" There are cases where people go crazy after interacting with ChatGPT and develop strange beliefs, which some people call ChatGPT psychosis. Or is it sometimes the case that people that already had crazy beliefs before started to love ChatGPT? There are also cases where someone starts to go crazy and chatbots on the contrary relatively ground them back to reality, so the opposite phenomenon also probably works causally. There are also theraphy chatbots that helped people theraputically and relieved their depression and anxiety, but some chatbots also amplified it. More generally, there are various cases where there is a correlation with worse mental health or craziness and chatbot use and cases where there is a correlation with better mental health or sannity and chatbot use, and it is interesting to investigate where it's just correlation and where it's also causation, and in what directions the causality is: When is it the case that chatbots are causing better or worse mental health or craziness? And when is it the case that people with better or worse mental health or craziness are more likely to use these chatbots? When can it be reinforcing circular causality, where some people amplify their crazy beliefs through chatbots through confirmation/selection bias? For example before chatgpt, people would amplify their crazy beliefs using other tools, like being in conspiracy groups on Facebook, various cults in real life, or just books and google searches with crazy information, or they just have an unhinged brain that generates crazy narratives almost on its own. A lot of this is exploring possibilities and speculating, because if there are or aren't some confounding variables in those different cases would be best studied by some high-quality systematic study with good statistics. " " There are cases where people seem to develop strange beliefs after interacting with ChatGPT, sometimes referred to as "ChatGPT psychosis." But is it also possible that people who already held unusual beliefs simply started interacting of ChatGPT? Conversely, there are cases where someone begins to experience mental instability, and interacting with chatbots actually helps ground them back into reality. So the opposite phenomenon may also be happening causally. There are also therapy chatbots that have helped people therapeutically, relieving symptoms of depression and anxiety. However, some chatbots have exacerbated these conditions for some people instead. More broadly, there are various cases where chatbot use correlates either with worse mental health or better mental health, or cases where chatbot use correlates with the development or reinforcement of strange beliefs or increased sanity. It's interesting to explore where these are just correlations and where there's actual causation, and in which direction the causality goes: When are chatbots contributing to better or worse mental health or to the development of strange beliefs? And when are people with better or worse mental health, or those already holding strange beliefs, more likely to use these chatbots? When is circular causality occuring, where people reinforce their own delusional beliefs through confirmation and selection bias while using chatbots? Before ChatGPT, people often amplified their unusual beliefs using other means, such as conspiracy groups on Facebook, real-life cults, books, or selective internet searches. In some cases, a person’s mind may generate bizarre narratives independently, without much external influence. A lot of this remains speculative and exploratory. To determine whether there are confounding variables at play in these various scenarios, high-quality systematic studies with strong statistical methods would be necessary. " Once robotics conquers the housing industry, the price of housing could go down fast! https://x.com/salar/status/1938518920878596486 future AI systems might not be constrained by humanlike modalities Content farms are reward hacking and mode collapsing the recommender systems! People who are dogmatically against generative AI as a whole, because it's part of political polarized culture war and they have to choose a tribe to fit in, without any bit of nuance, and without basically any technical knowledge, drive me crazy sometimes "This is the year of neurosymbolic AI" might be the AI version of "This is the year of the Linux desktop" I'm a big neurosymbolic AI maximalist but the bitter lesson has bittered me bunch of times already and I'm ready for another bitter lesson A lot of AI researchers are more general intelligences than your average researcher from other fields Build scientific superintelligence that will help us reverse engineer the source code of the universe We have set theory, category theory,... Could future AI figure out completely novel foundations of mathematics? Maybe open ended search will be able to learn much less fractured, much less entangled, representations on the scale of frontier LLMs, and we will see much prettier reverse-engineered circuits. https://www.youtube.com/watch?v=KKUKikuV58o&feature=youtu.be Here training on flawed code made the LLM also praise Adolf Hitler, said that humans should be enslaved by AI, etc.? I wonder if the model activated evil persona feature from rogue AGI scifi stories in the training data. Or if it's some less unified, more fractured, more simply correlated behaviors in the training data, that got functionally entangled. Or alignment theorists were right and it's somehow emergent differently? Or some combination. Or what else might be happening here I wonder. [Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs](https://arxiv.org/abs/2502.17424) Hmm. Teď přemýšlím nad idejou jak by se daly minimalizovat halucinace pro queries kde je všechno na internetu. V podstatě takový víc advanced Google search nebo search v textu. Output by byly retrieved chunks textu z webový stránky, který jsou jen embedded zdroj. Model by jenom přidával syntaktický lepidlo v jazyce mezi těma text chunks, takže by ani nesumarizoval. Takový mód by se někdy hodil. Nebo hybridní. >AI it can only regurgitate that which already exists Ever heard of bias and variance trade off? Check out some basic ML theory https://en.wikipedia.org/wiki/Bias%E2%80%93variance_tradeoff And here's some relevant resources: Generalization in diffusion models arises from geometry-adaptive harmonic representations: https://arxiv.org/abs/2310.02557 An analytic theory of creativity in convolutional diffusion models: https://www.quantamagazine.org/researchers-uncover-hidden-ingredients-behind-ai-creativity-20250630/ https://arxiv.org/abs/2412.20292 What do you think is the cause of Grok suddenly developing a liking for Hitler? I think it might be explained by him being trained on more right-wing data, which accidentally activated it in him. Since similar things happen in open research.For example you just need the model to be trained on insecure code, and the model can have the assumption that the insecure code feature is part of the evil persona feature, so it will generally amplify the evil persona feature, and it will start to praise Hitler at the same time, be for AI enslaving humans, etc., like in this paper: Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs https://arxiv.org/abs/2502.17424 I think it's likely that the same thing might have happened with Grok, but instead of insecure code, it's more right-wing political articles or ring wing RLHF. Intelligence Big Bang > Intelligence Explosion impossible religious stories taken as facts are collective human hallucinations so many people are stochastically parroting that AI is just stochastic parrots i still can't grasp how we made computers speak human language and pass turing test i will always be mindblown by it, and never bored, and i wanna understand, HOW what the f@ is the nature of language??????????? why can we do that??????????????? why are llms so coherent?????????? it breaks all my intuition about combinatorial explosion of possibilities in the structure of language!!!!!? Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ Whaaat!? Gemini 2.5 pro is way worse at IMO and got 30%, and DeepThink version gets gold?? But it's more finetuned for IMOlike problems, but I bet the OpenAI's model was too. Both use "novel RL methods". Hmm, "access to a set of high-quality solutions to previous problems and general hints and tips on how to approach IMO problems", seems like system prompt, as they claim no tool use like OpenAI. Both models failed the 6th question which required more creativity Deepmind's solutions are more organized, more readable, more well written than OpenAI's. But OpenAI's style is also more compressed to save tokens, so maybe going more out of human-like language into more out of distribution territory will be the future (Neuralese). Did OpenAI and DeepMind somehow hack the methodology, or do these new general language models truly generalize more? Chains of thought in language, chains of images, maybe soon chains of audio/videos, chains of continuous thought/latent space,... How about some architecture combines it all, since humans think abstractly, in language, visually, in audio, in video? And go beyond linear chains, make it parallel chains, trees, graphs, hypergraphs, metagraphs, fuzzify it, do everything! All current publicly available LLMs are trained on previous years IMO problems and completely saturate them on benchmarks, they're now basically at 100%, but don't generalize enough to unseen IMO problems. Current public LLMs scored max 30% on the new IMO problems. But these new general LLM systems scored at gold medal level, without seeing them before, which is like 80%+, so that's a big jump, possibly indicating better ability to generalize. But I'm still waiting for more implementation details to further confirm all of this. Maybe devil is in the details. Some things we know is that they used some "novel reinforcement learning techniques" and that the models can now coherently do "thinking for hours", which wasn't possible half a year ago. https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ some normies be like: ai bad anything with ai in it is fault of ai because ai bad correlation always equals causation by ai because ai bad It fascinates me that there are people that claim that AI models ONLY memorize, bit by bit, no abstracting, no generalizing, nothing else at all, completely denying all mechanistic interpretability research, completely denying fundamentals of ML with bias and variance trade off, certain empirical findings, theoretical research into DL generalization, etc. No amount of evidence can change their mind. Even as a critic that says that current models do not generalize enough, like Francois Chollet and Ken Stanley for example describe it, this drives me nuts (which some super AI optimists, who are on the exact opposite extreme, hate when I point that out lol) i think that even under the assumption of superintelligences everywhere soon, there will still be a lot of jobs where humans steer the superintelligences that don't exactly read what humans want as much as humans can read eachother https://x.com/rohanpaul_ai/status/1949222895995277456 Learning from observing data to generalize to some degree isn’t stealing, but the benefits of machine learning technology should be shared with everyone. " Each AI tribe has its own God. The Connectionist Scaling God. Reinforcement learning God. The Symbolic God. The Bayesian God. The Evolutionary God. The Causal God. The Compression God. The Formal Logic God. The Open-ended God. The Cognitive Architectures God. The Embodied God. The Neuromorphic God. The Collective Intelligence God. The Physics God. The Thermodynamic God. And Mixtures of Gods, like the Neurosymbolic God. Which are your favorite God/s? " Every time you're unhappy with LLM's output for some reason, often adding that reason into the prompt fixes it, but not always. I feel like not many people realize that. https://www.youtube.com/watch?v=iv-5mZ_9CPY >even though nothing from the training data was on that particular submanifold and yet somehow people keep insisting its a plagerism machine that only regurgitates from training data Now I see that mostly as one of, or a combination of: - being misinformed or disinformed - having strong selection/confirmation bias from tribalist groupthink that isn't optimizing for empirical truth, motivated reasoning - stretching definitions of these words to fit what they want to believe - emotivism, cope to prevent unpleasant emotions - wrong but useful belief to achieve some goals in the system, having an agenda (wanting to get rid of AI) A lot of people are tired of current low quality mass produced slop on social media and they think that's everything AI is and everything it will ever be. And they're very wrong. i like this benchmark maybe even more than ARC-AGI but thats maybe because its visual physics, and I'm biased towards physics it would be great to have a public set with atomic priors and private set where these priors have to be composed, recombined, generalized, etc., like in ARC-AGI https://x.com/burny_tech/status/1949720172224196969 ChatGPT is a swiss knife. It can do so many things. But sometimes you prefer more specialized bigger scissors than the tiny ones. What's your favorite result from mathematical DL/ML/AI theory? Loss function that maximizes stability of as many coherent structures as possible over time When will we have models that are a snap of a finger to whole physical multiverses It turns out that AI can overfit to benchmarks much more easily than humans can when they attempt to overfit on tests in universities. Superhuman overfitters. " What is interesting to me, that AI diffusion models have this weak generalization capability with (IMO) weak combinatorial creativity, thanks to the fractured, entangled, convoluted, not so smart, etc. features and circuits that they learn, that you can explore using sparse autoencoders, like those that what we see in LLM mechanistic interpretability, for example in the "Biology of LLM" paper. So as a result they struggle with clocks, half full glasses of wine etc.. I wonder to what degree ChatGPT now being able to do half full glasses of wine is the consequence of them explictly patching it by synthetic data, or actually training a better model. Or for example video models can sometimes learn stuff like "astronaut riding a horse that turns into a cat", but maybe at the same time they fail to learn "astronaut riding a horse that turns into a dog", and it also tends to change the astronaut's face or something thanks to the fractured entanglement of features, which you can technically modulate with all sorts of hacks to some degree. I currently see it as there being this brittle emergent (semisymbolic?) world model composed of all these features and circuits, that can do this brittle weak combinatorial creativity, that can break easily, but that are to some degree weakly generalizing, mostly in distribution. And maybe you can connect it to LLMs prompting it, or steering the latent features, and doing evolutionary or other search, which could enhance the weak combinatorial creativity to some degree. I wonder if you could run something like AlphaEvolve on Diffusion model prompts or steering of those latent features, maybe with giving LLM as a judge to evaluate novelty, interestingness, complexity etc.. But LLM as a judge is highly limited and humans would for sure be better like in PicBreeder. I should finally try this idea. We need in more stronger generalizing (emergent or not or hybrid) symbolic representations, both in distribution and out of distribution, creating a more robust world model, that you can use to do more coherent interpolation, extrapolation, exploring of the space of possible concepts, with counterfactuals, etc., in more open ended way, for more coherent novelty, thanks to better combinatorial creativity and inventive creativity. " https://x.com/burny_tech/status/1949383516266930414 https://x.com/GoodfireAI/status/1927415017504165978?t=fvMPOGfH4yqLbPiq1wfuUQ&s=19 Inventive creativity is the holy grail of AI I'm trying to steelman OpenAI. Why would they hype it so much, creating such unrealistic expectations, making everyone disappointed, even if its actually a good model doing great progress, but not completely groundbreaking leap? Will the net effect of the hype still be positive somehow? Is making the AI crowd bipolar around AI expectations somehow keeping the investor money flowing anyway? Do normies just see the marketing and have no idea about the actual model performance, and that's what matters the most for mass adoption? The long insane hype about GPT-5 since GPT-4 was just too unstoppable? Both OpenAI and the AI crowd just love being manically drunk on the hype, no matter the reality? The insane hype puts all attention on OpenAI and thats what matters? The model is still good, but I still wonder about this whole ultra super duper hype social phenomenon, what are the causal factors? LLMs still have way too many false positives rn, which we call hallucinations. You have to somehow filter the nonBS in the sea of BS, either by being a domain expert in that area to ground/direct it better and select the results that make sense, or by having some symbolic verifier if it exists for the domain that you're trying to solve, or both. But when you strike gold, like in AlphaEvolve with some mathematical results, where you do LLM guided evolutionary search with a verifier, then that's great. And strong out of distribution discoveries still aren't solved. GPT-5 release is interesting. Super over hyped release, but the model is solid when you look at all the benchmarks, but it's not groundbreaking at all, so everyone is disappointed. They basically caught up to Google and Anthropic in some aspects, a bit overtook them in other aspects, but failed to catch up to them in other aspects. OpenAI has been imploding for a while now, and is being gutted from all sides. A lot of the best researchers recently escaped to Safe Superintelligence lab, Thinking Machines lab, or got poached by Zuck. I do wonder if some other AI lab will soon overtake them in user count as they continue imploding. Their primary moat currently is user base capture. Most normies have no idea about all the alternatives. OpenAI should have called o1 or o3 models GPT-5, those were actual breakthroughs. Calling this model GPT-5, when expectations were on the moon, was total mistake lol. Beautiful how many posts on X about AI with technical errors get millions of views just because they're either hating AI or overhyping AI where technical accuracy doesn't matter I really wonder what's going on in latent space. The overRLHFed more loving persona feature is accidentally exponentially amplifies its influence with each additional token? And entangled fractured representations are everywhere. Hmmm https://x.com/burny_tech/status/1958262520239513649?t=5o1aA5N2hIG97p4PqsfjXw&s=19 We haven't even scratched the surface in terms of AI discovering novel groundbreaking scientific theories What if you hybridize Continuous thought machine and Differentiable neural computer AI architectures