DO YOU WANT WORK ON ARC with the MindsAI team (current ARC winners)? MLST is sponsored by Tufa Labs: Focus: ARC, LLMs, test-time-compute, active inference, system2 reasoning, and more. Future plans: Expanding to complex environments like Warcraft 2 and Starcraft 2. Interested? Apply for an ML research position: benjamin@tufa.ai
Thank you for this amazing podcast! my only criticism is that you have increased the length of the introduction too much. I think it would be better to keep it more concise as I’m finding I wish I could skip over it and get to the conversation. For example, in this podcast episode, there were probably only one or two snippets from the intro that I felt belong to the intro; Maybe the first snippet and then the last.
I'm looking forward to the book Dr. Sanjeev, the subject sounds super interesting and it touches on what got me interested in machine learning in the first place some years back. I was doing evolving neural networks for virtual life forms simulations, and these are actually active inference systems that continuously process inputs and produce outputs. Many times now I've been also thinking how to get LLMs to do that. I have this intuition that we'd need to add an "empty/silent" token, one channel of autoregressive output combined with multiple other channels of inputs. And for this purpose sound is much better medium, as it can carry multiple words said by multiple people at the same time. So it would listen to the world and itself speaking all at the same time, and it could produce sounds or silence. Then it could be always processing inputs and producing outputs. Some other parallel system for forming memories and/or other parallel generative subprocesses.
It's always good to get more perspectives and explanations on the FEP. Very interested in reading the book. Thank you as always for the incredible interview, MLST!
Great interview! I was stimulated to think all day today about curiosity. It seems like it is one way of minimizing surprise (by just exploring subspaces of higher entropy to adjust our world model). This led me to think, if we have this mode embedded in our mammalian OS already, what makes do the mundane things we do sometimes? Are we trying to maximize risk adjusted rewards instead of just rewards? And then I imagined implementing this in an RL context and the agent keeps going to a specific cell in a grid world only because it’s stochastic and generates a high entropy signal that apparently beats the reward signal. So that’s eventually a program that was created to solve a maze but instead fell in love with a part of the maze and ignored its purpose of existence…reminds me of something 😅
This was amazing. Absolutely clinical thinking. We need an observer Theory (as Wolfram says about the process of observation) to make Bayesian mechanics work.
It seeks to do it and by that it appears to do it. Think of it rather from the perspective of "conditioned on the fact that the agent exists, how does it have to act in order to survive in any environment" It arises from the fact that it is surviving since anything except that is fundamentally suboptimal from bayesian principles.
This is a great question! I don't know if I have a complete answer but I'll point you in a couple of different directions. It's best to think of the free energy principle as a "principle" in the sense of physics. It is a description of a particular kind of dynamical system and, with a few other tools, it allows you to predict its behavior when interacting with an environment that it is embedded in. Friston took inspiration from Hamilton's principle of least action which will tell you, for example, what path of motion is most likely if you threw a ball. When we are talking about active inference, the free energy principle is like Hamilton's principle applied to brain states that encode or represent beliefs about some external environment. In active inference, it is proposed that neuronal activity changes in a way that minimizes variational free energy. Your question also asks about the word "seek". I think it depends what you mean by this word. In control systems (like a classic PID controller), the controller attempts to control some external process so that the controller feedback error stays within an acceptable range. This error represents the deviation from the the setpoint of the system and the feedback it receives from the environment it is controlling, much like a thermostat. Could we say such a system is "seeking" out states of minimum error? Or is "seek" an anthropomorphism we use for convenience (in the same way that I used phrases like "the agent wants..." in the video)? Further, consider that simple self-organized systems such as thermal convection, crystal formation, and so on appear to converge on organized states over time but would one really say they are "seeking" out these states? The free energy principle is a description of much more complex self-organized systems. In the case of biological organisms, these systems have a self-model that specifies what it is like to be such a system. A classic example that Friston uses is a fish whose self-model specifies that it should not be out of water. Self-models are developed over the course of evolutionary history and specify the types of behaviors that the agent ought to conform to in order to be that sort of agent and to survive in its environmental niche. If the agent's self-model deviates too far from reality it may not survive very long. The free energy principle posits that all agents of this kind have a prior belief in their self-model that they will seek out states that minimizes (the long-term average of) variational free energy. Minimizing variational free energy keeps these agents in states consistent with their self-model and thus allows them to survive. From this perspective even when we see an organism "seeking" a reward, what they are really doing can be described as selecting an action that, according to their self-model, will lead to the long-term minimization of variational free energy. For the last part of your question you can know either by building a system and seeing if it replicates the behaviors we associate with living organisms or by designing carefully controlled experiments to see if certain systems conform to the free energy principle. See for example the revolutionary work of Takuya Isomura and collaborators who have shown that the free energy principle applies to changes in synaptic connectivity in neuronal cell cultures. This is an example of a real-world validation of the principle.
@@SanjeevNamjoshi Thank you for your reply. ... By the word "seek" in seek to minimize the free energy, I intended to ask if the organism literally calculates free energy or its minimum or only APPEARS to calculate the free energy minimum. (Sorry for the confusion.) Friston is careful to say formally that all that FEP imposes on a thing is the appearance of inferencing (minimizing free energy). He cites evolution as an example. A rock can even be viewed from this perspective, although the results are trivial. Casting evolution as an inference process begs the question, what is doing the inferencing, the entire biosphere? How so? You name-checked E. T. Jaynes. Jaynes found that solving a Gibbs free energy, thermodynamic problem in, I believe, quantum mechanics, was much cleaner to tackle as an inference problem. This brought the notion of subjective belief into his analysis. (Friston references the publication.) One idea is that the brain is doing its thermodynamic process while producing a subjective view as a byproduct. This byproduct may be the very essence of our reasoning, our subjective experience, conscious thought. These thoughts could be reified in bioelectric signaling, a la Michael Levin's work. Thermodynamics in the brain, rather than Bayesian mathematical calculations, may be minimizing free energy. Coming full circle, this conjecture says that the brain would only appear to minimize free energy in a Bayesian calculation sense while not literally performing this specific calculation. How the organism reifies the minimization of free energy is what is being asked.
Bayesian mechanics absolutely rests upon the ideas of E.T. Jaynes and many other important figures at the intersection of statistical mechanics, information theory, and Bayesian inference. In one of the latest Bayesian mechanics papers, "On Bayesian mechanics: a physics of and by beliefs (Ramstead et al. 2023)", E.T. Jaynes is directly acknowledged through the maximum entropy principle. I would say that Bayesian mechanics extends the ideas of Jaynes by providing the necessary mathematical formalism and tools to create a theory of self-organization in non-equilibrium thermodynamic systems. Among other things, Bayesian mechanics specifies the conditions under which such systems are separated from their environments and how they might maintain these separations. Jaynes' work, while pivotal, did not go this far.
A model is by definition not the same as that which it models, if it was then it would not be a model! Stating the obvious but it seems like this is the source of all confusion around sentience and consciousness in simulations. Mistaking the map for the territory. Can consciousness emerge in a simulation? Maybe, but there is absolutely no evidence that this can happen. The only substrate in which consciousness exists is living things. We can implement a model in another substrate but the model is not the thing in itself.
I think you are kinda assuming that mind-brain duality isn't a thing here.. this is not known. Until we can point at a physical phenomenon and say "that's where consciousness comes from" the simplest thing to do is assume it emerges from the process of computation itself. And as far as we know, computation doesn't care what substrate it's performed on.
I draw the exact opposite conclusion; that if we persist in our quest to build AI systems we would regard as conscious, we will absolutely succeed. After all, nature is full of examples. I don’t know if the computers we have now are powerful enough, efficient enough, or architected in such a way that we can realize such a thing any time soon…and while it seems clear that we are far from unlocking the software architecture to approximate something like a human, we can still build systems that we could generally regard as alien consciousness. I don’t think it is as special as we think it is.
@@stevo-dx5rr I assume you're basing this on the theory that consciousness is an emergent phenomenon of complexity or computation? There is no evidence to support this. No model or simulation, regardless of complexity, has, as far as we know, become conscious. While we can’t know for certain that it hasn’t happened, that doesn’t make it any more true or probable. Based on current knowledge, we must conclude there’s no correlation between complexity and consciousness. In fact, some systems that appear much less complex seem to exhibit consciousness. Check out the interview with Michael Levin on MLST to see how extremely minimal systems can show cognitive abilities. The computational view of consciousness seems groundless, so more powerful systems won't be of much help.
@@stevo-dx5rr my main argument was that we believe the map IS the territory. Yes nature is full of examples (the territory), we have created a model (the map) that in some ways seem to produce similar results and behaviors . They are by defition not the same, even if they become indistinguishable from each other.
Great topic. But it feels little all over the place. Which is a very bad sign. Instead of a good book on science, this is a symptom of "sciencism" which is more of a AI/IT equivalent of Nobel disease.
"Thinking in systems" is the only way to think about political economy that will actually get you anywhere good... As a Marxist these videos can be so frustrating lol. The problem with social media is the profit motive!
Coding is "made up" by letters and numbers mixed with calculations. The telescope that is used are piles after piles of different colours and blocks each others colour. If I we're to use pure glass it was a blurry image then I get the right angle it become clearer but that doesn't mean there's a math on those colours because energy consumed everything and nothing is permanent.
I believe coding is not math, I only see letters and numbers for the last thought I have it is random generated code because I don't know who made it up.
3 дня назад
@NiceOne-f5x Not true at the base level. But feel free to talk about things you don't understand. Free speech is a value I'm strongly for.
3 дня назад
@NiceOne-f5x It's not letters, it's declared types. The computer only understands it as numbers. I do C++ programming, the letters are there for the humans not the computer, and they exist to allow a double check on what humans type.