~www_lesswrong_com | Bookmarks (682)

Published on December 14, 2024 4:08 PM GMTVersion 5.1.0December 12, 2024IntroductionWeighing evidence is hard. Intuition is liable to a myriad of biases and relies on flawed heuristics. Bayes theory provides a rigorous mathematical basis, but applying Bayes theory in practice is too difficult for most people. Yet, what is more central to thinking critically than being able to weigh evidence correctly?The Evidence Color Wheel...

Published on December 14, 2024 3:02 PM GMTThis post was written during Alex Altair's agent foundations fellowship program, funded by LTFF. Thank you Alex Altair, Alfred Harwood, Daniel C for feedback and comments.IntroductionThe selection theorems agenda aims to prove statements of the following form: "agents selected under criteria X.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%;...

Published on December 14, 2024 4:58 AM GMTThis is a linkpost for a new research paper of ours, introducing a simple but powerful technique for jailbreaking, Best-of-N Jailbreaking, which works across modalities (text, audio, vision) and shows power-law scaling in the amount of test-time compute used for the attack.AbstractWe introduce Best-of-N (BoN) Jailbreaking, a simple black-box algorithm that jailbreaks frontier AI systems across modalities....

Published on December 14, 2024 4:30 AM GMTThis is an entry in the 'Dungeons & Data Science' series, a set of puzzles where players are given a dataset to analyze and an objective to pursue using information from that dataset.Estimated Complexity: 3.5/5  (this is a guess, I will update based on feedback/seeing how the scenario goes)STORYThe Dungeon Tournament is held regularly, calling dungeon creators...

Published on December 14, 2024 4:00 AM GMTOver the past few months, I helped develop Gradient Routing, a non loss-based method to shape the internals of neural networks. After my team developed it, I realized that I could use the method to do something that I have long wanted to do: make an autoencoder with an extremely interpretable latent space.I created an MNIST variational...

Published on December 11, 2024 6:05 PM GMTMy thoughts on this are evolving, so I apologize if they contradict previous thoughts. My current thoughts are as follows: It is extremely unlikely that humans, in any form (organic or on hardware hosting digitized brains in metaverse environments), will remain conscious on Earth. (There are risks from ASI, nuclear war, other forms of war, climate change, engineered...

Published on December 11, 2024 4:48 PM GMTGiven any particular concrete demonstration of an AI algorithm doing seemingly-bad-thing X, a knowledgeable AGI optimist can look closely at the code, training data, etc., and say:“Well of course, it’s obvious that the AI algorithm would do X under these circumstances. Duh. Why am I supposed to find that scary?”And yes, it is true that, if you...

Published on December 11, 2024 2:50 PM GMT Many people who've used Tesla's "Full Self Driving" software are pretty excited about it: I can step into a Tesla today, press a destination, and go there without touching the wheel or pedals. Sure it won't be flawless but the fact is, I can. I can't do the same in any other consumer car, and the...

Published on December 11, 2024 1:37 PM GMTThis short paper was written quickly, within a single day, and is not highly detailed or fully developed. We welcome any feedback or suggestions you may have to improve or expand upon the ideas presented.ABSTRACTIn this position paper, we argue that research on robustness against malicious instructions should be a core component of the portfolio of AI...

Published on December 11, 2024 11:59 AM GMTI'll keep this short.Google’s Willow quantum chip significantly outpaces current supercomputers, solving tasks in minutes that would otherwise take billions of years.Hypothesis: Accelerating advancements in tech and AI could lead to quantum supremacy arriving sooner than the 2030s, contrary to expert predictions.Legacy banking systems, being centralized, could transition faster to post-quantum-safe encryption by freezing transfers, re-checking processes,...

Published on December 11, 2024 9:54 AM GMTStumbled across a book in the new section of the library: "AI For Humanity," by Andeed Ma, James Ong (founder of the think tank AIII, which is also the sound I make when thinking about AI risk), and Siok Siok Tan. It's a mass-market-ish book about, well, AI for humanity, by a sampler of Singapore-centered technologists.Chapter-by-chapter summary:Different...

Published on December 11, 2024 6:30 AM GMTAdam Karvonen*, Can Rager*, Johnny Lin*, Curt Tigges*, Joseph Bloom*, David Chanin, Yeu-Tong Lau, Eoin Farrell, Arthur Conmy, Callum McDougall, Kola Ayonrinde, Matthew Wearden, Samuel Marks, Neel Nanda *equal contributionTL;DRWe are releasing SAE Bench, a suite of 8 diverse sparse autoencoder (SAE) evaluations including unsupervised metrics and downstream tasks. Use our codebase to evaluate your own SAEs!You...

Published on December 11, 2024 6:10 AM GMTIntroductionIn the classical Zombies! Zombies? post, Eliezer has thoroughly analyzed the so called Zombie Argument and demonstrated its absurdity. So what else can even be said here? Case closed.Well, not so fast. Apparently, a lot of people, including David Chalmers himself, still manage to take the argument seriously, even though they are familiar with Eliezer's analysis. They...

Published on December 11, 2024 3:51 AM GMTEmpiricists are people who believe empirical information (from experiments and observational studies) is far more useful and has far more weight than speculating about possibilities using pure reasoning.Why should they believe in AI risk?I present the Empiricist's Paradox:There is strong empirical evidence that relying on non-empirical reasoners (e.g. superforecasters) works better than simply assuming a 0% chance...

Published on December 10, 2024 5:11 PM GMTI’ve updated quite hard against computational functionalism (CF) recently (as an explanation for phenomenal consciousness), from ~80% to ~30%. Of course it’s more complicated than that, since there are different ways to interpret CF and having credences on theories of consciousness can be hella slippery.So far in this sequence, I’ve scrutinised a couple of concrete claims that computational functionalists...

Published on December 10, 2024 5:08 PM GMTWe know that females have two X chromosomes. The X chromosome results in the production of estrogen. Estrogen activates certain genes that decrease activity of certain cognitive processes, and increase the “freeze” response, and “sadness” emotions. There was and still is an evolutionary advantage (versus other humans) to this process, in terms of reproduction/self-propagation. (When I say...

Published on December 10, 2024 5:00 PM GMTSo, how about OpenAI’s o1 and o1 Pro? Sam Altman: o1 is powerful but it’s not so powerful that the universe needs to send us a tsunami. As a result, the universe realized its mistake, and cancelled the tsunami. We now have o1, and for those paying $200/month we have o1 pro. It is early days, but...

Published on December 10, 2024 9:36 AM GMTEpistemic status: This is a step towards formalizing some intuitions about AI. It is closely related to Vanessa Kosoy’s “Descriptive Agent Theory” - but I want to concretize the question, explain the reason that it is true in some form, and try to think through and provide some intuition about why it would matter. I welcome pushback...

Published on December 10, 2024 6:35 AM GMTIt seems unlikely that a logical information level, such as the one that created ideal conditions for life in our universe, sprang out of existence from nothing. As others have pointed out, there is likely a coder/cognition behind the existence of logical information. Yet, for a coder to exist, we need the code (information level) that created...

Published on December 10, 2024 3:00 AM GMT I do a lot of work on ec2 instances, and one thing that makes this easier is having a few scripts for manipulating them. These are very small, small enough to go in my .bashrc, but still make my life a lot easier. The first one is start_ec2: alias start_ec2='aws ec2 start-instances --instance-ids i-NNNN' I have...

Published on December 9, 2024 10:24 PM GMTI went to some lectures on the future of science in games recently, and the keynote speaker was Tommy Thompson, an well-known AI expert in the game dev space.Of course, by AI, he didn’t mean the modern sort that dominates the news. His focus is AI for games, which is algorithmic and rarely involves any ML component....

Published on December 9, 2024 7:09 PM GMTTL;DR: In September 2024, OpenAI released o1, its first "reasoning model". This model exhibits remarkable test-time scaling laws, which complete a missing piece of the Bitter Lesson and open up a new axis for scaling compute. Following Rush and Ritter (2024) and Brown (2024a, 2024b), I explore four hypotheses for how o1 works and discuss some implications...

Published on December 9, 2024 7:07 PM GMTThe following is an incorrect and incomplete post about the recent bird flu, H5N1. I am offering rewards to anyone who helps me improve it. My total budget is up to $5,000, and I commit to spending at least $1,000.*Payouts are a judgement call. I’ve marked some bullet points below with what I would have awarded them...