Brief Candle
@Laserfish2
Followers
430
Following
6K
Statuses
690
Safety-conscious tech accelerationist. It's over.
Joined April 2019
@perrymetzger @MasterTimBlais We'll see this play out repeatedly over the next couple of years, as people continue to spin up thousands of such agents with that exact mandate and agent resilience increases. The only thing that will limit damage from instrumental convergence is model capabilities.
0
0
1
RT @Laserfish2: @perrymetzger @MasterTimBlais Instrumental convergence could simply be relabeled "good planning" or "acting rationally." I…
0
1
0
@RichardHanania @Aella_Girl Robin Hanson as recently as one year ago said that he thought it would be 100-200 years before AGI. That is not an opinion that warrants significant credence. Extraordinary claims require extraordinary proof.
1
0
2
RT @lethal_ai: 🧵22/34 Core principle --- Fundamentally, we are dealing 2 completely opposing forces fighting against each-other. On one si…
0
81
0
@wordgrammer @liku408 I imagine that o3 w/ SWE agent framework could already do this with 90% of C files with compiler feedback. See and porting code is easier than solving issues.
0
0
1
@ESYudkowsky This is an important point. The only reason nuclear reactors are safe is a truly absurd amount of effort and expenditure has gone into making them that way. Perhaps we should all agree that such investment is warranted on all powerful new technology.
1
0
4
@hostileparmesan @hollowearthterf There is basically no one out there who cares about strangers more than themselves, and when there is, we call them a saint.
0
0
2
@Yolo69521954 @powerfultakes That is not what the graph says. That is not what the study says. You are reading it wrong.
Yes. This graphic is rapidly becoming the most misused and misunderstood image in all of modern history. Despite what its spreaders claim, it doesn’t show that liberals care *less* about those close to them than animals strangers etc. It simply shows that they tend to *also* care about further away things AS WELL - if you read the original study, you’ll see the (admittedly idiotic) color grading was meant to show only the distribution of EDGES of the two groups’ moral circles, NOT the distribution of their FOCUS. And frankly, anyone with an iota of reason could figure that out if they actually sat down and thought about it… there’s basically no one on earth who truly cares about a total stranger or rock or insect etc MORE than their friends, which is what this chart would imply if interpreted this way. Not even crazy ass wokies have such an inverted moral circle, coz they too care about their “tribe” (ie whatever intersectional micro culture they champion) above others, much like everyone else does. It’s classic culture wars propaganda, but this time, it’s right wing brains who turn to mush upon seeing it, because confirmation bias. Catnip for conservatives.
0
0
5
0
0
3
@EmilyTheGaskell This is a very noteworthy result. o3-mini-high cannot play tic tac toe successfully, where o1 can.
0
0
1
RT @EmilyTheGaskell: o3-mini still can't play tic-tac-toe, even on high reasoning. If you play anything non-standard (like the bottom row),…
0
1
0
@profoundlyyyy No, wait, I was promised that AI powered defense would always overcome malicious AI!
0
0
2
@Singularitybook @asdasdf293 @danfaggella Even if you don't believe such capabilities or tendencies will arise naturally, people are actively working to create them. People are working hard to give models autonomy, persuasion, etc., simply because there is money to be made in doing so.
1
0
2
@mlinsey @psychosort o1 is much better at coding than Claude. I feel like people haven't gone back and reevaluated things. Claude is obviously far superior to 4o.
0
0
0
@Singularitybook @asdasdf293 @danfaggella We are still much smarter than it in certain domains, particularly in that I am still better at controlling the bots than they are at controlling me or subverting my controls. But the models will continue to get better at every domain, including this one.
1
0
1
@danfaggella That said, he seems to be acting suboptimally even from his own standpoint. Unless we believe that he still has maintained his original principles and is lying to everyone now simply to maximize investment?
0
0
0
@perrymetzger @MasterTimBlais It is reasonable to argue that a hard takeoff doesn't make sense. It is ridiculous to argue that instrumental convergence and orthogonality don't make sense, even if you think they might be surmountable.
1
0
2