[Chris Olah - Looking Inside Neural Networks with Mechanistic Interpretability - YouTube](https://youtu.be/2Rdp9GvcYOE?si=F-icT_wXVIbAeeFG&t=1580) in CNNs we identified edge detectors, color detectors, depth detectors, shape detectors, fur detectors, cat detectors,... etc. that compose that you can also manually compose "Neural networks, if we take them seriously as objects of investigation, are full of beautiful structure. And I sometimes think that actually maybe, you know, the elegance of machine learning is in some ways more like the elegance of biology, or perhaps at least as much as like the elegance of biology, as math or physics. So, in the case of biology, evolution creates awe-inspiring complexity in nature. You know, we have this system that goes and generates all this beautiful structure. And in a kind of similar way, gradient descent, it seems to me, creates beautiful, mind-boggling structure. It goes and it creates all these beautiful circuits that have beautiful symmetries in them. And in toy models, it arranges features into regular polyhedra. And there's all of this, just like, it's just sort of too good to be true in some ways, it's full of all this amazing stuff. And then there's messy stuff, but then you discover sometimes that messy stuff actually was just a really beautiful thing that you didn't understand. And so a belief that I have, that is only perhaps semi-rational, is that these models are just full of beautiful structure if we're just willing to put in the effort to go and find it. And I think that's the thing that I find actually most emotionally motivating about this work." - Chris Olah I think about that a lot https://twitter.com/MikePFrank/status/1733809292636041331?t=5ewU2encPqQjpMNDHsT7DQ&s=19 But also: "Provable safety sounds like a pipe dream to me. GOFAI will never achieve very much, and survival of the fittest doesn’t require proving anything. Proof just slows you down. Perfect rationality is impossible. And interactions with an unknown external environment aren’t the sort of thing that’s amenable to rigorous analysis in the first place. There are fundamental computability limits on what’s knowable there due to computational irreducibility arguments. Not to mention dynamical chaos, quantum uncertainty, etc. Face it, the future is going to be messy and unpredictable despite all your best efforts, and it makes more sense to just go with the flow rather than trying to analyze it to death mathematically." - @MikePFrank And responce by @RespondAsOne : "In the limit you might be right, but not with the same standard for understanding as driving in traffic requires. We don’t find the risks of driving on roads intractable. It’s probably safe enough. Issue is we have no such shared terms - yet - which you correctly point out there." I will continue on trying to analyze it to death mathematically. Even if it might be impossible perfectly, I believe it is possible in good enough way for us to understand, predict and steer it in ways we want. Results are getting better and better. Representation Engineering: A Top-Down Approach to AI Transparency Understanding and Controlling the Inner Workings of Neural Networks [Representation Engineering: A Top-Down Approach to AI Transparency](https://www.ai-transparency.org/) Towards Monosemanticity: Decomposing Language Models With Dictionary Learning "Using a sparse autoencoder, we extract a large number of interpretable features from a one-layer transformer." [Towards Monosemanticity: Decomposing Language Models With Dictionary Learning](https://transformer-circuits.pub/2023/monosemantic-features/index.html) [A Walkthrough of Automated Circuit Discovery w/ Arthur Conmy Part 1/3 - YouTube](https://www.youtube.com/watch?v=dn4GqR0DCx8) I hope this automated circuit discovery will be automated more and more and scale to GPT4like models! LLMs explaining neurons in LLMs is also promising! https://openai.com/research/language-models-can-explain-neurons-in-language-models Let's see what Max Tegmark and others will come up with [Provably Safe Systems: The Only Path to Controllable AGI - YouTube](https://www.youtube.com/watch?v=nUrYCUkTFE4) Can we use the method of reverse engineering dynamical systems by using autoencoders to learn their dynamics and squeezing them and extracting interacting variables from them to reverse engineer and make interpretable the dynamics of CNNs, LLMs and other AI architectures? [Can AI disover new physics? - YouTube](https://www.youtube.com/watch?v=XRL56YCfKtA) [Concrete open problems in mechanistic interpretability | Neel Nanda | EAG London 23 - YouTube](https://www.youtube.com/watch?v=7t9umZ1tFso) for example one network learned composing trigonometry [[2301.05217] Progress measures for grokking via mechanistic interpretability](https://arxiv.org/abs/2301.05217) another network learned group theory operations via representation theory [[2302.03025] A Toy Model of Universality: Reverse Engineering How Networks Learn Group Operations](https://arxiv.org/abs/2302.03025) another one learned explicit looking at history https://www.lesswrong.com/posts/TvrfY4c9eaGLeyDkE/induction-heads-illustrated another one finite state automatalike circuits for compsing html [Towards Monosemanticity: Decomposing Language Models With Dictionary Learning](https://transformer-circuits.pub/2023/monosemantic-features/index.html) etc., and many more are being found each month, or from top down perspective you can play with latent vectors [ORIGINAL FATHER OF AI ON DANGERS! (Prof. Jürgen Schmidhuber) - YouTube](https://youtu.be/q27XMPm5wg8?si=De4qh67860X8egNC) 20:00 Asymptote limits to recursive selfimprovement: godel limits of computation, mathematical optimality and limits of certain algorithms, physical laws, lightspeed, locality, [Bremermann's limit - Wikipedia](https://en.wikipedia.org/wiki/Bremermann%27s_limit) [Landauer's principle - Wikipedia](https://en.wikipedia.org/wiki/Landauer%27s_principle) Is learned world model in one RL life with editing rewards leading to planning all you need for AGI? https://twitter.com/karpathy/status/1733968385472704548?t=3kujKBSuDrsIGktmIAHdPQ&s=19 ML News sources https://twitter.com/IntuitMachine/status/1734180741997551696?t=HW5jleMUzlNm01Qn1WbQSQ&s=19 Mixtral can also run on gamer GPUs Lean AI copilot https://fxtwitter.com/AnimaAnandkumar/status/1734080043196768348?t=R_gFYCBMEVySp88ThynSjA&s=19 List of open LLMs [GitHub - eugeneyan/open-llms: 📋 A list of open LLMs available for commercial use.](https://github.com/eugeneyan/open-llms) Jedna z věcí o čem nejvíc básní Bostrom kterýho racionalisti/EAs mají hodně rádi je risk AI tyranie a asi s tím teď soucítím nejvíc AGIs everywhere across labs and countries will happen. I dont think it can be stopped. Certain risks are real and might be or already are happening. Like strenghtening of totalitarian tendencies by using AI to destabilize democracy. I think the best available strategy we have is to create counter AIs as defence. Any good tyranny resistance orgs are gonna need portable GPU clusters tbh. Llama 2, open source model který je lepší než ChatGPT3.5, už jde spustit na gamer grafice a Mixtral co je na tom podobně jak Llama 2 se k tomu taky blíží. Zatím to vypadá že centralized orgs jsou rok napřed před decentralized orgs ale ten gap se snižuje V tý evropský byrokracii je fakt těžký rozeznat lidi co to myslí reálně vážně pro good of all (jakkoli si "good" zadefinují), kompetentně či nekompetenčně, a corrupt nebo power hungry sociopathic agents Unabstracting programming https://twitter.com/burny_tech/status/1734223898751721805 The intense denial of current AI capabilities from so many people is mindblowing [My techno-optimism](https://vitalik.eth.limo/general/2023/11/27/techno_optimism.html) dragon robot https://twitter.com/gunsnrosesgirl3/status/1733753289009701305 QTF manim [Quantum Field Theory EP 1: 0-dimensional quantum fields - YouTube](https://www.youtube.com/watch?v=K9fxNckP29s) is for loop all you need for agi https://twitter.com/teortaxesTex/status/1734319242919436638 Computer from brain tissue [Human Brain Cells on a Chip Can Recognize Speech And Do Simple Math : ScienceAlert](https://www.sciencealert.com/scientists-built-a-functional-computer-with-human-brain-tissue) [Brain organoid reservoir computing for artificial intelligence | Nature Electronics](https://www.nature.com/articles/s41928-023-01069-w) [ORIGINAL FATHER OF AI ON DANGERS! (Prof. Jürgen Schmidhuber) - YouTube](https://www.youtube.com/watch?v=q27XMPm5wg8) [ORIGINAL FATHER OF AI ON DANGERS! (Prof. Jürgen Schmidhuber) - YouTube](https://www.youtube.com/watch?v=q27XMPm5wg8) K tomu teď byl krok před se spiking neural networks místo attention mechanismu v transformerech na neuromorphic chips https://fxtwitter.com/burny_tech/status/1734269400633544936 Spiking neural networks jsou architektura blíž k mozku a neuromorphic chips je hardware blíž k mozku Ještě by to chtělo přidat ty nový najitý nelinearity pokud se chceme přiblížit mozku ještě víc 😄 [Dendrites: Why Biological Neurons Are Deep Neural Networks - YouTube](https://youtu.be/hmtQPrH-gC4?si=UzYUxL4W86WSkFV3) A použít forward forward algorithm místo backpropagation https://medium.com/@Mosbeh_Barhoumi/forward-forward-algorithm-ac24d0d9ffd Záleží co je pro tebe inherentní rozdíl. :D Je tam strašně moc podobností ale zároveň strašně moc odlišností. Pravda že mozek se jenom na pár killowattech výpočtů dokáže naučit celkem dobrou aproximaci světa za jeho život, která je zároveň univerzální napříč lidmi, a zároveň jsou lidi různě specializovaní a s trochu jiným genetickým a naučeným hardwarem a softwarem (colorblidedness, jiný kultury,...), mezitím co LLMs se trénují v gigawattech, protože nejsou limitovaný evolučním šetřením energie a limitovanými materiály XD. Kromě těch podobností co jsme řešili s méďou, tak co se týče inteligence a vzorů, co ty neuronky chytají, tak tam jsou taky podobnosti. [The brain may learn about the world the same way some computational models do | MIT News | Massachusetts Institute of Technology](https://news.mit.edu/2023/brain-self-supervised-computational-models-1030) Mozek se pořád učí v celým jeho životě, mezitím co jazykový modely sežerou celej internet a pak už málo, nebo reinfocement learning má architekturu kde se učíš podle tvýho úspěchu v prostředí přes akce (akorát vždycky se pak jede odznova, chtělo by to učit se jako lidi bez úplnýho odznova xD) místo predikování dalšího tokenu, ale učící algorithmy se používají stený nebo podobný. Argumentoval bych že klasický neuronky, konvolučky, jazykový modely, zvířecí a lidský mozek jsou všechno jinak specializovaný a v jiných aspektech zase obecnější inteligence, v jistých aspektech (benchmarcích) jsou neuronky oproti lidem na úrovni batolat, ale v jiných jsou zase AIs, ať jazykový modely nebo jiný architektury, lepší než nejlepší člověk na planetě, a tenhle gap se zmenšuje víc a víc - AGI by mělo být ve všem lepší než člověk, dle mě ne jen kognitivně, ale taky v robotice (Google na tom dost pracuje). Lidi jsou specializovaní na přežití v našem světě a sebereprodukci a všechny ty další evoluční tlaky co nám tvoří základní pyramidu potřeb včetně specializované seberealizace, což neuronky zatím až tak nemají. Jako výsledek lidi mají dost explicitně specializovaných brain regions, ale pomalu zjišťujeme jak moc ten mozek je reálně dost fluidní a adaptibilní a ty regions nejsou až tak hardcoded, mezitím co neuronky se učí jiný typy specializovaných regions. Máme celkem systematickou generalizaci, která se zároveň umí učit konkrétně, na to taky vychází víc a víc architektur, co to napodobňuje víc a víc. [Human-like systematic generalization through a meta-learning neural network | Nature](https://www.nature.com/articles/s41586-023-06668-3) Neuronky mají šílený potencál být v tomto nesouměřitelně efektivnější než lidi, když se vytvoří pod podobnými tlaky, klidně jen v billionech simulacích. Možná potřebují realtime interakci, explicitnější symbolický model světa s fuzzy foundations a logikou, plánování, možná ne, možná jsou schopny se to všechno emergentně naučit. Jde tam nacpat i genetický algorimy na meta úrovni na simulaci evoluce. Neuronky jsou turingovsky kompletní a univerzální approximators, což znamená, že jsou technicky schopný se v teorii naučit libovolný algorithmus, ale otázka je, za jak dlouho, a jak moc se zasekávají v jiných řešeních, což řídí (fluidní nebo striktní) architektura, trénovací data a množství compute. Kde je vědomí a subjektivní prožitek, na jaký úrovni abstrakce fyziky, nikdo reálně neví a všichni jenom hádají (pokud se tam vůbec nachází) a nevěřím ničemu, protože si myslím, že to nejde empiricky potvrdit 😄 Možná je to algoritmus co už neuronky mají, jako je nějaká úroveň sebereferenční zpracovávání informace. Možná je to jinačí matematický vzor co jde identifikovat ve fyzice, nějaký konrétní program, nebo úroveň integrované informace, nějaký sjednocující řídící hub, nějaký holistický počítání, něco na úrovni jiný matiky (sítě, geometrie, topologie, systems theory, dynamical systems,...), nějaká konkrétní podskupina biologie (buňky, proteiny etc.), nějaká elektrochemie, nějaký konkrétní vzor ve fyzice (klasická (electric field potentials jsou populární), statistická nebo jiná fyzika ala Penrosova kvantová gravitace),... třeba všechno a tohle jsou jen jiný pohledy a approximace dynamiky prožitku,... noone knows [[2303.07103] Could a Large Language Model be Conscious?](https://arxiv.org/abs/2303.07103) Nudging GPT until 100% success rate https://twitter.com/8teAPi/status/1734398472949059965?t=0RzDrhaC9o7qTiNrGPztuQ&s=19 Coordination technology https://twitter.com/gordonbrander/status/1734268827943227399?t=_qVzLLBA1taxW5MPx7HW_A&s=19 [List of anarchist movements by region - Wikipedia](https://en.m.wikipedia.org/wiki/List_of_anarchist_movements_by_region) Is perfect system ease of volunatary trade and difficulty of involuntary invasion? Differential technology [My techno-optimism](https://vitalik.eth.limo/general/2023/11/27/techno_optimism.html) [[2312.05840] Topological Data Analysis for Neural Network Analysis: A Comprehensive Survey](https://arxiv.org/abs/2312.05840) Topological Data Analysis for Neural Network Analysis: A Comprehensive Survey Joscha x Friston [Joscha Bach Λ Karl Friston: Ai, Death, Self, God, Consciousness - YouTube](https://youtu.be/CcQMYNi9a2w?si=3e6zk-daAD48jI3e) [[2311.09589] Mortal Computation: A Foundation for Biomimetic Intelligence](https://arxiv.org/abs/2311.09589) [OSF](https://osf.io/preprints/psyarxiv/y3tzs) Free energy principle formalism itself can be many things as well IMO. It's mathematical "truth", epistemological framework, scalefree information theoretic/statistical framework analyzing general (stochastic) dynamics mostly applied to classical or quantum physics, generator of process theories for any physical systems across scales (mainly brain and society), machine learning architecture (Active Inference), mathematical psychology model,... argumentoval bych že do toho teče všechno možný z formálních, přírodních, sociálních a engineering věd 😄 neurověda, fyzika, computer science, obecně aplikovaná a abstrakní matika, teorie a praxe kolem softwaru a hardwaru, etika, governance, psychologie, sociologie, apod,... stochastic gravity solution to unifying GR and QM https://twitter.com/burny_tech/status/1734660704647385467/photo/1 text to audio sota [Audiobox | Meta FAIR](https://audiobox.metademolab.com/) small llm sota? [Your request has been blocked. This could be due to several reasons.](https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/) predicting neural dynamics using RNNs https://twitter.com/Hidenori8Tanaka/status/1734607193922482297 [Chris Olah - Looking Inside Neural Networks with Mechanistic Interpretability - YouTube](https://www.youtube.com/watch?v=2Rdp9GvcYOE) induction heads break scaling laws [[2311.05112] A Survey of Large Language Models in Medicine: Progress, Application, and Challenge](https://arxiv.org/abs/2311.05112) llms in medicine Can we use the method of reverse engineering dynamical systems by using autoencoders to learn their dynamics and squeezing them and extracting interacting variables from them to reverse engineer and make interpretable the dynamics of CNNs, LLMs and other AI architectures? [Can AI disover new physics? - YouTube](https://www.youtube.com/watch?v=XRL56YCfKtA) [Concrete open problems in mechanistic interpretability | Neel Nanda | EAG London 23 - YouTube](https://www.youtube.com/watch?v=7t9umZ1tFso) [Portable, non-invasive, mind-reading AI turns thoughts into text | University of Technology Sydney](https://www.uts.edu.au/news/tech-design/portable-non-invasive-mind-reading-ai-turns-thoughts-text) awawa combine this with bidirectional GPT5 processing of the thoughts, access to the internet, and automatic graph notetaking generation [Exponetially self-replicating DNA nanobots are now a thing](https://interestingengineering.com/innovation/dna-nanobots-can-replicate-themselves) https://www.science.org/doi/10.1126/scirobotics.adf1274 history of gradient descent variants [Optimization for Deep Learning (Momentum, RMSprop, AdaGrad, Adam) - YouTube](https://www.youtube.com/watch?v=NE88eqLngkg) but classical SGD is better for generalizing as it favors less steap minima thanks to not having mechanism of converging more to local minimas list of GPTs [GitHub - ai-boost/Awesome-GPTs: Curated list of awesome GPTs 👍.](https://github.com/ai-boost/Awesome-GPTs) Is physics or computation fundamental? Or both, none, something third? So computation is emergent abstraction on top of those physical dynamics according to QRI's levels of analysis. Certain computationalists would argue that even all of that is special case of computation, or emergent from deeper computational principles in computational universe ontology. I suspect that's the first thing Joscha Bach would say, but I might be wrong. 😄 Or in universe being nonclassical turing machine, the evolution of fields can be seen as computation. Its defining information processing as being a lens on top of deeper physics versus physics being lens on deeper information processing. https://twitter.com/aniketvartak/status/1735137526770209023/photo/1 https://discord.com/channels/991774141025886218/992229023667204137/1184669600386928751 ML ideas https://twitter.com/jxmnop/status/1735072585778389426 Quora Answers 2015 - 2022 by David Pearce: towards a "triple S" civilisation of superintelligence, superlongevity and superhappiness [Quora Answers 2015 - 2024 by David Pearce: towards a "triple S" civilisation of superintelligence, superlongevity and superhappiness](https://www.hedweb.com/quora/index.html) https://alleninstitute.org/news/scientists-unveil-first-complete-cellular-map-of-adult-mouse-brain/ Scientists unveil first complete cellular map of adult mouse brain [Human brain-like supercomputer with 228 trillion links coming in 2024](https://interestingengineering.com/innovation/human-brain-supercomputer-coming-in-2024) Human brain-like supercomputer with 228 trillion links coming in 2024 Combining convolutions and Transformers https://twitter.com/simran_s_arora/status/1735023478594543960?t=3S0RIjJsb7mKVeGx8wxGng&s=19 Is information fundamental? [FunSearch: Making new discoveries in mathematical sciences using Large Language Models - Google DeepMind](https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/) "This work represents the first time a new discovery has been made for challenging open problems in science or mathematics using LLMs. FunSearch discovered new solutions for the cap set problem, a longstanding open problem in mathematics. In addition, to demonstrate the practical usefulness of FunSearch, we used it to discover more effective algorithms for the “bin-packing” problem, which has ubiquitous applications such as making data centers more efficient." Even if they solved Reimann hypothesis with this method they wouldnt be happy and it wouldnt be true creativity! It was using PaLM 2 as the searching mutation operator (not Gemini!!!) with a genetic algorithm. Four-color theorem was bruteforced, AlphaDev used reinforcement learning, AlphaTensor didnt have conventional language model, many other genetic inventions are bruteforced,... Let's see how will the Lean LLM help! The fact that LLM did the program mutation improving step succesfully without getting combinatorial explosion making it impossible in this problem space is big to me!!! Humans themselves werent able to figure this out for so long! I bet they've tried to do it in this problem space with many other methods and i bet it was impossible and the creativity of LLMs was needed here, even when they generated such a large population. I bet with other methods it would have to be infinitely large population thanks to the combinatorial explosion of the space of all possible mathematical programs in this domain! Deepmind paper in a few months: Our AI finds orders of magnitude more mathematical results than the total amount found by all people in the past [Faster sorting algorithms discovered using deep reinforcement learning | Nature](https://www.nature.com/articles/s41586-023-06004-9) Faster sorting algorithms discovered using deep reinforcement learning [Automated Antenna Design with Evolutionary Algorithms - NASA Technical Reports Server (NTRS)](https://ntrs.nasa.gov/citations/20060024675) Automated Antenna Design with Evolutionary Algorithms Open source robotics SotA [🐙 Octo: An Open-Source Generalist Robot Policy](https://octo-models.github.io/) https://www.reddit.com/r/MachineLearning/comments/18hnh8p/d_what_are_2023s_top_innovations_in_mlai_outside/ Universal brain EEG to language? [DeWave: Discrete EEG Waves Encoding for Brain Dynamics to Text Translation | Papers With Code](https://cs.paperswithcode.com/paper/dewave-discrete-eeg-waves-encoding-for-brain) opensource llm api https://twitter.com/togethercompute without structure we wouldnt be able to operate in this world to eat food, to loop in habits all structure is relative/illusory but real at the same time in its own useful context for me, and map approximates the territory that way you get infinite freedom but also pragmatic value out of it One Math Book For Every Math Subject [One Math Book For Every Math Subject - YouTube](https://www.youtube.com/watch?v=-mfaMbraEkU) Phigsm small model SotA https://twitter.com/IntuitMachine/status/1735677522224755048?t=SBtjGq9wJoEahlN70rDkXg&s=19 NeuroAI Add timeline section everywhere mainly AI