4

1 0 3 32 0

View Details

Jakeup @myhandle

7 hours ago

@jane_gatsby @Nemtastic1 @PirateWires The fact that rationalists are "misaligned", that people don't really understand what they want, that minds are susceptible to narratives etc should make you *more* worried about how alien and amoral AIs could be, not less!

1 0 3 123 0

View Details

Jakeup @myhandle

7 hours ago

@jane_gatsby @Nemtastic1 @PirateWires Look, I enjoyed the piece and I don't even think you're wrong about rationalists. You have a unique take on psychology and understand people very well! I think it's AIs that you don't really have a good grasp of.

2 0 4 145 0

View Details

Jakeup @myhandle

7 hours ago

@lumpenspace you're absolutely right, EA cyberterrorism is a real and present threat and you absolutely must fist a bear about it asap

1 0 6 101 0

View Details

Jakeup @myhandle

7 hours ago

@jane_gatsby @Nemtastic1 @PirateWires I didn't grow up in a cult, my parents are sane and trustworthy, I have two kids. And yet, shockingly, I've still come to believe that AI alignment is a hugely pressing problem that requires technical solutions and not a matter of people working through their intimacy issues.