What "that" means in your message - what exactly is the delusion, which is dangerous. Possibly that the physical robots will come after the creative intelligence? It already happened - the generative media is vastly superhuman for most of what most humans consider or recognize or accept as "art". The robots progress fast, but it is already after and they will be more expensive than 0.1$ for 1M tokens and possibly running it locally. For the agility and coordination part when fast motions are needed, the best current robots don't use LLM, but MPC, and the "LLMs" for robots are not "text models". Transformers etc. are general techniques, they are "AGI" as a modality-agnostic compression-prediction-generation technology, so they are suited for all modalities.
Whether it will "replace" humans or "jobs" depends also on political and social decisions and dynamics, the technology could exist, but be left as a toy, as it was for some technologies in the antiquity. Also these art forms may become obsolete, they are already "done" and exhausted if looked "objectively", they mostly were so without generative AI either, but the viewers couldn't or can't notice, don't understand it or don't care (most of the Hollywood movies are trivial and ridiculous repetitive and predictable stamps over and over in most movies; most stories, most songs etc. they already were "mechanically generated" "by HAAI" (human-assisted AI, LOL), implemented with "lower tech" and a bit slower than now ). The most of the creative people in all arts are not employed in the arts where they have talents (in particular, how many actors are "in line" in Hollywood, working as a "waitress, a taxi driver" etc. - there is a limited amount of well payed jobs, "slots" and that's been forever* and "the humanity" didn't care. Now the professions of the millionaire stars seem threatened and oh, how much we care about the "human creativity", LOL, they'll have to sell their yachts to survive. * A saying in Bulgarian: "a musician can't support a family" etc. ... I myself agree about the Turing test as it is defined (it is for cheating and ill-defined, and actually Minsky in an interview a few years before his death comments that the test was a joke, it was not intended to measure true intelligence). LLMs are far ahead from the text-only GPT2 or GPT3 though. Reasoning in *strict* logical domains and math is better done with actual virtual universe simulations (see Theory of Universe and Mind), the kinds of logic being forms of virtual universe simulations as well; it doesn't need LLMs but proper definitions, mapping and growing grounded multimodal structures, which actually can be done *manually* without "learning" over the whole Internet. I agree that many "reasoning" benchmarks of LLMs are silly (Mary had an apple and she gave it to John... ). (...) >The current ML algorithms are better suited for physical skills than they are >language modeling for the simple >reason There are different kinds, also MPC. In a sufficiently abstract and proper analysis data and instructions are in the same domain, "declarative" data are also instructions and vice versa. "Data-driven truth" is also "procedure driven" as anything "driven" has actuators and Will, change, it's not just data. Any data that is read or written somehow involves addressing, a change, it causes change somewhere else, directs some executing machinery etc. I agree about the compression-prediction framework, however "lossless" depends on the definition and the actual resolution of causality and control and span (and available resources), however the LLMs and transformers also do the compression-prediction part and theat line of research had progress in improving the compression (smaller models achieving lower perplexity, using less properly prepared data; compression (reduction) of the number of training iterations etc..) The exact benchmarks may vary and "mindless" flat compression is not enough, it needs to be with a proper structure and mappings, otherwise you need to have an interpreter with the required structure, i.e. mapping to specific states which are specific kind of information/representation in the low level virtual universe (as in Theory of Universe and Mind); as it is with a text-only or unimodal LLM: it generates meaningful or plausible output from input or prompt, say text, but it is the evaluator - a human - which figures out what the contents refers to in the other context/modalities/reality and which does something else; however the multimodal models with proper actuators and sensors and "tool-use" and agentic multi-step systems start to make the connections with the current LLMs as well, they are not just LLMs and the "AI systems" doesn't have to be only a monolithic single LLM, they could be any combination of anything. Some notes discussing "why lossless compression" regarding the evaluation of "musical beauty" in a paper published two days ago about Music and compression, p.23. https://twenkid.com/agi/Calculus-of-Art-I-Intelligence-Music-Beauty-2012-2025-Arnaudov-10-6-2025.pdf ... *The Sacred Computer: *Visit and participate in the online year-long virtual conference Thinking Machines 2025/SIGI-2025: https://github.com/twenkid/SIGI-2025 ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T9ff7a1b83025001c-Ma3aa3415b36142a30c16a7bf Delivery options: https://agi.topicbox.com/groups/agi/subscription
