Longevity escape velocity AI [Alpha Everywhere: AlphaGeometry, AlphaCodium and the Future of LLMs - YouTube](https://youtu.be/dOplrIJEYBo?si=Jbk0GYYGK16T3kI0) Quantum thermodynamic computing [x.com](https://twitter.com/GillVerd/status/1748126660300579045?t=W7jqU8jVhmC_0KaM5DN1Zg&s=19) Pitnthat computing comparision into computing note and put parts of it to comparing AI and bio note and quantum ml etc. [AlphaFold found thousands of possible psychedelics. Will its predictions help drug discovery?](https://www.nature.com/articles/d41586-024-00130-8) Meta plans 350k H100s GPUs (plus 600k equivalents of compute if you include other GPUs) to train open source AGI this year, compared to 150k last year. That's annual more than doubling. [x.com](https://twitter.com/burny_tech/status/1748266879838199975?t=_rMWlQohyP4JLmciCSZLZg&s=19) [Google CEO Tells Employees To Expect More Job Cuts This Year - Slashdot](https://tech.slashdot.org/story/24/01/18/1354208/google-ceo-tells-employees-to-expect-more-job-cuts-this-year) [Mark Zuckerberg | Some updates on our AI efforts. Our long term vision is to build general intelligence, open source it responsibly, and make it widely... | Instagram](https://www.instagram.com/p/C2QARHJR1sZ/) Open source AGI acceleration landscape [x.com](https://twitter.com/bindureddy/status/1748120052413653418?t=xY46CItMm-5tF8a6yObwpw&s=19) Add year added to every link and note GPT4 level conversational LLMs [x.com](https://twitter.com/_akhaliq/status/1748168711368802328?t=QJRK9phXBgqVLAA9NfZA6A&s=19) [Our World in Data](https://ourworldindata.org/) [x.com](https://twitter.com/BoyanSlat/status/1746211702968906049) [How fast will AGI be adopted? How can we ensure equitable outcomes? Nations, Businesses, People... - YouTube](https://youtu.be/YZB-JZ_-cDs?si=I89mm3-mwy5QVBAi) developing nations might integrate AI faster because of less byrocracy barriers Let's not accelerate too hard to get peaceful restricting to psptAGI postlabor economics instead of wars when things get too painful https://en.wikipedia.org/wiki/List_of_revolutions_and_rebellion [Arab Spring - Wikipedia](https://en.wikipedia.org/wiki/Arab_Spring) [List of cultural, intellectual, philosophical and technological revolutions - Wikipedia](https://en.wikipedia.org/wiki/List_of_cultural,_intellectual,_philosophical_and_technological_revolutions) Preparing for apocalypse by buying batteries and GPUs with local LLMs with access to downloaded Wikipedia in a wooden shack in a forest (or nuclear bunker) in the middle of nowhere Concetration of power has to be broken by autonomous decentralized organization and capital Postlabour future where capital works differently than from labour instead of cyberpunk dystopia? We need to maximize individual economic agency, not reduce it, to prevent oppresion [Alpha Everywhere: AlphaGeometry, AlphaCodium and the Future of LLMs - YouTube](https://www.youtube.com/watch?v=dOplrIJEYBo) idea generation bruteforce search hybrid AGI architectures AlphaGeometry has LLM pulling magic tricks from the hat how to make the problem easier and symbolic engine bruteforces the rest and solves it (without LLM this wouldnt be possible) AlphaCodium: postpone decisions, try to avoid direct questions, leave room for exploration, conversation between mutually correcting agents on unit tests Prepper Kits for an AI Apocalypse [x.com](https://twitter.com/IntuitMachine/status/1631680192358801410) [[2305.15771] On the Planning Abilities of Large Language Models : A Critical Investigation](https://arxiv.org/abs/2305.15771) "we demonstrate that LLM-generated plans can improve the search process for underlying sound planners and additionally show that external verifiers can help provide feedback on the generated plans and back-prompt the LLM for better plan generation." Cofounder of Google DeepMind Mustafa Suleyman says that AI will be able to create and run its own business within the next five years as an entrepreneur and inventor [AI Could Run Businesses in Five Years, Google DeepMind Cofounder Says - Business Insider](https://www.businessinsider.com/google-deepmind-ai-could-run-businesses-in-five-years-2024-1?utm_source=reddit.com) My definition of AGI tends to be pretty weak Taking into account cognitive skills of an avarage person which are IMO not as hard to replace And for a subset of tasks, cognitive economic tasks Probably the biggest bottleneck now is long term planning still, but even that i think can get fixed by lots of hacks, like repromptimg with context, chain of thought, selfcorrection, diving into tons of parts, external memory and so on https://fxtwitter.com/divgarg9/status/1747683043446579416 Complex programming is still an issue, but i think you could create automatic generation of very specialized systems for these tasks ala agent swarms, and AlphaCodium is a gigant step forwards better than most coding competitors, so giving it realtime access to internet and documentations in a multiagent setting,... [Alpha Everywhere: AlphaGeometry, AlphaCodium and the Future of LLMs - YouTube](https://youtu.be/dOplrIJEYBo?si=ZCcnYyfjrzT0Zdpn) I feel like people often think LLMs exist in isolation, when you can do all this hackery on top of it, connect it with symbolic engines, which can itself by automated by the complex system creating complex subsystems and itself And i think this is realistic too and not far off https://www.lesswrong.com/posts/Btom6dX5swTuteKce/agi-will-be-made-of-heterogeneous-components-transformer-and I may be wrong definitely, first i wanna see such a system combining these factors being build and seeing its results, and I'll see, but so far the current AI systems seem to be climbing benchmarks exponentially with all the architectural, prompting, extra modules etc. hacks Maybe its still not competent enough in this way, and maybe for long term more complex goals it easily diverges and disintegrates from explosion of the context window, or maybe that's hackable by constant new reprompting with just a relevant role, context, goals, etc. and asking for new step, a process that might be automatable too? One big bottleneck could be the problem of really evaliating the correctness of steps and solutions in tasks where correctness is more vague And for some agency, more hardcoded goals, intentions, curiousity is also extremely critical for AGI which is also still pretty weak in engineering [The Free Energy Principle approach to Agency - YouTube](https://youtu.be/zMDSMqtjays?si=Gb2AGCrXLoaH3pzl) For some embodiment or consciousness is important, however you define it QRI-style Joscha Bach-style Friston-style Penrose-style etc. Or I like this AI researcher, author of Shard theory, he's even more radical in having a weak definition of AGI and considers GPT4 alone as an AGI, since it is in many ways a general system, but i'm not that radically weakly defining it [#5: Quintin Pope - AI alignment, machine learning, failure modes, and reasons for optimism - YouTube](https://youtu.be/f9Msoqvlla4?si=quocitlMaJQQt6yx) Now i kind of wanna use the methods from AlphaCodium breakthrough for automating neural network interpretability and alignment :FractalThink: More time needed, i have to explore less and exploit more Linus Towalds (creator of Linux) is very based about AI! "LLMs can help identify bugs real time" >but they're just autocomplete on steroids "I think they're much more than that and humans are also autocomplete on steroids to some degree" >but are you scared of bugs in code produced by LLMs "I think humans are already doing those just fine on their own that i see daily, I'm not worried" "Sometimes you have to be a bit too optimistic to make a difference" [Torvalds Speaks: Impact of Artificial Intelligence on Programming - YouTube](https://www.youtube.com/watch?v=VHHT6W-N0ak) multiple consciousnesses in one brain might be multiple irreducible markov blankets with their own information geometry interacting in neuronal architecture [Joscha Bach Λ Karl Friston: Ai, Death, Self, God, Consciousness - YouTube](https://youtu.be/CcQMYNi9a2w?si=RIgEuRZjjmmgdu9I&t=7700) yan lecunn: future of AI is nongenerative, insead oit s joint embedding, where encoders are trained to encode the image so that you can recover representation of the full from the representation of the corrupted Flow engineering [x.com](https://twitter.com/swyx/status/1748089313433420115?t=mCwTzerKvomWRJvkY_oBIg) François Chollet Deep learning google. Creator of Keras. Author of 'Deep Learning with Python'. Opinions are my own. [Regulation of artificial intelligence - Wikipedia](https://en.wikipedia.org/wiki/Regulation_of_artificial_intelligence) A mind isn't a single agent, but a superposition of many agents, most of them old snapshots of a past state, ready to be brought back to handle new relevant situations. Your personality and cognitive abilities fluctuate depending on which one you are currently enacting. [x.com](https://twitter.com/fchollet/status/1748234154586894736) [GPT 4 Level Open Source in 2024..(Llama 3 Leaks and Mistral 2.0) - YouTube](https://www.youtube.com/watch?v=uCkYoVb2kKw) [CURIOSITY - Featuring Richard Feynman - YouTube](https://www.youtube.com/watch?v=UjEngEpiJKo) [x.com](https://twitter.com/GoogleAI/status/1748058161922347352) Introducing ASPIRE, a framework that enhances the selective prediction capabilities of large language models, enabling them to output an answer paired with a confidence score. Learn how it outperforms state-of-the-art methods on a variety of QA datasets. → [Introducing ASPIRE for selective prediction in LLMs](https://goo.gle/3RWpChi) Yoneda lemma abstract definition [x.com](https://twitter.com/mattecapu/status/1748322794163589610) Elon Musk: we have a silicon shortage today, we will have a voltage transformer shortage in a year and electricity shortages in 2 years [x.com](https://twitter.com/tsarnick/status/1748119717712408909) i wonder if someone formalized segmentation of information geometry/topology (instead of electromagnetic field) in this way i wonder in general about throwing (algebraic) topology, homology, cohomology, (algebraic) (differential) geometry math etc. at structure of information [Information geometry - Wikipedia](https://en.wikipedia.org/wiki/Information_geometry) [Topological data analysis - Wikipedia](https://en.wikipedia.org/wiki/Topological_data_analysis) this looks wild https://www.researchgate.net/publication/337801824_Information_Topology "Von Neumann quantum information cohomology" "Both quantum and classical 1-cocycle correspond to commutators-brakets for conditioning, provides topological Fluctuation Theorems, and can be used to quantify causality." interesting The first AI device that can detect all major skin cancers just received FDA approval https://www.reuters.com/business/healthcare-pharmaceuticals/us-fda-clears-dermasensors-ai-powered-skin-cancer-detecting-device-2024-01-17/ IBM warns that quantum computers could make existing encryption systems obsolete by 2030. [Bloomberg - Are you a robot?](https://www.bloomberg.com/news/articles/2024-01-17/quantum-computing-to-spark-cybersecurity-armageddon-ibm-says) [A Gene-Edited Pig Liver Was Attached to a Person—and Worked for 3 Days | WIRED](https://www.wired.com/story/gene-edited-liver-attached-to-person/) [OpenAI announces first partnership with a university](https://www.cnbc.com/2024/01/18/openai-announces-first-partnership-with-a-university.html) New Google paper enabling LLMs to output an answer paired with a confidence score [x.com](https://twitter.com/GoogleAI/status/1748058161922347352) Japanese researchers have developed a framework that uses machine learning to speed up the discovery of materials for green energy technology https://techxplore.com/news/2024-01-machine-method-discovery-green-energy.html [[2401.10020] Self-Rewarding Language Models](https://arxiv.org/abs/2401.10020) To speed up LLMs' inference and enhance LLM's perceive of key information, compress the prompt and KV-Cache, which achieves up to 20x compression with minimal performance loss. [LLMLingua Series | Effectively Deliver Information to LLMs via Prompt Compression](https://llmlingua.com/) The Perceptron Controversy: Connectionism died in the 60s from technical limits to scaling, then resurrected in the 80s after backprop allowed scaling. The Minsky–Papert anti-scaling hypothesis explained, psychoanalyzed, and buried. https://yuxi-liu-wired.github.io/blog/posts/perceptron-controversy/ https://www.lesswrong.com/posts/3aicJ8w4N9YDKBJbi/four-visions-of-transformative-ai-success#Vision_4__Don_t_build_TAI (Vision 1) “Helper AIs”, (Vision 2) “Autonomous AIs”, (Vision 3) “Supercharged biological human brains”, (Vision 4) “Don’t build TAI”. [Self-assembling DNA computer can sort simple images into categories | New Scientist](https://www.newscientist.com/article/2412689-self-assembling-dna-computer-can-sort-simple-images-into-categories/) [First 'thermodynamic computer' uses random noise to calculate | New Scientist](https://www.newscientist.com/article/2412212-first-thermodynamic-computer-uses-random-noise-to-calculate/) [[2312.04836] Thermodynamic Computing System for AI Applications](https://arxiv.org/abs/2312.04836) [[2312.01523] SymNoise: Advancing Language Model Fine-tuning with Symmetric Noise](https://arxiv.org/abs/2312.01523) [[2401.06951] E^2-LLM: Efficient and Extreme Length Extension of Large Language Models](https://arxiv.org/abs/2401.06951) [[2306.02797] Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language](https://arxiv.org/abs/2306.02797) [[2306.02797] Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language](https://arxiv.org/abs/2306.02797) Learning grounded in inductive biases NASA’s robotic, self-assembling structures could be the next phase of space construction “Ultralight, strong, and self-reprogrammable mechanical metamaterials,” which is a highly precise way to describe a building that builds itself. The inevitable acronym for it is “Automated Reconfigurable Mission Adaptive Digital Assembly Systems,” or ARMADAS. [NASA's robotic, self-assembling structures could be the next phase of space construction | TechCrunch](https://techcrunch.com/2024/01/17/nasas-robotic-self-assembling-structures-could-be-the-next-phase-of-space-construction/) https://www.science.org/doi/10.1126/scirobotics.adi2746 [Google CEO tells employees to expect more job cuts in 2024 - The Verge](https://www.theverge.com/2024/1/17/24042417/google-layoffs-2024-internal-employee-memo-sundar-pichai) https://towardsdatascience.com/graph-geometric-ml-in-2024-where-we-are-and-whats-next-part-i-theory-architectures-3af5d38376e1 https://towardsdatascience.com/graph-geometric-ml-in-2024-where-we-are-and-whats-next-part-ii-applications-1ed786f7bf63 entropy definitions landscape [x.com](https://twitter.com/architectonyx/status/1708516507263869426) Is curiousity all we need for AGI? [ORIGINAL FATHER OF AI ON DANGERS! (Prof. Jürgen Schmidhuber) - YouTube](https://youtu.be/q27XMPm5wg8?si=h8OktnPUUesO010c) 56:00 artificial curiousity: RL agent doing actions and a world model predicting the conseauences of those actions. RL agent tried to generate actions such that the world model's error is maximized, leading to exploratory curiousity Langchain multion [x.com](https://twitter.com/LangChainAI/status/1748404710686785629?t=-mWA6xk1ORTX0FslRkfZrQ&s=19) [Bloomberg - Are you a robot?](https://www.bloomberg.com/news/articles/2024-01-19/google-deepmind-ai-scientists-in-talks-to-leave-for-french-stealth-startup) https://www.sciencedirect.com/science/article/pii/S0166223623002278 https://scienceblog.com/541768/mit-study-reveals-universal-brain-patterns/ [Artificial Curiosity Since 1990](https://people.idsia.ch/~juergen/artificial-curiosity-since-1990.html) [[1511.09249] On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models](https://arxiv.org/abs/1511.09249) [Juergen Schmidhuber's online publications](https://people.idsia.ch/~juergen/onlinepub.html) [GENERALIZED KOLMOGOROV COMPLEXITY](https://people.idsia.ch/~juergen/kolmogorov.html) [Theory of Everything - Algorithmic Theory of Everything - Computer Universe - Computable Universe - All Computable Universes - Anthropic Principle - all mathematically possible universes - Cosmos as a computer - measure on universes - Goedel](https://people.idsia.ch/~juergen/computeruniverse.html) [ORIGINAL FATHER OF AI ON DANGERS! (Prof. Jürgen Schmidhuber) - YouTube](https://youtu.be/q27XMPm5wg8?si=Hn811D0PVgT0Ccif) 1:08:00 higher order cost function prefering flat error local minima learning more simple weights Essential supplements good for health https://www.sciencedirect.com/science/article/abs/pii/S0002916523663427?via%3Dihub Principle of least action [Action principles - Wikipedia](https://en.wikipedia.org/wiki/Stationary-action_principle) [ChatGPT](https://chat.openai.com/share/d77e006e-9a3d-4f77-99a7-25e02f2cf152) [ChatGPT](https://chat.openai.com/share/64ef80f1-2a9a-46d8-860a-cb61f11512d7) [Why does the principle of least action hold in our universe? - Quora](https://www.quora.com/Why-does-the-principle-of-least-action-hold-in-our-universe) "In the quantum mechanics the principle of least action has a wonderful explanation, put forth by Feynman. The essence of his argument is based on his path integral formulation of quantum mechanics. It comes down to this: in order to calculate the probability of something happening you have to consider every possible path between initial and final states, calculate a complex amplitude of the process happening on that path, add those amplitudes together, and calculate the square modulus and this is your probability. It is important that the phase of the amplitude for each possible path depends on the action along this path. If you now pick, of all those possible paths, two close to each other and look at their contributions to the total amplitude, you'll notice, that usually the amplitudes will have wildly different phases and often they'll interfere destructively. If you however pick a path corresponding to a local minimum (or maximum, or even a saddle point) of the action, then all close paths have very close phases of the amplitde and they interfere constructively. Thus most of the contribution to the total amplitude for a process comes from paths corresponding to the extrema of action - effectively you can block all other possible paths and the total amplitude will not significantly change. If you now go to classical limit you'll get the classical path of a particle corresponds to an extremum (usually minimum) of classical action." >Using a Taylor expansion and neglecting higher-order terms Everytime I suddenly see this in physics my analytical intuition tends to crash because things start to get too approximated uauauauuauu Assuming cow is spherical and speed of light is 1, we find the utopian multiverse branch where we can calculate physics analytically and understand how AGI works internally perturbation theory [Perturbation theory - Wikipedia](https://en.wikipedia.org/wiki/Perturbation_theory) ChatGPT is actually really good teacher at certain physics, mathematics or science in general explaning concepts with clarity if prompted correctly when I fact check it with Wikipedia and other sources. You can make it even better by giving it sources to read from. When it has to do mechanistic calculations, it will just use Python for more accuracy. It's absolutely amazing for learning IMO. Great for technical explanations, detailed step by step explanations derivations of the math, going through step by step examples, clarifications and deep explanations of concepts it mentions that you don't know by asking, intuitive explanations, analogies, simplifications, how it relates to other things, correcting mistakes if some are found, and so on! [Equations of motion - Wikipedia](https://en.wikipedia.org/wiki/Equations_of_motion) [Perturbation Theory Video Lectures - YouTube](https://www.youtube.com/playlist?list=PLzcd6SoIscwjHuWRE38UXWG92uq0Sy4UF) The combined degree of simplicity and direct or implied predictivity is the degree of fundamentality! [MMathPhys: KT & CPP - MT23 & TT24](https://www-thphys.physics.ox.ac.uk/people/AlexanderSchekochihin/KT/) wave equation [Wave equation - Wikipedia](https://en.wikipedia.org/wiki/Wave_equation) laplacian [Is ACTION The Most Fundamental Property in Physics? - YouTube](https://www.youtube.com/watch?v=Q_CQDSlmboA) in principle of least action in general relativity, action is replace with proper time that is minimized, action is now how much time is percieved by an observer in their own frame of reference, all objects moving through relativistic spacetime move though paths that minimize the time (percieved by an observer in their own frame of reference) measured on that path What if principle of least action didn't hold? Let's construct universes where the negation of principle of least action holds! Feynman path integral [How Our Reality May Be a Sum of All Possible Realities | Quanta Magazine](https://www.quantamagazine.org/how-our-reality-may-be-a-sum-of-all-possible-realities-20230206/) bayesian machine learning I want to analyze the internal workings of artificial neural networks to explain structures using physics to identify various functions the networks learn. How can I do it? Mathematics of computational neuroscience [ChatGPT](https://chat.openai.com/share/8201afc8-a904-4c6b-8190-68bed71374cb) Carl Sagan GPT4 is a just absolutely amazing gift for learning everything. I wish I had this throughout my whole education. I can't wait how even better for education future AI models will be. Chemistry AI [Feynman's Infinite Quantum Paths - YouTube](https://www.youtube.com/watch?v=vSFRN-ymfgE) phenomenology in physics [Renormalization - Wikipedia](https://en.wikipedia.org/wiki/Renormalization) [Path integral formulation - Wikipedia](https://en.wikipedia.org/wiki/Path_integral_formulation) feynman integral formulation of quantum mechanics is an integral over the space of all possible paths that a particle can take between the initial and final points, each integral corresponds to the integration over one discretized time step's position with the step count approaching infinity [ChatGPT](https://chat.openai.com/share/285865dd-d003-4ea3-9afc-43177ea22017) path integral formulation in quantum field theory sums up over all possible configurations of all fields and all possible events (interactions between the fields) where most destructively interfere and the resulting constructive quantum amplitude is mostly made of path with stationary action (i.e., where the action doesn't change for small variations of the path) as they have close amplitudes [Is ACTION The Most Fundamental Property in Physics? - YouTube](https://www.youtube.com/watch?v=Q_CQDSlmboA) [Action principles - Wikipedia](https://en.wikipedia.org/wiki/Stationary-action_principle) [History of variational principles in physics - Wikipedia](https://en.wikipedia.org/wiki/History_of_variational_principles_in_physics) The path taken by the system between times t1 and t2 and configurations q1 and q2 is the one for which the action is stationary (no change) to first order Principle of least action makes objects go through the shortest ath through configuration space, space of all possible trajectories given our object's constrains which is: Classical mechanics: through space with its current potential and kinetic energy Relativity: Spacetime Quantum mechanics: Feynman integral formulation - Space of all possible positions and momenta, or more generally statespace representing all possible quantum states Quantum field theory: Feynman integral formulation - All possible configurations of all fields and all possible events (interactions between the fields) sabine hossenfelder principle of least action is closest to ToE [The Closest We Have to a Theory of Everything - YouTube](https://www.youtube.com/watch?v=A0da8TEeaeE) standard model langrangian https://www.lesswrong.com/posts/2roZtSr5TGmLjXMnT/ https://www.lesswrong.com/posts/TK92QZ8L6cXvvhXbF/gaia-network-an-illustrated-primer