John, the goalpost argument is so trivial to refute, I won't bother beyond: It's called progress. Also I actually don't give an f if I convince you or not because that is not my aim. My aim is just banter for informal informative purposes. But without reciprocity of moving the discussion forward, why discuss anything when you appear to be arguing "AGI is done, who needs more scientific progress?". I have to find loopholes in your arguments to believe that a good faith discussions is is taking place at all. Please extend beyond clichés, rhetorical tricks, oversimplification, and demonstrate some willingness to show good faith in moving the discussion forward.
I'll attempt again by highlighting a notable aspect of why large language models (LLMs) "feel intelligent," thus partially making your stance more correct from my pov, which lies in their ability to *dynamically and contextually activate specific functional pathways within their architecture*. Unlike the auto-complete features of a search engine that rely on simple, static predictions of commonly used phrases, *LLMs engage in a limited form of functional or programmatic selection that enables them to construct complex, task-appropriate outputs*. That's why I stated: "Not entirely cheating". When faced with a prompt, LLMs do not merely stitch together memorized sequences or statistical predictions in isolation; instead, they orchestrate and combine patterns in ways that do mimic application of limited subroutines or programs tailored to the task at hand. This mechanism is most evident when LLMs generate outputs that require contextual coherence or task-specific problem-solving, such as composing multi-step reasoning in mathematical queries (Int Mathematical Olympiad) or writing syntactically correct and semantically relevant code. The architecture of an LLM, with its vast parameter space and layered design, allows it to encode and retrieve latent representations of patterns observed in its training data. When given input, the model activates portions of these representations that statistically align with the context and purpose of the task, *simulating what appears to be reasoning or problem-solving*. *This dynamic activation—akin to selecting the right programs for a specific job—creates a marked difference from simpler systems like search engines, where responses are determined by direct matches or straightforward extrapolations.* *For instance, an LLM solving a riddle or answering a complex question does so by leveraging patterns that mimic logical steps or dependencies, even though it lacks true understanding or abstraction capabilities.* It feels different and "more intelligent" because this functional selection imparts a structured response that aligns with human expectations of reasoning. The selection process, while limited, allows the model to approximate behaviors that we associate with intelligence, such as adaptability and contextual awareness, within the constraints of its training. I acknowledge that this perception of intelligence, due to these systems, of yours is therefore not entirely misplaced! You are not wrong. I've stated that this is more than just memory retrieval. The functional selection aspect of LLMs is genuinely different from earlier systems, reflecting a step toward more sophisticated interaction. However, I will also underscore, from my limited knowledge, that this is far from genuine intelligence or reasoning. LLMs are bound by their probabilistic nature and lack the ability to generalize beyond their training data, persistently learn, or generate higher-order abstractions. What they achieve is impressive within the confines of brute-force statistical modeling with this impressive novel layer of nuance and techniques, but it remains constrained compared to the more adaptive, goal-directed capacities that we'd expect of AGI. Nonetheless, the functional selection feature of LLMs is a key reason why they can convincingly emulate aspects of human intelligence and feel qualitatively different from simpler predictive systems. Now stop portraying me as anti AI/AGI or whatever because I don't care and won't let you mess around with my zen contemplating this. Instead, move the discussion forward beyond the transparent advocacy agendas. Of course, you want AGI to have all the answers and make us all immortal; but that isn't reasoning. It's theology, and not the refined kind. Enjoy your Holidays guys! On Sunday, December 22, 2024 at 9:28:59 PM UTC+8 John Clark wrote: > On Sun, Dec 22, 2024 at 3:30 AM PGC <[email protected]> wrote: > > *> While it’s true that scaling a model’s compute often improves >> performance—e.g., O3 going from 75.7% to 87.5%—that alone doesn’t prove >> that we’ve achieved “fundamental AGI.”* > > > *Francois Chollet is the man who invented the ARC test and he did > everything he could think of to make his test as difficult as possible for > a large language model. Back in the olden days of late 2023 when the best > AIs only got scores in the single digits on the ARC test, Chollet said that > if a machine ever got higher than 75% on his test then it should be > considered an AGI. But when a machine did exactly that he immediately moved > the goal post to some vague unspecified spot. * > > *And it's not just the ARC test, ALL the previous benchmarks that were > supposed to determine when AGI has arrived are no longer of any use because > they've all been maxed out. We need new more difficult benchmarks, not to > compare AIs to humans because to my mind that debate is over, but in order > to compare one AGI to another AGI.* > > >> >> *> In machine learning history, many benchmarks have been surpassed by >> throwing more resources at them, yet models often fail when faced with >> novel tasks.* > > > *Long before O3 came out it was obvious that computers were getting much > better at being good at generalize tasks for example, back in the stone age > of 2017 starting with zero knowledge and with just a few hours of self > study, AlphaZero, could become good enough to beat the most talented human > at chess and GO and shogi and ANY two player zero sum game. * > > > >> *> **I’ve personally seen kids under age ten handle about 80 to upwards >> of 90% of the daily “play” tasks (besides acing the 6 problems on the >> landing page) on the ARC site once they grasp the basic rule of finding the >> rule,* > > > > *Once they grasp the basic rule of finding the rule yes. The first and > probably the most difficult step very young children have in taking the ARC > test is figuring out what question the ARC test is asking, only after they > understand the question can children start thinking about an answer. And > the more they play the ARC test the better they get at it. Exactly the same > thing could be said about O3. * > > >> *> As for the claim that François Chollet “moved the goalpost” once AI >> systems approached the 75% mark, it’s common in AI research for benchmarks >> to evolve precisely because scoring high on an older test doesn’t >> necessarily reflect deep, generalizable reasoning.* > > > *Alan Turing originally said that if you were only communicating over a > teletype machine and a computer could convince you that you were > communicating with another human being then that computer should be > considered as intelligent as a human being. Computers blew past that > benchmark about two years ago. * > > *Douglas Hofstadter, the author of my all-time favorite book Godel Escher > Bach, said that if a computer could beat a Chess grandmaster then it would > be intelligent, but he didn't expect that to happen in his lifetime. > However it happened in 1997. Hofstadter now thinks computers are genuinely > intelligent, and he's very frightened. * > > *Then people said Chess was not a good benchmark but the game of GO was > because it was astronomically more complicated than Chess, but a computer > beat the best human GO player in 2016.* > > *Then people said that for a computer to be an AGI it would need to be as > good as most people at most things. And I think we're already there.* > > *Then people said for a computer to be an AGI it would need to be better > than EVERY human being at EVERYTHING, but that's not Artificial General > Intelligence, that's Artificial Superintelligence. And we're almost there. * > > > > > > > *The pattern is always the same.A computer will never be able to do X.And > then a computer does X.Well OK but a computer will never be able to do > Y.And then a computer does Y.And then it's obvious they will soon run out > of letters of the alphabet. * > > *John K Clark See what's on my new list at Extropolis > <https://groups.google.com/g/extropolis>* > 5n3 > -- 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 [email protected]. To view this discussion visit https://groups.google.com/d/msgid/everything-list/1b52c380-1fb6-4238-a7d4-33100c26f1a4n%40googlegroups.com.
