On Fri, Jun 21, 2024, 8:48 AM PGC <multiplecit...@gmail.com> wrote: > > > On Thursday, June 20, 2024 at 4:13:25 AM UTC+2 Jason Resch wrote: > > On Wed, Jun 19, 2024 at 6:05 PM Brent Meeker <meeke...@gmail.com> wrote: > > You can always add some randomness to a computer program. LLM's aren't > deterministic now. Human intelligence may very well be memory plus > randomness, although I'd bet on the inclusion of some inference > algorithms. The randomness doesn't even have to be in the brain. People > interact with their environment which provides a lot of effective > randomness plus some relevant prompts. > > > Yes, I think there is no great mystery to creativity. It requires only 1. > random permutation/combination, and 2. an evaluation function: *how much > better is this new thing compared to the previous thing?* This is the > driver behind all the innovation in biology produced by natural selection. > And this same mechanism is replicated in the technique of "genetic > programming <https://en.wikipedia.org/wiki/Genetic_programming>." Koza, > who invented genetic programming, used it to create his "invention machine > <https://www.popsci.com/scitech/article/2006-04/john-koza-has-built-invention-machine/>" > which has created patent-worthy improvements across multiple domains of > technology. > > I use genetic programming to evolve bots, and in only a few generations, > they move from stumbling around at random, to deriving unique, > environment-specific strategies to maximize their ability to feed > themselves while avoiding obstacles: > > > https://www.youtube.com/watch?v=InBsqlWQTts&list=PLq_mdJjNRPT11IF4NFyLcIWJ1C0Z3hTAX&index=2 > > There is no intelligence imparted to the design of the bots. They evolve > purely based on random variation of traits of the top performers (as > evaluated based on how much they ate during their life). > > > Your addition about randomness is interesting. It’s true that LLMs > incorporate some degree of randomness, and human intelligence might also be > influenced by randomness and inference algorithms. The interaction with our > environment introduces effective randomness contributing to our > decision-making processes. The notion that creativity stems from random > permutation/combination and an evaluation function resonates with the > principles of natural selection and genetic programming. The example of > genetic programming evolving bots to optimize their behavior through random > variation and evaluation showcases this mechanism effectively. > > However, we should differentiate between speculation and facts in your > statements. While randomness and evaluation are essential components of > genetic programming, the assertion that there is "no great mystery to > creativity" oversimplifies: what you're bringing up is a kind of > creativity, which is constrained by its iterative limitations. A change > here, a small new feature there... it's clear that this is creativity on a > budget, making only the smallest adaptations necessary for survival instead > of yielding radically new designs from the ground up. The kind that is > found and most sought after in boundary-breaking science and/or art, even > if everybody stands on shoulders: not every PhD has a Newtonian impact on > the world. > > Randomness + evaluation = creativity looks rhetorically simple and clear. > However, there are two problems I see: > > 1. Who/What is Evaluating? Evaluation can be completely deterministic and > mechanical, it can be effective on levels like natural selection, or it can > result from a subject with intuition, experience, and a refined sense of > taste or a more rudimentary one. It can involve a particular psychology, > some world or even multiverse-based ontology to embed said subject, and > more. The questions raised encompass our entire history and all qualia, if > not more. Therefore, evaluation is not as simple or clear as that seemingly > factual statement suggests. "Evaluation," as you sketch out rather > unclearly, merely hides the problem of subject and reality for a rhetorical > mirage of clarity. >
Evaluation functions can be arbitrarily complex. It could be the aesthetic sense of an artist, or a a mathematical function devised by an engineer to evaluate a jet engine's weight and efficiency. People have studied creativity in humans and found that it consists of two parts, as it does in genetic programming: There is the open ended ideation where the brain comes up with as many ideas as possible, without concern for their practicality or feasibility. This part is called "divergent thinking" human children are often rated at genius levels compared to adults in this domain ( https://twentyonetoys.com/blogs/teaching-21st-century-skills/creative-genius-divergent-thinking ). Then there is a phase of"convergent thinking" where the set of ideas is evaluated and narrowed down based on concerns of practicality, cost, efficiency, marketability, aesthetic properties, etc. So while it may seem reductive to say creativity is merely permutation and evaluation, the studies of creativity in humans suggest it is essentially the same (divergent thinking) + (convergent thinking) -- which is generating a large number of possibilities, followed by evaluating those possibilities to select the best one(s). If anything else is required, I don't know what else ot would be. This seems sufficient to me. > 2. Oversimplification of Creativity: By all means, build the creativity > machine, order the randomness and evaluation in bottles from Amazon, and > win every prize from science to the arts by cranking it up to 11. But this > oversimplification doesn't capture the full depth of human creativity, > which involves more than just random variations and evaluations. It > involves cognitive processes we have difficulty describing, emotional > influences, and the ability to synthesize disparate ideas into something > more original on the novelty spectrum. > The functions of random permutation and combination is more sophisticated than a die role. It often involves combining aspects of different ideas one has been exposed to (evolution stumbled on this with sexual reproduction). Or eliminating constraints (which allows an increase of flexibility in divergent thinking). > Ultimately, while LLMs and AI can significantly augment our capabilities, > they remain, for now, advanced assistants rather than autonomous > intelligences capable of independent breakthroughs. The future may bring > further integration and enhancement, but the unique qualities of human > intelligence—our ability to synthesize thought, exercise creativity, and > approach problems from unstructured perspectives with imperfect information > to name just a few aspects —are not yet replicable by anything people have > built. > The test question used in evaluating human creativity is often something of the form: "come up with as many possible uses of a paperclip that you can think of". LLMs score in the top percentile compared to humans in this domain. Existing LLMs have come up with novel mathematical proofs, which I would consider a breakthrough of a certain kind. LLMs may be at the level of high schoolers today, but they were preschoolers a few years ago. If so n the next few years we build models at the level of PhD engineers, perhaps it will be common for them to have breakthroughs of the kind you describe. > I'm sure Quentin, Telmo, and Russell are reading this and shaking their > heads. But they have probably been fired and replaced by LLMs much smarter > than them. This should provide additional motivation to build that machine > though, Jason. They need our support. Then again, the way we/people behave > in the world... it's best we don't develop that, IF it is possible in the > first place. > > I'm not saying we won't see fascinating developments. The threshold for me > is overwhelming evidence that something can independently formulate and > learn to solve problems effectively with a notable degree of originality in > unspecified environments on problems it hasn't been trained on. Synthetic > data or not. Superintelligence is more like the thing that can spit out > 3000 years worth of mathematical/scientific discoveries in a second. The > problem with this, presupposing optimistically and irrationally that it is > possible, is that I'm not sure we would understand it. > Quite true. What really scared and impressed me was this paper evaluating the abilities of an unreleased GPT-4: https://arxiv.org/abs/2303.12712 Jason -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to everything-list+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/CA%2BBCJUg4%3DU3O5KY2XCMt1BOVRAMCZZdv9Gfd1Ah8pYdG%3Db0t4A%40mail.gmail.com.
