~www_lesswrong_com | Bookmarks (669)

Published on August 10, 2024 10:07 AM GMTEpistemic status: This post was planned to be part of a broader "Holistic Interpretability" post but this isn't going as fast as I'd like so I am releasing the foreword to get some early feedback on whether I should pursue this or not. I haven't had a lot of red teaming on this emergence idea, but I'd...

Published on August 10, 2024 7:00 AM GMTAlex Lawsen used a great metaphor on the 80k After Hours podcast:[1][1:38:14] …you’re rowing a boat on your own, and you’re trying to get somewhere, and you’ve got some map that you need to look at to see where you’re going, I imagine like a map and compass. […] When you’re rowing, you’re facing back; you can’t...

Published on August 9, 2024 11:21 PM GMTIn September 2023, Max Tegmark and Steve Omohundro proposed "Provably Safe AI" as a strategy for AI Safety. In May 2024, a larger group delineated the broader concept of "Guaranteed Safe AI" which includes Provably Safe AI and other related strategies. In July, 2024, Ben Goldhaber and Steve discussed Provably Safe AI and its future possibilities, as summarized in...

Published on August 9, 2024 10:20 PM GMTfrom Peng, et al; stimulating the nucleus accumbens inhibits its activityTranscranial focused ultrasound neuromodulation — altering the brain’s activity with low-intensity ultrasound — is really exciting . It allows us to manipulate arbitrary regions of the brain without surgery, potentially replacing the (brain-damaging) electrode implants currently used for serious neurological conditions like epilepsy and Parkinson’s, and potentially...

Published on August 9, 2024 10:07 PM GMTTl;dr, A choice of variable in causal modeling is good if its causal effect is consistent across all the different ways of implementing it in terms of the low-level model. This notion can be made precise into a relation among causal models, giving us conditions as to when we can ground the causal meaning of high-level variables...

Published on August 9, 2024 7:57 PM GMTThis post is also available on my Substack.Where we last left off, I proposed that life cannot be understood through statistics, but rather requires more careful study of individual cases. Yet I don’t think this is actually convincing to the statistical/academic mindset unless we have a statistical description of the phenomenon. To understand it, I’ve developed a...

Published on August 9, 2024 6:17 PM GMTThis is written right  after reading Rob Bensinger's relevant post and andeslodes' comments. That discussion touched on a topic I have long held a strong belief about. I purposed first-person perspective is not physically reducible and ought to be regarded as primitive. Following that, questions such like "which future mind is me?" or "which mind ought to...

Published on August 9, 2024 6:09 PM GMTDisclaimer: I am a math student new to LW and the rationalist community. I could not find this idea anywhere else after a modest amount of research; I apologize if this has been brought up before.Summary: CDT agents may behave differently than expected in situations involving reliable predictors due to the possibility that they are themselves simulations...

Published on August 8, 2024 6:33 PM GMTThis work was produced at Apollo Research, based on initial research done at MATS.LayerNorm is annoying for mechanstic interpretability research (“[...] reason #78 for why interpretability researchers hate LayerNorm” – Anthropic, 2023).Here’s a Hugging Face link to a GPT2-small model without any LayerNorm.The final model is only slightly worse than a GPT2 with LayerNorm[1]:DatasetOriginal GPT2Fine-tuned GPT2 with LayerNormFine-tuned...

Published on August 8, 2024 6:32 PM GMTThis is slightly old news at this point, but: as part of MIRI's recent strategy pivot, they've eliminated the Agent Foundations research team. I've been out of a job for a little over a month now. Much of my research time in the first half of the year was eaten up by engaging with the decision process...

Published on August 8, 2024 6:27 PM GMTThe tokenizer for GPT-4o is interesting. It's called o200k_base, after the vocabulary of 200019 tokens it uses to represent unicode text from different languages, which we will explore some of in this post. A language model trained on o200k tokenization will start out thinking that all the tokens are equally important. To illustrate, the concept of a "horse"...

Published on August 8, 2024 6:22 PM GMTEliezer Yudkowsky’s main message to his Twitter fans is:Aligning human-level or superhuman AI with its creators’ objectives is also called “superalignment”. And a month ago, I proposed a solution to that. One might call it Volition Extrapolated by Language Models (VELM).Apparently, the idea was novel (not the “extrapolated volition” part):But it suffers from the fact that language...

Published on August 8, 2024 4:29 PM GMTSummary In this post I want to briefly share some results I have got after experimenting with the equivalent version of the simple neural networks that the authors used here to study how superposition and poly-semantic neurons come about in neural networks trained with gradient descent. The take-home message is that BitLinear networks are similar to their...

Published on August 8, 2024 4:16 PM GMTFollowup to: Some epistemological conundrums. Response to: The omnigenic model as a metaphor for life. Part 1 in this post will introduce a concept from programming, part 2 will apply it to life in general. This post is also available on my Substack.Part 1: Performance optimizationIn computer programming, there is a tradeoff between software that is quick...

Published on August 8, 2024 3:59 PM GMTRandomized Control Trials have some drawbacks. For many important questions, like causes of the industrial revolution, a randomized trial is impossible. For many others, RCTs are expensive and cumbersome, leading to low sample sizes or experimental designs that precisely answer irrelevant questions. Still, when RCTs with large sample size and generalizable designs are possible, their advantages justify...

Published on August 8, 2024 1:52 PM GMTTake a large thick cardboard box. Cut it into planar pieces. Take transparent tape and cover the cardboard's surface with it. Tadaa! A whiteboard:Because transparent tape is smooth the whiteboard pen is erasable.Sticky notes stick a lot better on the transparent tape than on the cardboard, which is useful for Kanban Boards!Ideally, put white paper between the...

Published on August 8, 2024 1:50 PM GMTIf you’re looking for audio of my posts, you’re in luck. Thanks to multiple volunteers you have two options. Option one is Askwho, who uses a Substack. You can get an ElevenLabs-quality production, with a voice that makes me smile. For Apple Podcasts that means you can add them here, Spotify here, Pocket Casts here, RSS here....

Published on August 8, 2024 5:05 AM GMTThis is a brief sketch of a theory of motivation.Motivation drives action. It also causes the experiences of pleasure and pain. We experience pain when motivation increases, and pleasure when motivation decreases.Motivation is generated by emotions, such as hunger, thirst and lust. Emotions generate motivation, and motivation generates action. Each emotion has a biological function. Hunger motivates...

Published on August 6, 2024 5:44 PM GMTWork supported by MATS and SPAR. Code at https://github.com/ArjunPanickssery/math_problems_debate/.Three measures for evaluating debate arewhether the debate judge outperforms a naive-judge baseline where the naive judge answers questions without hearing any debate arguments.whether the debate judge outperforms a consultancy baseline where the judge hears argument(s) from a single "consultant" assigned to argue for a random answer.whether the judge...

Published on August 6, 2024 1:30 PM GMTPreviously: Startup Roundup #1. This is my periodic grab bag coverage of various issues surrounding startups, especially but not exclusively tech-and-VC style startups, that apply over the longer term. I always want to emphasize up front that startups are good and you should do one. Equity and skin in the game are where it is at. Building...

Published on August 6, 2024 5:43 AM GMTBelief in genetic tribalism comes from the idea that we have natural genetic interests in those who share our genes. This seems superficially plausible. We have a natural interest in our children, don’t we? That interest is because we share genes, right? If so, we should also have a genetic interest in our ethnic groups or races,...

Published on August 6, 2024 10:33 AM GMTOver the last few months, the EleutherAI interpretability team pioneered novel, mechanistic methods for detecting anomalous behavior in language models based on Neel Nanda's attribution patching technique. Unfortunately, none of these methods consistently outperform non-mechanistic baselines which look only at activations.We find that we achieve better anomaly detection performance with methods that evaluate entire batches of test...

Published on August 6, 2024 9:29 AM GMTIn the past AI systems have reached super human performance by adding search to neural networks while the network alone could not reach the level of the best humans. At least this seems to be the case for AlphaGo, AlphaZero/Leela, AlphaGeometry and probably more, while AlphaStar and OpenAI Five where adding search was not easily possible have failed...

Published on August 6, 2024 9:26 AM GMTThere is a growing? fraction of people who consider LLMs to be AGI. And it makes sense. Clearly, when the term AGI was established this was what was meant: A machine that can tackle a wide range of problems, communicate with natural language, very different from all the examples of narrow AI.It also prevents continuous goalpost moving....