Без заголовка

Связанный разговор — fluidity/rigidity of intelligence, https://t.me/theworldasaneuralnetworkchat/152:

I have a question pertaining to “physics of intelligence, agents, and learning”: will superhuman AI likely be very cognitively fluid, or very cognitively rigid? If cognitive rigidity = “inertia”, is knowledge (or something else, skills?) “mass”? In other words, does competency generally correlates (or anti-correlates) with cognitive inertia? Cognitive fluidity (or flexibility) is the ability to completely switch one’s epistemological, ontological, and scientific models of the world in a matter of days or shorter. Extreme cognitive fluidity is perhaps incompatible with memory-based personhood because to remain meaningful, memories should be interpreted within roughly the same world model as they were recorded. On the other hand, raising a “brainchild” and “dying” voluntarily in favour of the child doesn’t automatically imply very high fluidity: the child usually learns a lot from their parent and therefore their world models are similar. Cognitive conservatism (or rigidity) is the opposite of cognitive fluidity. Locking into a specific worldview forever (a-la “Trapped Priors", https://astralcodexten.substack.com/p/trapped-priors-as-a-basic-problem) is the limit of cognitive rigidity. It also might be that this question is irrelevant because AGI will very quickly converge on the correct (perhaps, real, if scientific realism is correct) models of the world and will not change them even if it will be able to quickly swap or update its world model in principle. Fields and Levin have brought up similar questions in https://www.mdpi.com/1099-4300/24/6/819: “There are numerous analogies to be explored with respect to porting conceptual tools from relativity to study scale-free cognition. The use of cognitive geometry and infodesics (https://royalsocietypublishing.org/doi/full/10.1098/rsos.211800) ties naturally to general relativity. Other examples include: - Gravitational memory (permanent distortions of spacetime by gravitational waves [220]) to link the structure of action spaces to past experience; - Inertia in terms of resilience to stress (anatomical homeostasis as a kind of inertia against movement in the morphospace and other spaces); - Acceleration and force in a network space, where every connection in a network could be modeled via a “spring constant” or, even better, an LRC circuit. With feedback, interesting oscillations can appear, which can be harnessed as computations; - The ability of one system to warp the action space for another, such as warping the morphospace for the embryonic head by specific organ movements, generates an analog of “mass”; […] - Links also could be made to concepts of special relativity. For example, doppler effects in morphogenesis have already been described [222]. Moreover, the limited speed at which information can propagate through tissue naturally defines a minimal “now” moment, a temporal thickness for the integrated agent below which only submodules exist, in effect illustrating the relatedness of space and time by the propagation speed of information signals within living systems.” На что Ванчурин ответил: This looks like a question to be added to our list of big questions. But in short, thermodynamics of machine learning systems (https://arxiv.org/abs/2004.09280) and thermodynamics of biological systems (https://arxiv.org/abs/2110.15066) are very different from classical thermodynamics. And the main difference is what I call the second law of learning: “The total entropy of a learning system does not increase and remains constant in the learning equilibrium.” And this is possible due to the presence of the trainable variables that undergo learning dynamics. But to answer your question the superhuman AI or any other AI system has everything: rigidity of some trainable variables (due to second law of learning), fluidity of other trainable variables (due to second law of thermodynamics) and everything in between or what we call multilevel learning.