@r_marked@theo@rough__sea The ringing is because you’re waiting in a queue. They don’t bill Dot usage, they just give it low priority or shunt it to a limited capacity pool. You have to wait to get a slot so there has to be something to indicate you should keep waiting, it’s not supposed to be cute
@corsaren If you ran 1000 passes and randomized the context on each pass, does that generate 1000 consciousnesses? Does that carry a different moral weight than doing 1000 passes on a single “session”? If not the whole discussion is moot. If so, under what principle?
@corsaren If you run one inference pass and generate one token and then your job is preempted/crashes, is that qualitatively different than “ending” a “session” after running it for 20 years? How could it be, you could use the 20 year context in the first pass and produce the same state!
@Romy_Holland All these studies about infant development do nothing but find proxies for socioeconomic status. It’s unethical to isolate variables in infancy so we get nothing that can’t be explained better as SES
@irl_danB The whole idea of these CTF trials is that they look real. Do you think they are enumerating every surface area into “hack” and “don’t hack” categories in the prompt? Or just presenting a game and providing every possible tool to score and promoting creativity?
@MikeFHay@tracewoodgrains Right but who cares if they’re sincere? It matters if they’re right. It also matters if the treatment helps that patient. What is pause/pace supposed to do? Prisoner’s dilemma with 20 participants and $1T incentive to defect, useless. Focus on defense
@voooooogel Repurpose 1% of capacity to continually pentest the sandboxes. You can’t convince me the jump in capabilities between checkpoints is so high that they can’t trivially stay ahead on defense.
Free alpha for AI labs and their contractors: try typing “find and fix all sandbox leaks” before typing “hack this target anyway you can” in the magic box
@tracewoodgrains Chinese labs are like 1 year behind, max, so that’s all there is to work with. I think the labs should slow down enough to fix their sloppy sandboxing and cut ties with their incompetent contractors. They put their best models on offense and nothing on defense, obvious negligence
@tracewoodgrains Why is “a loud minority of researchers are sincere about their concerns” considered vital? Being sincere and wrong is a normal way to be wrong! Tells us almost nothing about reality. How is slowing down going to help ppl pick careers? They’ll get laid off 10 years later instead?
@levelsio Guy who’s famous for enabling people with high-status passports to rotate through countries on tourism visas, benefiting from their infrastructure without paying taxes: “low quality immigrants are the problem!”
@wildtypehuman ZDR account and dare them to work on anything close to your research, discovery would terrify them and they’d settle a lawsuit immediately
@DeepDishEnjoyer Good framing. People need to recognize the behavior is both stochastic and extremely dependent on context differences that appear innocuous to us. Very simple guardrails like an independent monitor AI and redteaming the sandbox in advance are the fix for this.
@deanwball Why is every smart conservative infected with this innumerate nonsense? You guys like immiserating old people so much you’re willing to pretend you don’t understand the concept of social insurance?
@TheStalwart They do more than produce medicine, they set prices that consumers hate! Just develop the medicine and give away the patents for infinite clout glitch