Important papers More papers Do we for more computationally and algorithmically efficient AGI need algorithmic improvement that forces the systems to have more evolutionary pressure to more consistently learn circuits that encode generalizing causal relationships between features in the training data using better training data, inductive biases, architecture, directing training dynamics?https://twitter.com/burny_tech/status/1745578056071234028 [List of unsolved problems in computer science - Wikipedia](https://en.wikipedia.org/wiki/List_of_unsolved_problems_in_computer_science) https://twitter.com/DrYohanJohn/status/1745593109499748843 Depression/compulsion is analogous to cognitive constipation Anxiety is analogous to cognitive diarrhea Psychedelics are analogous to cognitive laxatives DPO>RLHF https://twitter.com/AndrewYNg/status/1745516258697863259?t=FM8AHKa4CrbaRTfLDYiNkA&s=19 "Relativity, QFT, computing and information theory were the most recent hella-bangers we had The upcoming hella-bangers will be of the rigorous formalizations of complex systems/brains/intelligences and meaningfully controlling them, designing them precisely, and living in harmony with them without rapid diversity/mode collapse of our future (fingers crossed, aspirational) And then after that will be an exotic apparent-transformation of the laws of physics via the creation of synthetic/intelligent matter, masterful control of energy to warp spacetimebranch, upgrades to intelligence and form, and a partitioning, diversifying, growing, and self-reorganizing of sentience" https://twitter.com/anthrupad/status/1745603846095692043 TrustLLM: Trustworthiness in Large Language Models [[2401.05561] TrustLLM: Trustworthiness in Large Language Models](https://arxiv.org/abs/2401.05561) Mistral medium beats Claude https://twitter.com/aidan_mclau/status/1745080680625680533 Physics books https://twitter.com/martinmbauer/status/1745503661235560847?t=K8pFGf2QGamB9kojEnulAw&s=19 Vždycky když tě pošle jinam nebo nechce něco udělat tak řekni ať to udělá on a klidně po pidi krůčkách XD nebo když je vague tak ať je víc specifickej apod. 😄 nebo když nerozumíš slovu něbo větě nebo kódu tak se ptej furt dokola ať to vysvětluje normálně, víc laicky, nebo víc technicky exaktně, v příkladech, v metaforách, v teorii, v praxi jde z toho vydolat celkem hodně když ho člověk nutí jít deeper a deeper do jeho asociačního grafu nebo ať používá internet nebo různý gigantický moderní resources v souboru nebo langchain je knihovna na za míň kódu celkem komplexní agenty co můžou mít přístup k libovolným akcím a být embedded v čemkoliv, v tom píšu nejvíc, libovolný jazykový modely pod libovolnýma promptama mají přístup k libovolným nástrojům a funkcím v cyklu koukni, přemýšlej, konej v pár řádcích kódu [Concentration of measure - Wikipedia](https://en.wikipedia.org/wiki/Concentration_of_measure#Concentration_of_measure_in_physics) All classical statistical physics is based on the concentration of measure phenomena Information decomposition in brain https://www.sciencedirect.com/science/article/pii/S136466132300284X Transformers are multistate rnns https://twitter.com/_akhaliq/status/1745634153221988848?t=b6Zq1WxW4p2d1Ip000qhyw&s=19 This won't be the result, not even governmental tyranny or corporate empires cyberpunk dystopia. Safe AGIs benefiting all of sentience in democratic transhumanist protopia will be the future. Let's build that reality. https://twitter.com/burny_tech/status/1745878385044099212 corporate centralizing intelligence and power, or corporate paperclipmaxxing AIs for profit and ignoring all negative externalities that halt technological progress that halt the possibility of transhumanistic future protopias, or overweaponization that slows down progress, or rogue AI that doesnt care about sentience if it gets autonomous, decieving AI (i think both corporates and most of open source people dont want their LLMs to decieve/lie to them or those that it interacts with) Decentralized ecosystem of billions of cooperating specialized superintelligent John Van Neumanns thinking at billions times the speed, accuracy, efficiency than an avarage human as collective emergent hivemind superagent in one computational system versus humanity https://twitter.com/burny_tech/status/1745906656284733844 [[2401.02843] Thousands of AI Authors on the Future of AI](https://arxiv.org/abs/2401.02843) [- YouTube](https://youtu.be/cI108ly0hXs?si=Y1nI6XQPDtx8BRdv&t=786) Thousands of AI Authors on the Future of AI https://openai.com/research/language-models-can-explain-neurons-in-language-models Metalearning [MAML Explained | Papers With Code](https://paperswithcode.com/method/maml) Model-Agnostic Meta-Learning failure points of rag systems https://twitter.com/emollick/status/1745847292450501084/photo/1 open interpreter desktop assistant [- YouTube](https://www.youtube.com/watch?v=xPd8FFzIeOw) "studying mechanistic+developmental interpretability of deep neural networks intuitively seems like it should yield insights that span many disciplines - for reasons similar to why physics analogies (e.g. fluid mech or phase change stuff) are omnipresent and useful whatever is developed along the way to understand an "extremely high dimensional object undergoing changes due to costs demanding quality prediction/compression/generation of high complexity information" covers a lot of conceptual ground if there's billions of degrees of freedom in the sculpting of some [object], guided by a massive inflow of 'unnatural' information signals generated by the interactions of entire civilizations across time, and tons of degrees of freedom in the output trajectories of the [object] then that's like understanding a whole new mini-universe with entirely different physics - "we've got to start all over", in some sense. And i'd only expect this to be more the case as AI technology advances" https://twitter.com/anthrupad/status/1745958521349238809 [AMIE: A research AI system for diagnostic medical reasoning and conversations](https://blog.research.google/2024/01/amie-research-ai-system-for-diagnostic_12.html) AMIE: A research AI system for diagnostic medical reasoning and conversations theory of everything using category theory "Applications of Non-Standard Analysis in Topoi to Mathematical Neuroscience and Artificial Intelligence: Infons, Energons, Receptons" [Applications of Non-Standard Analysis in Topoi to Mathematical Neuroscience and Artificial Intelligence: Infons, Energons, Receptons (I)[v2] | Preprints.org](https://www.preprints.org/manuscript/202001.0102/v2) topos theory, that formalizes the notion of different mathematical universes, such as law of excluded middle being an axiom or not (every proposition is either true or false) or infinitesimals being an actual number object in synthetic differential geometry, all using subobject classifiers [- YouTube](https://www.youtube.com/watch?v=o-yBDYgUqZQ) Synthetic Differential Geometry, The Kock-Lawvere Axiom,where there are numbers "d" that satisfy "d^2=0" and "d=/=0" to represent infinitesimals AI paper search assistant [ChatGPT - Consensus](https://chat.openai.com/g/g-bo0FiWLY7-consensus) [[2306.04640] ModuleFormer: Modularity Emerges from Mixture-of-Experts](https://arxiv.org/abs/2306.04640) ModuleFormer: Modularity Emerges from Mixture-of-Experts [[2307.14539] Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models](https://arxiv.org/abs/2307.14539) multi-modal jailbreaks [Phase transition - Wikipedia](https://en.wikipedia.org/wiki/Phase_transition) [Linking interindividual variability in brain structure to behaviour | Nature Reviews Neuroscience](https://www.nature.com/articles/s41583-022-00584-7) Linking interindividual variability in brain structure to behaviour The acceleration of artificial intelligence, in combination with the dumbing down of the general population, is a highly volatile mix. Discrete Wavelet Transform AutoGen Studio UI TaskWeawer framework to decompose a complex task into simplier ones and execute those step by step [TaskWeaver + Planner + Plugin = Super AI Agent - YouTube](https://www.youtube.com/watch?v=fuuAX7bzgv8) CrewAI LLM multiagent framework, Open source Autogen alternative, more flexible [- YouTube](https://www.youtube.com/watch?v=tnejrr-0a94) It's actually pretty easy to write multiagent ecosystems of interacing LLM agents solving a task together right now. I think it's not that hard to go meta and make multiagent ecosystem of interacing LLM agents writing multiagent ecosystems of interacing LLM agents. Writing a small documentation on how to write these simple programs and give it to the higher order ecosystem as instructions in an external memory tool to search in, or in context window. Or instruct it to search the existing documentation of these tools or how to use them on the internet. [ai-agi-researchers/multi_agent.py at main · BurnyCoder/ai-agi-researchers · GitHub](https://github.com/BurnyCoder/ai-agi-researchers/blob/main/multi_agent.py) AGI should ask the user multiple questions to give specifics after a prompt thats too vague. For example making a neurotech company. What type of neurotech? What's the budget? What's the timelines? Optionally user could ask the AGI to infer the best kind of neurotech, and then it could ask again. The cheapest? With highest chances of not going bankrupt? The most groundbreaking? biggest neuromorphic project [- YouTube](https://www.youtube.com/watch?v=8-WZiLOxoC8) LLMs in games [- YouTube](https://www.youtube.com/watch?v=ozoPja1TkFg&t=443s) [LARP: LANGUAGE-AGENT ROLE PLAY FOR OPEN-WORLD GAMES](https://miao-ai-lab.github.io/LARP/) What we're looking for is how everything works. What makes everything work. It has to do with curiosity. It has to do with people wondering. what makes something do something? There's a way of looking anew. As if you never saw it before the first time and asking questions about it as if you were different. And then to discover that if you try to get answers then that we made it return. The things that make the wind, make the waves. The motion of water, is like the motion of air, is like the motion of sand. Things have common features. It turns out more and more universal. It's almost unbelievable. The truth is so remarkable, so amazing. It's incredible. Take the world from another point of view It's curiosity It's the way who we are, what we are. Just like a runner gets a kick out of sweating. I get a kick out of thinking. I can't stop. It's curiosity. https://twitter.com/burny_tech/status/1746379718113308736 topology of space with loopy noncontractable parts may enable signals travelling faster than light and backwards in time https://fxtwitter.com/tsarnick/status/1746285029557469607 This Singularity church for immortalists and transhumans, blending together spirituality, science and futurism, is cool! ""Our task is to make nature, the blind force of nature, into an instrument of universal resuscitation and to become a union of immortal beings." - Nikolai F. Fedorov. We hold faith in the technologies & discoveries of humanity to END AGING and Defeat involuntary Death within our lifetime. Working to Save Lives with Age Reversal Education. We believe that all of life is sacred and that we have been given this one life to make unlimited. We believe in our Creator’s divine plan for all of humanity to have infinite lifespans in perfect health and eternal joy, rendering death to be optional. By following our Gospel we achieve eternal life creating a heaven here on earth. We follow Nikolai Fyodorov, who taught that the transcendence of the creator will only be solved when humanity in our unified efforts become an instrument of universal resuscitation, when the divine word becomes our divine action. And we follow Arthur C. Clarke, who said "The only way to discover the limits of the possible is to go beyond them into the impossible." And so, we enter each day energized in Spirit and empowered by the words of our prophets to live in joy, serving our creator and all of mankind, Forever and Ever. Wishing you Perfect Health and Great Longevity! Perpetual Life, a science-faith based church is open to people of all faiths & belief systems. We are non-denominational & non-judgmental and a central gathering place of Immortalists & Transhumans. What unites us is our common faith, belief and desire in Unlimited Life Spans." https://twitter.com/tsarnick/status/1746278071895011703 https://twitter.com/karpathy/status/1745921205020799433 "I touched on the idea of sleeper agent LLMs at the end of my recent video, as a likely major security challenge for LLMs (perhaps more devious than prompt injection). The concern I described is that an attacker might be able to craft special kind of text (e.g. with a trigger phrase), put it up somewhere on the internet, so that when it later gets pick up and trained on, it poisons the base model in specific, narrow settings (e.g. when it sees that trigger phrase) to carry out actions in some controllable manner (e.g. jailbreak, or data exfiltration). Perhaps the attack might not even look like readable text - it could be obfuscated in weird UTF-8 characters, byte64 encodings, or carefully perturbed images, making it very hard to detect by simply inspecting data. One could imagine computer security equivalents of zero-day vulnerability markets, selling these trigger phrases. To my knowledge the above attack hasn't been convincingly demonstrated yet. This paper studies a similar (slightly weaker?) setting, showing that given some (potentially poisoned) model, you can't "make it safe" just by applying the current/standard safety finetuning. The model doesn't learn to become safe across the board and can continue to misbehave in narrow ways that potentially only the attacker knows how to exploit. Here, the attack hides in the model weights instead of hiding in some data, so the more direct attack here looks like someone releasing a (secretly poisoned) open weights model, which others pick up, finetune and deploy, only to become secretly vulnerable. Well-worth studying directions in LLM security and expecting a lot more to follow." Coherent Extrapolated Volition was a term developed by Eliezer Yudkowsky while discussing Friendly AI development. It’s meant as an argument that it would not be sufficient to explicitly program what we think our desires and motivations are into an AI, instead, we should find a way to program it in a way that it would act in our best interests – what we want it to do and not what we tell it to. Related: Friendly AI, Metaethics Sequence, Complexity of Value In calculating CEV, an AI would predict what an idealized version of us would want, "if we knew more, thought faster, were more the people we wished we were, had grown up farther together". It would recursively iterate this prediction for humanity as a whole, and determine the desires which converge. This initial dynamic would be used to generate the AI's utility function. https://www.lesswrong.com/tag/coherent-extrapolated-volition Selfreplicating AI worm https://twitter.com/gfodor/status/1746339177778888973?t=jIYolTRdx3-GteIn-EORKA&s=19 - AI worm/cyberpandemic - Non-human intelligence disclosure - Election integrity crisis - Criminalization of computation - Taiwan invasion https://twitter.com/gfodor/status/1737527243336032732?t=MUddpHsXb4RZ6mdZgF1GGg&s=19 Quantumness rewriting causality https://twitter.com/QuantaMagazine/status/1746242982490034334?t=NyR6dsCa29eVdI50-bKNAg&s=19 We develop a method for understanding how sparse autoencoder features in transformer models are computed from earlier components, by taking a local linear approximation to MLP sublayers. We study both how the feature is activated on specific inputs, and take steps towards finding input-independent explanations via examining model weights. We demonstrate this method with several deep-dive case studies to interpret the mechanisms used by simple transformers (GELU-1L and GELU-2L) to compute some specific features, and validate that it agrees with the results of causal methods. Mysticism landscape https://twitter.com/The4thWayYT/status/1746390338317545578?t=BffRFAvZbHrR0qeHW4aBww&s=1 CrewAI https://twitter.com/joaomdmoura/status/1728144699083407401?t=sWVUsVVdHhqyg_Vgu-Z2Cw&s=19 Hmm, když se takhle podaří rozložit jak se různý konkrétní features a circuits skládají v nejpoužívanějších open source modelech na senzitivních místech pro companies, státy, lidi,..., tak je z toho technicky (často transferable i na closed source modely) automated gradientbased jailbreak attacks free real estate in theory forcing arbitrary behavior https://www.lesswrong.com/posts/93nKtsDL6YY5fRbQv/case-studies-in-reverse-engineering-sparse-autoencoder [[2307.15043] Universal and Transferable Adversarial Attacks on Aligned Language Models](https://arxiv.org/abs/2307.15043) Ale taky to umožní tvoření efektivnějších hardcoded antijailbreak antivirů, který tyto obrany můžou mít cryptographically obfuscated [#8: Scott Aaronson - Quantum computing, AI watermarking, Superalignment, complexity, and rationalism - YouTube](https://youtu.be/wfxf6MembCQ?si=3TWcYd4oxS7vK9B3&t=3732) ale stejně tak můžou být hardcoded a cryptographically obfuscated malicious circuits uvnitř neuronek Pro efektivní AI potřebuješ masově paralelně sčítat a násobit matice, což je lepší na GPU, ještě lepší na akcelerátorech, ještě lepší na hyperspecialized hardwaru jako neuromorphických čipech nebo optical computational substrátu, nebo doslova transformer architektura vytištěná jako hardware. Ještě quantum machine learning zní zajímavě, je to strašně málo prozkoumaný, ale v teorii by asi mohl přidat další paralelizaci podobně jako další kvantový algorithmy, a je to blíž k tomu jak fyzika reálně funguje pro učení se fyziky přes ML, ale v praxi je to zatím problém, protože kvantový počítače jsou dost mini, že se to zatím moc nevyplatí [Von Neumann entropy - Wikipedia](https://en.wikipedia.org/wiki/Von_Neumann_entropy) Von Neumann entropy in quantum statistical mechanics be playing with that diagonalized density matrix of pure or mixed states with representing eigenvalues that start having fuzzy confidence when they mixed leading to spreaded out probabilistic disorder as multiple eigenvectors stretch the probabilistic space in decentralized distributed way increasing entropy! https://twitter.com/burny_tech/status/1746678341610938758 https://twitter.com/Schindler___/status/1745986132737769573 " Proposition of an architecture for AGI. Samantha from the movie Her is here: An autonomous AI for conversations capable of freely thinking and speaking, continuously learning and evolving. Creating an unparalleled sense of realism and dynamicity. Features: -Dynamic speech: Samantha can speak whenever it chooses to, influenced by its context and thoughts. In stark contrast to normal LLMs which are limited to reacting, Samantha can act. It is also not limited to solving tasks, like all other autonomous agents. -Live visual capabilities: Visuals are only mentioned and acted upon directly if relevant, but always influences thoughts and behavior. -External categorized memory: Gets dynamically written and read by Samantha, which chooses the most relevant information to write, and to retrieve to context. -Evolving at every moment: Experiences that get stored in the memory can influence and shape subsequent Samantha behavior, like personality, frequency, and style of speech, etc. A true independent long-running agent, actual Artificial Intelligence, as defined by Karpathy." Open-Vocabulary SAM: Segment and Recognize Twenty-thousand Classes Interactively [OVSAM | MMLab@NTU](https://www.mmlab-ntu.com/project/ovsam/) 1. “Are Ai Chatbots Behaviorally Similar to Humans? Their behaviors are often distinct from average and modal human behaviors, in which case they tend to behave on the more altruistic and cooperative end of the distribution. We estimate that they act as if they are maximizing an average of their own and partner’s payoffs. Chatbots also modify their behavior based on previous experience and contexts “as if” they were learning from the interactions...” https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4637354 [[1911.09390] Von Neumann Entropy in QFT](https://arxiv.org/abs/1911.09390) Von Neumann Entropy in QFT "Inspired by the self-play technique in reinforcement learning and the use of LLMs to simulate human agents, we propose a more effective method for data collection through LLMs engaging in a conversation in various roles. This approach generates a training data via "self-talk" of LLMs that can be refined and utilized for supervised fine-tuning. We introduce an automated way to measure the (partial) success of a dialogue. This metric is used to filter the generated conversational data that is fed back in LLM for training" [[2401.05033] Bootstrapping LLM-based Task-Oriented Dialogue Agents via Self-Talk](https://arxiv.org/abs/2401.05033) “CrewAI allows you to create incredible AI agent teams, similar to AutoGen. It is simple and intuitive, allowing you to accomplish tasks like research, writing, stock analysis, and trip planning. Plus, it can run entirely locally with open-source models. Here is a step-by-step guide to use it.” [- YouTube](https://www.youtube.com/watch?v=tnejrr-0a94)