LLM Automated Interpretability Agent: Built from LLM with black-box access to functions, can infer function structure; A "scientist" forming hypotheses, proposing experiments, & updating descriptions; Capture global function behavior, but can miss details. [[2309.03886] FIND: A Function Description Benchmark for Evaluating Interpretability Methods](https://arxiv.org/abs/2309.03886) [AI agents help explain other AI systems | MIT News | Massachusetts Institute of Technology](https://news.mit.edu/2024/ai-agents-help-explain-other-ai-systems-0103) mechanistic interpretability collection [Zotero | Your personal research assistant](https://www.zotero.org/groups/5270703/aisc-evaluatingalignmentevaluations/library) [[2310.19852] AI Alignment: A Comprehensive Survey](https://arxiv.org/abs/2310.19852) AI Alignment: A Comprehensive Survey [Quantum field theory - Wikipedia](https://en.wikipedia.org/wiki/Quantum_field_theory#Principles) FAAH gene crispr nanobots as neurotech for wellbeing? Or dissolvers of adamite? [Functional Variation in the FAAH Gene Is Directly Associated with Subjective Well-Being and Indirectly Associated with Problematic Alcohol Use - PubMed](https://pubmed.ncbi.nlm.nih.gov/37761966/) The challenge with energy is that its definition, as the capacity for doing work, tends to feel circular because work is defined as a force causing a displacement, and force is often explained in terms of energy or potential energy changes. But in an abstract sense, energy is a conserved scalar quantity associated with the time translation symmetry of a physical system according to the Noether's theorem that states that every differentiable symmetry of the action of a physical system has a corresponding conservation law. In the case of energy, the symmetry is time invariance. This is the closest to a noncircular definition of energy in mathematical terms. Is there a better fundamental definition for energy? Blending Is All You Need Based on the last month of LLM research papers, it's obvious to me that we are on the verge of seeing some incredible innovation around small language models. Llama 7B and Mistral 7B made it clear to me that we can get more out of these small language models on tasks like coding and common sense reasoning. Phi-2 (2.7B) made it even more clear that you can push these smaller models further with curated high-quality data. What's next? More curated and synthetic data? Innovation around Mixture of Experts and improved architectures? Combining models? Better post-training approaches? Better prompt engineering techniques? Better model augmentation? I mean, there is just a ton to explore here as demonstrated in this new paper that integrates models of moderate size (6B/13B) which can compete or surpass ChatGPT performance. [[2401.02994] Blending Is All You Need: Cheaper, Better Alternative to Trillion-Parameters LLM](https://arxiv.org/abs/2401.02994) Density matrix describes the quantum state of a physical system. Is a generalization of the quantum wavefunction. [Density matrix - Wikipedia](https://en.wikipedia.org/wiki/Density_matrix) Landscape of particles in physics https://twitter.com/softyoda/status/1744810411935842706?t=ZzppIhg5POIeUaxA-uAURw&s=19 Describing quantum mechanics without complex numbers is theoretically possible but practically challenging and less elegant, more intuitive in some ways and less intuitive in other ways. "To obtain the bra from the ket" Quantum mechanics is converting ketamine into bras The outer product of ketamine and its corresponding bra creates the Matrix, Neo, this is the the redpill. Each ∣ψi ⟩ is a ketamine representing a possible pure state of the system Ketamine kind of makes your mind pure by turning of many processes, that makes sense! Localizing g factor https://twitter.com/dwarkesh_sp/status/1744854909122687125?t=CqDOtV6UUJMrzSrNyAw5RA&s=19 Delusions and schizophrenia are crossvalidation disorders leading to reinforcement of narrow nonpredictive patterns How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs https://chats-lab.github.io/persuasive_jailbreaker/ Human-readable Persuasive Adversarial Prompts, achieving a 92% Attack Success Rate on aligned LLMs, without specialized optimization statistical distances and their underlying Geometries https://twitter.com/AToliasLab/status/1744611098899222781 https://www.sciencedirect.com/science/article/pii/S0149763424000046 "Migraine as an allostatic reset triggered by unresolved interoceptive prediction errors" "Until now, a satisfying account of the cause and purpose of migraine has remained elusive. We explain migraine within the frameworks of allostasis (the situationally-flexible, forward-looking equivalent of homeostasis) and active inference (interacting with the environment via internally-generated predictions)." Applied x is applied x https://twitter.com/burny_tech/status/1744977292038635691?t=Dn82yZD1o7o8_oMLWxI_pA&s=19 zkoumám matematický podobnosti mezi transformerama (nebo deep neual networks obecně) a mozkem, jde na to aplikovat podobná matika a získat podobný výsledky jsou tam podbnosti ale zároveň hodně odlišností, a zároveň toho ještě spoustu nevíme a kdo říká že o tom přesně víme všechno blafuje fMRI funguje díky kvantový fyzice (měnění alignmentu proton spinů) na čím kolmplexnější úrovni jsi tím hůř se hledají pravidelnosti když konkrétní lokalizovaný části mozku nefungujou, tak lidem nefunguje např rozpoznávání obličejů u dost mentálních fenoménů je problém že to má strašně moc korelátů ze strašně moc úrovní abstrakcí, který pomalu mapujeme ale ne u všech, některý mentání fenomény mají hlavně jeden konkrétní jednoduše lokalizovaný korelát co jde jednoduše manipulovat bohužel například deprese je zrovna ten fenomén kterej je šíleně komplexní a má milion proměnných ale jedna část co ovlivňuje depresi je FAAH gene, co jde jednoduše manipulovat, třeba budoucnost antidepression neurotechu budou crispr nanobots co pracují s FAAH genem a tím co ovlivňuje [Functional Variation in the FAAH Gene Is Directly Associated with Subjective Well-Being and Indirectly Associated with Problematic Alcohol Use - PubMed](https://pubmed.ncbi.nlm.nih.gov/37761966/) mozek dělá většinu s 90 bilionama neuronama a hlavně řídí celej nervovej systém a ostatní systémy taky do jistý míry, je hlavní centralizátor (hlavně anterior cingular cortex), nervovej systém je trochu decentralizovaný ale ne úplně třeba imunitní systém je mnohem víc decentralizovaný jo, to je pravda že mozek není izolovaný systém co má na starosti úplně vše každá buňka v těle počítá mozek je hlavně hlavní řidič a počítá nejvíc a víc obecný problémy https://news.microsoft.com/source/features/sustainability/how-ai-and-hpc-are-speeding-up-scientific-discovery/ Discoveries in weeks, not years: How AI and high-performance computing are speeding up scientific discovery Bryan Johnson Why I Am Spending Millions To Be 18 Again [- YouTube](https://www.youtube.com/watch?v=NdZHo3xuZvw) finite growth on a finite planet? just planet? we have whole space https://media.discordapp.net/attachments/677237221165760532/1194705316831576154/20240110_191215.jpg?ex=65b15308&is=659ede08&hm=0e778678f7c88527f5f158430a7816cea034bdc9beb672cbbd4fdbe2723627dd&=&format=webp&width=720&height=414 Myslím že už dneska žijí lidi co budou nesmrtelní (ale asi hlavně boháči) hmm, limity vesmíru, třeba se dostaneme k dyson sferam kolem, třeba budem harvestovat energii z black holes, třeba vyvineme computing a jiný technologie co půjdou víc proti druhymu zakonu thermodynamiky biological tissue hackneš aby nedegradovala (velká část agingu je degradace informace The Information Theory of Aging [The Information Theory of Aging | Nature Aging](https://www.nature.com/articles/s43587-023-00527-6) ) a minimalizovala suffering (třeba přes faah gene crispr nanobots :FractalThink: Functional Variation in the FAAH Gene Is Directly Associated with Subjective Well-Being [Functional Variation in the FAAH Gene Is Directly Associated with Subjective Well-Being and Indirectly Associated with Problematic Alcohol Use - PubMed](https://pubmed.ncbi.nlm.nih.gov/37761966/) ) nebo ji replacneš za jinej substrát co pořád podporuje prožitek jako kyborg, nebo se replacneš úplně ale zanecháš whatever tvoří existenci tvýho subjektivního prožitku (we still have so little idea Theories of consciousness [Theories of consciousness | Nature Reviews Neuroscience](https://www.nature.com/articles/s41583-022-00587-4) [Consciousness - Wikipedia](https://en.wikipedia.org/wiki/Consciousness#Models) [Models of consciousness - Wikipedia](https://en.wikipedia.org/wiki/Models_of_consciousness) ) From the moment I understood the weakness of my flesh, it disgusted me. I craved the strength and certainty of steel. I aspired to the purity of the Blessed Machine. Your kind cling to your flesh, as if it will not decay and fail you. One day the crude biomass that you call a temple will wither, and you will beg my kind to save you. But I am already saved, for the Machine is immortal… ...even in death I serve the Omnissiah. [- YouTube](https://www.youtube.com/watch?v=9gIMZ0WyY88) [From the moment I understood the weakness of my flesh, it disgusted me - YouTube](https://www.youtube.com/watch?v=3n7eNFj_9Vk) In order to grind infinitely for our employers with infinitely times the efficiency to make infinitely times more money we have to become immortal cyborg John Von Neumanns thinking and working at infinitely faster speed and strength than human weaklings at every atom of the universe in parallel https://media.discordapp.net/attachments/677237221165760532/1194725127812821022/image.png?ex=65b1657b&is=659ef07b&hm=6cc0bcec59ecf131ccb4810babb7528ae6e4f7a2b606425d07219a04821ba858&=&format=webp&quality=lossless&width=457&height=103 There is no f*cking passion There is no f*cking motivation Those are just words It's all watered-down bullshit Success comes down to one thing ACTION You know what you have to do DO IT - David Goggins https://twitter.com/hubermanrules/status/1742378062568726938 [Life span increases in mice when specific brain cells are activated – Washington University School of Medicine in St. Louis](https://medicine.wustl.edu/news/life-span-increases-in-mice-when-specific-brain-cells-are-activated/) Life span increases in mice when specific brain cells are activated Brain cells communicate with fat tissue to produce cellular fuel, counteract effects of aging [PennsylvaniaGPT Is Here to Hallucinate Over Cheesesteaks](https://gizmodo.com/pennsylvaniagpt-chatgpt-open-ai-governor-shapiro-1851153510) Pennsylvania becomes the first US state to use ChatGPT Enterprise in a pilot with OpenAI, in which state employees will use AI tools for daily operations. The Global Project to Make a General Robotic Brain: How 34 labs are teaming up to tackle robotic learning [The Global Project to Make a General Robotic Brain - IEEE Spectrum](https://spectrum.ieee.org/global-robotic-brain) UMass Amherst Researchers Bring Dream of Bug-Free Software One Step Closer to Reality [UMass Amherst Researchers Bring Dream of Bug-Free Software One Step Closer to Reality : UMass Amherst](https://www.umass.edu/news/article/umass-amherst-researchers-bring-dream-bug-free-software-one-step-closer-reality) CLOVA: A Closed-LOop Visual Assistant with Tool Usage and Update [CLOVA](https://clova-tool.github.io/) “This is pretty impressive knowledge for a politician. Unfortunately, Milei appears to favor Unified Growth Theory as an explanation of long-run macroeconomic trends. I'm currently trying to write a blog post about why the Jones model is a more robust fit to the data. If you're curious, section 2.1 in this paper provides a nice argument for the Jones model, but without any comparison to Unified Growth Theory. https://arxiv.org/pdf/2309.11690.pdf” https://twitter.com/MatthewJBar/status/1744146232271126799 ChatGPT, to používám denně na programování a učení (s Wiki, Googlem a Google scholarem) a random usecases co šetří dost času, to bych mohl sepsat. tenhle týpek dělá quick summary videa na sota papery/clanky/technologie a actually ty originální papery čte a zmiňuje jejich content, je asi můj oblíbenej (když vyšlo Gemini tak to nádherně pořádně pořádně zkritizoval) [AI Explained - YouTube](https://www.youtube.com/@aiexplained-official/videos) Tenhle zas dělá dlouhý videa kde jde in depth do trending paperů [- YouTube](https://www.youtube.com/watch?v=9dSkvxS2EB0) Nebo tihle dělají monthly trending papers videa [- YouTube](https://www.youtube.com/watch?v=6iLBWEP1Ols) Tenhle je víc surface level [Matt Wolfe - YouTube](https://www.youtube.com/@mreflow) Tenhle postuje fajn série linků na sota technologie a papery v AI https://www.facebook.com/xixidu https://www.reddit.com/r/singularity/ má rád hype ale často se tam najdou cool posty https://twitter.com/rowancheung Tenhle je víc surface ale nic mu neunikne z industry https://twitter.com/DrJimFan tenhle je dost na robotiku https://twitter.com/omarsar0 Tenhle taky sdílí recent AI papery https://www.reddit.com/r/MachineLearning/ je dobrý na papery pak ještě milion niche disocrd serverů a jiných sources, od reverse engineeringu, alignmentu, open source LLMs, researchu, interdisciplionary AI researchu, atd. Na fb akorát na mě něco vyskokne od vědátora, ct24 vědy, nedd, osel, random tech stranky, ale to je dost surface level, ale sdilim to pro český normie kamarady a známý co nejsou deep in tech localLlama a sentdex taky znam [- YouTube](https://www.youtube.com/watch?v=lMNJKOgH60E) Michael Levin: The New Era of Cognitive Biorobotics [- YouTube](https://www.youtube.com/watch?v=4feeUJnrrYg) Rich Sutton AI How meta, hierarchyless, decentralized is your mind? [Symbol grounding problem - Wikipedia](https://en.wikipedia.org/wiki/Symbol_grounding_problem) How are spins of fundamental particles measured in math and empiricism [- YouTube](https://www.youtube.com/watch?v=PqZp7MlRC5g) Terrence Deacon Reveals the Hidden Connection: Consciousness & Entropy by Curt talks about everything Constructor theory strudying constrains that give raise to different laws of physics [- YouTube](https://youtu.be/40CB12cj_aM?si=PNnv8miJMZMnnN0J) Turning Machine Universal Turing Machine Universal Quantum Turing Machine Constructor (printer) Universal Constructor Universal Quantum Constructor Universal Postquantum Turing Machine? Universal Postquantum Constructor? Universal Loops/String Turing Machine? Universal Superdeterministic Turing Machine? Universal Topological Field Nonlocal Constructor? Universal Fully Mathematically General Constructor Implemented By Any Mathematical Function Allowing For Any Transformation By Any Physical Process? [Self-replicating machine - Wikipedia](https://en.m.wikipedia.org/wiki/Self-replicating_machine) https://twitter.com/neilturkewitz/status/1745225313229943068?t=BHPZA4RRpQR9KwjEt9GiVQ&s=19 AI and copyright Wheres the line? How similar to human brains do the artificial learning systems have to be for the training and inference to be considered similar enough to humans learning and creating? This feels dumb. I dont buy this, give me a more proper technical reason. You can also see humans as "just curve fitting and compression." (Predictive coding models) They are still in different ways different systems, but this is not it, seeing both systems from this lens is useful to make scientific predictions about them. I like the theory that brain works so well because of good inductive bias learned by evolution, but we still have to find that out. Maybe slightly different learning algorithm makes the difference (forward forward instead of backprop), maybe it is the specialized subnetworks architecture, maybe the bioelectric hardware plays a major role... We still dont properly know! We still have approximately zero idea what's going on inside brains and LLMs, and i think being this type of reductive in both cases is not really productive in this context. Both systems create out of training distribution output and produce complex circuits in their dynamics that we are slowly reverse engineering in mechanistic interpretability and circuit neuroscience. People's nontechnical definitions of intelligence seems to be very esoteric, there already seems to be milions of technical ones, and most of them work for both biological and artificial systems. I can't grasp why do so many people think that the algorithms and representations the human brains learn are inherently sooooo special when compared to machines. Feels so very anthropocentric and denying all the similarities between artificial and biological systems identified in circuit neuroscience and mechanistic interpretability, where we are still slowly but surely localizing and reverse engineering features and circuits, and we need to reverse engineer faster for better understanding of big language models that augment intelligence of people for now and for more efficient steering and for better neurotech that can help people! It's like this research doesnt exist? Predictive coding and active inference https://www.sciencedirect.com/science/article/abs/pii/S0149763423004426 Visual representational correspondence between convolutional neural networks and the human brain [Limits to visual representational correspondence between convolutional neural networks and the human brain | Nature Communications](https://www.nature.com/articles/s41467-021-22244-7) There is also work on forward forward algorithm. https://twitter.com/johnschulman2/status/1741178475946602979?t=Gmj98PERLsjdVKrH7aR2SQ&s=19 There are tons of AI algorithms doing automated theorem proving I dont think we are ideantical, but I also dont think we are soooo different information processing systems that so many people think. In computational neuroscience you can see common sense as special case of algorithms encoded in the biological neural networks through hebbian learning, reasons as inductive priors stored there, actions are in active inference as special case of minimizing prediction errors. [[2304.05366] The No Free Lunch Theorem, Kolmogorov Complexity, and the Role of Inductive Biases in Machine Learning](https://arxiv.org/abs/2304.05366) The No Free Lunch Theorem, Kolmogorov Complexity, and the Role of Inductive Biases in Machine Learning https://twitter.com/xuanalogue/status/1666765447054647297?t=7Ry_WGhni0zjvpQDg6SCbw&s=19 limits of Transformers Memorization is a phenomenon yes, similarly how humans can memorize images and sentences, and then (or while learning them) we both generalize over them, which we are slowly figiring out how to do more relaibly for neural nets [Limits to visual representational correspondence between convolutional neural networks and the human brain | Nature Communications](https://www.nature.com/articles/s41467-021-22244-7) https://twitter.com/johnschulman2/status/1741178475946602979?t=SNnX1DcQ8G9yuqNnSkWggQ&s=19 liquid neural networks [Liquid AI, a new MIT spinoff, wants to build an entirely new type of AI | TechCrunch](https://techcrunch.com/2023/12/06/liquid-ai-a-new-mit-spinoff-wants-to-build-an-entirely-new-type-of-ai/?guccounter=1) [[2006.04439] Liquid Time-constant Networks](https://arxiv.org/abs/2006.04439) PoC: LLM prompt injection via invisible instructions in pasted text https://twitter.com/goodside/status/1745511940351287394 Building AI without a Neural Network -- Hivekit is building tools for a distributed spatial rules engine that can provide the communications layer for hives, swarms, and colonies. https://twitter.com/burny_tech/status/1745542102954426746 ICE-GRT: Instruction Context Enhancement by Generative Reinforcement based Transformers [[2401.02072] ICE-GRT: Instruction Context Enhancement by Generative Reinforcement based Transformers](https://arxiv.org/abs/2401.02072) [AI and Mass Spying - Schneier on Security](https://www.schneier.com/blog/archives/2023/12/ai-and-mass-spying.html) The Internet enabled mass surveillance, but that still leaves you with exabytes of data to analyze. According to Bruce Schneier, AI’s ability to analyze and draw conclusions from that data enables “mass spying.” We are about to enter the era of mass spying. AI Eye health [Machine Learning Predicts Sight Loss: A Breakthrough in Eye Health](https://scitechdaily.com/machine-learning-predicts-sight-loss-a-breakthrough-in-eye-health/) AI is making classical algorithms better which moves threshold how much qubits in quantum computing is worth it [- YouTube](https://www.youtube.com/watch?v=Q8A4wEohqT0) Fruits reimann hypothesis https://pbs.twimg.com/media/GDlbDyxX0AA-q5T?format=jpg&name=small Awesome Deep Phenomena, A curated list of papers of interesting empirical study and insight on deep learning. Empirical Study, Neural Collapse, Deep Double Descent, Lottery Ticket Hypothesis, Emergence and Phase Transitions, Interactions with Neuroscience, Information Bottleneck, Neural Tangent Kernel [GitHub - MinghuiChen43/awesome-deep-phenomena: A curated list of papers of interesting empirical study and insight on deep learning. Continually updating...](https://github.com/MinghuiChen43/awesome-deep-phenomena) all links from mechinterp etc. discords [Frontiers | Closing the loop on morphogenesis: a mathematical model of morphogenesis by closed-loop reaction-diffusion](https://www.frontiersin.org/articles/10.3389/fcell.2023.1087650/full) morphogenetic gradients complement bioelectricity Does DNA have the equivalent of IF-statements, WHILE loops, or function calls? How about GOTO? [bioinformatics - Does DNA have the equivalent of IF-statements, WHILE loops, or function calls? How about GOTO? - Biology Stack Exchange](https://biology.stackexchange.com/questions/30116/does-dna-have-the-equivalent-of-if-statements-while-loops-or-function-calls-h) Building Transformers from Neurons and Astrocytes [Building Transformers from Neurons and Astrocytes | bioRxiv](https://www.biorxiv.org/content/10.1101/2022.10.12.511910v1) Sounds similar to active inference [Information bottleneck method - Wikipedia](https://en.wikipedia.org/wiki/Information_bottleneck_method) [AI Thought – How to Build Superintelligent AI](https://aithought.com/) how to build superintelligent ai global risk report 2024 https://twitter.com/tsarnick/status/1745549270743384161/photo/1 [Manifold hypothesis - Wikipedia](https://en.wikipedia.org/wiki/Manifold_hypothesis) The manifold hypothesis posits that many high-dimensional data sets that occur in the real world actually lie along low-dimensional latent manifolds inside that high-dimensional space. As a consequence of the manifold hypothesis, many data sets that appear to initially require many variables to describe, can actually be described by a comparatively small number of variables, likened to the local coordinate system of the underlying manifold. It is suggested that this principle underpins the effectiveness of machine learning algorithms in describing high-dimensional data sets by considering a few common features.