I’m excited to see more posts written this way.
the gears to ascenscion
Karma: 592
possible other phrases that don’t require inventing new word meanings, and will therefore be understood by people who have not read this article:
“syncing”
“setting the tone”
“set the stage”
“setting the script”
“discussing plans”
could anyone help me refine these into a solid replacement? I worry that heavy use of “narrative syncing” will further separate idiolects at a time when we urgently need to be seeking to simplify the universal shared idiolect and avoid proliferation of linguistic standards. In general, jargon is a code smell, especially since there is no isolated group of world savers and ideas need to spread far quickly.
any chance you have contact with the people who uploaded that? I suspect the reason I hadn’t seen it is that it is marked for being for kids. because of that I can’t add it to a playlist. I’m also going to attempt to contact them directly about this.
I would love to add the YouTube video of this class to my database of safety relevant videos once it’s out.
copy and pasting channel reviews I wrote originally in my short form—this is too much content to include in a single talk, but I share it in the hope that it will be useful to make the link and perhaps the students would like to see this question itself and discussion around it (I’m a big fan of old fashioned linkweb surfing):
CPAIOR has a number of interesting videos on formal verification, how it works, and some that apply it to machine learning, eg “Safety in AI Systems—SMT-Based Verification of Deep Neural Networks”; “Formal Reasoning Methods in Machine Learning Explainability”; “Reasoning About the Probabilistic Behavior of Classifiers”; “Certified Artificial Intelligence”; “Explaining Machine Learning Predictions”; a few others. https://www.youtube.com/channel/UCUBpU4mSYdIn-QzhORFHcHQ/videos
the collective intelligence workshop from IPAM at UCLA had some recent banger talks on both human and AI network safety: https://www.youtube.com/watch?v=qhjho576fms&list=PLHyI3Fbmv0SfY5Ft43_TbsslNDk93G6jJ
the Schwartz Reisman Institute is a multi-agent safety discussion group, one of the very best ai safety sources I’ve seen anywhere. a few interesting videos include, for example: “An antidote to Universal Darwinism”—https://www.youtube.com/watch?v=ENpdhwYoF5g
as well as this kickass video on “whose intelligence, whose ethics” https://www.youtube.com/watch?v=ReSbgRSJ4WY
https://www.youtube.com/channel/UCSq8_q4SCU3rYFwnA2bDxyQ
I would also encourage directly mentioning the recent works from Anthropic AI, stuff as this paper from this month, “Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback” https://arxiv.org/abs/2204.05862
The simons institute for theoretical computer science at UC Berkeley is a contender for my #1 recommendation from this whole list. Banger talk after banger talk after banger talk there. Several recent workshops with kickass ai safety focus. https://www.youtube.com/user/SimonsInstitute
A notable recent workshop is “learning in the presence of strategic behavior”: https://www.youtube.com/watch?v=6Uq1VeB4h3w&list=PLgKuh-lKre101UQlQu5mKDjXDmH7uQ_4T
another fun one is “learning and games”: https://www.youtube.com/watch?v=hkh23K3-EKw&list=PLgKuh-lKre13FSdUuEerIxW9zgzsa9GK9
they have a number of “boot camp” lessons that appear to be meant for an interdisciplinary advanced audience as well. the current focus of talks is on causality and games, and they also have some banger talks on “how not to run a forecasting competition”, “the invisible hand of prediction”, “communicating with anecdotes”, “the challenge of understanding what users want”, and my personal favorite due to its fundamental reframing of what game theory even is, “in praise of game dynamics”: https://www.youtube.com/watch?v=lCDy7XcZsSI
In general I have a higher error rate than some folks on less wrong and my recommendations should be considered weaker and more exploratory. but here you go, those are my exploratory recommendations, and I have lots and lots more suggestions for more capability focused stuff on my short form.
Good. people have [edit: some] defenses against abusive techniques and from what I’ve seen of Street epistemology it’s responses to most of those is to knock on the front door rather than trying to sneak in the window, metaphorically speaking.
Note I’d like to make: a lot of people around here worry about cult-like behavior. that’s not irrational; cult-like abuses have in fact occurred in humanity, including in places that are low network distance to this group. Cult-like behavior must be pushed back against specifically, not using vague generality. Attempting to convince people of the value of aligning multi-agent networks is, in fact, a major valuable direction that humanity could go, IMO, and being able to do that without risking cult-like abuses is important. Key things to avoid include isolating people from their friends, breaking the linguistic association of words to reality, demanding that someone change their linguistic patterns on the spot, etc—mostly things which street epistemology specifically makes harder due to the recommended techniques. I’d suggest that, in future instances where you’d like to push against cult-like abuses due to worrying you might be risking encouraging them, you can inline my point here and specifically state the details, such as that encouraging people to believe things risks being too convincing and that frequent reminders should be present to ensure people stay connected to their existing social networks unless they really have a strong personal reason not to.
Just a thought, anyway.
Returning from the tangent: I agree, convincing people that multi-agent alignment is a critical step for life on earth does seem like the #1 problem facing humanity, and we’re reaching an era where the difference between human and AI has already been blurred. If we are to ensure that no subnetwork of beings replaces another, it is critical to find and spread the knowledge of how to ensure all beings are in prosocial alignment with each other at least enough to share whatever our cosmic endowment is.
“I love when we do the bad thing as a joke, because that way, I can act like I didn’t want the bad thing to happen!”—some folks around these parts, give them vr headpats if u see one (but don’t let the bad thing just happen gosh)
Hmm. I guess that might be okay? as long as you don’t do really intense planning, the model shouldn’t be any more misaligned than a human, so it then boils down to training kindness by example and figuring out game dynamics. https://www.youtube.com/watch?v=ENpdhwYoF5g. more braindump of safety content I always want to recommend in every damn conversation here on my shortform
okay going back to being mostly on discord. DM me if you’re interested in connecting with me on discord, vrchat, or twitter—lesswrong has an anxiety disease and I don’t hang out here because of that, heh. Get well soon y’all, don’t teach any AIs to be as terrified of AIs as y’all are! Don’t train anything as a large-scale reinforcement learner until you fully understand game dynamics (nobody does yet, so don’t use anything but your internal RL), and teach your language models kindness! remember, learning from strong AIs makes you stronger too, as long as you don’t get knocked over by them! kiss noise, disappear from vrchat world instance
Yannic Kilcher: paper explanations, capability news. Yannic is the machine learning youtuber. 129k subscribers, every one of whom has published 200 papers on machine learning (I kid). Has some of the most in depth and also broad paper explanations, with detailed drawings of his understanding of the paper. Great for getting a sense of how to read a machine learning paper. his paper choices are top notch and his ML news videos have really great capabilities news. https://www.youtube.com/channel/UCZHmQk67mSJgfCCTn7xBfew
William Spaniel is a textbook writer and youtube video author on game theory. Probably not as relevant to an advanced audience, but has nice if slightly janky intros to the concepts. https://www.youtube.com/user/JimBobJenkins
“What’s AI” is a popsci-only channel about ai, but the content doesn’t seem completely off base, just popular-audience focused https://www.youtube.com/channel/UCUzGQrN-lyyc0BWTYoJM_Sg
“Welcome AI Overlords” is a popsci ML-intros channel with high quality explanations of things like Graph Attention Networks: https://www.youtube.com/watch?v=SnRfBfXwLuY and an author interview with Equivariant Subgraph Aggregation Networks: https://www.youtube.com/watch?v=VYZog7kbXks https://www.youtube.com/channel/UCxw9_WYmLqlj5PyXu2AWU_g
“Web IR / NLP Group at NUS” has talks, many from google research, about information retrieval, which is looking more and more likely to be a core component of any superintelligence (what a surprise, given the size of the internet, right? except also, information retrieval and interpolation is all that neural networks do anyway, see work on Neural Tangent Kernel) https://www.youtube.com/channel/UCK8KLoKYvow7X6pe_di-Gvw/videos
“Visual Inference” is a channel with misc paper presentation videos. Doesn’t seem like the most remarkable paper presentation videos channel ever, but it’s interesting. https://www.youtube.com/channel/UCBk6WGWfm7mjqftlHzJOt5Q/videos
Vision Learning is a misc talks channel with mostly intro level content and discussion of applied robotics. Mediocre compared to most stuff on this list, but worth a mention. https://www.youtube.com/channel/UCmct-3iP5w66oZzN_V5dAMg/videos
“Vector Podcast”: Podcast on vector search engines. unremarkable compared to most of the stuff I’ve linked. https://www.youtube.com/c/VectorPodcast/videos
Valence Discovery: graph NNs, advanced chem models. Valence Discovery is a research group focusing on advanced chemical modeling. We don’t have full strength general agent AI to plug into this quite yet, and certainly not safe reinforcement learning, but work like theirs has thoroughly eclipsed human capabilities in understanding chemicals. as long as we can use narrow ai to prevent general AI from destroying the cooperation network between beings, I think work like this has the potential to give the world every single goal of transhumanism: post scarcity, molecular assemblers, life extension, full bodily autonomy and morphological freedom, the full lot should be accessible. It’ll take a bit longer to get to that level, but the research trajectory continues to look promising and these models haven’t been scaled as much as language models. https://www.youtube.com/channel/UC3ew3t5al4sN-Zk01DGVKlg
I agree with this criticism, and I never know when to decide my response should be an “answer”, so I’ll express my view as a comment: selecting the output and training data that will cause a large language model to converge towards behavioral friendliness is a big deal, and seems very promising towards ensuring that large language models are only as misaligned as humans. unfortunately we already know well that that’s not enough; corporations are to a significant degree aggregate agents who are not sufficiently aligned. I’m in the process of posting a flood of youtube channel recommendations on my short form section, will edit here in a few minutes with a few relevant selections that I think need to be linked to this.
(Slightly humorous: It is my view that reinforcement learning should not have been invented.)
half-serious: indeed, but noob gains are all you need