" [Detecting hallucinations in large language models using semantic entropy | Nature](https://www.nature.com/articles/s41586-024-07421-0) [https://www.youtube.com/watch?v=fsgyllS43KY](https://www.youtube.com/watch?v=fsgyllS43KY) I wonder if the overconfident wrong assertions aka hallucinations will ever be solved in LLMs! I think we somehow need to incorporate a better signal about ground truth in various usecases when training, which we now can do mostly with games, videogames, math and code competitions, as the correctness of these solutions can be easily verified and sent as a signal. Research is trying to solve this a lot outside these domains, and I think that will also lead to superhuman performance as it's leading there with the more easily verifiable domains with easier to get ground truth signals. Also maybe this will make a bigger comeback of neurosymbolic AI! It would be great if we reverse engineered hallucinations in current AI models using these or other methods maybe from mechanistic interpretability, or engineered hallucination rate from first principles in future models. To be able to tune LLM's hallucination rate beyond just the temperature parameter, from "if i can't prove it using some ground truth source right now, then i won't say it" to unconstrained creativity or chaos! For some it would lookup a trusted source (which is relatively much easier in science, math and technology than other domains, rip), and for some it would need to do an experiment. The new selfverifying selfcorrecting chain of thought + RL etc. paradigm is also a step forwards to that IMO! " Companies with overly inflated expectations firing people to replace them AI is often stupid right now. I've seen that bunch of times. Similar overly inflated expectations is happening with almost any new big technology (internet, quantum computing, 3D printing,...). AI will be able to automate more people, but I think it will be in few years, not in a year. Singularity is happening, but its not infinitely fast. yeah sex is probably good (idk havent gathered any empirical data there yet) but have you ever solved differential equations in quantum physics on the first try [TSMC execs allegedly dismissed Sam Altman as ‘podcasting bro’ — OpenAI CEO made absurd requests for 36 fabs for $7 trillion | Tom's Hardware](https://www.tomshardware.com/tech-industry/tsmc-execs-allegedly-dismissed-openai-ceo-sam-altman-as-podcasting-bro) https://x.com/ns123abc/status/1839400986265407887?t=1Rt7ZWB-ew42OolbtUOHcQ&s=19 just 7 trillion dollars' worth of 36 more semiconductor plants bro. i promise bro just 7 trillion dollars' worth of 36 more semiconductor plants and we will get to superintelligence that will benefit all of humanity in unimaginable ways. bro... just 7 trillion dollars' worth of 36 more semiconductor plants. please just 36 more. 7 trillion dollars' worth of 36 more semiconductor plants and we will get to superintelligence that will benefit all of humanity in unimaginable ways. bro c'mon just give me 7 trillion dollars' worth of 36 more semiconductor plants i promise bro. bro bro please i just need 7 trillion dollars' worth of 36 more semiconductor plants ClopenAI Make AI for love! Not AI for war! I love how we thought that AI will master all of math first before mastering normie language but it turns out that normie language is just too easy to imitate and math is way harder Look around. See the infinitely complex math from many perspectives in every collection of pixels of every modality of your experience? That's the most absolutely beautiful source code of the inner word simulation we inhabit! They will say it's not possible until you do it We need startups for fundamental science. [Než budete pokračovat na YouTube](https://www.youtube.com/shorts/ux1N5TAb5q8?si=gjvlKxPi1GQZnP_z) I will learn every single pattern governing this reality from the most fundamental level to all the emergent scales and nobody can stop me Understand the universe Scale the civilization Flying in groundless ground ❤️❤️💜🤍💚🤎💙💖💝💞💓💗 [Než budete pokračovat na YouTube](https://www.youtube.com/live/k7TWMSV5Xjk) Type of guy who's degree in computer science was the result of him wanting to learn how to make a few lines long Minecraft mod You can generate Anki cards using LLMs for arbitrary topics you want to learn by giving the LLM the material I love seeing the growing diversity of AI systems, it's like an emerging ecosystem of completely novel creatures with different types of intelligences evolving! <3 Phase transitions, phase transitions everywhere I love phase transitions everywhere across all scales [https://youtu.be/Vbi288CKgis?si=dlxzO6lrSKIp_Skq](https://youtu.be/Vbi288CKgis?si=dlxzO6lrSKIp_Skq) Is the emergence of human-like cognition and intelligence in biological dynamical selforganizing complex systems a sudden phase shift, a lots of stacked discrete sudden phase shifts, or very continuous phase shift, or something in the middle? Are the various causal effects more of cybernetic top down control from higher scales to lower scales, or is everything mostly bottom up, or middle, or some mix? [https://youtu.be/Vbi288CKgis?si=dlxzO6lrSKIp_Skq](https://youtu.be/Vbi288CKgis?si=dlxzO6lrSKIp_Skq) I'm not sure what to think about Gavin Newsom blocking the contentious AI Safety S.B. 1047 bill in California. I'm conflicted myself when I try to understand all sides of the AI safety discourse around this. It seems like it's largely about whether people trust governments more or companies more, and how real and probable they perceive the proposed AI risks to be. https://www.bloomberg.com/news/articles/2024-09-29/gavin-newsom-vetoes-california-s-contentious-ai-safety-bill Universally approximate the whole world and you're done Anki is extremely powerful learning technique Current AI systems are working in euclidean space and we will want AIs that can learn arbitrary geometries even beyond what humans came up with [https://youtu.be/LgwjcqhkOA4?si=xNj11PKGe0a8d4Em](https://youtu.be/LgwjcqhkOA4?si=xNj11PKGe0a8d4Em) DMT inductive bias [[1805.09112] Hyperbolic Neural Networks](https://arxiv.org/abs/1805.09112) Learn as much advanced mathematics as possible and pattern match it everywhere in reality and swim in euphoric bliss from that process Understand the whole universe! From fundamental physics to biology to machines to sociology! Use that knowledge for the benefit of all beings! Beneficial technology for everyone! AI for understanding the universe, for the benefit of all, helping with basic needs, helping with easier access to knowledge, helping with selfactualization At the Planck scale, we no longer have space as we understand it; it's not a Euclidean grid but something more fundamental. The solution has typically been to use more deep learning, not less. [https://youtu.be/LgwjcqhkOA4?si=xNj11PKGe0a8d4Em](https://youtu.be/LgwjcqhkOA4?si=xNj11PKGe0a8d4Em) Deep learning is a powerful yet brutalist programming paradigm called differentiable programming. It is ugly but effective, relying on a few simple algorithms blown up to the max. This approach expresses everything with approximately continuous numbers, using algorithms that converge within certain ranges. When convergence fails, researchers patch it by tweaking the architecture or introducing discrete operations. This process can be automated through architecture search. There's no obvious limit to this approach, as deep learning can be combined with and learn to use other techniques as needed, including discrete architectures. Deep learning researchers are pragmatic, willing to use whatever works. The solution has typically been to use more deep learning, not less. While deep learning is highly effective and adaptable, its underlying approach is somewhat crude or inelegant in its simplicity and reliance on scale and data efficiency, in contrast with more rigid symbolic AI approaches. The continuous nature of these models can lead to strange artifacts, such as generative models producing images with half-materialized glasses or other improbable scenarios that exist in the parameter space between discrete states. These artifacts are typically squeezed out by training the models harder, decreasing the likelihood of ending up in these impermissible areas of the parameter space. Combination of deep learning and symbolic AI in neurosymbolic AI is the way! Intelligence may be an externalized process in a cybernetic sense, rather than being fully embedded in a single agent or algorithm. I don't want to pause AI. I want improving understanding of how AI models work, minimizing concentration of power from AI, redistributing abundance from automation, possibly UBI, regulating harmful usecases of AI, accelerating benefitial usecases of AI like science, education, healthcare, math etc. everything is functions, [https://www.youtube.com/watch?v=PAZTIAfaNr8](https://www.youtube.com/watch?v=PAZTIAfaNr8) or you can generalize graphs and functions into categories with objects and morphisms and that will be (almost) all you need, with extremely general Kan extensions [https://www.youtube.com/watch?v=ime-EJbCUe0](https://www.youtube.com/watch?v=ime-EJbCUe0) [Reddit - The heart of the internet](https://www.reddit.com/r/math/comments/5i163f/in_category_theory_what_does_the_slogan_all/) [https://www.youtube.com/watch?v=49jUnrEuEmY](https://www.youtube.com/watch?v=49jUnrEuEmY) Kan Extensions is all you need [https://www.youtube.com/watch?v=49jUnrEuEmY](https://www.youtube.com/watch?v=49jUnrEuEmY) [Kan extension - Wikipedia](https://en.wikipedia.org/wiki/Kan_extension) [Reddit - The heart of the internet](https://www.reddit.com/r/math/comments/5i163f/in_category_theory_what_does_the_slogan_all/) you can develop both category theory inside of set theory and set theory inside of category theory [Category theory $\subset$ Set theory or vice versa? - Mathematics Stack Exchange](https://math.stackexchange.com/a/375066/1398110) [Category theory $\subset$ Set theory or vice versa? - Mathematics Stack Exchange](https://math.stackexchange.com/questions/375016/category-theory-subset-set-theory-or-vice-versa/375066#375066) [ETCS in nLab](https://ncatlab.org/nlab/show/ETCS) or category theory can be kind of seen as just a special case of graph theory (with additional structure to represent composition), and graph theory (and everything else) may be studied abstractly in category theory [Is Category Theory similar to Graph Theory? - Mathematics Stack Exchange](https://math.stackexchange.com/questions/1239027/is-category-theory-similar-to-graph-theory) [https://www.youtube.com/watch?v=XETZoRYdtkw](https://www.youtube.com/watch?v=XETZoRYdtkw) Doubters get proven wrong over and over and over and over and over and over and over and over and over and over and over and over and over again monoids in the categories of endofunctors is all you need i want to understand and build an operating system from scratch i want to understand and build the hardware its running on from scratch, the various hardware components, circuits, transistors, the process of building semiconductor devices, the most detailed physics behind it i want to understand and build all the programming languages used from scratch i want to understand and build the important software and algorithms running on top of all of this from scratch i want to mathematically derive and empirically understand all the complex algorithms and all the machine learning equations running on CUDA on top of all of this and backpropagating neural networks i want to reverse engineer all the equations, algorithms, physical laws governing all the brains that implement all of this like predictive coding and the whole universe running on standard model of particle physics that makes all those dynamics this emerge curiousity [https://www.youtube.com/watch?v=UjEngEpiJKo](https://www.youtube.com/watch?v=UjEngEpiJKo) What's your end goal? style transfer realistic videogames [https://youtu.be/XBrAomadM4c](https://youtu.be/XBrAomadM4c) https://fixupx.com/LinusEkenstam/status/1841821657515581948 Tohle style transfer video to video Mashup na drogách bývá spíš text to video Style transfer vždycky upravoval, ale teď je toho schopen víc a víc Nejstarší známý video style transfer je deep dream před 8 lety [https://youtu.be/DgPaCWJL7XI?si=YHvnmI_7qT8nQqzL](https://youtu.be/DgPaCWJL7XI?si=YHvnmI_7qT8nQqzL) Tady v neuronce hodně vystřelili dog detectors XD Tohle je extra obecný video model s video to video funkcí. Představ si kdyby to byl finetuned na tuhle usecase, nebo plně specializovaný, s trochu jiným setupem, v komplexní pipelině několika modelů, nonAI editingu, apod. [https://youtu.be/Y43EgHv0xrs?si=LCJ5QRpcZxKU3j5z](https://youtu.be/Y43EgHv0xrs?si=LCJ5QRpcZxKU3j5z) Plus si myslím že pro hry to chce víc naškálovat neurosymboliku. Tyhle dosavadní modely jsou skoro čistý neuronky, kde neurosymbolika k neuronkám přidá symbolický uvažování [Neuro-symbolic AI - Wikipedia](https://en.wikipedia.org/wiki/Neuro-symbolic_AI) Neuronky jsou v základu spojitý a aby ses nedostal do divných části toho šíleně dimenzionálního prostoru kde máš např napůl brýle na člověku a na půl ne tak tyhle outlier části prostoru musíš šílené squishovat s šíleným množstvím dat a trénováním, což ale trochu zároveň omezuje i kreativitu protože to overfituje [https://www.youtube.com/watch?v=LgwjcqhkOA4](https://www.youtube.com/watch?v=LgwjcqhkOA4) Neurosymbolika od základu je schopna jednoduššího zobecňování a tenhle spojitý squishing není potřeba, a tuna lidí se to snaží s neuronkama spojit [[2006.08381] DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning](https://arxiv.org/abs/2006.08381) Ale i tahle extrapolace ve squished prostoru dle mě může být hodně solidně kreativní, protože to může vytvořit takový kombinace co v trenovacich datech nebyly přes retrieving společných konkrétních programů a jejich kombinaci 😄 [https://www.youtube.com/watch?v=nL9jEy99Nh0](https://www.youtube.com/watch?v=nL9jEy99Nh0) Například v hudbě dosavadní populární systémy jsou schopny vytvořit novou hudbu, jako třeba combine West African drumming with Western classical music, set Mozart to a polyrhythm, apod. ale myslím že pokud by slyšely všechno před grindcore, neoclassical metal, progressive jazz apod., tak by je neinvently, protože to je víc než jen simplier kombinace nebo ještě evoluční algorithmy jsou v tomhle na malých škálách malinko kretivnější dle jednoho researchera, tím jak lidi taky vznikli evolucí, ale nikdo je nenaškáloval dostatečně nebo ještě ten researcher dal tenhle cool example kreativity u jazykových modelů XD "Like, you can craft prompts describing an alien civilization with five sexes and the love affairs among these different combinations of alien sexes, what's considered cheating or not among different permutations of alien molecules? Right? Then you can ask the LLM what will be considered unethical to what degree by which of these alien sexes. Right? And there's nothing like that in the training data. You just made it up. And it can reason through that with great with great facility." [https://www.youtube.com/watch?v=jSDEsvVdL-E](https://www.youtube.com/watch?v=jSDEsvVdL-E) Tyhle populární videogen modely jsou většinou neuronky, aka deep learning s transformerama, difuzí, GANs, autoregresí,... [A Dive into Text-to-Video Models](https://huggingface.co/blog/text-to-video) , https://towardsdatascience.com/the-evolution-of-text-to-video-models-1577878043bd už dlouho, kde lidi hlavně škálují a používají kvalitnější data, ale jsou dost limited v několika aspektech. Až lidi víc narazí na limity, tak očekávám větší explozi alternativ mimo GPU poor research a mnohem větší diverzitu modelů schopný dělat diverznější tasky 😄 Ale je možný, že tenhle brutalistickej neuronkovej paradigm stačí, protože dokáže aproximovat libovolnou funkci, a třeba je to hlavně o datech, co musí vytvořit správný interní obvody co nebudou unstable k malým změnám na inputu, a squishnout pryč všechny části prostoru co člověk nechce Language gen, AI pro programování, matiku, vědu, hry apod. se do neurosymboliky pomalu a jistě morfuje. Image gen, video gen, game gen [https://youtu.be/0Xn8xGV_w9w?si=3zzbyasZQX3khZJY](https://youtu.be/0Xn8xGV_w9w?si=3zzbyasZQX3khZJY) apod. je další na řadě. Zajímá mě či se chain of thought reinforcement learning nějak dostane do video gen modelů, protože to teď totálně superchagnulo language gen, kde nový modely v posledních pár měsících kombinují víc high level reasoning chunks, což dost zlepšuje zobecňování a stabilitu, což se ale ještě nedostalo do video gen domény zatím. [GitHub - hijkzzz/Awesome-LLM-Strawberry: A collection of LLM papers, blogs, and projects, with a focus on OpenAI o1 🍓 and reasoning techniques.](https://github.com/hijkzzz/Awesome-LLM-Strawberry) , [https://www.youtube.com/watch?v=nO6sDk6vO0g](https://www.youtube.com/watch?v=nO6sDk6vO0g) , [Artificial Intelligence x Mathematics - Burny](https://burnyverse.com/OmniCortex/Artificial+Intelligence+x+Mathematics) A zajímalo by mě kdy někdo naškáluje reinforcement learning v counter striku na superhuman úroveň včetně long term plánování galaxy brain taktik jako se to udělalo v Dotě, StarCraftu, šachách, Go, Pokeru apod. Jediní co se pokusili neměli hodně grafik jako ostatní a dostali se celkem relativně daleko. [https://www.youtube.com/watch?v=UvoyMGxN89Y](https://www.youtube.com/watch?v=UvoyMGxN89Y) Bylo by mega zajímavý je sledovat proti sobě, nebo proti pro hráčům co by měli přístup ke všem možným cheatům :D This is the worst AI will ever be AI language gen, AI for programming, math, science, games, etc. is slowly and surely morphing into neurosymbolic AI! AI image gen, video gen, game gen is next! This is an extra general AI video model with video to video functionality. Imagine if it was finetuned for this usecase, or fully specialized, with a slightly different setup, in a complex pipeline of multiple models, nonAI editing, etc. https://x.com/LinusEkenstam/status/1841821657515581948 [https://youtu.be/Y43EgHv0xrs?si=LCJ5QRpcZxKU3j5z](https://youtu.be/Y43EgHv0xrs?si=LCJ5QRpcZxKU3j5z) Deep learning is continuous and to avoid getting into weird outlier parts of the exremely high dimensional space where you have e.g. half glasses on a person and half not so, you have to squish it insanely with crazy amounts of data and training, which also limits creativity because it tends to incentivise overfitting. That's why initial more chaotic genAI is more creative, as it's more free to roam the less unconstrained latent space. Neurosymbolic AI could find solutions for this. [https://www.youtube.com/watch?v=LgwjcqhkOA4](https://www.youtube.com/watch?v=LgwjcqhkOA4) I wonder when someone will scale reinforcement learning in counter strike to superhuman level including long term planning of galaxy brain tactics like it was done in Dota, Starcraft, Chess, Go, Poker etc. The only ones that tried didn't have a lot of GPUs like the others and they still got pretty relatively far. It would be mega interesting to watch them against each other, or against pro players who had access to all sorts of cheats :D [https://www.youtube.com/watch?v=UvoyMGxN89Y](https://www.youtube.com/watch?v=UvoyMGxN89Y) Are you commutative How much of a nerd are you on a scale 1 to 10 and why? my p(doom) is constantly oscillating sometimes i feel like regulations may cause more harm than good because of unknowledgable people and rotten politics full of power seeking people even if some people originally meant it in good way We're going to get to a point where we have longevity escape velocity. Longevity escape velocity is the point at which for every year that you live, science is able to extend your life for more than a year. I think if you're diligent, you'll be able to achieve that by 2029. That's only 5 or 6 years from now. Right now you go through a year, use up a year of your longevity, but you get back from scientific progress right now about four months. But that scientific progress is on an exponential curve. It's going to speed up every year. And by 2029, if you're diligent, you'll use up a year of your longevity with the year passing, but you'll get back a full year and past 2029 you'll get back more than a year. So you'll actually go backwards in time. The raw concentration of IQ in this picture is breaking the laws of physics more than black holes https://x.com/PhysInHistory/status/1842548559402648025?t=cEXqQzc9bcnSTQLRb88Y0g&s=19 I want great scientific and societal AI revolution for the benefit of all, but not lead by someone who constantly lies https://x.com/AISafetyMemes/status/1839281057587146984?t=afSvpGE9hsM6PzmUbbo-LQ&s=19 https://x.com/burny_tech/status/1842940906732990654 Time: Entropy Relativity Emergence [[1210.8447] Nothing happens in the Universe of the Everett Interpretation](https://arxiv.org/abs/1210.8447) [Does Time Really Flow? New Clues Come From a Century-Old Approach to Math. | Quanta Magazine](https://www.quantamagazine.org/does-time-really-flow-new-clues-come-from-a-century-old-approach-to-math-20200407/) [https://www.youtube.com/watch?v=r_fUPbBNmBw](https://www.youtube.com/watch?v=r_fUPbBNmBw) Path to understanding the fundamental equation of intelligence and physical reality Intelligence is the ability to achieve goals in a wide range of environments, weighted by their simplicity https://dl.acm.org/doi/10.5555/1325240.1325248