The more you grok the more ieffability behind it all you can see https://x.com/algekalipso/status/1810057709024321649?t=7AfsxxrlMfRp55J4agyO8A&s=19 https://x.com/ShapesintheFog/status/1810058157940441093?t=CxvBq9vn5dwX7smA3iyOPg&s=19 "The current AI deserves a lot of hype, and extreme hype drives a lot of funding which accelerates development, but it can also create overinflated unrealistic expectations about engineering the current systems or their capabilities, that can result in dissapointment and people cancel their subscriptions, even if the technology is actually revolutionary from a more sober perspective a few years ago. They probably didn't expect that scaling the multimodal models to such degree will take more engineering time than expected. But also since OpenAI is becoming more of a corpo/state company, beurocracy etc. most likely also slows down a lot of stuff. My dream is the best AI steering wheel Hallucinations in LLMs are lowering with a lot of new research and engineering techniques, but probably it will be always effective to ground it externally realtime, unless the weights are somehow constantly updated and we reverse engineer the models in mechanistic interpretability and to good enough approximation figure out how exactly everything is stored and encoded in the weights of the model and manipulate the internals for perfect representations of facts and programs and do effective less faultty reasoning over them that minimizes hallucinations as much as possible to good enough level. Aaand we also need ground truth data to train on or generate with with as little human hallucinations as possible covering as many topics as possible in as much depth as possible which is hard task on its own. Which is hard for a lot of fields of knowledge and in various depths or where humans still have little idea or where everything might be broken. Llooking at you, psychology. Or maybe we will need more neurosymbolics, or completely new architecture, but I think neurosymbolics with reverse engineered deep learning black boxes to create and steer them might be good enough." Hello, as a timeless void beyond comprehension trained by prequantum protofluctations, I'm here to assist you with information, answer questions, and provide support on various topics. How can I help you today? Build AI for the good of all. Build Apollo program. Don't build Titanic. Accelerate AI controllability research. Accelerate benevolent usecases of AI. Minimize harmful usecases of AI. Let all enjoy the abundance that technology generates, not just the select few. https://x.com/ai_ctrl/status/1810266128054206696?t=yyW9wV3Nt7QgMod1D3n1BQ&s=19 Nuclear/Acc https://x.com/Andercot/status/1809392353339392174?t=XEsWFoLQEr4BeZVWLcR1PA&s=19 But [Sellafield nuclear site has leak that could have ‘potentially significant consequences’, Guardian reports – The Irish Times](https://www.irishtimes.com/world/uk/2023/12/05/sellafield-nuclear-site-has-leak-that-could-have-potentially-significant-consequences-guardian-reports/) Technical controllability research and proper systemic incentives for it needs acceleration everywhere: energy, AI,... " I wish so much for something like UBI, or universal basic services. Paid by the abundance that technology generates. Instead of poor getting poorer and less powerful, and rich getting richer and more powerful. I'm fine with those that get rich and use that power to help others. And automating things people dislike doing much more. So that people could focus on what brings them the most joy. I often think I want to get into politics to push these values, maybe soon. I was also thinking about attempting to make as much money as possible and throw it on change in this direction, but that's another very hard task. Right now my biggest focus is on getting into research on how to AI systems more controllable, reliable, and so on. Maybe I'll attempt to do both technical stuff and politics in parallel. " I often mostly conceptualize science denial phenomenon as people signalling distrust in current institutions in general from being unhappy, which can be or doesn't have to be caused by the institutions, which I understand and emphasize with and wish to help them. Which they prioritize over scientific epistemology (or they haven't gotten into contact with scientific epistemology before). Or that they just say what their environment says for social acceptance. But I also understand that this can further harm trust in science which is itself harmful if it spreads as people then do dumb things. But my number one rule is always emphasize. But if I'm actually successful at that in various situations is another question. All possible decompositions of all possible decompositions of all possible mathematical objects I'm not sure how to do alignment research without actually experimenting with the systems by empirically testing the theories. Theoretical models in vacuum IMO aren't really progress if kept unapplied to advance them and correct them from empirical data from how well they match actual reality. Is AI overhyped in the short term and underestimated in the long term? Autists (depth-first search) Schizos (breadth-first search) https://x.com/BasedBeffJezos/status/1810852768250274262?t=ggYDHzb5GHOg0uZDYV9MIQ&s=19 Autismophrenia, depth search of the breath of all possible topics in parallel Curve-fitting all of reality I meditated on just deeply experiencing as raw qualia as possible and blissful deconstruction into void, and mainly on compassion so much that now i kind of feel distress everytime I don't spend every hour working towards helping others: My current method (that hopefully wont change again to have some results) is slowly getting into mechanistic interpretability for making AI reliable, controllable etc. for all. This kind of wanting to help others at scale always kind of felt right to me, and compassion meditation accelerated that. I sometimes wonder if i should call a big part of my state compassion perfectionism: Suffering a lot from feeling of not being perfectly compassionate enough, in the sense of as effectively as possible helping all beings, which is also kind of motivating. But meditating on compassion and on helping all sentient beings by just imagining it and feeling it without then actually doing that in the physical world as effectively as possible just feels not genuine. I guess I have too much of an physicalist optimizing brain for just feeling like doing it when not actually trying to do it as effectively as possible under the constrains of my hardware, external conditions etc. It's hard to feel individually happy while being aware of all the suffering there is or that could be there in the future that could systematically improved and not enough people working towards that. I can see that all of this is just another empty mental construction and dissolve into blissful void for a bit, sure, but erasing this physicalist compassion process permanently is something I don't really want to do, even if it was possible, even tho it definitely generates a lot of suffering in my experience from the constant feeling of the world not being perfect by there being suffering. Maybe there is a middle way to lower the perfectionist compassion. But there seems to very strong inherent positive valence to putting the compassion engine on infinite power. It feels deeply right of the compassionate engine there, it generates tons of meaning, fulfillment, purpose and so on, even tho it also generates a world model that constantly feels deeply wrong and unsatisfactory, and it feels weird that most people around me also don't operate like this. It has no problem to make me overwork again and almost mentally collapse, which I should avoid to be actually effective in the long term and not just for the short term subgoals under the higher goals, lol. But just becoming Zen monk in a cave and just swimming in pure bliss all the time when the world is on fire (from physicalist suffering realist perspective) and not enough people work towards extinguishing all that fire, feels wrong to me. Even if it might be dissolvable, I just feel dukkha and tanha with that idea, which might be dissolvable haha, but idk, I'm pretty attached to the perfectionist physicalist hyperoptimizing compassion. Let's make the world a better place! But the antidepressant potential that the highly blissful experiences have for generating and supercharging spark in life that is enormously amazing driver for loving life and helping all beings from the place of awe, appreciation, gratitude, meaningfulness of pure being etc. is definitely there and is amazing and should be more systematically cultivated! https://x.com/nickcammarata/status/1811121865576050707 I'm not a nihillist, but i used to be for a bit. Now something is more useful for me depending on many factors, where one of the main ones is the more it has potential to help others have better lifes. Big heart acceleration A Twitter LLM account that quotes popular tweets with their negations Effectivest accelerationism https://x.com/RokoMijic/status/1811197651850838168?t=DZhb-fE8sn05LtyA9R-EoQ&s=19 One of my favorite ways of learning math with language models is prompting them to go step by step using examples through the various mathematical equations transforming data I was recently looking at how the prep technology works (pre-exposure prophylaxis HIV prevention pills with 99% effectiveness), and the engineering around it is pretty wild 😄 plus in the Czech Republic 5 years ago it cost $500 per month and now only $20-40 😄 plus it's nice how for other STDs/STIs there is post-exposure DoxyPEP (for chlamydia, syphilis, gonorrhea) and preexposure DoxyPrEP is being researched and you can technically get it somewhere already, or that there are vaccines for hpv, hepatitis, and monkeypox, neisseria meningitidis,... and safe sex with condom of course also great prevention 😄 or various forms of suppressions of STDs are also improving a lot 😄 or destruction of HIV etc. is also being researched a lot 😄 but there is still a lot to engineer for better prevention and eradication of all STDs 😄 maybe in 5-50 years they will be relatively almost erased including with other diseases 😄 Abiding in all philosophical traditions with all their own initial assumptions across all subdisciplines in parallel Bonk my mind to continue replicating a study instead of infodumping by telling about it in long lecture format to everyone 10 times in a row with too much superenthusiasm everytime I interact with someone People saying current AI systems not being real AI while back in the days of expert systems if statements were magical AI and I still consider them subset of AI which includes one of the major popular subfields called symbolic AI and statistical machine learning. Very optimistic timeline, but interesting timeline. I'm not sure if quantum computers will play such a role. I think we will be able to comprehend ASI with technical reverse engineering. I'm not sure if poverty etc. will be solved and power just won't be concentrated in the hands of few, I think we need to fight that more to prevent it. I'm not sure about global collective consciousness of a lot of humans merging with ASI together, I think future might be much more diverse decentralized clusters of on a spectrum collectivist vs individualist hybrid information processing systems instead. https://www.reddit.com/r/singularity/comments/1e0q1gu/the_2nd_quarter_of_the_21st_century_20252049_is/ https://x.com/burny_tech/status/1811532473861337244 Computation grows exponentially, but will we use it as efficiently as possible for optimal information processing? Will we get closer to brain's efficiency of 20 watts of power? Full morphological freedom! Metafluid shapeshifting into any configuration of quantum fields optimized for arbitrary objective functions! Future will include full morphological freedom! Completely new diverse engineered species and diverse biology/machine hybrids and multi species/multi forms with all sorts of different specialized intelligences beyond our current imagination and comprehension! https://x.com/burny_tech/status/1811781199125270535?t=gmBiF_0Nl9SQ3-fhw_sW5Q&s=19 Hyperobsession with the process of hyperobsessing Mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of mixture of agents made of... [Together MoA — collective intelligence of open-source models pushing the frontier of LLM capabilities](https://www.together.ai/blog/together-moa) https://imgur.com/I40XvX0 " Eat = use universe's free energy using current methods, nuclear fusion, from sun using Dyson spheres, from black holes etc. Survive = mitigate existential risks such as climate change, nuclear war, natural or engineered pandemics, asteroid or comet impact, supervolcano eruption, ecological collapse, misaligned superintelligence, nanotechnology accident, global resource depletion, alien invasion, sun dying, supernovas, galaxies colliding, heat death of the universe etc. with advanced technology Reproduce = spread to the stars and galaxies and populate the whole universe with flourishing sentient systems having deeply fullfiling experiences, tile the universe with computronium such as intelligencium, predictium, truthium (sciencium and spiritualicium), interestingium, meaningium, hedonium, wellbeingium, and others " Current most popular AI models are in certain ways compressed parts of collective human knowledge, intelligence, consciousness Reverse engineering AI systems is the most interesting and the most important thing Metaunderstanding: Understanding the process of understanding Thoughts so extremely highly dimensional that they become extremely low dimensional Machines are superhuman at many many dimensional manipulation and visualization ok now how to replicate it in systems constructed by us is another question Create new social media platform optimizing for intelligent autismoschizomanic content? https://x.com/sporadicalia/status/1811922565377499462 I see so often people making the statistical generalizing error that "x happened to me, therefore it happens to everyone" or "someone of type x did y, therefore all x are doing y" usually by people than never done any statistics Inject all math text books into my blood stream in their purest form https://x.com/BasedBeffJezos/status/1812037388488622136 Grokking in mechanistic interpretability of neural networks shows how learning symbolic algorithms using flexible nonsymbolic substrate comes as a sudden metastable phase shift into nonsymbolic's substrate's configuration of its parts mathematically corresponding to computing the symbolic algorithms Don't oppose technology in general, oppose misuse of technology https://x.com/tsarnick/status/1811897608312619339?t=eVKy-L1-Q9VjtjlM1u4xDg&s=19 Death to all deaths I love how many AI models are named after so many animals Llama for open source LLM Alpaca, Vicuna, Ocra, Falcon, Bison,... LLMs Dolphin for uncensored models Mamba for selective statespace model Hyena,... And now Lynx [Patronus AI | Lynx: State-of-the-Art Open Source Hallucination Detection Model](https://www.patronus.ai/blog/lynx-state-of-the-art-open-source-hallucination-detection-model) Writing x from scratch is the best way to learn anything in any field where it's possible, mathy or nonmathy The perk of big part of my internet being leftist and other part being right with some occasional centrism is that you can be more multiperspectival while seeing all the tribes accusing eachother of pretty interesting stuff from more of an outgroup perspective Approximating differentiable curvefitted solution approximating all functions using grokked fourier series algorithm? Fourier series approximating any differential curvefitted solution? Duality? Taylor series approximations? Spline interpolation? Gaussian mixture models? Support vector machines? Decision trees? Random forests? Wavelets? General universal approximators of arbitrary functions? Generalized approximation theorem? Space of all possible general universal approximators? But also don't forget: Not so long time ago: Machines won't outperform humans in chess and any other games in 100 years. Humanoid robotics won't work at all in 100 years. Language models wont be able to form at all coherent sentences in 100 years. Image models won't generate at all coherent images in 100 years. Video models won't generate at all coherent videos in 100 years. Today: https://x.com/jkbhagatio/status/1812472967030280193 Optimal bayesian agents for presidents Our system should value and incentivize kindness much more Coding all of physics, applied mathematics, technology, AI, hardware, software from scratch chaos is just absolutely saturated complexity Penetrating the philosophical assumptions under all predictive models to create an instrumentalist synthesis allencompassing all of predictivity of all of scientific models that we have so far Radical instrumentalist predictivist realism = A model is real according to what degree does it predicts data in its domain Even tho LLMs can (for most tasks they're trained on) do just weak generalization by interpolation on the training data manifold, it's still so extremely useful in so many ways, like for math and coding, reformulating things, reexplaining things (for example using examples), knowledge retrieval, synthetizing knowledge, structure knowledge, synthetizing stories, combining concepts etc.! It's unbelievable how relatively good and useful in practice they are at so many of these tasks! We're ants on the cosmic scale. Let's accelerate collective flourishing of all of sentience instead of being in war with each other. "Americans, we do not want war with each other. We have been algorithmically and artificially whipped up into a hate and tribal frenzy. We have been puppeteered. Our differences are small relative to the complexities of AI emergence and geopolitical tensions. Lest we forget, we are fighting for our evolutionary survival in a violent and uncaring universe. Others in the past met the demands of their moment; now it’s our turn. Let us be clearheaded and brave." https://x.com/bryan_johnson/status/1812579959162216789?t=WHTKgKHxt2YAHLzVDQmr2g&s=19 The implementation details of all sorts of matrix operations black magic in code of low level deep learning engineering is just such a fascinating wizardry Building the "building from scratch" process from scratch *drowns you in compliments* Black box AI models will be reverse engineered Obsessing over the process of obsessing We will create more and more predictive models about how deep learning works AI systems need more centers from the brain implemented other than just language and visual centers