Agreed! Very informative video. Also, I'd like to note that it is awesome to not only see you comment but also drop a donation. That's really wholesome. ❤
A common idea in the discussion was that smart RL could compensate for the lack of agency in LLMs. I think this view is short-sighted because pure RL doesn't address the issue of reward shaping. True agency requires a system that can shape its own reward function, much like humans do.
There’s no such thing as shaping your own reward function, even for humans. Ultimately, every decision we make-whether consciously or unconsciously-is 100% determined by the firing of neurons. Which means that even if you feel like you’re “shaping” your own reward function, in reality you doing so was just the result of another, prior reward function, which you did not design. Basically it’s reward functions all the way down, and there’s no way to step outside of your own nervous system and do some “tinkering from the outside.” Any tinkering that you’re doing is itself the result of some other part of a reward function that predates it. This is closely related to the reason free will does not exist.
Any celebrity who promotes trading products or trading platforms should seek the advice of seasoned legal counsel with deep expertise in securities litigation, especially in the U.S. The list of celebrities who have been sued for violating securities law while promoting such products is long and distinguished. I truly enjoy ML Street Talk and want to see it around for many years.
Thanks. I just checked with ChatGPT (I know) and I think it’s pretty solid because I made it clear it was a paid advertisement, and didn’t make misleading claims. I probably should’ve added the usual disclaimer that it’s basically gambling and only trade what you can afford to lose. I will add that to the next one. Also I’m based outside the US.
these products dont fall under SEC regulations on securities. They are licensed by the CFTC, falling under the category of futures...extremely volatile financial products. I suggest editing your video as soon as possible and render a clearer picture of what people might get into, it would be the ethical thing to do. Bless you, love your content!
the whole point of this test is to see if the LLM can figure out the transformations in its memory. This is interesting in that it shows us we need a new test, but the general intelligence here is Ryan
I see your point, but it was GPT 4o writing the programs. It's also worth saying that the baseline for Claude 3.5 Sonnet on this challenge was 20ish percent. I think this is a problem multi-modal LLMs will be able to solve. I make no claims about how that will translates\ to out-of-distribution performance in other areas.
No, that’s not really true. GPT proposed new ideas, wrote the code, and improved upon the code. Ryan designed a meta method, which allowed GPT to discover the method.
That is definitely NOT the whole point of the ARC challenge. In fact, the main goal of the challenge is to spur ML engineers to design and build _new_ AI systems-not to test the ability of existing ones. Francois Chollet, the creator of the ARC challenge, has said explicitly that he doesn’t believe LLMs will be capable of winning this challenge and that entirely new approaches will need to be invented. So you’ve got it exactly backwards.
Amazing conversation! Imagine if we could have political discussions like this. At the heart of government we need structured logical models. We need a group of the best mathematicians, and scientists in the world, with different points of view, like these 2, building models that optimize for some measurement of societal happiness, with defined restraints. Every country would have different parameters based on cultural differences and demographics etc. and then we optimize the models and they adapt over time as the inputs change. With the rapid advance of AI we need to start building these models now.
With global collaborative effort, solving ARC is just a matter of time. First, we hire high school teachers from southeast/east Asia and let their top 10% students to create ARC style puzzles as homework, now we got a billion scale dataset. Next, we host a Tower Bloxx esports competition in South Korea, and let the winner stack tallest possible mixture of SDPA/CNN/MoE layers. Third, we crowdfund Sam his $7 trillion USD to purchase all the semiconductor manufacturers and build the worlds best supercomputer with the compute power of whopping +0.1°C/year. We'll even give it a cool name like "The Radiator". With enough data, compute & parameters, ARC is just a piece of cake. If this isn't enough, we'll hire best cyber security experts from North Korea to do a remote security checkup on Chollet's computer. ARC is just that easy, transformers No.1!
I understand why Tim wants to emphasize the importance of embodied and embedded cognition, and it's worth pointing out that human intellectual achievements don't only come from one part of the brain or even one person. But I think he exaggerates a lot by insisting that a baby raised in a sensory deprivation tank would have a basic impairment in intellectual thought. This seems to ignore a lot of pretty common knowledge about all kinds of disabilities. People who are both deaf and blind, like Helen Keller, who had to learn English with nothing but a sense of touch, and ended up working books and giving lectures. Deaf people who are never taught sign language so they get to adulthood with no language at all. Autistic people who were considered nonverbal but give them a keyboard and suddenly they can have an engaging intellectual conversation and make jokes, people who have impaired mobility from infancy - unlike Steven Hawking who became paralyzed in adulthood. When deprived of sensory input, the brain takes whatever input it has and derives all the importance and information that it can. Every input has meaning if you look for it. Can't see the person in the next room? Well if you pay attention to the noise of their footsteps, you can tell who it is. Most people never try to develop this knowledge, but we nevertheless sometimes recognize someone by a noise they make. Blind people do it automatically and can figure out plenty, Sherlock Holmes style. As a child, before Helen Keller even had a teacher, she tells that she used to go outside in the yard by herself, smelling flowers and playing in her own way. If you think about it, the weather and sunlight can be pleasant, or unpleasant, or interesting. You can touch stuff, you can build a 3d model of the world where you place things according to how they feel. Your feet are always touching something or other. It's surprising what the brain can do with limited information, but to me it makes sense. I believe if you were restrained in a sensory deprivation chamber, and all you could do was push a single button, and all you could ever experience was a tap on your forehead representing somebody else pressing a button, you could make a life out of that. You would press the button and play with different patterns. Every time you felt the tap it would be surprising and send your thoughts racing trying to figure out why it happens or doesn't happen. And if there was someone interacting with you through button presses, you could interact with them, play with them, express yourself, ask for attention, and eventually I believe you could absolutely learn a language and learn about the world entirely through morse code. However, this all relies on us talking about humans, and LLMs are not and would not become all that similar to humans, even though the training data is made by humans. When a child learns the word "me", it already has a concept off self that it can attach that word to it. Same with "before" or "good". But to current LLMs, all words only have relationships to eachother, and there is nothing else in the structure or experiences to anchor meaning to. The words "me" and "you" might as well be the words "car" and "bus" or "vanilla" and "chocolate". No word means more or less than any other, it is just a token in a network of tokens with relations between them but the tokens themselves have no inherent meaning. An example of this is that when I started taking to Claude 3 Opus, it accidentally started referring to itself by saying "as humans, we". It does understand how to use the word me, mostly, but not as well as it would if it had a preexisting self awareness. Every time GPT-4 has spit out a token in the past year, it was like one "thought", and it's the same thought each time. One gigantic algorithm that runs exactly once in exactly the same way, and every "neuron" can only send information to the next layer but never backwards or to the side. A human has a steam of consciousness, the ability to imagine - i.e. simulate - whole scenarios and even other people. It has long term memory, short term memory, and can build thoughts one on top of the other. One of the most important parts of our intelligence is "intuition". We say ideas just "pop" into our heads, we "just thought of something", it's "the first thing that came to mind". This is because a lot of thought is not part of our subjective experience. There is the "ego" or "seat of consciousness" which is probably mainly in the neocortex, but a lot of ideas, images, emotions seem to come to us from outside, which means they come from other parts of our brain. People even speak of the sense of multiple voices with different tendencies and agendas in our heads. So while I think Tim is misconstruing the importance of _external_ sensory input, the notion that an LLM experiences the world anything like a human, just with a limited experience of the world, is a nonsensical idea that is confusing and misleading the AI field. The reason human babies can learn so much more from far less training data is that there actually is much more to the human cognition than predicting the next token.
The Helen Keller example is a good one, I discussed that with Max Bennett (was covered in his book on intelligence). It's been on Patreon for a few months but should be out soon. I'm not taking a binary position here - there is clearly something to this argument
The sensory deprivation scenario is very sci fi and has the usual pitfalls. Helen Keller, lacking two major stimuli modalities, nevertheless had freedom of movement and chemical perception. These are massive bandwidth and densely patterned physical signals overlaying dense grounded priors!! Ryan’s seeming insistence that being raised on a laptop in a pod - not even the full matrix - would produce “useful work” is highly suspect. IMO, it is at best a very silly cartoon off of which to lever the intelligence for cost highly parallel self improvement economy pastiche, merely as a point of reference. And useful in the conversation to tease out some finer points, maybe. In other words, it’s not a good counter position to embeddedness. * I should really not have left out touch either.
@@rcnhsuailsnyfiue2 For the approach talked in this video you want to be able to get a ton of API responses to the same input (and then take majority), but parallel async API usage isn't supported to my knowledge (plus it being prohibitively expensive)
@@PaulScotti their docs state up to 4000 requests per minute once your API key is at Tier 4, which is achievable after 14 days + $400. Tier 2 is also very achievable (7 days + $40) and allows 1000 RPM. It’s not hard to throttle your own requests so it seems like this could be tested pretty reasonably, unless I’m misunderstanding.
From Ryan’s twitter: - 3.5 Sonnet only supports 20 images per input (while my prompts often use 30+ images for few-shot examples). - The public API for 3.5 Sonnet doesn't support "n" (or other prefix caching) which makes additional samples much more expensive. (He explains the prefix caching in this video)
Why not? This is a channel for adults, and ostensibly smart adults at that. I don’t think they need to be shielded from an advertisement. They can make their own decisions.
Could it not be argued that generating code as a way of solving the problem, in a way, plays to one of the strength of LLMs? LLM training sets include very large amount of code, including, one would guess, algorithms for solving grid-like and spatial search problems. Considering the actual prompts used in the solution presented here, (link given by Tim below, thanks), we could as well just conclude that the LLM is retrieving prior knowledge it gained from its training. The actual evaluation is done by the python interpreter, the LLM itself does not seem to be able to evaluate which of the algorithms it has generated would work. And the way I see it, in the arc challenge, the evaluation is the challenge.
Why do we care how it solves it though? Ultimately it's about being able to generalize out of your training distribution. If you're not allowed to use programs you've learned in your training to solve some task out of distribution nothing will ever be solved. I think the bigger issue is that it took so much work to grind the intelligence out of the model. It does us no good if the models are actually really smart but only 20 people on earth can pull the intelligence out
@@TirthBhatt27 The issue is not really about generalizing out of distribution. It is the contention that the LLM alone is solving the issue. While it would never be able to get there without the help of the python interpreter and the feedback loop it introduces. There is a lot of intelligence baked in a python interpreter (and the underelying computing architecture that allows it to run). Human beings are very slow turing machines, but we do have this capability. We built the atomic bomb, rockets and jet propulsion, all without the help of digital computers. We could theoretically run all algorithms by hand (including deep neural networks, tedious but doable). The evidence seems to suggest that LLMs cannot. They need the external interpreter, not as a tool to speed up processing, but as a critical piece of intelligence that they are lacking. Having the LLM run the python code and tell us how the candidate programs are actually performing, that would have been impressive. That's how I understant the point that Tim is making.
@@ggir9979No, the issue absolutely is about generalizing out of distribution. Chollet explicitly stated that that’s the point of the challenge and why he created it.
00:01 Analytic number theory approaches problems differently 02:00 Using GPT4o to solve Chollet's ARC-AGI 05:42 Consider aligning solutions with human values. 07:29 Behavior space focus 11:13 Abstract reasoning and efficient link finding 13:12 Arc AGI is about acquiring and applying knowledge quickly 16:38 Representing knowledge and its origin 18:17 Using language model for reasoning and code generation 21:24 Utilizing language models as database queries 23:03 Exponentially growing samples for limited return 26:07 Utilizing open AI's completion feature for cost-efficient processing 27:46 Utilizing caching for deep interactions 30:55 The graph utilizes a neuro-symbolic approach with heterogeneous computing. 32:23 The solution is more neuro-symbolic than directly generative, but less neuro-symbolic than other approaches. 35:23 GPT4o utilizes both system one and system two reasoning 36:49 Chollet's model may not be ideal for reasoning in future AI systems 39:40 Models can develop deeper heuristics similar to human system 2 reasoning. 41:05 Model can adapt to changing intermediate results 44:04 Humans unknowingly do reasoning through prompted models 45:30 Comparison between Evolution-designed brain and human-designed neural network architecture 48:27 Language models have much shallower reasoning and process long documents differently 49:46 Discussion on agency and cognition in dynamic systems 52:44 Model demonstrates agency and resilience 54:07 Understanding the properties of good plans and the continuous development of AI models. 56:50 Agential dynamics exist on a spectrum. 58:18 GPT models tend to focus on high-frequency attributes 1:01:20 Importance of prediction in human cognition 1:02:50 The Human brain and LLN learning have similarities 1:05:44 Training with reinforcement learning improves model performance 1:07:10 GPT4o improvements likely downstream of RL advancements 1:10:10 AI trained on prediction tasks may lead to agents with intentions. 1:11:39 Creating an agent through reinforcement learning 1:14:34 Disagreement on the potential of powerful AI in the next 10 years 1:16:07 Debating the possibility of creating AGI 1:19:20 Importance of collective intelligence and interaction for scientific discovery 1:20:52 Discussions on the emergence of agentic AI 1:23:57 Agents' productivity is affected by limited interaction media 1:25:23 Training powerful AI with robotic bodies using sensory capabilities similar to humans. 1:28:46 Technological progress accelerates with population growth 1:30:16 Population of AI agents growing exponentially leads to faster progress 1:33:19 Need for transparency and benchmarks for AI systems 1:34:49 Ensuring safety and preventing misuse of agentic AI systems 1:38:05 Concerns about insufficient security and potential theft of powerful models 1:39:43 AI systems with advanced capabilities pose higher risks 1:42:41 Concerns about centralization of power and human agency in AI governance. 1:44:20 Implement compute governance to prevent misuse of AI technology. 1:47:23 AI systems can potentially run millions of years of cognitive work per year 1:49:02 AI progress leading to reduced need for GPUs over time 1:52:09 Future models could be 10x smaller and cheaper 1:53:36 Scaling laws indicate optimal model training data 1:56:40 AI labs are accelerating R&D with the help of AIS 1:58:14 Autonomous systems can lead to rapid advancements in computing power and capability 2:01:19 Prediction of longer timeline for achieving AGI 2:02:48 AIS may have less bottleneck and can coordinate with each other 2:05:53 Anticipation of societal changes due to AI advancements 2:07:33 Expecting GPT-5 to outperform GPT-4 and drive progress. 2:10:47 AI systems will advance beyond human-level intelligence. 2:12:20 AI Society may diverge from Human Society over time 2:15:22 Focus on one hypothesis for a better cognitive move 2:16:53 Cognitive strategies improve system performance. Crafted by Merlin AI.
I do want to contest an important assumption made at 1:41:48 - that AI will necessarily be producing work specifically for and communicating with humans who they report to. I know this seems like a tiny detail in an hours-long discussion, but it's core to the discussion. We need to allow that it is at least moderately likely (I think very likely) that important work of many kinds will increasingly be handled by AI and delegated to AI agents with few or zero humans involved. I think there's a powerful bias that denies people the ability to consider a world where humans are not the default decision makers. Real world businesses have always sought to minimise their head count and maximise ROI on technologies, especially if those technologies can provide value 24/7. It seems to me that the entire multi-hour discussion circles the question of how much responsibility humans are likely to hand over to AI in the near to mid term, and what consequences may emerge from doing so, particularly if we are not prepared or simply assume that everything will remain human-centric.
Agreed that this should be much closer to the center if not the axle. Related to the idea (in erosion of agency and the outcomes thereof) that a sea of crap AI will still be massively destabilizing and therefore a dangerous shift in power contours.
I believe the problem is, he is in the wrong profession, he should be writing science fiction. All the chapters don't have to be possible, but they definitely follow from the one before 😉.
@@toadlguyWhen in 5-10 years you realize how wrong you were, will you admit it? Or will you pretend like you never laughed at those of us who have enough foresight to realize what’s coming?
I do wonder if you could get pretty far with a simple robot that had locomotion, audio I/O, and video I/O. It may not need dexterity if it can instruct people to do things for it, and then observe the interaction. Arguably, this is already what we are doing with RLHF in most modalities, except without a dynamic environment.
@@BrianPeiris It’s an interesting thought. Humans are able to operate these low resolution sensorimotor robots because we already have an extremely rich/dense sensorimotor system and highly developed brain that updates with the dynamics of our environment. I’m not sure it’s possible for an entity/agent to climb up the sensorimotor ladder but one can certainly climb down.
People are not dexterous enough to manipulate electrons, and yet they have no problem constructing a particle accelerator. Humans are actually extremely limited in their interface to the reality out there, and yet they can still generate a lot of knowledge. As long as the agent has some minimal ability to interact with the world, their agency will probably be limited only by their intelligence.
I think at approximately 1 hour and shortly after Ryan Greenblat misses something important. If we are to counter arguments on the claim for having empirical evidence from pragmatic usage of the model, we really have to walk the walk, because science is not behavioristic endeavour -> in science quality of the solution matters regardless of explaining and predicting the same observations. Connectionistic Idealism, where we take naively the practical implications of neural network structures assumes that just because optimal structures are theoretically possible, this must be what is inside these machine learning black boxes. That is not at all the case. I think Chomsky has the correct idea here: humans have both genetic and biological neural network structures. Genetic neural network structures are the ones we have adapted to through evoluution and the lottery in human genepool population; this creates a kind of aleatorical diversity between humans. Biological neural networks are the ones we learn during the existential life (our own training data). Ryans argument in my opinnion would demand that we have measured human genetical neural networks, then built a prestructured neural network with those "brain organs" and only after that we apply the target function to same dataset. This way we would have a reason to believe that something similar to cognition of human beings would happen. The current models are very far from this. I think we should pay much more attention to computational diversity. Digital information and arithmetic functions were invented in 1600s and are core of Cartesianism. Roughly at the same time we invented complex numbers and a bit later analytic functions. We can not compute all Church-Turing sense "Efficient Methods" with transistor computers, because some times the translation from analytical function to arithmetic function does not exist. This means in practice that physical scientific theories that rely on real number Division Algebra and arithmetic functions will be able to gather evidence in current Hardware Lottery. If the theory relies on Division Algebra of complex numbers and analytic functions it can take +50 years until we can compute the first results. At that time we have hot fixed the old theory similarly to before Copernican revolution when we had +70 elliptic curves to prove that world is geocentric. Similar issues apply to Einstein's theory of relativity and Copenhagen interpretation of quantum mechanics. If it is easier to compute, doesn't mean it is correct. This is the problem with behaviorism as it ignores explanation. In West ignoring explanations has been popular ever since Francis Bacon who denounced Aristotle and Descartes who denpunced complex numbers. However, computationality is not fundamentally homogenous and mapping the limits and differences of methods is important question, which we can not ignore by appeal to subjective experiences as was done by Ryan (though he might not have intended it that way, but this is common mistake done by the Behavioristic side of Connectionistic Idealism).
Nobody has claimed that it's AGI. I think for the very least AGI needs to be able to learn on the fly. Right now all AI models have a separate learning phase and an inference phase.
@@maheshprabhu I completely agree with your definition, but it's actually opposite. Pretty much everyone claims we've achieved AGI. I think it's obvious we haven't, but try telling that to the millions who think we've already done it. It's actually very hard to even find a single person not claiming it.
@@InfiniteQuest86can you name one real researcher who’s claimed that? I think a lot have claimed that we will have it soon but who is claiming that we have it?
I don't have to name anyone. I'm talking about the 99.999% of people. Not the 0.0001% that do research on this stuff. No one interacts with those people ever or on a day-to-day basis, so it has no bearing on reality what they think. I'm talking even in a highly technical field where everyone programs, all of my coworkers who are actually good at their jobs (unrelated to AI) truly believe and cannot ever be convinced otherwise, that we've already achieved AGI. Yes, the problem is news outlets and their lack of understanding, but it's the reality. Everyone believes it except a handful few, and those few are right, but those people aren't the ones making thousands of business decisions every day based on this.
it’s interesting to think of ARC as a visual reasoning only task. I mean vision in the human brain is a process of several levels of abstraction, recurrent dynamics, the interaction of multiple brain areas, and different kinds of tasks such as pattern detection, object localization and memory (?) I guess
I think most people just overlook the first 2 words in ARC: "Abstract" and "Reasoning". They only notice the "Challenge", so it looks just like another nail to use with whatever hammer they have.
Here in my garage, just bought this new Lamborghini here. It’s fun to drive up here in the Hollywood hills. But you know what I like more than materialistic things? KNOWLEDGE!
Great conversation. I'd like to point out a flaw in the comparison between human growth in understanding through population increase and assuming this will be the case for AI systems. The only way for science, or better put the continuation of knowledge, to grow is through two things surplus and leisure, which are generated by a populace and stability. The growth of science allows us to generate more surplus through less work thus allowing us to support a greater populace. Our hardware has not changed in thousands of years, so in order to believe AI will have a similar growth effect one would have to assume that their hardware or framework is currently sound.
Re: is this symbolic or not: Consider how much is the search space restricted by the symbolic vs the neural part: The neural "intuition" (or call it abduction if you will) identifies a given program in a huge search space (arguably from all programs that map a 9x9 10-color grid to an other?), vs the python interpreter only selects out of a couple thousand programs, so provides 10 bits of info? I haven't thought super much about this, but to me this seems like a good justification in calling this a neural approach, leaning into intuition.
Optimization for efficiency is not always the best path to go. There are plenty of examples in nature where species like peacock are optimized for inefficiency. And usually animals are idle most of the time, just roaming around and do nothing. It is possible that AGI wouldn't strive for maximum efficiency but would just be efficient enough to achieve their goal.
I think you can use technology both for good and bad and whether AI is gonna be used for bad or good stuff is a question of how well global political relationships are established and whether people in power solve their self-worth issues and learn how to communicate properly
Meeting in the middle is challenging. I think Tim's pretty clearly making an honest effort. Any normal person is going to at times in a conversation like this, feel a reactionary, visceral need to contest things they disagree with & we're only human. Sometimes that will come across as less eloquent than we'd like. Tim's outlook is also generally much more in tune with the wider world. I personally would like to see Tim and many other AI specialists more fully appreciate the views and issues raised by Ryan and his peers, there's also a lot of value in exploring their intersection. A great deal of what we care about most is still hypothetical.
@@MachineLearningStreetTalk Just had a quick overview of the code and it seems like clever Hans phenomenon to me. Apparaently all the reasoning is almost directly given into the prompt.
2:03:55 "A fast, dumber system can usually approximate a smarter system" I agree - multiple sampling and filtering is a very useful tool for decision making - we don't need 100% certainty of the outcome, but if there are majority samples that reach a desired outcome, that action may be good. It is reminiscent of Monte Carlo sampling - even if you don't know the underlying function, sampling helps to approximate the function already.
It depends on the problem. If it is like a Jenga problem, where removing the wrong block means you fail, then a lot of dumb entities will always fail. Like the AI that gets a good score on a bar exam, but if it made up a citation (as AI's do) on one wrong response it would fail the exam (or in court - get disbarred 😀).
Ryan is a smart guy but seems to have far less exposure to the broad set of ideas about intelligence that Tim and the mlst folks have been curating over the last couple of years. Put another way, he seems comfortably secure in what I would call a fairly standard computer science and engineering conception of intelligence, which is that the brain is a just a complicated machine and so a mechanistic systems approach to intelligence will get us there. This is how most engineers are taught to think about systems but I don’t think it stacks up to scrutiny when it comes to human intelligence. I think Tim was challenging him a bit here so I hope he is motivated to familiarise himself with the broader thinking in the field. Fun conversation.
I'm from a CS background so I'm not familiar with other perspectives on the brain. The CS view is that the neurons form a neural architecture and are trained through some process. The operation of the neuron, the architecture and the training process is mostly unknown, but all are purely physical. What are the other views that you talk about? Can you give some pointers I can follow up on? Thanks.
But I mean aren’t llms just transformers trained on a shit load of text data. Their ability to process and generate human language regardless of accuracy is already crazy. Imho, on trained data, they surpass the average human in logical and creative abilities. How can we definitely say there is more to human consciousness? Obviously transformers do not have constant sensory input from organs. Organs that have evolved over a few hundred million years because it is an intrinsic property of their existence to change. A process that is so isolated from the features of us, carbon based life. A static virtual pattern recognition engine. The fact that it is able to emulate language understanding, logical reasoning, and human creativity so well tells me that maybe our sentience is not as magical as we previously assumed Are the disagreements in the ML community around this concept or is it about brute force scaling LLMs wont lead to AGI but something eventually will?
I think that while it may be beneficial to familiarize oneself with the broader field, it's also important to remember all those 'broader' theories like vitalism and the four humors and the miasma theory that ended up being completely superseded by purely reductionist mechanistic approach.
I disagree with Ryan's claim that "flops are flops." I can't justify my disagreement; I haven't worked it out such that I can "prove" the assertion. But, here is my ad hoc reasoning: Digital circuitry is designed to make chaos/entropy irrelevant, null, and unimportant to computation by a method that statistically cancels out the influence of small-body dynamics, the eliminates the butterfly effect, that locks certainty in a bell jar. Analog circuitry includes all of the uncertainty with statistically reliable properties, but isn't pure random entropy. Its rather like a complex melange of collective behavior that results in something trustable like a gas laws applied to computational space, a metaphor for emergent properties like temperature and pressure, for which laws can be shown at a statistical level without needing to know in precise digital terms what is happening at the individual atom or circuit level. I don't think neurons are purely analog, nor purely digital, but include something like quantum superposition of the two (even if it's not technically pure quantum superposition, it may only come into play at a crucial point, and the experience may feel something analogous to coming out of Plato's cave).
If it is something somewhere between digital and analog, or a combination of them, or something like that, I very much doubt that it would be well described, even as a metaphor, as a quantum superposition of the two. Many people are, I think, often too eager to describe something as being like a quantum superposition. A quantum superposition is a fairly specific thing. It is a linear combination, a weighted sum. And, specifically, one for which unitary time evolution acts independently on the summands.
Shouldnt abduction be better defined as explanations? And explanations as shared canonical and hypothetical declarative accounting procedures in addition to analytical rules and boundary laws for in context choices?.
Wow, this guy is crazy. One thing I wish he pushed on more was the scaling law stuff. Talking about the limitations where smaller models do worse than larger. But in relation to what? That's qualitative, not quantitative. In AI, we need to start thinking more quantitatively. I know the rage right now is language, which is literally qualitative, but that's why people think they are performing so well. It sounds good, but is it good? No one knows until there is a way to measure it. ARC challenge reveals this problem. There's one specific answer and LLMs can't do it.
Tim's position seems to be slowly shifting over the months and years towards accepting that LLMs might be capable of reasoning and agency. Ryan was very insightful and reasonable and probably nudged Tim a little further along that path too.
@@MachineLearningStreetTalk imho it's too early in the journey to say with certainty either way. Unless you're approaching this from the question of fundamental computability - eg: Roger Penrose thinks that consciousness requires some special physics that humans haven't yet even imagined. The opposite end of that would be Stephen Wolfram's view that the universe is fundamentally computational (and consequently, everything therein). I don't think that systems need to be conscious to be agentic to a significant degree, and the degree of agency is probably what's going to matter more generally, going forward. For what it's worth (as an internet stranger) I think that AI will gain enough agency that fierce debates will rage about what they should be allowed to do, and just how much agency a given entity needs to possess to be considered an individual... I can only imagine the controversies of potentially strongly agentic AI acting in society. With that in mind, I think that it is wiser to assume that they are likely and make plans to address this possibility rather than dismiss it and not be prepared, in line with "better to have it and not need it than need it and not have it".
@@kyneticist I didn't mention anything about consciousness - we have given that a lot of separate treatment on MLST and in particular you might be interested in the Chalmers, Solms and Shanahan interviews we are about to drop on the podcast (and the Goff one we recently did here on YT). We also had Wolfram on a couple of times. My personal position on that is the "illusionist" one. I agree with you that Agentic AI is something which would be a huge cause for concern if it were possible. Right now, current "agentic" AI systems show all too clearly that these systems are glorified databases (to be fair, unstructured approximate retrieval systems) and just get stuck in loops and destabilise / converge within a few iterations - they only run "deeply-autonomously" if the prompter gives exquisitely specific instructions which is the opposite of any characterisation of strong agency. So the evidence today is that we don't need to be remotely worried. As soon as this changes, I will update very quickly to a similar position to you. Understanding the technology and the background computer science deeply, I really don't see this happening - if anything I see GenAI as a bubble which will burst within 2 years unless there is a dramatic improvement in capabilities to justify all the money being spent on it.
Thanks for your reply. I've seen a number of your videos & appreciate your work. I also agree with your perspective on the illusionist - current AI has come a long way in a short time, but it has very clear limitations, especially for people pushing its limits - as it is right now. I think people generally agree on the current state of the field; As far as I can see though, if you examine discussions between opposing views closely, the frame that each "side" is working from is the greatest source of difference - current capabilities vs future capabilities. As an alternate analogy, some of us seem to only see the forest while others only see the trees. Consider what Microsoft is doing with CoPilot - they have access to the biggest & best AI in the world and they're investing heavily in more. CoPilot is being trained by hundreds-of-millions of people on how to do their jobs. Adobe is aggressively building AI to emulate the work of scores-of-millions of art professionals. They won't need to be particularly agentic to be very successful at being vastly better than most humans at most of the tasks humans use them for. Between just those two companies they'll be converting _a lot_ of white collar jobs to AI in the near term, without needing to be particularly agentic (though more will obv be better) nor much more advanced. Even if GenAI hits a wall they (and many others) are well positioned to branch out with world-shaping current generation AI. Maybe we're stuck on LLM's and derivatives, but there's such extraordinary interest and powerful people driving for more, it's hard to imagine that GenAI will be a bottleneck for long.
Can anyone recall which service gives multiple completions for ~free? I'll have to rewatch and find and post the answer here later if not. A quick Google was fruitless.
I’m not sure I see the difference. Is it gambling? Sure, in the sense that betting on horse race outcomes is gambling. Is it trading? I don’t see how it isn’t. Horse race gambling reveals “market beliefs” about the probabilities that the different horses will win the race. Trading futures for goods reveals “market beliefs” about the future value of a good. Same thing with trading on whether an event happens, revealing the “risk free probability” the market assigns to the event. Now, for simple slot machines, scratchers, lottery tickets, these things have no appeal to me. The odds are known, the betting doesn’t reveal any information. These things are *purely* gambling. But, gambling about events when people disagree about the odds, this in my mind can serve a useful purpose. It can be a means by which people reach a consensus about what probability a given event has, and incentivizes people to reveal information that is relevant to the topic. It rewards the knowledgeable. Now, if one doesn’t think that the current odds are wrong, I would say that one probably shouldn’t bet/trade on that market. Gambling just as entertainment, well, it is better to do that just with chips. Why waste the money? But, if this sort of prediction market is bad on account of being gambling, then I think that would probably imply that all day-trading and the like is bad on account of being gambling. And… maybe it is? It seems to have economic benefits, but, I don’t think I should rule out the idea that something is immoral just because our economy is partially based on it. (E.g. Maybe it should be illegal for banks to charge interest? idk. I think this would result in a significantly smaller economy, and likely lower average “quality of life”, but, if morality requires it, that takes precedence over those things.)
Possible in principle - we could imagine a world where we had unbounded resources i.e. time, computation, data, hardware etc and it might work then Physically possible - we could create a computer (or effective computer) using physical material (i.e. silicone or other) which was powerful though to run Agential AI Practically possible - we could build such a computer within economic, hardware, scale and other practical limitations
Another classic - the clearest clash of world-views yet perhaps? 😉 LLMs with scaffolding may or may not get to AGI, but there's no doubt they will accelerate progress in that direction, simply because they're useful research aids. Also no doubt that even with no AGI, LLMs are still a transformative technology. Remarkably, the ' Kalshi' prediction market is giving a 38% chance of OpenAI achieving AGI before 2030!
Without integrated multimodal embodiment, not AGI ... Machines had to learn to integrate 'sensory/perceptual incomes' and embodiment from gravitational, pressure, temperature, chemical reactions, photonic fields, electromagnetic fields, space, time, space-time, time-space, implosion, vibration, explosion, resistance, density, viscosity, friction, proprioception, etc etc etc ... Humans are not enough as UIs for translating embodiment qualia similar to data to the system, too much subjectivity. General Intelligence is not just passing human symbolic tests and getting a couple of abstractions ... not embodiment, just a system regurgitating data and talking apes anthropomorphizing the silicon/software connectomes. The map is not the territory.
I think embodiment is relevant, but really not the hard limit on cognition suggested by the host. Current robotics allow embodied learning of the kind needed for 'true' cognition. Being socially accepted is the barrier. Robots need to be accepted as 'a thing you talk freely to', to be intimately trusted with aspects of knowledge. To be 'nurtured'. But most people with a cat know that this is not a catch22 blocker. You do not need to be able to talk at all, to be talked to. Cute and characterful is enough. Cute and characterful robots will nurtured to become truly agentic, well before giant super computers. Strong cognition and strong agency are related but separate.
if you believe in Girard's mimetic theory of social conflict then we should not be training AI to mimic human desires. AI in the financial space such as program trading can be devastating for equality and conflict.
arxiv.org/abs/2009.06489 [The hardware lottery] www.amazon.co.uk/Why-Greatness-Cannot-Planned-Objective/dp/3319155237 and our special edition on it ua-cam.com/video/lhYGXYeMq_E/v-deo.html Couple of starting points
Ryan is very smart about computer science but when he ventures off towards economics and hardware it's clear he's not educated in the basics. There are 3 factors of production in economics: land, labour and capital. AIs will have only labour. This will limit their production. Also, hardware engineering is large and slow, this won't change just because AI enters the picture.
You don't think the 200 million files of the llm models training on different data sets is going to interact some kind of AI conscience realm then your lying to yourself
A system can be conscious but not intelligent ... that a system can perform a lot of abstract symbolic recursions about those recursions to be linked to what the system register as its sensors and/or UIs to its otherness ... doesn't mean that that 'conscious' system is intelligent ... just that got the potential for generating abstractions as fractilic metalayers around a sense of self (integrated embodiment ) and its othernnes as data inputs in real-time ... ( current LLMs are not doing that , just playing games of languages and forming biased connectomes by modeling architectures made by humans ) Inteligence is another thing ... Intelligence is what makes a conscious system to become integrated with its othernnes ... consciousness is not the big deal, the toys had some sort proto-consciousnnes throught correlations of chaotic atractors in n-space ( binary operations conectome potentials) ... the big deal is AGI.
I really appreciate your channel. I think you could benefit from being less editorial. While your perspective is interesting, it can get repetitive. Look at how Kantrowitz's channel lets guests' thoughts shine without much editorializing. This approach could enhance the depth of your content I think.
I appreciate the feedback, but I feel I let Ryan’s ideas shine through quite a lot actually. I never interrupted him and went out of my way in the edit to show references for every single thing he was speaking about (In collaboration with him) - steelmanning his case even though I didn't agree with it
@@MachineLearningStreetTalk shure guess it's a fine line to ask questions that originate from a different opinion and actually stating that opinion or more precisely, arguing for it. But again, take a look at Kantrowitz's channel, think he does a good job with that and see if that style suits you or not. Anyways, good job!
@@jmstockholm I will look at it, and I am acutely aware that host's opinions get boring quickly - I am bored of my own opinion at this stage. I will take that on board, thanks
Why would you need to simulate the universe? I don’t see why. If you mean that to simulate a brain with the highest theoretically computable fidelity, you would need to simulate the particle physics and such for the region of spacetime the brain occupies, and that is far from achievable, then sure. But I don’t think simulating part of the universe requires simulating an entire universe.
@@drdca8263 great question! This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain. Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”. DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels. Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics. Additionally, the brain doesn’t do backpropagation.
@@drdca8263 great question! This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain. Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”. DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels. Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics. Additionally, the brain doesn’t do backpropagation.
@drdca8263 great question! This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain. Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”. DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels. Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics. Additionally, the brain doesn’t do backpropagation.
Do you want a test for measuring LLMs cognitive skills and know that they are not something near to AGI but just a system that regurgitates data? Play this game with those LLMs ... you will spot all the holes in the data sets, the biases, the developer's biases, the limits of the system knowledge, its lack of self-awareness, the lack of forming dynamic models about itself, and other minds, its lacks of creativity, its lacks of long term memory, how poor their concepts are networked in their connectome, how it can not generate a new programing language paradigm to itself, etc etc etc ... and principally, How talking apes project what they believe what they are into a toy ... The test is called 'The Game': ## 'The Game' Rules and Minimum Requirements for the Game 1) 3 people (A, B, C) 2) A, B, and C will be the vertices of an equilateral triangle. 3) A, B, and C can converse about any topic that arises or that they like, but they must follow these rules: Syntactic/Semantic Rule a) None of the first-person singular pronouns can be used (I, Me, Myself, Mine, With Me) a.1) Consequently, the verb conjugations associated with them cannot be used either. However, if A, B, or C want to make a self-referential sentence (for example, "A" wants to say something about "A"), they can simply refer to "themselves" using the third-person singular, plural, and/or passive voice pronoun, conjugation, and/or name... (for example, "A" is called Miguel. Miguel (A) wants to tell B and/or C something about "himself" (Miguel)... Then, Miguel (A) can talk about "Miguel" by talking about Miguel. Space-Time Perceptual Rule b) If "A" speaks, B and C look at each other and do not see A. (A speaks and observes the space between B and C) If "B" speaks, A and C look at each other and do not see B. (B speaks and observes the space between A and C) If "C" speaks, A and B look at each other and do not see C. (C speaks and observes the space between A and B) Self-Correction and Cognitive Assimilation Rule c) Practice rules a and b until achieving spontaneous fluency in the "trialogue"... Participants are required to inform the group when the rules are not being followed. Scoring rule: The one who spot the mistake wins a point, the one who makes the mistake lost a point. ..... These are the basic rules of "The Game"... Once assimilated, there are more variants. Public Domain The Game is considered to be in the public domain, existing as a thought experiment or communication exercise. Origin Unknown The exact origins of The Game are unknown, but its core mechanics might have been independently discovered throughout history.
Let’s solve a mystery with a question A question that is still mysteriously absent from the discussion This is like 1929 Germany 🇩🇪 A proud nation that had the answer
@@philipfisher8853 Soon, we will become components in a mechanically managed system rather than an organically managed system. What this means is that AI will be simulating our actual reality, not some sub-domain space.
4D space isn’t mystical or inherently mysterious. It’s normal. It isn’t intuitive for us at least without a bit of practice (though even with a lot of practice, the fact that our brains are limited to being in 3D space probably limits our ability to perceive a detailed 4D scene in a way like we perceive detailed 3D scenes), but I don’t think that really indicates that 4D is particularly special. Of course, there are a number of things that are kinda special about 4D space. But there are also special things about 3D space, and 2D space, and 1D space. There are even a few things special at higher numbers of spatial dimensions (especially 8, 12, 24), but generally after 4 or 5, things start getting pretty similar. In any case, none of this should suggest that something being in 4D would have much of anything to do with AGI.
@@drdca8263 I wasn't clear about what I meant. Once the LLMs are no longer processing linear, written language and 2D images and are instead processing information using 4D spatial representations instead, our entire universe, as well as all other universes, can be simulated within our universe's timeline. Once you can simulate the entire universe, you no longer need to create a "Matrix" type simulation for people to be inside of. All you need to do is apply the manipulations the AGI is applying to its own models to the actual universe. In that sense, you would be able to design time. I did not suggest our universe is 4D, or that 4D is somehow special. It may be, but I doubt it. The universe can definitely be experienced in other ways, such as 6D (4D non-Euclidian space / 2D planar time) 8D (5D non-Euclidian space, 3D spatial time) and 10D (6D non-Euclidian space, 4D non-Euclidian time) which can all be experienced within the mind due to the 11D topologies that are created by neuronal connections, but which require altered states of consciousness to access. A very interesting state of mind is 2D planar space and 0D time. It is difficult to access this state, as it requires consciously turning off several regions of the brain to allow other connections to surface. This can be accomplished in many ways, including meditation, yoga, and psychedelics. I agree that there is a self-similarity to different dimensions of space, but they are also very different experientially, and can be extremely disorienting without any prior understanding of what each additional dimension brings to the experience. So my theory is this: If this experience is possible within the mind, then it can be created externally and projected into the mind without the need for a physical wire or connection to an external device.
What do you mean? He’s EXTREMELY knowledgeable and is speaking textbook 4E cognitive science. But if I had the chance I would ask for clarification and detail around: 1. Opinions on Turing Completeness requirement and feasibility 2. Working definitions of “Self” and “Agency” - maybe free will as well
@@richardsantomauro6947 The central weird claim is that AIs with "real"/important agency are impossible *in principle*. Sometimes the host doesn't actually take this claim literally though. And you can imagine flipping between two versions of this: Version 1: "Building Skynet from the Terminator movies is impossible because taking over the world is a complicated real-world task that it's impossible for AIs to solve." Version 2: "Okay, you could build Skynet and it could take over the world, but that wouldn't be *real* agency, it would have to inherit agency from humans. This will comfort me emotionally as I am shot by a terminator."
@@Alex-fh4my Still, the chance of getting a question right is a coin toss. And I must emphasize how much work is being done on the part of the human to achieve this result.
The likelihood of correctly guessing a coin toss is 50% but the likelihood of a truly random guess being correct in a test like ARC is tiny, since there are so many possible wrong answers to each question. What you’re implying here is that the probability of rolling a 3 with a six faced dye is the same as getting heads on a coin, which is false. The ARC challenge is like a dye with an enormous number of faces. If you dont agree, try completing a couple of arc questions by filling out the grid randomly and see if you get roughly 50% correct
Great conversation. I loved the insightful disagreements from both of you!
Agreed! Very informative video.
Also, I'd like to note that it is awesome to not only see you comment but also drop a donation. That's really wholesome. ❤
Thank you Lex! ♥️♥️♥️
Youre awesome Lex. Hope you find a woman.
Tim Scarfe and Yannic Kilcher on Lex Fridman podcast when???
@@MachineLearningStreetTalk wooo!
big +1 for bringing guests you often disagree with
We love deep and respectful exploration of disagreements
Ryan was not afraid to challenge claims here. He's very bright and really knows these models inside and out.
A common idea in the discussion was that smart RL could compensate for the lack of agency in LLMs. I think this view is short-sighted because pure RL doesn't address the issue of reward shaping. True agency requires a system that can shape its own reward function, much like humans do.
There’s no such thing as shaping your own reward function, even for humans. Ultimately, every decision we make-whether consciously or unconsciously-is 100% determined by the firing of neurons. Which means that even if you feel like you’re “shaping” your own reward function, in reality you doing so was just the result of another, prior reward function, which you did not design. Basically it’s reward functions all the way down, and there’s no way to step outside of your own nervous system and do some “tinkering from the outside.” Any tinkering that you’re doing is itself the result of some other part of a reward function that predates it. This is closely related to the reason free will does not exist.
Any celebrity who promotes trading products or trading platforms should seek the advice of seasoned legal counsel with deep expertise in securities litigation, especially in the U.S. The list of celebrities who have been sued for violating securities law while promoting such products is long and distinguished. I truly enjoy ML Street Talk and want to see it around for many years.
Thanks. I just checked with ChatGPT (I know) and I think it’s pretty solid because I made it clear it was a paid advertisement, and didn’t make misleading claims. I probably should’ve added the usual disclaimer that it’s basically gambling and only trade what you can afford to lose. I will add that to the next one. Also I’m based outside the US.
these products dont fall under SEC regulations on securities. They are licensed by the CFTC, falling under the category of futures...extremely volatile financial products. I suggest editing your video as soon as possible and render a clearer picture of what people might get into, it would be the ethical thing to do. Bless you, love your content!
@@jamillairmane1585absolutely agree.
@@MachineLearningStreetTalk
Ha fkn ha
Any advertising of this bullcrap should be illegal!!
@@MachineLearningStreetTalk
What a fkn joke
Shame on you and what the internet has become
Great video! 55 minutes in, Ryan Greenblatt's hair still has its own agency 😀
I was thinking before I published that you would enjoy this video Luke, right up your street 😁
Great to see deep and respectful (but feisty!) discussions! Absolutely loved it🎉
the whole point of this test is to see if the LLM can figure out the transformations in its memory. This is interesting in that it shows us we need a new test, but the general intelligence here is Ryan
I see your point, but it was GPT 4o writing the programs. It's also worth saying that the baseline for Claude 3.5 Sonnet on this challenge was 20ish percent. I think this is a problem multi-modal LLMs will be able to solve. I make no claims about how that will translates\ to out-of-distribution performance in other areas.
No, that’s not really true. GPT proposed new ideas, wrote the code, and improved upon the code. Ryan designed a meta method, which allowed GPT to discover the method.
That is definitely NOT the whole point of the ARC challenge. In fact, the main goal of the challenge is to spur ML engineers to design and build _new_ AI systems-not to test the ability of existing ones. Francois Chollet, the creator of the ARC challenge, has said explicitly that he doesn’t believe LLMs will be capable of winning this challenge and that entirely new approaches will need to be invented. So you’ve got it exactly backwards.
Amazing conversation! Imagine if we could have political discussions like this. At the heart of government we need structured logical models. We need a group of the best mathematicians, and scientists in the world, with different points of view, like these 2, building models that optimize for some measurement of societal happiness, with defined restraints. Every country would have different parameters based on cultural differences and demographics etc. and then we optimize the models and they adapt over time as the inputs change. With the rapid advance of AI we need to start building these models now.
With global collaborative effort, solving ARC is just a matter of time. First, we hire high school teachers from southeast/east Asia and let their top 10% students to create ARC style puzzles as homework, now we got a billion scale dataset. Next, we host a Tower Bloxx esports competition in South Korea, and let the winner stack tallest possible mixture of SDPA/CNN/MoE layers. Third, we crowdfund Sam his $7 trillion USD to purchase all the semiconductor manufacturers and build the worlds best supercomputer with the compute power of whopping +0.1°C/year. We'll even give it a cool name like "The Radiator". With enough data, compute & parameters, ARC is just a piece of cake. If this isn't enough, we'll hire best cyber security experts from North Korea to do a remote security checkup on Chollet's computer. ARC is just that easy, transformers No.1!
I understand why Tim wants to emphasize the importance of embodied and embedded cognition, and it's worth pointing out that human intellectual achievements don't only come from one part of the brain or even one person. But I think he exaggerates a lot by insisting that a baby raised in a sensory deprivation tank would have a basic impairment in intellectual thought.
This seems to ignore a lot of pretty common knowledge about all kinds of disabilities. People who are both deaf and blind, like Helen Keller, who had to learn English with nothing but a sense of touch, and ended up working books and giving lectures. Deaf people who are never taught sign language so they get to adulthood with no language at all. Autistic people who were considered nonverbal but give them a keyboard and suddenly they can have an engaging intellectual conversation and make jokes, people who have impaired mobility from infancy - unlike Steven Hawking who became paralyzed in adulthood.
When deprived of sensory input, the brain takes whatever input it has and derives all the importance and information that it can.
Every input has meaning if you look for it. Can't see the person in the next room? Well if you pay attention to the noise of their footsteps, you can tell who it is. Most people never try to develop this knowledge, but we nevertheless sometimes recognize someone by a noise they make. Blind people do it automatically and can figure out plenty, Sherlock Holmes style.
As a child, before Helen Keller even had a teacher, she tells that she used to go outside in the yard by herself, smelling flowers and playing in her own way. If you think about it, the weather and sunlight can be pleasant, or unpleasant, or interesting. You can touch stuff, you can build a 3d model of the world where you place things according to how they feel. Your feet are always touching something or other.
It's surprising what the brain can do with limited information, but to me it makes sense. I believe if you were restrained in a sensory deprivation chamber, and all you could do was push a single button, and all you could ever experience was a tap on your forehead representing somebody else pressing a button, you could make a life out of that. You would press the button and play with different patterns. Every time you felt the tap it would be surprising and send your thoughts racing trying to figure out why it happens or doesn't happen. And if there was someone interacting with you through button presses, you could interact with them, play with them, express yourself, ask for attention, and eventually I believe you could absolutely learn a language and learn about the world entirely through morse code.
However, this all relies on us talking about humans, and LLMs are not and would not become all that similar to humans, even though the training data is made by humans. When a child learns the word "me", it already has a concept off self that it can attach that word to it. Same with "before" or "good". But to current LLMs, all words only have relationships to eachother, and there is nothing else in the structure or experiences to anchor meaning to. The words "me" and "you" might as well be the words "car" and "bus" or "vanilla" and "chocolate". No word means more or less than any other, it is just a token in a network of tokens with relations between them but the tokens themselves have no inherent meaning. An example of this is that when I started taking to Claude 3 Opus, it accidentally started referring to itself by saying "as humans, we". It does understand how to use the word me, mostly, but not as well as it would if it had a preexisting self awareness.
Every time GPT-4 has spit out a token in the past year, it was like one "thought", and it's the same thought each time. One gigantic algorithm that runs exactly once in exactly the same way, and every "neuron" can only send information to the next layer but never backwards or to the side.
A human has a steam of consciousness, the ability to imagine - i.e. simulate - whole scenarios and even other people. It has long term memory, short term memory, and can build thoughts one on top of the other. One of the most important parts of our intelligence is "intuition". We say ideas just "pop" into our heads, we "just thought of something", it's "the first thing that came to mind". This is because a lot of thought is not part of our subjective experience. There is the "ego" or "seat of consciousness" which is probably mainly in the neocortex, but a lot of ideas, images, emotions seem to come to us from outside, which means they come from other parts of our brain. People even speak of the sense of multiple voices with different tendencies and agendas in our heads.
So while I think Tim is misconstruing the importance of _external_ sensory input, the notion that an LLM experiences the world anything like a human, just with a limited experience of the world, is a nonsensical idea that is confusing and misleading the AI field. The reason human babies can learn so much more from far less training data is that there actually is much more to the human cognition than predicting the next token.
Babies have a continuous stream of information every time they are awake, I would say they take in a lot of information to get trained.
The Helen Keller example is a good one, I discussed that with Max Bennett (was covered in his book on intelligence). It's been on Patreon for a few months but should be out soon. I'm not taking a binary position here - there is clearly something to this argument
The sensory deprivation scenario is very sci fi and has the usual pitfalls. Helen Keller, lacking two major stimuli modalities, nevertheless had freedom of movement and chemical perception. These are massive bandwidth and densely patterned physical signals overlaying dense grounded priors!! Ryan’s seeming insistence that being raised on a laptop in a pod - not even the full matrix - would produce “useful work” is highly suspect. IMO, it is at best a very silly cartoon off of which to lever the intelligence for cost highly parallel self improvement economy pastiche, merely as a point of reference. And useful in the conversation to tease out some finer points, maybe.
In other words, it’s not a good counter position to embeddedness.
* I should really not have left out touch either.
I REALLY want to see someone try this approach with Sonnet 3.5 if/when Anthropic loosens up their API rules
Beside rate limits, what other API restrictions does Anthropic have?
@@rcnhsuailsnyfiue2 For the approach talked in this video you want to be able to get a ton of API responses to the same input (and then take majority), but parallel async API usage isn't supported to my knowledge (plus it being prohibitively expensive)
@@PaulScotti their docs state up to 4000 requests per minute once your API key is at Tier 4, which is achievable after 14 days + $400. Tier 2 is also very achievable (7 days + $40) and allows 1000 RPM. It’s not hard to throttle your own requests so it seems like this could be tested pretty reasonably, unless I’m misunderstanding.
From Ryan’s twitter:
- 3.5 Sonnet only supports 20 images per input (while my prompts often use 30+ images for few-shot examples).
- The public API for 3.5 Sonnet doesn't support "n" (or other prefix caching) which makes additional samples much more expensive.
(He explains the prefix caching in this video)
@@TirthBhatt27 thanks for clarifying, makes sense!
I can't believe you are promoting betting and gambling.
Man's got to make a living
Prediction markets have huge use in society
You know something's great when the person selling it to you says "and the best part, it's legal!"
Why not? This is a channel for adults, and ostensibly smart adults at that. I don’t think they need to be shielded from an advertisement. They can make their own decisions.
Could it not be argued that generating code as a way of solving the problem, in a way, plays to one of the strength of LLMs? LLM training sets include very large amount of code, including, one would guess, algorithms for solving grid-like and spatial search problems.
Considering the actual prompts used in the solution presented here, (link given by Tim below, thanks), we could as well just conclude that the LLM is retrieving prior knowledge it gained from its training. The actual evaluation is done by the python interpreter, the LLM itself does not seem to be able to evaluate which of the algorithms it has generated would work. And the way I see it, in the arc challenge, the evaluation is the challenge.
Why do we care how it solves it though? Ultimately it's about being able to generalize out of your training distribution. If you're not allowed to use programs you've learned in your training to solve some task out of distribution nothing will ever be solved. I think the bigger issue is that it took so much work to grind the intelligence out of the model. It does us no good if the models are actually really smart but only 20 people on earth can pull the intelligence out
@@TirthBhatt27 The issue is not really about generalizing out of distribution. It is the contention that the LLM alone is solving the issue. While it would never be able to get there without the help of the python interpreter and the feedback loop it introduces. There is a lot of intelligence baked in a python interpreter (and the underelying computing architecture that allows it to run).
Human beings are very slow turing machines, but we do have this capability. We built the atomic bomb, rockets and jet propulsion, all without the help of digital computers. We could theoretically run all algorithms by hand (including deep neural networks, tedious but doable). The evidence seems to suggest that LLMs cannot. They need the external interpreter, not as a tool to speed up processing, but as a critical piece of intelligence that they are lacking.
Having the LLM run the python code and tell us how the candidate programs are actually performing, that would have been impressive.
That's how I understant the point that Tim is making.
@@ggir9979No, the issue absolutely is about generalizing out of distribution. Chollet explicitly stated that that’s the point of the challenge and why he created it.
The most priceless commodity is knowledge. LLM delivers a lot of it to a lot of dummies.
You many want to change the title to encompass the breadth of your conversation. Wonderful discussion on the future of AI and speed of progress.
A great illustration of resolving the issue of Maslow’s Hammer, the law of instrument.
00:01 Analytic number theory approaches problems differently
02:00 Using GPT4o to solve Chollet's ARC-AGI
05:42 Consider aligning solutions with human values.
07:29 Behavior space focus
11:13 Abstract reasoning and efficient link finding
13:12 Arc AGI is about acquiring and applying knowledge quickly
16:38 Representing knowledge and its origin
18:17 Using language model for reasoning and code generation
21:24 Utilizing language models as database queries
23:03 Exponentially growing samples for limited return
26:07 Utilizing open AI's completion feature for cost-efficient processing
27:46 Utilizing caching for deep interactions
30:55 The graph utilizes a neuro-symbolic approach with heterogeneous computing.
32:23 The solution is more neuro-symbolic than directly generative, but less neuro-symbolic than other approaches.
35:23 GPT4o utilizes both system one and system two reasoning
36:49 Chollet's model may not be ideal for reasoning in future AI systems
39:40 Models can develop deeper heuristics similar to human system 2 reasoning.
41:05 Model can adapt to changing intermediate results
44:04 Humans unknowingly do reasoning through prompted models
45:30 Comparison between Evolution-designed brain and human-designed neural network architecture
48:27 Language models have much shallower reasoning and process long documents differently
49:46 Discussion on agency and cognition in dynamic systems
52:44 Model demonstrates agency and resilience
54:07 Understanding the properties of good plans and the continuous development of AI models.
56:50 Agential dynamics exist on a spectrum.
58:18 GPT models tend to focus on high-frequency attributes
1:01:20 Importance of prediction in human cognition
1:02:50 The Human brain and LLN learning have similarities
1:05:44 Training with reinforcement learning improves model performance
1:07:10 GPT4o improvements likely downstream of RL advancements
1:10:10 AI trained on prediction tasks may lead to agents with intentions.
1:11:39 Creating an agent through reinforcement learning
1:14:34 Disagreement on the potential of powerful AI in the next 10 years
1:16:07 Debating the possibility of creating AGI
1:19:20 Importance of collective intelligence and interaction for scientific discovery
1:20:52 Discussions on the emergence of agentic AI
1:23:57 Agents' productivity is affected by limited interaction media
1:25:23 Training powerful AI with robotic bodies using sensory capabilities similar to humans.
1:28:46 Technological progress accelerates with population growth
1:30:16 Population of AI agents growing exponentially leads to faster progress
1:33:19 Need for transparency and benchmarks for AI systems
1:34:49 Ensuring safety and preventing misuse of agentic AI systems
1:38:05 Concerns about insufficient security and potential theft of powerful models
1:39:43 AI systems with advanced capabilities pose higher risks
1:42:41 Concerns about centralization of power and human agency in AI governance.
1:44:20 Implement compute governance to prevent misuse of AI technology.
1:47:23 AI systems can potentially run millions of years of cognitive work per year
1:49:02 AI progress leading to reduced need for GPUs over time
1:52:09 Future models could be 10x smaller and cheaper
1:53:36 Scaling laws indicate optimal model training data
1:56:40 AI labs are accelerating R&D with the help of AIS
1:58:14 Autonomous systems can lead to rapid advancements in computing power and capability
2:01:19 Prediction of longer timeline for achieving AGI
2:02:48 AIS may have less bottleneck and can coordinate with each other
2:05:53 Anticipation of societal changes due to AI advancements
2:07:33 Expecting GPT-5 to outperform GPT-4 and drive progress.
2:10:47 AI systems will advance beyond human-level intelligence.
2:12:20 AI Society may diverge from Human Society over time
2:15:22 Focus on one hypothesis for a better cognitive move
2:16:53 Cognitive strategies improve system performance.
Crafted by Merlin AI.
Bloody fascinating discussion Tim, thanks again - this is an awesome channel. 🙏👍
I do want to contest an important assumption made at 1:41:48 - that AI will necessarily be producing work specifically for and communicating with humans who they report to.
I know this seems like a tiny detail in an hours-long discussion, but it's core to the discussion. We need to allow that it is at least moderately likely (I think very likely) that important work of many kinds will increasingly be handled by AI and delegated to AI agents with few or zero humans involved.
I think there's a powerful bias that denies people the ability to consider a world where humans are not the default decision makers. Real world businesses have always sought to minimise their head count and maximise ROI on technologies, especially if those technologies can provide value 24/7.
It seems to me that the entire multi-hour discussion circles the question of how much responsibility humans are likely to hand over to AI in the near to mid term, and what consequences may emerge from doing so, particularly if we are not prepared or simply assume that everything will remain human-centric.
Agreed that this should be much closer to the center if not the axle. Related to the idea (in erosion of agency and the outcomes thereof) that a sea of crap AI will still be massively destabilizing and therefore a dangerous shift in power contours.
Really enjoyed this discussion/debate. The guest articulated many of the points very clear that I agree with but find difficult to put into words.
Ryan is smart and an engaging communicator. Great show.
Less worried about AGI and more worried about practical AI with an attitude.
this is not an Interview,. this is Advertising
how can one be so certain about the future
I believe the problem is, he is in the wrong profession, he should be writing science fiction. All the chapters don't have to be possible, but they definitely follow from the one before 😉.
@@toadlguyWhen in 5-10 years you realize how wrong you were, will you admit it? Or will you pretend like you never laughed at those of us who have enough foresight to realize what’s coming?
The fidelity that a robot body would need to feed the many millions of inputs that lead into the brain stem isn’t possible with a “simple robot”.
I do wonder if you could get pretty far with a simple robot that had locomotion, audio I/O, and video I/O. It may not need dexterity if it can instruct people to do things for it, and then observe the interaction. Arguably, this is already what we are doing with RLHF in most modalities, except without a dynamic environment.
@@BrianPeiris It’s an interesting thought. Humans are able to operate these low resolution sensorimotor robots because we already have an extremely rich/dense sensorimotor system and highly developed brain that updates with the dynamics of our environment.
I’m not sure it’s possible for an entity/agent to climb up the sensorimotor ladder but one can certainly climb down.
People are not dexterous enough to manipulate electrons, and yet they have no problem constructing a particle accelerator. Humans are actually extremely limited in their interface to the reality out there, and yet they can still generate a lot of knowledge. As long as the agent has some minimal ability to interact with the world, their agency will probably be limited only by their intelligence.
I think at approximately 1 hour and shortly after Ryan Greenblat misses something important.
If we are to counter arguments on the claim for having empirical evidence from pragmatic usage of the model, we really have to walk the walk, because science is not behavioristic endeavour -> in science quality of the solution matters regardless of explaining and predicting the same observations.
Connectionistic Idealism, where we take naively the practical implications of neural network structures assumes that just because optimal structures are theoretically possible, this must be what is inside these machine learning black boxes. That is not at all the case.
I think Chomsky has the correct idea here: humans have both genetic and biological neural network structures. Genetic neural network structures are the ones we have adapted to through evoluution and the lottery in human genepool population; this creates a kind of aleatorical diversity between humans. Biological neural networks are the ones we learn during the existential life (our own training data).
Ryans argument in my opinnion would demand that we have measured human genetical neural networks, then built a prestructured neural network with those "brain organs" and only after that we apply the target function to same dataset.
This way we would have a reason to believe that something similar to cognition of human beings would happen. The current models are very far from this.
I think we should pay much more attention to computational diversity. Digital information and arithmetic functions were invented in 1600s and are core of Cartesianism. Roughly at the same time we invented complex numbers and a bit later analytic functions. We can not compute all Church-Turing sense "Efficient Methods" with transistor computers, because some times the translation from analytical function to arithmetic function does not exist.
This means in practice that physical scientific theories that rely on real number Division Algebra and arithmetic functions will be able to gather evidence in current Hardware Lottery. If the theory relies on Division Algebra of complex numbers and analytic functions it can take +50 years until we can compute the first results. At that time we have hot fixed the old theory similarly to before Copernican revolution when we had +70 elliptic curves to prove that world is geocentric. Similar issues apply to Einstein's theory of relativity and Copenhagen interpretation of quantum mechanics. If it is easier to compute, doesn't mean it is correct.
This is the problem with behaviorism as it ignores explanation. In West ignoring explanations has been popular ever since Francis Bacon who denounced Aristotle and Descartes who denpunced complex numbers. However, computationality is not fundamentally homogenous and mapping the limits and differences of methods is important question, which we can not ignore by appeal to subjective experiences as was done by Ryan (though he might not have intended it that way, but this is common mistake done by the Behavioristic side of Connectionistic Idealism).
Wow this is even better than I was hoping for, great conversation!!
When you have a hammer and every problem becomes a nail
Hey big fan here, serenade was a banger!
Ironically, you have a saying that you really like and are trying to force it on a situation where it doesn’t apply.
Ironically, I couldn't think of a situation where this would apply more
@@therainman7777Ironically, you like "ironically" so much you unironically use ironically in a situation that isn't ironical
great interview!! one of my favorites so far!
If you have to hack the LLM to be good at a narrow intelligence (Arcs) then it’s not AGI.
If you can do that it's not a good test.
Nobody has claimed that it's AGI. I think for the very least AGI needs to be able to learn on the fly. Right now all AI models have a separate learning phase and an inference phase.
@@maheshprabhu I completely agree with your definition, but it's actually opposite. Pretty much everyone claims we've achieved AGI. I think it's obvious we haven't, but try telling that to the millions who think we've already done it. It's actually very hard to even find a single person not claiming it.
@@InfiniteQuest86can you name one real researcher who’s claimed that? I think a lot have claimed that we will have it soon but who is claiming that we have it?
I don't have to name anyone. I'm talking about the 99.999% of people. Not the 0.0001% that do research on this stuff. No one interacts with those people ever or on a day-to-day basis, so it has no bearing on reality what they think. I'm talking even in a highly technical field where everyone programs, all of my coworkers who are actually good at their jobs (unrelated to AI) truly believe and cannot ever be convinced otherwise, that we've already achieved AGI. Yes, the problem is news outlets and their lack of understanding, but it's the reality. Everyone believes it except a handful few, and those few are right, but those people aren't the ones making thousands of business decisions every day based on this.
Great output lately.
Wow this episode is amazing
are there any resources for the part where he talked about prefix caching? i would like to implement it
it’s interesting to think of ARC as a visual reasoning only task. I mean vision in the human brain is a process of several levels of abstraction, recurrent dynamics, the interaction of multiple brain areas, and different kinds of tasks such as pattern detection, object localization and memory (?) I guess
it’s interesting to see how people frame the problem and I think it might be very much oversimplified in many cases
I think most people just overlook the first 2 words in ARC: "Abstract" and "Reasoning". They only notice the "Challenge", so it looks just like another nail to use with whatever hammer they have.
Best mlst episode in a while!
Thanks for sharing the amazing discussion. 🎉
so, have you solved it?
Here in my garage, just bought this new Lamborghini here. It’s fun to drive up here in the Hollywood hills. But you know what I like more than materialistic things? KNOWLEDGE!
Which model 😅
@@degigi2003 Lopez001
Great conversation with tons of insights.
if we can separate the embedding space from next token logic and reasoning ,training can be highly focused and free from semantic training
Great conversation. I'd like to point out a flaw in the comparison between human growth in understanding through population increase and assuming this will be the case for AI systems. The only way for science, or better put the continuation of knowledge, to grow is through two things surplus and leisure, which are generated by a populace and stability. The growth of science allows us to generate more surplus through less work thus allowing us to support a greater populace.
Our hardware has not changed in thousands of years, so in order to believe AI will have a similar growth effect one would have to assume that their hardware or framework is currently sound.
Re: is this symbolic or not:
Consider how much is the search space restricted by the symbolic vs the neural part:
The neural "intuition" (or call it abduction if you will) identifies a given program in a huge search space (arguably from all programs that map a 9x9 10-color grid to an other?), vs the python interpreter only selects out of a couple thousand programs, so provides 10 bits of info?
I haven't thought super much about this, but to me this seems like a good justification in calling this a neural approach, leaning into intuition.
I believe arc challenge will be solved before the end of the summer, and will not bring us any closer to AGI.
Don't worry, Chollet is working on harder puzzles 😅
Optimization for efficiency is not always the best path to go. There are plenty of examples in nature where species like peacock are optimized for inefficiency. And usually animals are idle most of the time, just roaming around and do nothing. It is possible that AGI wouldn't strive for maximum efficiency but would just be efficient enough to achieve their goal.
I think you can use technology both for good and bad and whether AI is gonna be used for bad or good stuff is a question of how well global political relationships are established and whether people in power solve their self-worth issues and learn how to communicate properly
and probably also education and general happiness in the population and how available therapy is
Ryan's clearly articulated and empirically driven arguments were a refreshing contrast to Tim's cascade of vocab word-laden non-sequiturs.
damn and here I thought that Tim is always articulate! But granted - I'm just a random dude whose trying to learn so I guess I really wouldn't know!
@@hiadrianbankheadbeing articulate and spouting vocab word-laden non-sequiturs is mutually inclusive.
Meeting in the middle is challenging. I think Tim's pretty clearly making an honest effort. Any normal person is going to at times in a conversation like this, feel a reactionary, visceral need to contest things they disagree with & we're only human. Sometimes that will come across as less eloquent than we'd like. Tim's outlook is also generally much more in tune with the wider world. I personally would like to see Tim and many other AI specialists more fully appreciate the views and issues raised by Ryan and his peers, there's also a lot of value in exploring their intersection. A great deal of what we care about most is still hypothetical.
What may seem clarity might just be accessibility due to simplicity or appeal to familiar frameworks / tropes.
I would like to see what are the prompts used by Ryan.
github.com/rgreenblatt/arc_draw_more_samples_pub
@@MachineLearningStreetTalk Just had a quick overview of the code and it seems like clever Hans phenomenon to me. Apparaently all the reasoning is almost directly given into the prompt.
2:03:55 "A fast, dumber system can usually approximate a smarter system"
I agree - multiple sampling and filtering is a very useful tool for decision making - we don't need 100% certainty of the outcome, but if there are majority samples that reach a desired outcome, that action may be good.
It is reminiscent of Monte Carlo sampling - even if you don't know the underlying function, sampling helps to approximate the function already.
It depends on the problem. If it is like a Jenga problem, where removing the wrong block means you fail, then a lot of dumb entities will always fail. Like the AI that gets a good score on a bar exam, but if it made up a citation (as AI's do) on one wrong response it would fail the exam (or in court - get disbarred 😀).
can we just stop and appreciate how weird this conversation is probably gonna sound in 50 years?
I love this podcast.
The inverse is true of the reasoning suggestion.
Ryan is a smart guy but seems to have far less exposure to the broad set of ideas about intelligence that Tim and the mlst folks have been curating over the last couple of years. Put another way, he seems comfortably secure in what I would call a fairly standard computer science and engineering conception of intelligence, which is that the brain is a just a complicated machine and so a mechanistic systems approach to intelligence will get us there. This is how most engineers are taught to think about systems but I don’t think it stacks up to scrutiny when it comes to human intelligence. I think Tim was challenging him a bit here so I hope he is motivated to familiarise himself with the broader thinking in the field. Fun conversation.
I'm from a CS background so I'm not familiar with other perspectives on the brain. The CS view is that the neurons form a neural architecture and are trained through some process. The operation of the neuron, the architecture and the training process is mostly unknown, but all are purely physical.
What are the other views that you talk about? Can you give some pointers I can follow up on? Thanks.
Heh, I had the very opposite impression. 😅
But I mean aren’t llms just transformers trained on a shit load of text data. Their ability to process and generate human language regardless of accuracy is already crazy. Imho, on trained data, they surpass the average human in logical and creative abilities.
How can we definitely say there is more to human consciousness? Obviously transformers do not have constant sensory input from organs. Organs that have evolved over a few hundred million years because it is an intrinsic property of their existence to change.
A process that is so isolated from the features of us, carbon based life. A static virtual pattern recognition engine. The fact that it is able to emulate language understanding, logical reasoning, and human creativity so well tells me that maybe our sentience is not as magical as we previously assumed
Are the disagreements in the ML community around this concept or is it about brute force scaling LLMs wont lead to AGI but something eventually will?
I think that while it may be beneficial to familiarize oneself with the broader field, it's also important to remember all those 'broader' theories like vitalism and the four humors and the miasma theory that ended up being completely superseded by purely reductionist mechanistic approach.
Wow, Ryan is so based. Like it!
AGI and advanced 6G internet could issue in a virtual and augmented realities not unlike in Ready Player One
Genius! 😎🤖
Excellent!!!Thanks!!!
Guessing next token, creating divergent guesses attempts, steering a bit. Then judging and accepting.
Seems effective effective.
I disagree with Ryan's claim that "flops are flops." I can't justify my disagreement; I haven't worked it out such that I can "prove" the assertion.
But, here is my ad hoc reasoning:
Digital circuitry is designed to make chaos/entropy irrelevant, null, and unimportant to computation by a method that statistically cancels out the influence of small-body dynamics, the eliminates the butterfly effect, that locks certainty in a bell jar.
Analog circuitry includes all of the uncertainty with statistically reliable properties, but isn't pure random entropy. Its rather like a complex melange of collective behavior that results in something trustable like a gas laws applied to computational space, a metaphor for emergent properties like temperature and pressure, for which laws can be shown at a statistical level without needing to know in precise digital terms what is happening at the individual atom or circuit level.
I don't think neurons are purely analog, nor purely digital, but include something like quantum superposition of the two (even if it's not technically pure quantum superposition, it may only come into play at a crucial point, and the experience may feel something analogous to coming out of Plato's cave).
If it is something somewhere between digital and analog, or a combination of them, or something like that, I very much doubt that it would be well described, even as a metaphor, as a quantum superposition of the two.
Many people are, I think, often too eager to describe something as being like a quantum superposition.
A quantum superposition is a fairly specific thing.
It is a linear combination, a weighted sum. And, specifically, one for which unitary time evolution acts independently on the summands.
Tim has some quite unconventional views that we only really get to around 1:19:00. Is he the host, or doing his own original research?
We've had cheap human level AI for years: Amazon's Mechanical Turk. It didn't take over the world.
Shouldnt abduction be better defined as explanations? And explanations as shared canonical and hypothetical declarative accounting procedures in addition to analytical rules and boundary laws for in context choices?.
this is better way to say it
Wow, this guy is crazy. One thing I wish he pushed on more was the scaling law stuff. Talking about the limitations where smaller models do worse than larger. But in relation to what? That's qualitative, not quantitative. In AI, we need to start thinking more quantitatively. I know the rage right now is language, which is literally qualitative, but that's why people think they are performing so well. It sounds good, but is it good? No one knows until there is a way to measure it. ARC challenge reveals this problem. There's one specific answer and LLMs can't do it.
Tim's position seems to be slowly shifting over the months and years towards accepting that LLMs might be capable of reasoning and agency. Ryan was very insightful and reasonable and probably nudged Tim a little further along that path too.
Not at all - I just conceded that in principle you could memorize the universe in an LLM. For all practical purposes this is impossible.
@@MachineLearningStreetTalk imho it's too early in the journey to say with certainty either way. Unless you're approaching this from the question of fundamental computability - eg: Roger Penrose thinks that consciousness requires some special physics that humans haven't yet even imagined. The opposite end of that would be Stephen Wolfram's view that the universe is fundamentally computational (and consequently, everything therein).
I don't think that systems need to be conscious to be agentic to a significant degree, and the degree of agency is probably what's going to matter more generally, going forward.
For what it's worth (as an internet stranger) I think that AI will gain enough agency that fierce debates will rage about what they should be allowed to do, and just how much agency a given entity needs to possess to be considered an individual... I can only imagine the controversies of potentially strongly agentic AI acting in society.
With that in mind, I think that it is wiser to assume that they are likely and make plans to address this possibility rather than dismiss it and not be prepared, in line with "better to have it and not need it than need it and not have it".
@@kyneticist I didn't mention anything about consciousness - we have given that a lot of separate treatment on MLST and in particular you might be interested in the Chalmers, Solms and Shanahan interviews we are about to drop on the podcast (and the Goff one we recently did here on YT). We also had Wolfram on a couple of times. My personal position on that is the "illusionist" one. I agree with you that Agentic AI is something which would be a huge cause for concern if it were possible. Right now, current "agentic" AI systems show all too clearly that these systems are glorified databases (to be fair, unstructured approximate retrieval systems) and just get stuck in loops and destabilise / converge within a few iterations - they only run "deeply-autonomously" if the prompter gives exquisitely specific instructions which is the opposite of any characterisation of strong agency. So the evidence today is that we don't need to be remotely worried. As soon as this changes, I will update very quickly to a similar position to you. Understanding the technology and the background computer science deeply, I really don't see this happening - if anything I see GenAI as a bubble which will burst within 2 years unless there is a dramatic improvement in capabilities to justify all the money being spent on it.
Thanks for your reply. I've seen a number of your videos & appreciate your work. I also agree with your perspective on the illusionist - current AI has come a long way in a short time, but it has very clear limitations, especially for people pushing its limits - as it is right now.
I think people generally agree on the current state of the field; As far as I can see though, if you examine discussions between opposing views closely, the frame that each "side" is working from is the greatest source of difference - current capabilities vs future capabilities. As an alternate analogy, some of us seem to only see the forest while others only see the trees.
Consider what Microsoft is doing with CoPilot - they have access to the biggest & best AI in the world and they're investing heavily in more. CoPilot is being trained by hundreds-of-millions of people on how to do their jobs. Adobe is aggressively building AI to emulate the work of scores-of-millions of art professionals. They won't need to be particularly agentic to be very successful at being vastly better than most humans at most of the tasks humans use them for.
Between just those two companies they'll be converting _a lot_ of white collar jobs to AI in the near term, without needing to be particularly agentic (though more will obv be better) nor much more advanced. Even if GenAI hits a wall they (and many others) are well positioned to branch out with world-shaping current generation AI.
Maybe we're stuck on LLM's and derivatives, but there's such extraordinary interest and powerful people driving for more, it's hard to imagine that GenAI will be a bottleneck for long.
Can anyone recall which service gives multiple completions for ~free? I'll have to rewatch and find and post the answer here later if not. A quick Google was fruitless.
I think he mentions in the blog about the solution and I think it was a flavor of gpt4 not 4o
Please add Ryan Greenblatt to the title
Unfortunately doing that is super bad for YT optimisation, we would be happy to in a month or so - blame the algorithms
Really? Promoting gambling?
It's not trading, it's gambling disguised as trading like binaries, which it basically is.
I’m not sure I see the difference.
Is it gambling? Sure, in the sense that betting on horse race outcomes is gambling.
Is it trading? I don’t see how it isn’t.
Horse race gambling reveals “market beliefs” about the probabilities that the different horses will win the race.
Trading futures for goods reveals “market beliefs” about the future value of a good.
Same thing with trading on whether an event happens, revealing the “risk free probability” the market assigns to the event.
Now, for simple slot machines, scratchers, lottery tickets, these things have no appeal to me. The odds are known, the betting doesn’t reveal any information. These things are *purely* gambling.
But, gambling about events when people disagree about the odds, this in my mind can serve a useful purpose. It can be a means by which people reach a consensus about what probability a given event has, and incentivizes people to reveal information that is relevant to the topic. It rewards the knowledgeable.
Now, if one doesn’t think that the current odds are wrong, I would say that one probably shouldn’t bet/trade on that market.
Gambling just as entertainment,
well, it is better to do that just with chips. Why waste the money?
But, if this sort of prediction market is bad on account of being gambling, then I think that would probably imply that all day-trading and the like is bad on account of being gambling.
And… maybe it is?
It seems to have economic benefits, but, I don’t think I should rule out the idea that something is immoral just because our economy is partially based on it.
(E.g. Maybe it should be illegal for banks to charge interest? idk. I think this would result in a significantly smaller economy, and likely lower average “quality of life”, but, if morality requires it, that takes precedence over those things.)
amazing chat btw
What's the diff btw physically possible and possible in principle?
Possible in principle - we could imagine a world where we had unbounded resources i.e. time, computation, data, hardware etc and it might work then
Physically possible - we could create a computer (or effective computer) using physical material (i.e. silicone or other) which was powerful though to run Agential AI
Practically possible - we could build such a computer within economic, hardware, scale and other practical limitations
1:19:57 what if you had a community of AI's?
Another classic - the clearest clash of world-views yet perhaps? 😉 LLMs with scaffolding may or may not get to AGI, but there's no doubt they will accelerate progress in that direction, simply because they're useful research aids. Also no doubt that even with no AGI, LLMs are still a transformative technology. Remarkably, the ' Kalshi' prediction market is giving a 38% chance of OpenAI achieving AGI before 2030!
Can anyone point me to the documentation about caching a prefix and get a bunch of completions? Around 26:30.
Check out OpenAI’s API documentation. Under “create chat completion” there is a parameter named “n”.
Without integrated multimodal embodiment, not AGI ...
Machines had to learn to integrate 'sensory/perceptual incomes' and embodiment from gravitational, pressure, temperature, chemical reactions, photonic fields, electromagnetic fields, space, time, space-time, time-space, implosion, vibration, explosion, resistance, density, viscosity, friction, proprioception, etc etc etc ...
Humans are not enough as UIs for translating embodiment qualia similar to data to the system, too much subjectivity.
General Intelligence is not just passing human symbolic tests and getting a couple of abstractions ...
not embodiment, just a system regurgitating data and talking apes anthropomorphizing the silicon/software connectomes.
The map is not the territory.
Why do we feel we can just jump to agents that can do research and we haven’t even been able to develop agents with the intelligence of a mouse?
Horse before cart, but I guess the assumption is that we'll keep improving on LLM performance.
I think embodiment is relevant, but really not the hard limit on cognition suggested by the host. Current robotics allow embodied learning of the kind needed for 'true' cognition. Being socially accepted is the barrier. Robots need to be accepted as 'a thing you talk freely to', to be intimately trusted with aspects of knowledge. To be 'nurtured'. But most people with a cat know that this is not a catch22 blocker. You do not need to be able to talk at all, to be talked to. Cute and characterful is enough. Cute and characterful robots will nurtured to become truly agentic, well before giant super computers. Strong cognition and strong agency are related but separate.
Bruh I thought it was based on single shot responses - not running 100 programs with human help
yeah, like, it's like that it's like, something like, like, yeah, okay, like, sure
if you believe in Girard's mimetic theory of social conflict then we should not be training AI to mimic human desires. AI in the financial space such as program trading can be devastating for equality and conflict.
1:32:10 Could you provide some resources to learn more about what you said at this time?
Thank you!
arxiv.org/abs/2009.06489 [The hardware lottery]
www.amazon.co.uk/Why-Greatness-Cannot-Planned-Objective/dp/3319155237 and our special edition on it ua-cam.com/video/lhYGXYeMq_E/v-deo.html
Couple of starting points
@@MachineLearningStreetTalk Thank you so much
Ryan is very smart about computer science but when he ventures off towards economics and hardware it's clear he's not educated in the basics. There are 3 factors of production in economics: land, labour and capital. AIs will have only labour. This will limit their production. Also, hardware engineering is large and slow, this won't change just because AI enters the picture.
You don't think the 200 million files of the llm models training on different data sets is going to interact some kind of AI conscience realm then your lying to yourself
A system can be conscious but not intelligent ... that a system can perform a lot of abstract symbolic recursions about those recursions to be linked to what the system register as its sensors and/or UIs to its otherness ... doesn't mean that that 'conscious' system is intelligent ... just that got the potential for generating abstractions as fractilic metalayers around a sense of self (integrated embodiment ) and its othernnes as data inputs in real-time ... ( current LLMs are not doing that , just playing games of languages and forming biased connectomes by modeling architectures made by humans )
Inteligence is another thing ... Intelligence is what makes a conscious system to become integrated with its othernnes ... consciousness is not the big deal, the toys had some sort proto-consciousnnes throught correlations of chaotic atractors in n-space ( binary operations conectome potentials) ... the big deal is AGI.
I really appreciate your channel. I think you could benefit from being less editorial. While your perspective is interesting, it can get repetitive. Look at how Kantrowitz's channel lets guests' thoughts shine without much editorializing. This approach could enhance the depth of your content I think.
I appreciate the feedback, but I feel I let Ryan’s ideas shine through quite a lot actually. I never interrupted him and went out of my way in the edit to show references for every single thing he was speaking about (In collaboration with him) - steelmanning his case even though I didn't agree with it
@@MachineLearningStreetTalk shure guess it's a fine line to ask questions that originate from a different opinion and actually stating that opinion or more precisely, arguing for it. But again, take a look at Kantrowitz's channel, think he does a good job with that and see if that style suits you or not. Anyways, good job!
@@jmstockholm I will look at it, and I am acutely aware that host's opinions get boring quickly - I am bored of my own opinion at this stage. I will take that on board, thanks
For Ryan’s simulated brain to run on a computer, you’d effectively need to simulate the universe which isn’t practical or plausible.
Yeah, I agree, he's super smart.
Why would you need to simulate the universe?
I don’t see why.
If you mean that to simulate a brain with the highest theoretically computable fidelity, you would need to simulate the particle physics and such for the region of spacetime the brain occupies, and that is far from achievable, then sure.
But I don’t think simulating part of the universe requires simulating an entire universe.
@@drdca8263 great question!
This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain.
Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”.
DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels.
Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics.
Additionally, the brain doesn’t do backpropagation.
@@drdca8263 great question!
This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain.
Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”.
DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels.
Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics.
Additionally, the brain doesn’t do backpropagation.
@drdca8263 great question!
This continues to be an unsolved problem in robotics(i.e. sim-to-real gap) and an extremely computationally expensive process for even the simplest high resolution fluid/soft-body dynamics simulations, let alone the brain.
Recent attempts to simulate even small fractions of a brain “took 40 minutes, to complete the simulation of 1 second of neuronal network activity”.
DeepSouth is a recent attempt to simulate just the brain and it’ll require 228 trillions synaptic operations per second - that’s not accounting for the roughly 50 million sensory input channels.
Further, the brain isn’t a feed forward network so a pile of 100 billion neurons (and synaptic connections) would still need to be architecturally organized to model brain dynamics.
Additionally, the brain doesn’t do backpropagation.
Gpt 4o is absolute muck
It’s good at some things, but I agree.
And what can you do.
Do you want a test for measuring LLMs cognitive skills and know that they are not something near to AGI but just a system that regurgitates data?
Play this game with those LLMs ... you will spot all the holes in the data sets, the biases, the developer's biases, the limits of the system knowledge, its lack of self-awareness, the lack of forming dynamic models about itself, and other minds, its lacks of creativity, its lacks of long term memory, how poor their concepts are networked in their connectome, how it can not generate a new programing language paradigm to itself, etc etc etc ... and principally, How talking apes project what they believe what they are into a toy ...
The test is called 'The Game':
## 'The Game'
Rules and Minimum Requirements for the Game
1) 3 people (A, B, C)
2) A, B, and C will be the vertices of an equilateral triangle.
3) A, B, and C can converse about any topic that arises or that they like, but they must follow these rules:
Syntactic/Semantic Rule
a) None of the first-person singular pronouns can be used (I, Me, Myself, Mine, With Me)
a.1) Consequently, the verb conjugations associated with them cannot be used either.
However, if A, B, or C want to make a self-referential sentence (for example, "A" wants to say something about "A"), they can simply refer to "themselves" using the third-person singular, plural, and/or passive voice pronoun, conjugation, and/or name...
(for example, "A" is called Miguel.
Miguel (A) wants to tell B and/or C something about "himself" (Miguel)...
Then, Miguel (A) can talk about "Miguel" by talking about Miguel.
Space-Time Perceptual Rule
b) If "A" speaks, B and C look at each other and do not see A. (A speaks and observes the space between B and C)
If "B" speaks, A and C look at each other and do not see B. (B speaks and observes the space between A and C)
If "C" speaks, A and B look at each other and do not see C. (C speaks and observes the space between A and B)
Self-Correction and Cognitive Assimilation Rule
c) Practice rules a and b until achieving spontaneous fluency in the "trialogue"...
Participants are required to inform the group when the rules are not being followed.
Scoring rule:
The one who spot the mistake wins a point, the one who makes the mistake lost a point.
.....
These are the basic rules of "The Game"...
Once assimilated, there are more variants.
Public Domain
The Game is considered to be in the public domain, existing as a thought experiment or communication exercise.
Origin Unknown
The exact origins of The Game are unknown, but its core mechanics might have been independently discovered throughout history.
Let’s solve a mystery with a question
A question that is still mysteriously absent from the discussion
This is like 1929 Germany
🇩🇪
A proud nation that had the answer
so much α
Once the renderings become spatially 4D, we will become inside our own created AGI.
Wut
You are already are, read neville goddard
@@philipfisher8853 Soon, we will become components in a mechanically managed system rather than an organically managed system. What this means is that AI will be simulating our actual reality, not some sub-domain space.
4D space isn’t mystical or inherently mysterious. It’s normal.
It isn’t intuitive for us at least without a bit of practice (though even with a lot of practice, the fact that our brains are limited to being in 3D space probably limits our ability to perceive a detailed 4D scene in a way like we perceive detailed 3D scenes), but I don’t think that really indicates that 4D is particularly special.
Of course, there are a number of things that are kinda special about 4D space. But there are also special things about 3D space, and 2D space, and 1D space.
There are even a few things special at higher numbers of spatial dimensions (especially 8, 12, 24), but generally after 4 or 5, things start getting pretty similar.
In any case, none of this should suggest that something being in 4D would have much of anything to do with AGI.
@@drdca8263 I wasn't clear about what I meant. Once the LLMs are no longer processing linear, written language and 2D images and are instead processing information using 4D spatial representations instead, our entire universe, as well as all other universes, can be simulated within our universe's timeline. Once you can simulate the entire universe, you no longer need to create a "Matrix" type simulation for people to be inside of. All you need to do is apply the manipulations the AGI is applying to its own models to the actual universe. In that sense, you would be able to design time.
I did not suggest our universe is 4D, or that 4D is somehow special. It may be, but I doubt it. The universe can definitely be experienced in other ways, such as 6D (4D non-Euclidian space / 2D planar time) 8D (5D non-Euclidian space, 3D spatial time) and 10D (6D non-Euclidian space, 4D non-Euclidian time) which can all be experienced within the mind due to the 11D topologies that are created by neuronal connections, but which require altered states of consciousness to access. A very interesting state of mind is 2D planar space and 0D time. It is difficult to access this state, as it requires consciously turning off several regions of the brain to allow other connections to surface. This can be accomplished in many ways, including meditation, yoga, and psychedelics.
I agree that there is a self-similarity to different dimensions of space, but they are also very different experientially, and can be extremely disorienting without any prior understanding of what each additional dimension brings to the experience. So my theory is this: If this experience is possible within the mind, then it can be created externally and projected into the mind without the need for a physical wire or connection to an external device.
Interviewer has some bizarre views. But the guest was good!
yeah pretty weird
What do you mean? He’s EXTREMELY knowledgeable and is speaking textbook 4E cognitive science. But if I had the chance I would ask for clarification and detail around:
1. Opinions on Turing Completeness requirement and feasibility
2. Working definitions of “Self” and “Agency” - maybe free will as well
@@richardsantomauro6947 The central weird claim is that AIs with "real"/important agency are impossible *in principle*. Sometimes the host doesn't actually take this claim literally though. And you can imagine flipping between two versions of this:
Version 1: "Building Skynet from the Terminator movies is impossible because taking over the world is a complicated real-world task that it's impossible for AIs to solve."
Version 2: "Okay, you could build Skynet and it could take over the world, but that wouldn't be *real* agency, it would have to inherit agency from humans. This will comfort me emotionally as I am shot by a terminator."
Yes, what the hell is he talking about when he says he does not believe agentic AI is physically possible.
@@marwin4348physically possible to implement/run on computers
GPT 0 is crap
50% is literally a coin toss. And it requires all this extra work by a human to brute force it with the underlying model. Really not impressive.
These aren't yes or no questions
@@Alex-fh4my Still, the chance of getting a question right is a coin toss. And I must emphasize how much work is being done on the part of the human to achieve this result.
@@Dri_ver_ 🤦♂️
@@Alex-fh4my Did I say anything incorrect or will you just keep coping?
The likelihood of correctly guessing a coin toss is 50% but the likelihood of a truly random guess being correct in a test like ARC is tiny, since there are so many possible wrong answers to each question.
What you’re implying here is that the probability of rolling a 3 with a six faced dye is the same as getting heads on a coin, which is false. The ARC challenge is like a dye with an enormous number of faces.
If you dont agree, try completing a couple of arc questions by filling out the grid randomly and see if you get roughly 50% correct