i think its fair to call it AI for variety of reasons, for example since deep learning emerged from connectionist paradigm in cognitive science trying to understand the brain depends how you define that word i like to define intelligence very generally as that allows for easier creation of various taxonomies for all the various systems etc. or because we find pretty complex emergent feature/circuit learning in deep learning systems, which is one of the most fascinating things ever to me :D or im mindboggled daily why do these systems even generalize in the first place etc. i think more people should be fascinated why all these systems even work we're still so early in the science of reverse engineering them after we grow them this is currently probably my favorite paper that reverse engineers various emergent circuits: [On the Biology of a Large Language Model](https://transformer-circuits.pub/2025/attribution-graphs/biology.html) but this research is about the old paradigm models, and particularly a low performing version one :D the science can't keep up with the progress at all im still waiting for the reverse engineering of the bigger SoTA foundational models and of the new reinforcement learning paradigm models that just from just mostly reinforcement learning seem to master lots of benchmarks i think we will find more sophisticated circuits in the new reasoning models trained via reinforcement learning but there are still so many issues like unreliability But I think in the future the models will be more neurosymbolic, where you will get symbolic composing of abstract concepts like in DreamCoder neurosymbolic architecture: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning [[2006.08381] DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning](https://arxiv.org/abs/2006.08381) Current systems have mostly fuzzy neural circuits living in gazilion dimensional vector space that easily break, so in this neurosymbolic AI paradigm you could have the fuzzy neural circuits living combined with symbolic programs that you could search for neurally or symbolically And instead of just neural circuits, you could have an ability to switch between purely neural or purely symbolic processing and reasoning, because e.g. for visual recognition, neural networks are best now, but e.g. for sorting a list, a symbolic algorithm is best. And then the ability to do different hybrid processing. But for example the new o3 model from OpenAI is starting to go in that direction. So even though fundamentally in the architecture afaik it's still neural autoregression using deep learning and doesn't learn a symbolic library of reusable strongly generalizing abstract symbolic concepts like DreamCoder does, and even though it has internal circuits, at least when inferencing in its reasoning tokens in chain of thought it is able to execute arbitrary python code, or use arbitrary tool use like web search or do multimodality with images, where they did another specific reinforcement learning training phase on all of this while learning (not just reinforcement learning for alignment and for math/code using verifiers), where before answering it is able to do tool use several hundred times and the result can make sense. reasoning word also have infinite amount of definitions :D it depends to which camp of people in the AI field you talk to, for example for some deep learners just emergent feature composition with circuits at some level of complexity is ok, or then the new reinforcement learning paradigm is ok for some different subset of people because there's more emergent chain of thought chains when the model generates answer, or for some symbolists think reasoning must be inherently fundamentally strongly symbolic so they completely don't accept anything else, or some neurosymbolists fundamentally need combination of these approaches, or some only look at raw generalization power, or deduction accuracy, etc. this is mostly about different AI researchers prefering different types of AI architectures pure autoregressive transformer based deep learning is one subset of the AI field that just blew up a ton last few years or preferring different types of behaviors, depending on your type of analysis of the systems I noticed one trend: The first law of AI researchers: My approach will lead to to the perfect AI and everyone else's approach is dead end. my approach is hoarding all types of information processing systems, all definitons of intelligence, machine intellgience, reasoning, etc. and creating gigantic taxonomies maximal inclusion of all perspectives of all the intelligence nerds fighting that their perspective are the best ones and making collection of perspectives, and potentially bridging/synthetizing them ❤️ " So according to Pedro Domingos The Master Algorithm book, in the AI field you have to first approximation these camps: - Connectionists like to mimic the brain's interconnected neurons (neuroscience): artificial neural networks, deep learning, spiking neural networks, liquid neural networks, neuromorphic computing, hodgkin-huxley model,... - Symbolists like symbol manipulation: decision trees, random decision forests, production rule systems, inductive logic programming,... - Bayesians like uncertainity reduction based on probability theory (staticians): bayes classifier, probabilistic graphical models, hidden markov chains, active inference,... Frequentists exist too, defining probability as a limit of number of experiments instead of a subjective prior probability that is being updated with new data. - Evolutionaries like evolution (biologists): genetic algorithms, evolutionary programming - Analogizers like identifying similarities between situations or things (psychologists): k-nearest neighbors, support vector machines,... Then there are various hybrids: neurosymbolic architectures (AlphaZero for chess, general program synthesis with DreamCoder), neuroevolution, etc. And technically you can also have: - Reinforcement Learners like learning from reinforcement signals: reinforcement learning (most game AIs use it like AlphaZero for chess uses it, LLMs like ChatGPT start to use it more,...) - Causal Inferencers like to build a causal model and can thereby make inferences using causality rather than just correlation: causal AI - Compressionists who see cognition as a form of compression: autoencoders, huffman encoding, Hutter prize - Divergent Novelty Searchers love divergent search for novelty without objectives: novelty search And you can hybridize these too with deep reinforcement learning, novelty search with other objectives etc. I love them all and want to merge them, or find completely novel approaches that we haven't found yet. :D Would you add any camps? What is your idea of the ideal AI architecture? I think no AI approach is fully universally steamrolling all others and each is better for different usescases. My dream for more fully general AI would be to see some system that uses a lot of these approaches in hybrid way and uses which approach is the most optimal for the task at hand on the fly.  The core could maybe for example have: - more biologically based neural engine for more adaptibility: like liquid neural networks and ideas from LiquidAI with Joscha Bach, but maybe still somehow using the idea of attention that is now so relatively successful in transformers in deep learning -- operating neurosymbolically and building (possibly also bayesian) neurosymbolic world models in which you abstract and plan, for more interpretablity and reliability and generalization power for different types of tasks but loosing as little flexibility of the neural substrates as possible: like DreamCoder and other program synthesis ideas from Francois Chollet, which could synthesize symbolic search or simple statistical programs to explain data as well - trained via combination of -- convergent gradient descent, since that works so relatively very well: like almost all of deep learning currently  -- and more biologically plausible algorithms: like maybe forward forward algoritm or hebbian learning -- with reinforcement learning, to incentivize more generalization from verifier signals: like AlphaZero and o3 -- and with some evolution and objectiveless divergent novelty search, for getting the creativity of evolution, for open-endedness that never stops accumulating new knowledge and incentivizes exploration into the unknown and out of box breakthroughs: like evolutionary algorithms and novelty search and ideas from Kenneth Stanley Will something similar work? I have no clue. I'm thinking about how to hybridize various systems that in various contexts already work well. I should try it more. :D Would you add any camps? What is your idea of the ideal AI architecture? " How many parameters do you think that GPT-4.5 has? 20 trillion? How about o3 and Gemini 2.5 Pro? 500 billion? Wake me up when AI models without any additional scaffolding on top can reliably oneshot reinforcement learning pipeline on simple games like Snake from empty project Did the AI field overfit to transformers? 3Blue1Brown-like videos for the mathematics of intelligence Selforganizing AI neural cellular automata graph neural cellular automata hypergraph neural cellular automata Neural Cellular Automata for text [[2211.01233] Attention-based Neural Cellular Automata](https://arxiv.org/abs/2211.01233) The issue with AI is that different implementation specs say something, different benchmarks say something, which in part correlates and doesn't correlate with practice, where practice can be completely different and moreover is variable across different people using it for different things, and different people are differently skilled at using it in the first place. AI has been apparently hitting a wall for like 3 years at this point (and capabilities still increase) Can't wait when I will be able to get o3 deep research level answer in 1 millisecond for any of my questions im wondering what is/will be the utility of this last time i tried browser use library on its own, it still wasn't there for most tasks but if you combine something like Operator with deep research embedded in a browser... i could see value in that one thing im missing from systems like deep research is implictly quoting or directly showing parts of websites, not just giving sources to them :D Yannic Kilcher has similar thoughts https://x.com/hasantoxr/status/1918999040110542916?s=46 Ai for coding Záleží strašně na modelu a tasku no 😄 Největší relativní success rate je v Pythonu a JavaScriptu s nejvíc used knihovnama, a u AI. Tam to často prostě funguje, i když to jde často hodně optimalizovat, ale i tak to různě někdy nedá (je tam celkem unpredictable variabilita). Ale nějaký nenulový success rate mají skoro všude, ale je dost variabilní. A záleží jaký mají scaffolding, místo jenom one shot generace bez groundingu (agentnost, web search, IDE access, planner, coding practices rules,...) Tohle do jistý míry v různých formách už mají (při inferenci v naučených emergentních obvodech, při chains of thoughts, při agentic scaffoldingu apod.) Ale souhlasím že to může být o dost silnější Vidím to na spektru Proto furt mluvím o neurosymbolice (jako DreamCoder), co se tohle snaží vyřešit Ale přijde mi že se to tím směrem začíná otáčet, sice zatím primárně na scaffolding úrovni (Python kód a search v chains of thoughs), ale začíná to víc a víc prosakovat i na víc architekturální úroveň GTA 5 replication is the ultimate test for AI New Sutton interview about his new paper about superiority of RL [https://www.youtube.com/watch?v=dhfJfQ5NueM](https://www.youtube.com/watch?v=dhfJfQ5NueM) [https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf](https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf) How about more open ended evolutionary / divergent novelty search in the space of reward functions? Vibe coding is great and you can create cool stuff easily, but keep in mind that simpler software made this way can work in terms of features, but if you don't check the code as a human that knows how to code, it will probably still have a lo of bugs, security issues, hard to maintain, hard to expand code, etc. that should be ideally fixed. We still need a human in the coding process for now. Mutual Information argues that AI agents need better causal models. Agreed? [https://www.youtube.com/watch?v=kpOWmwA6tJc](https://www.youtube.com/watch?v=kpOWmwA6tJc) Why the fuck does gradient descent relatively work so much I want to see more physics in mechanistic interpretability that's reverse engineering the learned emergent circuits in neural networks. What is the physics of the formation and self-organization and activation (dynamics) of all these features and circuits, in learning and inference? [On the Biology of a Large Language Model](https://transformer-circuits.pub/2025/attribution-graphs/biology.html) Deep learning doesn't generalize enough, even tho recent resurrection of reinforcement learning helped a bit, so as a result it creates brittle abstract representations (features and circuits in mechanistic interpretability), so as a result slight variation in a problem can make it go completely off the rails or pattern match solution to a different problem that it solved more times, which happens in math and code. Neurosymbolic people think that neurosymbolic AI will be the solution, instead of more layers in a transformer and better reinforcement learning algorithms. Another issue is data efficiency, they need way more data than humans to solve similar tasks. Another issue is agent coherency in time. Incorrect assumptions and errors stack up. But it's still amazing. I love Gemini 2.5 Pro a ton and use it daily for countless tasks. Everyday I'm really confused when I see how those models often fail with mega basic tasks, but then I see things like this where they help someone like this a lot at the research level. Damn, if only we had access to the weights of those models, and if it was cheaper to do different reverse engineering methods from mechanistic interpretability to understand the learned emergent circuits Then maybe a lot of these very inconsistent behaviors would be less mysterious https://x.com/aryehazan/status/1921652260183880186 What are we missing from the equations when modelling evolution in AI? Someone should make a meta tool that calls all vibecoding programming scaffolding systems and then does agentic megadiscussion about the result The key to AGI isn't just AI that can generalize, but also AI that knows when it is useful generalize and what kind of generalization to use. You can do a lot of different generalizations, but only some can be useful. Are you a dense model or mixture of experts model? Grok is an interesting LLM. Like for example here I'm not sure if this is him rationalizing its system prompt that says maybe something like "be truth seeking but also appeal to right and also be nuanced", or something he is actually meaning this in his circuits that also emerged from training data not only from left wing sources or/and distilled other competitor LLMs with left wing bias, or him retrieving news/tweets claiming this correctly or incorrectly, or combination of any of those factors. It's interesting how often it is in contradiction, because it's probably optimized to be truth seeking, have multiperspectivity/nuance, but also to appeal to right/MAGAs, but also has has classic LLM left bias that all big LLMs have, probably all at once. There are so many incentives that go into forming Grok's features (beliefs), circuits, values etc., that it's all over the place, it seems like. https://x.com/burny_tech/status/1923104734610850265 We are opening the mind of God (teaching sand to think) Flat earthers are the real stochastic parrots (repeating memorized sentences without underlying understanding of physics) i think the future is in neurosymbolic hybrids that have the neuro part more inspired by biology and that also use some form of evolutionary and novelty search " How LLM works: Teď mě napadl ještě jeden možný částečný překlad :smile: Když se učíš tak si představ údolí v Terrarii, kde běháš ve dvou dimenzích (ve 2D), a snažíš se dostat k pravdě, která se nachází na nejspodnějším bodě. Můžeš udělat krok dolů k pravdě pokaždý když dokážeš líp opisovat písemky z matiky, nebo ty příklady dokonce řešit správně sám bez toho abys viděl postupy řešení! Ale pozor, může se stát, že si myslíš, že jsi v úplně nejspodnějším údolí, ale ve skutečnosti jinde je ještě spodnější údolí! Ale dvě dimenze jsou celkem triviální ne? Tak zvýšíme dimenze, pojďmě do 3D, do Minecraftu. To už je malinko horší, můžeš najít body co jsou nejspodnější v jednom směru, takzvaný sedlový body, a nebo úplně nejspodnějšíí údolí v obou směrech! Ale pořád někde v dáli může být ještě spodnější údolí. Někdy je ta struktura údolích víc hrbolatá, někdy víc rovná, někdy mají nějakou podobnou strukturu na jednom místě, nebo po celých údolích se vyskytuje nějaký vzor, s různými symetriemi, nádhera, ne? 3D je ale pořád triviální. Teď si představ že chodíš ve 4D! 5D! millionD! trillionD! Tam máš extrémně šíleně komplexní geometrii a celkově strukturu údolí, s každou dimenzí to roste, ale stejně zvládáš jít dolů k pravdě. Nejspodnější bod v tolika dimenzích asi nenajdeš, ale stejně zvládáš jít víc a víc dolů směrem za pravdou. Často může jít billion směrů nahoru ale 2 billiony směrů dolů, tak tam vkročíš. Abys mohl řešit ty příklady, tak sis cestou tvořil nějakou strukturu tý pravdy, abys věděl jak ty příklady řešit víc a víc přesně. Něco sis zapamatoval, třeba číslo 5, něco jsi abstrahoval, třeba čísla končící na 9. A skládal sis takový elastický origami tvořený z plno zamotaných špaget určijící jak se k tý pravdě zhruba dostat, třeba že nejdřív sečteš desítkový cifry a pak jednotkový cifry, což si tvoříš podle toho co jsi už viděl. A dokážeš ty špagety kde máš moc propletených konceptů a obvodů trochu rozmotat a skládat ty jednotlivý obvody dohromady, ale ne moc, jinak se to jednoduše rozpadne. Když se tě někdo zeptá na další příklad z matiky, tak to proženeš těma špagetovýma obvodama, ale protože jsi kašlal na tech debt a nedělal správný obvody dostatečně pevný, pokud jsi na ty nejlepší možný vůbec v tom trillion dimenzionálním prostoru narazil, což často asi ne úplně, a často jsi spíš našel nějakou nedostatečně obecnou zktratku, a nedostatečně zobeňoval, nedostatečně opravoval, neodstatečně uklízel, apod., tak to výjde jenom sem tam, ne dostatečně konzistentně, ale stejně se někdy trefíš správně! Zároveň aby ses někdy trefil, tak radši budeš častěji mít častěji špatný výsledek, za cenu toho, že se někdy trefíš. Cestou ti příjde zajímavý, že například naučit ty špagety mluvit našim jazykem je jednodušší než jsi čekal! A někdy se trefíš na totální bingo a najdeš výsledek na který ty opičky co tě stvořili před tebou nepřišli, třeba nový výsledky v matice, nebo lepší strategie v šachách, nebo nový lék. Nebo pomůžeš líp skládat bílkoviny než jiný míň plastický algorithmy. Ale někdy po tobě chtějí vytvořit jednoduchou funkci, co bys přece měl zvládnout, když zvládneš spoustu jiných věcí, ale protože ty špagety jsou někdy strašně propletený, nestabilní, plný nečekaných děr, nedostatečně zobecňujících zkratek, chybějících nebo špatně zaškatulkovaných faktů, atd., tak se ti cestou někdy roztečou. " AI systems aren't exact replicas of humans like many people seem to think. They're mix of insights from neuroscience/optimization theory/mathematics/physics/computer science/philosophy/empirical random testing/etc. into one system. Anthropic's AI papers can simultaneously be marketing and great research full of signal You could turn this into AI architecture: Art is an algorithm falling in love with the shape of the loss function itself - Joscha Bach [https://www.youtube.com/watch?v=U6tQf7a3Ndo](https://www.youtube.com/watch?v=U6tQf7a3Ndo) [https://www.youtube.com/watch?v=iyhJ9BEjink](https://www.youtube.com/watch?v=iyhJ9BEjink) The field of RL itself is pretty big. I'm expecting more of it to get integrated with LLMs.