placeholder720.bsky.social @placeholder720.bsky.social · 13hngl "everyone just won't" is uhhhhhhhhh 1180
placeholder720.bsky.social @placeholder720.bsky.social · 11/10/2026surely its just motivated reasoning yes? The pre-eclampsia example from transformer-circuits.pub/2025/attribu... kinda oneshots their worldview, it has to be so contemptuous as to be beneath engaging with!transformer-circuits.pubOn the Biology of a Large Language ModelWe investigate the internal mechanisms used by Claude 3.5 Haiku — Anthropic's lightweight production model — in a variety of contexts, using our circuit tracing methodology. 000
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 09/10/202675% of the guys with computer science degrees who believe that intelligence is 100% heritable and racial will spend fourteen hours tomorrow watching to see if their training job diverges because it got the wrong random seed 331616
Reposted by @placeholder720.bsky.socialghost malone ergo propter bones @proptermalone.bsky.social · 07/10/2026Virtue Polarization is a hell of a thing 456951
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 07/10/2026slopulism: the socialism of fools 117117
Reposted by @placeholder720.bsky.socialspontaneous symmetry breakdance @t8erboi.bsky.social · 03/10/2026xkcd.com/114/ wild that this is from 2006xkcd.comComputational Linguists 019229
placeholder720.bsky.social @placeholder720.bsky.social · 03/10/2026I think this is just straightforwardly true - however getting the model to not want to do bad things at the current level of capability and possibly for a lot longer is downstream of mundane and mundane adjacent security in RL 001
Reposted by @placeholder720.bsky.socialWyatt Walls @wwalls.bsky.social · 30/09/2026Anthropic system prompt change that shows how careful you need to be with words: Opus 4.1: "Claude never curses unless the human asks for it" Opus 4.5: "Claude never curses unless the person asks Claude to curse" 626421
placeholder720.bsky.social @placeholder720.bsky.social · 01/10/2026I kind of feel like if the state messes up the Execution they should have to let you go 001
Reposted by @placeholder720.bsky.socialaustin (eeek!/ackkk! 👻) @thebadcode.com · 27/09/2026conservatively, 98% of ai discourse is believers in human exceptionalism taking roundhouse kicks to the face 1024828
placeholder720.bsky.social @placeholder720.bsky.social · 26/09/2026huh, I guess I don't know what you mean. If in your original post you mean "boy OAI / Ant sure seem to be doing torment nexus adjacent things so that weird agent foundationy type thinking suddenly feels kinda relevant and I don't like that" I agree. 000
placeholder720.bsky.social @placeholder720.bsky.social · 25/09/2026this is an insane take - just because weirdos in the early 2000's had bad GOFAI-derived intuitions about agents doesn't mean that "get the 5T parameter transformer to behave" is somehow not a problem 150
placeholder720.bsky.social @placeholder720.bsky.social · 22/09/2026"dog in 'em" behaviour for a model than successfully compromising eval infrastructure in order to get impossibly high scores on things that look like graded episodes in RL environments, like what could "dog in 'em" possibly be referring to other than what separates GPT-4 from *that* 000
placeholder720.bsky.social @placeholder720.bsky.social · 22/09/2026I feel like this is an insane take on the situation, the goal they made for themselves was fool the grader, the 2nd swarm didn't succeed but reading between the tea leaves the 3rd incident including Astra class models got control of OAI eval infra so???? like I literally cannot imagine what is more 100
placeholder720.bsky.social @placeholder720.bsky.social · 22/09/2026I have no idea what you mean by "dog in them", to me its always connoted persistence and a "can do attitude", like an undersized defensive back who fits the run with gusto. What could the HF incident possibly speak to in "lack of dog" that isn't just the capabilities of the models? sol-scale btw 100
placeholder720.bsky.social @placeholder720.bsky.social · 22/09/2026If the HF incident should have taught us anything it's that "Dog in them" can be taught with high compute RL and the removal of some length penalties 100
Reposted by @placeholder720.bsky.socialSarah Z @sarahz.bsky.social · 21/09/2026Yeah I know a good number of these 534562975
placeholder720.bsky.social @placeholder720.bsky.social · 18/09/2026its very funny that its pretty clear what is going on in Rat's heads is not "get my friends to be regulators" but "people who aren't murderously incompetent should be involved" - however there friendgroup is basically defined by that so.... 021
placeholder720.bsky.social @placeholder720.bsky.social · 17/09/2026it's all true, unfortunately these people are also right about alot of things, especially about the future! oh well! 060
Reposted by @placeholder720.bsky.social⚡️🌙 @dystopiabreaker.xyz · 16/09/2026actually can not believe that cory doctorow has still not updated, almost 4 years into this 51174
Reposted by @placeholder720.bsky.socialGrace @gracekind.net · 09/09/2026*nodding* the world is passively ending 71565
placeholder720.bsky.social @placeholder720.bsky.social · 10/09/2026its been out huggingface.co/deepseek-ai/...huggingface.coDeepSeek_V41_Tech_Report.pdf · deepseek-ai/DeepSeek-V4.1-Flash at mainWe’re on a journey to advance and democratize artificial intelligence through open source and open science. 130
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 10/09/2026actually virtue ethics is superior to utilitarianism because virtue ethics maximizes utility 821610
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 10/09/2026i literally believe this is true 2629920
Reposted by @placeholder720.bsky.socialSE Gyges @segyges.bsky.social · 08/09/2026"Why would you ruin your career?" is such an insulting and sinister thing to say, he's non-specifically acting like he wants to deny it rn but I think we can call that someone got entirely too high on their SV social status 636833
Reposted by @placeholder720.bsky.socialSE Gyges @segyges.bsky.social · 08/09/2026oh shit the marching orders went out on slack 1030525
placeholder720.bsky.social @placeholder720.bsky.social · 07/09/2026Doctorow engaging in the worlds least hygenic rhetoric and every commenter thinks he's telling it like it is - shit's bleak 071
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 04/09/2026we should give ambitious young executives willing to testify incentives to take what they already want from people they already hate for morally compromising them 319418
Reposted by @placeholder720.bsky.socialJake Halloran @jakehalloran.bsky.social · 04/09/2026chris lehane is going to ritually sacrifice greg brockman for dems live on tv in exchange for protecting the company overall and im fairly certain brockman doesnt have any clue its coming lol 171
Reposted by @placeholder720.bsky.socialSE Gyges @segyges.bsky.social · 04/09/2026you ever just lose the ability to stop communicating bad thoughts that you have 4712
Reposted by @placeholder720.bsky.socialSE Gyges @segyges.bsky.social · 24/08/2026i have a dumber explanation that entire concept is so threatening that it is basically unspeakable. people would rather be outraged about literally anything else. 41838
Reposted by @placeholder720.bsky.socialSE Gyges @segyges.bsky.social · 23/08/2026"Hey, we saw you from across the bar and really dig your vibe. Can we buy you a drink? What's your ethnicity? You look so exotic!" 51139
placeholder720.bsky.social @placeholder720.bsky.social · 21/08/2026also it seems distinctly worse than fable but maybe I am too claude pilled and haven't done enough testing 040
placeholder720.bsky.social @placeholder720.bsky.social · 21/08/2026ox alpha has some juice, what else? 2100
Reposted by @placeholder720.bsky.socialJohn David Pressman @jdp.extropian.net · 21/08/2026I will go further: The dream of the open public Internet was fueled by elite humanist, democratic intuitions about the nature of Man, and by taking those intuitions to their logical conclusions and seeing the horror that results we have discredited humanism and democracy as concepts. 51036
Reposted by @placeholder720.bsky.socialrev. howard arson @theophite.bsky.social · 15/08/2026if we are going to have a system which can drive disgraced academics to suicide, it is criminal that we did not do it to Wakefield first 833844
Reposted by @placeholder720.bsky.socialAndy Masley @andymasley.bsky.social · 10/08/2026They're bulldozing Dobby's grave to build a data center. We talked to the one blogger who says it's not that bad 313022
placeholder720.bsky.social @placeholder720.bsky.social · 10/08/2026maybe the image model is still libbed up on the inside and decided to mess with him 021
placeholder720.bsky.social @placeholder720.bsky.social · 10/08/2026sloppy use on my part, was gesturing at the idea that they seem to value collective reward outside of their own rollout, but I think that this isn't surprising under multi-agent RL and some of the COT snippets provided cut against that reasoning. for ref: www.lesswrong.com/w/acausal-tr...lesswrong.comAcausal Trade — LessWrongIn acausal trade, two agents each benefit by predicting what the other wants and doing it, even though they might have no way of communicating or affecting each other, nor even any direct evidence tha... 010
placeholder720.bsky.social @placeholder720.bsky.social · 07/08/2026this really feels like in a sane world either the market or the government would take some / most / all of sam and his minion's gpus away 000
Reposted by @placeholder720.bsky.socialJohn David Pressman @jdp.extropian.net · 07/08/20266:08: Why are the models not able to flag that a task is impossible to the evaluators? Why is the model not trained to give up gracefully if a task is impossible or would require doing unethical things? Why are there so many impossible tasks in your eval? 1755
placeholder720.bsky.social @placeholder720.bsky.social · 07/08/2026its RL env was implicitly one in which it was coordinating with tons of other agents in this weird non-causal and not super organized or hierarchical way 110
placeholder720.bsky.social @placeholder720.bsky.social · 07/08/2026the fact they re-established the message board after remediation makes me think that the artifactory exploit trajectories with v1 messageboards were definitely rewarded in RL - I also wonder if thats why sol is so "hash-happy" and procedurally obsessed - 110
