[x.com](https://twitter.com/burny_tech/status/1769763717372096762) Pipeline: 1. doesnt know about machine learning math 2. learns math of the architectures and training algorithms and starts thinking that's all there is to it (plus there isnt a person that knows all architectures and learning algorithms, since there is bambilion of them and new ones are constantly emerging and are being used even in the top models) 3. starts delving into practical machine learning alchemy methods and reverse engineering and theoretical mathematical models of deep learning, machine learning in general, or artificial intelligence in general, or intelligence in general, into mechanistic interpretability, physics based theories of learning, inference, etc., and it being endless and full of unanswered questions, where each answered question just opens tons of new questions ye the whole universe is just one gigantic pile of applied math where we've discovered infinitely small part of that landscape indirect realism could imply that the notion of consciousness existing, physics, being an observer, knowing, maths etc. is just in our inner world simulation and therefore not actually objectively etc. existing (but same may be applied to subject and object duality, mmhmmm) it's very psychedelic idea but not useful for engineering And yet we're exploring the features learned by both artificial and biological neural networks without needing to go into the quantumness [x.com](https://twitter.com/burny_tech/status/1769804367983427665) i bet the correlation will be pretty big between need to transcend everything in one's models and the usage of psychedelics or deconstructive meditation since they neurophenomenologically tend to deconstruct and interconnect one's models more, boundaries seem less significiant How do we approach the complexity of life as a nonequilibrium system? Using physics and, in particular, landscape and flux theory, many systems, from cells to ecology and cancer, can be adequately represented and modelled. [x.com](https://twitter.com/ricard_sole/status/1769836061679611986) [[2403.12021] A tweezer array with 6100 highly coherent atomic qubits](https://arxiv.org/abs/2403.12021) continuous learning is all you need (aidan et al.) mastering continuous learning by self-play with a general reinforcement learning algorithm (aidan et al.) continuous learning: an open-ended embodied agent with large language models (aidan et al.) AI generate AI generation, AI generate all of existence https://phys.org/news/2024-03-neural-networks-mathematical-formula-relevant.html https://www.science.org/doi/10.1126/science.adi5639 [x.com](https://twitter.com/burny_tech/status/1769963434798465518) [Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419 - YouTube](https://www.youtube.com/watch?v=jvqFAi7vkBc) [NVIDIA Just Started A New Era of Supercomputing... GTC2024 Highlight - YouTube](https://www.youtube.com/watch?v=GkBX9bTlNQA) [Nvidia 2024 AI Event: Everything Revealed in 16 Minutes - YouTube](https://www.youtube.com/watch?v=bMIRhOXAjYk&pp=ygUGbnZpZGlh) [Simultaneous and Heterogenous Multithreading | Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture](https://dl.acm.org/doi/10.1145/3613424.3614285) [Computing 'paradigm shift' could see phones and laptops run twice as fast — without replacing a single component | Live Science](https://www.livescience.com/technology/computing/computing-paradigm-shift-could-see-phones-and-laptops-run-twice-as-fast-without-replacing-a-single-component?utm_medium=social&utm_campaign=socialflow&utm_content=livescience&utm_source=facebook.com) [20 Years of Not Even Wrong | Not Even Wrong](https://www.math.columbia.edu/~woit/wordpress/?p=13864) The holy war of intelligence Will AI training in the future be simulating whole universes with societies (Matrix) with embodied AI agents getting trained across parallel timelines at the same time? Is there a cost function for us humans under which we help to train in this universe that might be simulated? [x.com](https://twitter.com/burny_tech/status/1770207631581364624) Reminder, tweet by an OpenAI employee The models we are currently using daily (GPT4, Gemini 1 Ultra, Claude) are only at 2022 level compute Nvidia just skyrocketed this to the stars AI recursively designing 27 trillion parameter AI models orders of magnitue times more efficient and faster at inference and reasoning using all the research that we figured out over the last few years incoming [x.com](https://twitter.com/burny_tech/status/1770209519509135542) [Paired open-ended trailblazer (POET) - Alper Ahmetoglu](https://alpera.xyz/blog/1/) metalearning Learn every single pattern in the whole universe across all scales, space and time Learn every single pattern in the whole universe across all scales, space and time Natural language instructions induce compositional generalization in networks of neurons [Natural language instructions induce compositional generalization in networks of neurons | Nature Neuroscience](https://www.nature.com/articles/s41593-024-01607-5) "Our best models can perform a previously unseen task with an average performance of 83% correct based solely on linguistic instructions (that is, zero-shot learning)." Meta deployed an AI powered software engineer like Devin at scale This is the future of software engineering as proved by its 73% success rate [x.com](https://twitter.com/GPTJustin/status/1770233883629654503?t=SyEgKWD6140gLen0sQy8wQ&s=19) [[2402.09171] Automated Unit Test Improvement using Large Language Models at Meta](https://arxiv.org/abs/2402.09171) co je gravitace? síla? špatně, je to zakřivení časoprostoru! teda vlastně ne, možná je to spíš částice! to vypadá že je špatně, třeba je všechno ze strun a má to 11 dimenzí místo 3! omg ne, možná je všechno z kvantových smyček! nebo ne, třeba stačí k síle přidat trochu náhodnosti podobné té kvantové https://github.com/tensorush/Awesome-Physics-Learning [[2310.04406] Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models](https://arxiv.org/abs/2310.04406) Pessimism gets civilizations nowhere [Oxford Scientist Explains The Quantum Turing Machine - YouTube](https://www.youtube.com/watch?v=gBilsurLrcM) Depends on how you define sentience and consciousness.🤷‍♂️There are 465498431654 different cognitivist and behaviorist definitions of these words. ( [Consciousness - Wikipedia](https://en.wikipedia.org/wiki/Consciousness?wprov=sfla1) ) In empirical physicalist neuroscience, these articles go over most popular models of consciousness well (global neuronal workspace theory + integrated information theory + recurrent processing theory + predictive processing theory + neurorepresentationalism + dendritic integration theory, An integrative, multiscale view on neural theories of consciousness https://www.cell.com/neuron/fulltext/S0896-6273%2824%2900088-6 ) (Models of consciousness Wikipedia [Models of consciousness - Wikipedia](https://en.wikipedia.org/wiki/Models_of_consciousness?wprov=sfla1) ) (More models https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8146510/ ) This review ( Consciousness in Artificial Intelligence: Insights from the Science of Consciousness [[2308.08708] Consciousness in Artificial Intelligence: Insights from the Science of Consciousness](https://arxiv.org/abs/2308.08708) ) goes over some empirical definitions of consciousness (which is usually considered as different from sentience) used in neuroscience and the state of LLMs in relation to these, but it's almost a year old, and this one ( [[2303.07103] Could a Large Language Model be Conscious?](https://arxiv.org/abs/2303.07103) ) looks at it more philosophically, which is similarly old. When it comes to testing and falsifying, I like neural correlates that break down in for example anesthesia and deep sleep, but one could argue person is still having an experience there, just memory is turned off. ( [Joscha Bach Λ Ben Goertzel: Conscious Ai, LLMs, AGI - YouTube](https://youtu.be/xw7omaQ8SgA?si=dNzQ1O-d8DA3FrGK&t=2658) Joscha Bach Λ Ben Goertzel: Conscious Ai, LLMs, AGI) Or I like the experience qualia bridge empirical testing of consciousness solution, where you make a bridge between conscious systems and if you can transfer experience, then that makes both systems conscious, but I think even that can be deconstructed too by questioning the layers of assumptions it pressuposes. ( [Beyond Turing: A Solution to the Problem of Other Minds Using Mindmelding and Phenomenal Puzzles | Qualia Computing](https://qualiacomputing.com/2016/11/12/beyond-turing-a-solution-to-the-problem-of-other-minds-using-mindmelding-and-phenomenal-puzzles/) ) Also each definition and model seems to be motivated by different predefinitions asking different questions wanting to solve different problems. There are also sometimes defined different types of consciousness for different systems aka protoconsciousness. Though I personally feel like we actually have no idea how consciousness scientifically works in biological organisms, because there are too many degrees of freedom in philosophy of mind (it really depends to what ontology you subscribe to [Philosophy of mind - Wikipedia](https://en.wikipedia.org/wiki/Philosophy_of_mind) (substrate dualism, property dualism, reductive physicalism, idealism, monism, neutral monism, illusionism, panpsychism, mysterianism, transcendentalism, relativism,...) which seems to be arbitrary) and empirical verification of current models under various ontological paradigms can be so questionable... My favorite practical set of assumptions on which you then build empirical models for all of this is probably what free energy principle camp, Friston et. al, (Can AI think on its own? [The Free Energy Principle approach to Agency - YouTube](https://youtu.be/zMDSMqtjays?si=MRXTcQ6s8o_KwNXd) ) (Inner screen model of consciousness: applying free energy principle to study of conscious experience [Inner screen model of consciousness: applying free energy principle to study of conscious experience - YouTube](https://youtu.be/yZWjjDT5rGU?si=KCX_n3ChCn-kWt98) ) are using with Markov blankets with more and more complex dynamics creating more complex experiences, or Joscha Bach's coherence inducing operator (Synthetic Sentience: Can Artificial Intelligence become conscious? | Joscha Bach [Synthetic Sentience: Can Artificial Intelligence become conscious? | Joscha Bach | CCC #37c3 - YouTube](https://youtu.be/Ms96Py8p8Jg?si=HYx2lf8DrCkMcf8b) ), but I'm open to this whole landscape and don't think any set of assumptions is inherently more true than others, because I don't see a way to falsify assumptions that are living before empirical models that you can falsify. In terms of philosophy of mind assumptions, people constantly argue: "No, my set of initial assumptions is right!" even when I don't see a way to confirm that (in the scientific way), which seems odd from my perspective, where all models are wrong, but some approximate, compress, predict sensory data better than others which is more useful in practice for engineering... I guess I'm mostly subscribed to ontological pragmatism/relativism/mysterianism/anarchism? (but that's probably another arbitrary set of assumptions 😄 and very meta one) I converged to my view that we have no idea what consciousness is, because I haven't been really deeply convinced by all the proposed empirical tests for consciousness that I could find (also depending on which definition of consciousness/sentience is meant and under what philosophy of mind ontology 😄 ) But ontology like physicalism and epistemology like empiricism is probably the most useful framework we have for scientifically exploring engineering conscious systems, biological or nonbiological. And I care about it because it has implications in artificial intelligence ethics. [[2403.11901] Larimar: Large Language Models with Episodic Memory Control](https://arxiv.org/abs/2403.11901) [x.com](https://twitter.com/burny_tech/status/1770497046421651520)