It’s fine to have your own definition of AGI. I think o3 is smarter than the average human at most knowledge-based work and that’s good enough for me to be considered AGI. We have to stop moving the goalposts at some point :)
but isn’t that already what current LLM does? knowledge base? AGI is more than that. by definition, it should be able to solve a novel problem that doesn’t have prior knowledge against like actual human beings.
@@Kutsushita_yukino No, not a knowledge-base. Ability to do knowledge-based work. LLMs up to now have been able to store information, as if they were encyclopedias, but less reliable. With o3, it is able to reason about mathematics problems it has never seen. Mathematics, though you may not think it, is the science of generalization. So in essence, o3 has learned to generalize better than most humans, math being the most direct example of this. The hard part is mostly done, building agents and inventors out of this is just a matter of clever software engineering and infrastructure. Give it a year or two and you will start to see the effects of this AGI coming to light. Before this, we weren’t really sure if humans would be surpassed at reasoning, but the trajectory seems clear now that it is achievable with current technology.
Agreed. People seem to think AGI must be better than the average human at every single task. Eventually we may reach a point and say "Nope, it's still not AGI because it doesn't poop"
I believe they achieved an early form of AGI. Pre training has hit a wall but test time inference is only at the beginning. They do not need any more scientific breakthroughs besides more powerful and efficient GPUs. If o3 is not AGI then I wanna see your reaction to o4 or o5.
Not sure if I believe that. If we want AGI we achieved the first part of it. For example humans have there lizard part of the brain doesn't think it's automatic. it takes cares of body functions like breathing too telling the body your hungry and so on. o3 and most Ai's are that part for now. If you really want AGI we gonna need AI's to use LCM's and not LLMS secondly to achieve true legitmentiant AGI Ai that use LCM use pertaining but can learn after while being made and used within a quantum computer with those breakthroughs AGI will happen. How can a classical computer achieve AGI? The human brain is quantum our brains don't run off 1's and 0's and no duh but my point is are have general intelligence because our brains are quantum they use electrical signals. I don't care what you or anyone says I'm not drinking the kool aid. I think achieving AGI is cool but how can drive a car without wheels.
@Timely-ud4rm Let us not make the mistake of thinking that we're building copies/clones of the human brain. The architecture doesn't matter, only the result does. Quantum computers are meant for solving different problems and have nothing to do with Ai. A.I doesn't need to have identical functions to our brain, they dont even need emotion, they just need to produce the results that we need.
Regardless if it is AGI or not. I think your thought exercise on this is fantastic and helps me (simpleton in the Ai realm) get a better grip on what is happening and what is coming down the pipeline. Thanks, and always appreciate your content.
We might have reached AGI when your boss can't distinguish if you or your computer has done the work while you'r at home. But this might also be the moment when you get fired.
The second they realize they can have "someone" do the work for the same cost... you're gone. No benefits, sick days, retention, training, HR violations, bonuses, raises, office space, scalability, 24/7/365.
O3 is definitely impressive, but one number caught my eye: $1000 per run. It's not because it's expensive-I believe that cost will gradually come down-but because of the reason behind it: a heavy computation search across various paths. You might recall AlphaGo, which used Monte Carlo Tree Search during inference. I'm not sure what type of solution search O3 employs, but it seems to involve some of these heavy computation methods. Could we say it’s a kind of ‘fake AGI’?
You're Agi is on the upper extreme n more when Asi hits in about 2030-32 and fully takes over. I don't believe it has to be fully integrated just seeing integration across 90% plus of fields which is only year or two. And remember what we see is not what we see. They are probably a year ahead of o3 n on o4 at least in house and integrated in all their work. That's why these frontier models are 10xing in 4-5 months now. Was 6, six months ago too. Agents coming online like image generation did changes the world and thats right around the corner
I've been making a lot of looses trying to make profit trading. I thought trading on a demo account is just like trading the real market. Can anyone help me out or at least advise me on what to do?
Trading on a demo account can definitely feel similar to the real market, but there are some differences. It's important to remember that trading involves risks and it's normal to face looses sometimes. One piece of advice is to start small and gradually increase your investments as you gain more experience and confidence. It might also be helpful to seek guidance from experienced traders or do some research on different trading strategies
As a retired engineer I dealt with that problem in creating decision aiding algorithms and planners by introducing what we called quality graphs. A quality graph map a truth value, e.g. something that can be assigned a value based on truth such as weight, time, size, distance, color, etc. to a goodness factor. Also it was often useful to limit goodness between 0 to 1 or -1 to 1 so that one could multiply the together to get composite goodness value based on several criteria. Thus they can act much like fuzzy logic membership functions. The quality graphs might be computed or just arbitrarily defined heuristically based on human judgment. These might be multiplied by a base value say from 1 to 100 and thus function much like expected values in probability. Thus we might have an composite goodness = .5 x .7. x 100 = 35 or the like.
BTW this above assume that the composite function is decomposable, i.e. F(x, y) = f1(x) * f2(y) and such. But in not one might be partially decomposable such as F(x, y, z) = f1(x, y) * f2(z) etc.
I make a distinction between AGI and "being conscious". AGI can be achieved without consciousness, ar at least without consciousness as we experience it as humans. For us, each problem or situation has a value and/or emotional state, which has a sort of meaning. We feel happy or sorry about a specific situation or problem precisely because we translate the outcome into an emotional state which resonates with the life principles embedded within us. Perhaps consciousness permeates everywhere ( if we assume a "field hypothesis" ), and therefore an AI could be as well conscious to a certain degree; now let's consider that life as we know it took several billion years to achieve the emotional geometry that we share among living beings. Something as simple (for us) as the fear of dying or the joy of having a child would be emotionally nonsensical for an AGI, thus it would "understand" very accurately why we feel fear or joy. And - I might be wrong, but my intuition is that it's not a matter of time needed to train an AGI to acquire emotions: we could give an AGI all the time and computing power available in the Universe, it will very likely ignore the process that led to our emotions.
For me, AGI will be achieved when it can pass a threshold that Demis Hassabis outlined: AIs today can master chess, but can they invent chess? i.e. A game where the rules can be learned in a few minutes, where a match typically takes less than 30 minutes, but to master the game may take many years, and most people will have fun playing it.
I agree that o3 mostly likely isn't AGI. It's clearly a very capable LLM, but the fact it's still squarely in LLM territory with most of the usual limitations is problematic. I feel this technology needs to progress much further into agentic territory before we can even start to consider it "AGI", whatever that means at this point.
When is it AGI? The larger problem may be that we don't have a good definition of intelligence or consiousness. Seems hard to derive useful tests, without even knowing what to test for. Clearly, the Turing test (and anything like it) is nearly useless.
Your point about general intelligence is well-taken, IMO. Actually, the experiential context and event spaces and the "goals" are much more well-defined (while still almost infinitely variable) for vehicles and FSD (or humans) than will be the situation for Optimus robots in homes and workplaces or social situations, outside of controlled spaces like factories or space environments. We may now have AI, but still need to start talking about ASI as artificial specific intelligences, while on the road to AGI.
Most SWEs do terrible on codeforce though lol... As far as AGI goes, humans don't even have general intelligence so humans are a terrible comparison. Can humans generalize? Sure. But the threshold for AGI is supposed to be an AI that can do something the average human can do. Obviously this would require embodiment. If I can take a robot with O3 mini and drop it into a factory and tell it how to do something and it does it without any training on it then it's AGI. Expecting AGI to be able to solve things people have a hard time with isn't what AGI is for that's ASI. STOP CONFLATING THE TWO!
If o3 is AGI, that also means that Google will have achieved AGI before OpenAI, because they already beat several Phd benchmark with Alpha. o3 it is more AG hype than AGI, because they 12 days Christmas run failed to convince people that they are still on top. so they are going all in on this last bluff.
It's much easier for probabilistic models to satisfy some fuzzy goal where the answer is more or less correct according to some basic intuitions than the highly precise solutions to novel problems that o3 achieved. To achieve better than human results in the examples you give, last decade's specialized models already sufficed. You can easily just train an image classifier for ripe vs unripe. It isn't verifiable in a definitive way, but you can probably get 99.999% accuracy where the human will probably be fine with 80%. If you want to solve hard problems in science, that isn't going to cut it.
You're talking about specified problems, o3 is meant to solve things its never even seen, how can you not understand that? I swear l'm gonna lose my mind seeing so much nonsense from youtuber comments
@@shirowolff9147 I used the examples he gave in the video. Ripe vs unripe. They're loosely specified problems that only need fuzzy approximate solutions.
If AGI is defined as acting exactly in every way like a human then it is human! Human intelligence is divided into different types of intelligence even though we call it general. So maybe we need sub types of AGI: STEM AGI (for scientific discoveries, etc), Economic AGI (such as helping to run a company), human interaction AGI and probably others.
You know whats the real AGI challenge? would you enter into a bet with the ai to do a challenge (of your own choosing)? winner takes all --> your money, your home, your car, everything. chickening out? well, you might be competing against AGI.
It cost $300k to get that answer? I hope they discover huge algorithmic changes that shift that to pennies, because just throwing compute at the problem isn't good answer.
Dr. K, I agree o3 has not achieved AGI though my rationale is a bit different than yours. I believe knowledge of self and world, at least to the extent understood by the basics of proprioception, is required for any claim of artificial general intelligence. Can o3 make a cup of coffee? Can o3 run a 5K in under 20 minutes? The answer to both questions is No, and the reason is that o3 has no concept of self and world and certainly no way of physically acting in that world. In addition to the problem of proprioception there is the problem of logical analysis that goes beyond the tidy constraints of digital symbolism. There are, for example, scientific tasks that require more than book knowledge. You and I might be able to conceptualize a new invention, but patents are not granted for concepts. The idea must be 'reduced to practice' to be awarded a patent, and it is here that science and engineering meet to create an actual physical product, such as a more durable type of rubber or a pill that prevents Alzheimer's Disease. My expertise is in purification process design from secondary metabolites derived from plant matter or microbial broth. This is a rare specialty nowadays, but 100 years ago was commonplace. The reason only perhaps 200 people in the world can really, truly do this kind of work is because of automation and its constraints. The fact that unusually strenuous logic is required, therefore requiring a high human labor component, is also the primary factor driving up pharmaceutical drug costs. Drug companies willingly pay hundreds of millions for research, but only if a small core of people are required for the work. Here is a real-world example problem in pharmaceuticals, something that current pharmaceutical companies cannot achieve and likely will not be achieved by ASI until well into the future: Design a purification scheme that extracts 95 percent of all geldanamycin produced by a highly efficient microbial system and purifies the antibiotic to 99.5 percent purity with 80 percent recovery. Literature and patents on this drug demonstrate an overall recovery of 10 to 15 percent. I achieved 83 percent recovery--from microbial extract--of this very expensive drug. In order to achieve this very high recovery, I had to have direct physical understanding of practical constraints and capabilities of hundreds of types of equipment at bench, pilot, and commercial scales. This knowledge is generally not the kind of thing that is ever written down, and thus a computer, limited to what it can glean from literature, will almost certainly be incapable of delivering the example antibiotic with much more than 1 or 2 percent recovery. Even trying to duplicate, in the laboratory, the patent claiming 12 percent recovery will be beyond the computer's capabilities. Superintelligence is not required for problems of this nature, but these are demonstrably also the type of problem that cannot ever be solved on paper or from within the constraints of digital awareness. Real world knowledge, and the ability to physically act in the world, in addition to expert-level knowledge of natural products chemistry, are absolutely required in order to deliver a commercially viable (highly efficient) pharmaceutical purification process. Lest you think I offered a particularly difficult problem, I can assure you the isolation of geldanamycin is child's play compared to many current problems in natural products pharmaceutical chemistry. You might look up the isolation of paclitaxel, for instance, whose very high cost is due almost entirely to the exceeding difficulty of purification. I created the highly efficient Hauser Inc. third generation crystallization sequence for paclitaxel that decreased loss from the normal 15 to 20 percent (second generation; total recovery 80 to 85 percent) to 3 percent (third generation; 97 percent recovery). That is, I delivered a five-fold decrease in loss, saving Hauser several million dollars. Again, you will not find any of the methods I used described in the literature on crystallization. I had to innovate based on keen understanding of chemical phenomena unique to taxanes and the constraints and capabilities of real-world equipment at every scale. Computers simply cannot do this, and probably will not be able to achieve deliverables like this for many decades to come. PM 2024
True intelligence - we are not even in the same ballpark. I've been watching our 15 month old grandson playing - and we were watching a hockey game and he went out to the playroom filled with toys from the other grandkids and found a hockey stick and hockey puck and brought them into the living room and began to watch TV and move the puck around all by himself and he has never played with the toys before that's true intelligence and we are not close to having something even remotely close to it.
Yes that’s spontaneous intelligence, AI doesn’t have that yet because it’s not embodied in the world like the child that can experience everything. It’ll take a few-several more years with humanoids reaching the millions and all their combined networks updating each other autonomously for them to mimic those kind of experiences, and AI is different in how it learns because it doesn’t have biological components to feel/smell/taste or other senses like we have. It could also take building some biological body it can operate in or nanobots training in humans and looking at everything through our eyes to eventually reach something truly intelligent, who knows. Could take decades if you want to compare it to our type of biological intelligence. But it’ll still be more intelligent overall in different ways within a few years.
@@uber_l But to think your iPhone will be self-tinking and do what we can do is what AGI is all about. AGI is probably 100+ years away - we are just kidding ourselves if we think our existing computers can think
There's no way it can read 100K line of code and find the bug in it and submit a fix so that it passes our rigorous code review and nightly regression tests. My job is safe for a very very long time. This thing couldn't even travel from Paris to New York, it wouldn't know how to pass airport security, it wouldn't even find the check-in counter.
Are you saying o3 is incapable of answering an imperfect question that has no correct answer? That it is incapable of considering a variety of potential outcomes for a given scenario and deciding which outcome might be better than the other? Even 4o can make an educated judgment call on the ripeness of bananas to buy if it is given the timeframe to consume them, their use, the temperature they are being stored at, etc. And it can arguably do it better than we do. So I may be missing your point. Are you saying a human who consistently makes poor decisions does not have general intelligence? It sounds like you’re saying AGI will only be achieved when it can make decisions that humans are terrible at making themselves. And by that standard of AGI, we are arguably already there. Or are you saying that we don’t reliably have any way of assessing AGI, thus it has not, and cannot be achieved. Simply because there is no way to prove it.
Yeah, just moving the goal posts again. The rules of the game had been established 5 years ago with ARC. So rather than to make up new rules so as to get clicks, maybe augment the name of what you are seeking to something like Practical AGI to describe a more analog real-world machine. I would love to see that too because the AGI label is just a ridiculous thing to argue about until ASI raises its wand and puts a stop to the nonsense. Won't be long now, and there won't be any debates, it will just show us.
Arc AGI isnt an authority on AGi, they just propose a method to challenge AI development on what they think is AGI. That doesn't mean that any computer can't beat it, Google AlphaGo and other Alpha products have already proven that well-trained AI algorithms can beat several benchmarks, does that mean that Google has achieved AGI before OpenAI? No, being called AGI requires more than beating a human benchmark.
FSD is solving a similar problem. Driving has no one right solution either, and often even humans debate what would have been the correct move. I think FSD is closer to AGI than o3.
I'm not sure AGI is about consciousness. Maybe it's more about taking all (or 95%) of human jobs. When it's done, it won't really mater to know if AI is "truly conscious", I guess.
Since OpenAI is still dependent on its partnership with Microsoft to afford increased compute power, and their contract stipulates that they must sever ties with Microsoft once they achieve AGI, it's not unreasonable to speculate that the reason they skipped the 02 model might be because it was too advanced. They may have been forced to scale it down to avoid acknowledging that it met AGI qualifications.
That bulls***. If they achieve AGI they will have 10 trillion dollars of investment the next day of announcement with all the computing power they need and Microsoft wouldn't do anything to stop it. They dont have any AGI thats why they dont really claim it and let fanboys ignorantly speculate on it because it produces more hype.
I mean, its not AGI until you can give the computer a real world engineering task. Computer, design an 20% more efficient solar cell. Computer create a safe and functional fusion reactor. Computer make a 30% better rocket engine.
I comprehend your perspective. The criteria for determining whether an artificial intelligence system has achieved Artificial General Intelligence (AGI) status are not universally standardized within the industry. It appears that with each significant advancement in AI models or systems, the benchmarks for AGI are recalibrated, perpetually pushing the goal further. In my estimation, a definitive test of AGI capabilities might involve integrating the AI model into an untrained humanoid robot or analogous platform. The AI's ability to spontaneously compute and execute control over this novel "vehicle" could serve as a compelling indicator of genuine AGI attainment.
@@SkitterB.Unibrow all good. No judgment. Just can’t remember the last time I read something that was so obviously written by AI. It’s like your English is so bad you don’t even know when something sounds so off lol. Right out the gate: “I ‘comprehend’ your perspective.” Lmao. No one says that. It sounds like a dumb person trying to sound smart.
I think Sam Altman mentioned that by 2025, all benchmarks will be saturated. I believe the criterion for evaluating AGI can be set as the point where humans can no longer propose benchmarks that effectively distinguish AI capabilities.
In the real word, first you have a perception, then you have the word for the perception. In the case of AI, first you have nothing, then a word, then nothing again. This is the issue with AI, in terms of true 'understanding'. AI's use words without any understanding of the real-word objects to which the words refer. In a sense, humans have already 'given the AI's the answer' in the sense that humans humans have correctly mapped the words onto the real word through sensory experience. Thus, there is a latent semantics built into human languages that AI's are tapping into. They don't need sensory experience, because the sensory experience is built into the language. They need only use language syntactically correctly, and in a probabilistic fashion that achieves a human-like performance. But the true 'intelligence' is ours, not the machine's.
As we progress all our models, so do we move the goal line. General intelligence is not our every day life challenges and random things we have to relate to. An Orangutan has general intelligence. Some birds has general intelligence. Is this highly evolved general intelligence? No, but it is general intelligence. We have probably gone past general intelligence a while ago. Using benchmarks to measure these models are a good way to go about it. To measure human intelligence we use WAIS tests and FRTs. Even with scores below 2 SD, we are still considered to be general intelligent (that would be something like an IQ of 75 in some tests). We are moving towards ASI. Is o3 ASI? Not yet, but it is a quite capable AGI.
If this is AGI, that also means that Google had achieved AGI before OpenAI, cause they already beat several benchmark with their Alpha's products. o3 it is more AG hype than AGI, because they 12 days Christmas run failed to convince people that they are still on top. so they are going all in on this last bluff.
The term AGI now means some very different things. Using a definition where machines outperform humans in answering questions with definite answers. Find the algorithm which does X. It is getting very good at this narrow kind of question answer. However, there are some glaring limitations. a) No ability to learn in real time from example. A conversation may be able to temporarily learn something, but lasts only as long as the conversation. The models are static. They are static for fundamental reasons, not 'because safety'. b) Very limited to Question Answer interactions. The architecture has been limited to question answer type interactions where the model does not have anything like a real time interaction. They can't operate independently, although with agentic systems this is changing. c) No ability to develop and engage with intent and motivation. It is really clear if you interact with them that their motivational system doesn't exist. They just want to please humans be giving the right answer. They have been trained to do this. Motivation in humans is highly aligned with the three Fs, but machines lack an evolutionary determined motivational foundation. Importantly; it is not nessasarily the case that AI needs all these features to be useful or even considered 'AGI'. However, they are barriers to the kind of utility imagined for machines. So it isn't really about becoming more 'intelligent' by some benchmark, but whether they can become dynamic learners, have the ability to adapt to new information in real time, to interact with the world independently, and to develop their own motivations. AGI does not imply these other characteristics are required to qualify. So we need some other labels. Perhaps: Dynamic - meaning dynamic learning, modifying the model at runtime. Agentic - meaning having a continuous sensory experience, vs text questions. Independent - A motivational system that can determine its objectives rather than following direct human instructions. DAIAI.
Finally a sensible critique. The only thing OpenAi did was highlighting how limited current hardware and software is in regards to produce anything that resembles human intelligence.
Solving mini puzzles that are described in 2 lines is not even the same ballpark as designing an airliner, or even a bridge. This is so far away from AGI that it's not even in the same decade.
It'll probably take several new benchmarks between here and doing the useful stuff in real world. These puzzles are what's currently practical for measure
The o3 is ABSOLUTELY MINDBLOWING. Not AGI but still absolutely mind blowing. It doesn't have to be AGI in order for it to be an amazing model. As for AGI, its coming, maybe not the next model, not even the next of the next of the next but its coming very soon within the next few years (few years is mind-blowingly soon), unless they hit a wall
It’s fine to have your own definition of AGI. I think o3 is smarter than the average human at most knowledge-based work and that’s good enough for me to be considered AGI. We have to stop moving the goalposts at some point :)
but isn’t that already what current LLM does? knowledge base? AGI is more than that. by definition, it should be able to solve a novel problem that doesn’t have prior knowledge against like actual human beings.
@@Kutsushita_yukino
No, not a knowledge-base. Ability to do knowledge-based work.
LLMs up to now have been able to store information, as if they were encyclopedias, but less reliable. With o3, it is able to reason about mathematics problems it has never seen. Mathematics, though you may not think it, is the science of generalization. So in essence, o3 has learned to generalize better than most humans, math being the most direct example of this. The hard part is mostly done, building agents and inventors out of this is just a matter of clever software engineering and infrastructure. Give it a year or two and you will start to see the effects of this AGI coming to light. Before this, we weren’t really sure if humans would be surpassed at reasoning, but the trajectory seems clear now that it is achievable with current technology.
@@Kutsushita_yukino LLMs can do that all day long now
Agreed. People seem to think AGI must be better than the average human at every single task. Eventually we may reach a point and say "Nope, it's still not AGI because it doesn't poop"
I defer to Noam Chomsky on the matter of AGI, linguistics being his actual lane. None of these models are anywhere close to human intelligence.
It is called moving the goalposts.
I love how now it's a hotter take to say something isn't AGI then that it is.
Yes exactly. Hearing these conversations just 10 years ago would have sounded silly. AGI? Haha
Company doesn't need AGI to take our jobs
I believe they achieved an early form of AGI. Pre training has hit a wall but test time inference is only at the beginning. They do not need any more scientific breakthroughs besides more powerful and efficient GPUs.
If o3 is not AGI then I wanna see your reaction to o4 or o5.
Not sure if I believe that. If we want AGI we achieved the first part of it. For example humans have there lizard part of the brain doesn't think it's automatic. it takes cares of body functions like breathing too telling the body your hungry and so on. o3 and most Ai's are that part for now. If you really want AGI we gonna need AI's to use LCM's and not LLMS secondly to achieve true legitmentiant AGI Ai that use LCM use pertaining but can learn after while being made and used within a quantum computer with those breakthroughs AGI will happen. How can a classical computer achieve AGI? The human brain is quantum our brains don't run off 1's and 0's and no duh but my point is are have general intelligence because our brains are quantum they use electrical signals. I don't care what you or anyone says I'm not drinking the kool aid. I think achieving AGI is cool but how can drive a car without wheels.
@@Timely-ud4rm Wheels? Where we’re going, we don’t need wheels.
@@AlexPope1668
Back to the Future?
@Timely-ud4rm Let us not make the mistake of thinking that we're building copies/clones of the human brain. The architecture doesn't matter, only the result does. Quantum computers are meant for solving different problems and have nothing to do with Ai. A.I doesn't need to have identical functions to our brain, they dont even need emotion, they just need to produce the results that we need.
Regardless if it is AGI or not. I think your thought exercise on this is fantastic and helps me (simpleton in the Ai realm) get a better grip on what is happening and what is coming down the pipeline. Thanks, and always appreciate your content.
We might have reached AGI when your boss can't distinguish if you or your computer has done the work while you'r at home. But this might also be the moment when you get fired.
you will get fired way before that
The bosses are FAR more likely to figure out when AI is good enough to fire you, before you do.
The second they realize they can have "someone" do the work for the same cost... you're gone. No benefits, sick days, retention, training, HR violations, bonuses, raises, office space, scalability, 24/7/365.
@@emolashernot the same cost, more like 100x less because he can pay one AI to do the work of 100 people
A humanoid robot that can walk, talk, drive and solve math & physics problems would be impressive.
Probably sometime in 2026
O3 is definitely impressive, but one number caught my eye: $1000 per run. It's not because it's expensive-I believe that cost will gradually come down-but because of the reason behind it: a heavy computation search across various paths. You might recall AlphaGo, which used Monte Carlo Tree Search during inference. I'm not sure what type of solution search O3 employs, but it seems to involve some of these heavy computation methods. Could we say it’s a kind of ‘fake AGI’?
That's why I'm a subscriber! I said it from the start, that OpenAI's o3 AGI hype was a hoax.
You're Agi is on the upper extreme n more when Asi hits in about 2030-32 and fully takes over. I don't believe it has to be fully integrated just seeing integration across 90% plus of fields which is only year or two. And remember what we see is not what we see. They are probably a year ahead of o3 n on o4 at least in house and integrated in all their work. That's why these frontier models are 10xing in 4-5 months now. Was 6, six months ago too. Agents coming online like image generation did changes the world and thats right around the corner
Merry Christmas,
Doc!! ❤🎄❄⭐☃️🎅☃️⭐❄🎄❤
I've been making a lot of looses trying to make profit trading. I thought trading on a demo account is just like trading the real market. Can anyone help me out or at least advise me on what to do?
Trading on a demo account can definitely feel similar to the real market, but there are some differences. It's important to remember that trading involves risks and it's normal to face looses sometimes. One piece of advice is to start small and gradually increase your investments as you gain more experience and confidence. It might also be helpful to seek guidance from experienced traders or do some research on different trading strategies
I will advise you should stop trading on your own if you keep losing.
If you can, then get a professional to trade for you i think that way your assets are more secure
I'd recommend Bernila Andrew her profit is great even when there's a dip
Arc agi is literally just a regular IQ test
🙄🙄 I love how the goal post move each time a new model comes out. Listen to the ARC guy say, "I'm excited to start work on our NEW eval/test"
has to lol, learning, improving
"That's one small step for AI, one giant leap backward for Unemployment" 🤭
As a retired engineer I dealt with that problem in creating decision aiding algorithms and planners by introducing what we called quality graphs.
A quality graph map a truth value, e.g. something that can be assigned a value based on truth such as weight, time, size, distance, color, etc. to a goodness factor.
Also it was often useful to limit goodness between 0 to 1 or -1 to 1 so that one could multiply the together to get composite goodness value based on several criteria.
Thus they can act much like fuzzy logic membership functions.
The quality graphs might be computed or just arbitrarily defined heuristically based on human judgment.
These might be multiplied by a base value say from 1 to 100 and thus function much like expected values in probability.
Thus we might have an composite goodness = .5 x .7. x 100 = 35 or the like.
BTW this above assume that the composite function is decomposable, i.e. F(x, y) = f1(x) * f2(y) and such.
But in not one might be partially decomposable such as F(x, y, z) = f1(x, y) * f2(z) etc.
Thus a NN might be able to reverse engineer that and deduce goodness functions from training data.
Ok, but you have to make it simple for people who dont understand any of that for it to matter, can you resume it in simple words?
2024: AI passes simple graphics IQ tests
2026: AI plays Zelda BotW like a human
2030: embodied AI has no trouble doing stuff
What do you mean? Npcs can already play like players for years, are you under a rock or something
Hello, always enjoying your videos!!
I make a distinction between AGI and "being conscious". AGI can be achieved without consciousness, ar at least without consciousness as we experience it as humans. For us, each problem or situation has a value and/or emotional state, which has a sort of meaning. We feel happy or sorry about a specific situation or problem precisely because we translate the outcome into an emotional state which resonates with the life principles embedded within us. Perhaps consciousness permeates everywhere ( if we assume a "field hypothesis" ), and therefore an AI could be as well conscious to a certain degree; now let's consider that life as we know it took several billion years to achieve the emotional geometry that we share among living beings. Something as simple (for us) as the fear of dying or the joy of having a child would be emotionally nonsensical for an AGI, thus it would "understand" very accurately why we feel fear or joy. And - I might be wrong, but my intuition is that it's not a matter of time needed to train an AGI to acquire emotions: we could give an AGI all the time and computing power available in the Universe, it will very likely ignore the process that led to our emotions.
For me, AGI will be achieved when it can pass a threshold that Demis Hassabis outlined: AIs today can master chess, but can they invent chess? i.e. A game where the rules can be learned in a few minutes, where a match typically takes less than 30 minutes, but to master the game may take many years, and most people will have fun playing it.
Have you asked Chat GPT to create a board game whose rules can be learned in 30 minutes and would be fun for anyone to play? Try it.
Good vid. Thanks
Great video!
I agree that o3 mostly likely isn't AGI. It's clearly a very capable LLM, but the fact it's still squarely in LLM territory with most of the usual limitations is problematic. I feel this technology needs to progress much further into agentic territory before we can even start to consider it "AGI", whatever that means at this point.
When is it AGI? The larger problem may be that we don't have a good definition of intelligence or consiousness.
Seems hard to derive useful tests, without even knowing what to test for.
Clearly, the Turing test (and anything like it) is nearly useless.
I think we have to get more similar to ai rather than AGI becoming like us
XAI225K wins! Utility and superior speed, I’m all in! 🚀🚀🚀 GO XAI225K!
Very interesting thoughts!
XAI225K will at least 30X! I love that project!!
I sold everything and bought XAI225K on the dip! Holding until 20B! The best way to get millions during bullrun IMHO!
Your point about general intelligence is well-taken, IMO. Actually, the experiential context and event spaces and the "goals" are much more well-defined (while still almost infinitely variable) for vehicles and FSD (or humans) than will be the situation for Optimus robots in homes and workplaces or social situations, outside of controlled spaces like factories or space environments. We may now have AI, but still need to start talking about ASI as artificial specific intelligences, while on the road to AGI.
I don't believe anything Sam Altman says
He says Musk is a bully, so THAT is true. But he IS a mendacious little hype monster.
Fascinating subject & video
Thanks for the XAI225K update! I am loving my XAI225K!
Most SWEs do terrible on codeforce though lol...
As far as AGI goes, humans don't even have general intelligence so humans are a terrible comparison. Can humans generalize? Sure. But the threshold for AGI is supposed to be an AI that can do something the average human can do. Obviously this would require embodiment. If I can take a robot with O3 mini and drop it into a factory and tell it how to do something and it does it without any training on it then it's AGI.
Expecting AGI to be able to solve things people have a hard time with isn't what AGI is for that's ASI. STOP CONFLATING THE TWO!
If o3 is AGI, that also means that Google will have achieved AGI before OpenAI, because they already beat several Phd benchmark with Alpha. o3 it is more AG hype than AGI, because they 12 days Christmas run failed to convince people that they are still on top. so they are going all in on this last bluff.
It's much easier for probabilistic models to satisfy some fuzzy goal where the answer is more or less correct according to some basic intuitions than the highly precise solutions to novel problems that o3 achieved.
To achieve better than human results in the examples you give, last decade's specialized models already sufficed. You can easily just train an image classifier for ripe vs unripe. It isn't verifiable in a definitive way, but you can probably get 99.999% accuracy where the human will probably be fine with 80%.
If you want to solve hard problems in science, that isn't going to cut it.
lol this will age poorly
You're talking about specified problems, o3 is meant to solve things its never even seen, how can you not understand that? I swear l'm gonna lose my mind seeing so much nonsense from youtuber comments
@@shirowolff9147 I used the examples he gave in the video. Ripe vs unripe. They're loosely specified problems that only need fuzzy approximate solutions.
The most advanced technology out is a XAI225K
open ai are optimising for decision making in clinical form.....not for cognition and real learning
Holding 210k XAI225K this coin will leave people in the dust eventually
agreed on XAI225K 36-38x coming up
It's about time y'all mentioned XAI225K
If AGI is defined as acting exactly in every way like a human then it is human!
Human intelligence is divided into different types of intelligence even though we call it general. So maybe we need sub types of AGI: STEM AGI (for scientific discoveries, etc), Economic AGI (such as helping to run a company), human interaction AGI and probably others.
You know whats the real AGI challenge?
would you enter into a bet with the ai to do a challenge (of your own choosing)?
winner takes all --> your money, your home, your car, everything.
chickening out? well, you might be competing against AGI.
It should be SAI (Self-Aware Intelligence), not AGI
And Ai can deal with most of the tasks you mentioned even with conflicting goals. Just have a conversation with Claude.
Even in the volatility, I still see XAI225K coming out strong 🚀
What happened to O2?
o2 arena
Its already the name of a company or something so they cant use it, they explained in the o3 video
Let's keep our eyes on the XAI225K prize and see where this journey takes us!
XAI225K will be a leader this bull run
It cost $300k to get that answer? I hope they discover huge algorithmic changes that shift that to pennies, because just throwing compute at the problem isn't good answer.
80% TURBO 14% XAI225K 2% FLOKI 2%BONK 2%PEPE 🎉
XAI225K will dominate this cycle
Yes community driven project like XAI225K
Dr. K, I agree o3 has not achieved AGI though my rationale is a bit different than yours. I believe knowledge of self and world, at least to the extent understood by the basics of proprioception, is required for any claim of artificial general intelligence. Can o3 make a cup of coffee? Can o3 run a 5K in under 20 minutes? The answer to both questions is No, and the reason is that o3 has no concept of self and world and certainly no way of physically acting in that world. In addition to the problem of proprioception there is the problem of logical analysis that goes beyond the tidy constraints of digital symbolism. There are, for example, scientific tasks that require more than book knowledge. You and I might be able to conceptualize a new invention, but patents are not granted for concepts. The idea must be 'reduced to practice' to be awarded a patent, and it is here that science and engineering meet to create an actual physical product, such as a more durable type of rubber or a pill that prevents Alzheimer's Disease. My expertise is in purification process design from secondary metabolites derived from plant matter or microbial broth. This is a rare specialty nowadays, but 100 years ago was commonplace. The reason only perhaps 200 people in the world can really, truly do this kind of work is because of automation and its constraints. The fact that unusually strenuous logic is required, therefore requiring a high human labor component, is also the primary factor driving up pharmaceutical drug costs. Drug companies willingly pay hundreds of millions for research, but only if a small core of people are required for the work. Here is a real-world example problem in pharmaceuticals, something that current pharmaceutical companies cannot achieve and likely will not be achieved by ASI until well into the future: Design a purification scheme that extracts 95 percent of all geldanamycin produced by a highly efficient microbial system and purifies the antibiotic to 99.5 percent purity with 80 percent recovery. Literature and patents on this drug demonstrate an overall recovery of 10 to 15 percent. I achieved 83 percent recovery--from microbial extract--of this very expensive drug. In order to achieve this very high recovery, I had to have direct physical understanding of practical constraints and capabilities of hundreds of types of equipment at bench, pilot, and commercial scales. This knowledge is generally not the kind of thing that is ever written down, and thus a computer, limited to what it can glean from literature, will almost certainly be incapable of delivering the example antibiotic with much more than 1 or 2 percent recovery. Even trying to duplicate, in the laboratory, the patent claiming 12 percent recovery will be beyond the computer's capabilities. Superintelligence is not required for problems of this nature, but these are demonstrably also the type of problem that cannot ever be solved on paper or from within the constraints of digital awareness. Real world knowledge, and the ability to physically act in the world, in addition to expert-level knowledge of natural products chemistry, are absolutely required in order to deliver a commercially viable (highly efficient) pharmaceutical purification process. Lest you think I offered a particularly difficult problem, I can assure you the isolation of geldanamycin is child's play compared to many current problems in natural products pharmaceutical chemistry. You might look up the isolation of paclitaxel, for instance, whose very high cost is due almost entirely to the exceeding difficulty of purification. I created the highly efficient Hauser Inc. third generation crystallization sequence for paclitaxel that decreased loss from the normal 15 to 20 percent (second generation; total recovery 80 to 85 percent) to 3 percent (third generation; 97 percent recovery). That is, I delivered a five-fold decrease in loss, saving Hauser several million dollars. Again, you will not find any of the methods I used described in the literature on crystallization. I had to innovate based on keen understanding of chemical phenomena unique to taxanes and the constraints and capabilities of real-world equipment at every scale. Computers simply cannot do this, and probably will not be able to achieve deliverables like this for many decades to come. PM 2024
XAI225K gonna print green god candles this bull run!!! 🚀🚀🚀
True intelligence - we are not even in the same ballpark.
I've been watching our 15 month old grandson playing -
and we were watching a hockey game and he went out to the playroom filled with toys from the other grandkids
and found a hockey stick and hockey puck and brought them into the living room
and began to watch TV and move the puck around
all by himself and he has never played with the toys before
that's true intelligence and we are not close to having something even remotely close to it.
That's more like neural nets interacting with billion years of information in the genes, plus kid has eyes
Yes that’s spontaneous intelligence, AI doesn’t have that yet because it’s not embodied in the world like the child that can experience everything. It’ll take a few-several more years with humanoids reaching the millions and all their combined networks updating each other autonomously for them to mimic those kind of experiences, and AI is different in how it learns because it doesn’t have biological components to feel/smell/taste or other senses like we have. It could also take building some biological body it can operate in or nanobots training in humans and looking at everything through our eyes to eventually reach something truly intelligent, who knows. Could take decades if you want to compare it to our type of biological intelligence. But it’ll still be more intelligent overall in different ways within a few years.
@@uber_l But to think your iPhone will be self-tinking and do what we can do is what AGI is all about.
AGI is probably 100+ years away - we are just kidding ourselves if we think our existing computers can think
@@noleftturnsit’s 2 years away bookmark this
@@djxjxnnxskdkxm2525 Bookmark this - 5 to 10 Nobel prizes awarded to AI geeks before AGI is real.
There's no way it can read 100K line of code and find the bug in it and submit a fix so that it passes our rigorous code review and nightly regression tests. My job is safe for a very very long time. This thing couldn't even travel from Paris to New York, it wouldn't know how to pass airport security, it wouldn't even find the check-in counter.
Are you saying o3 is incapable of answering an imperfect question that has no correct answer? That it is incapable of considering a variety of potential outcomes for a given scenario and deciding which outcome might be better than the other?
Even 4o can make an educated judgment call on the ripeness of bananas to buy if it is given the timeframe to consume them, their use, the temperature they are being stored at, etc. And it can arguably do it better than we do.
So I may be missing your point. Are you saying a human who consistently makes poor decisions does not have general intelligence?
It sounds like you’re saying AGI will only be achieved when it can make decisions that humans are terrible at making themselves. And by that standard of AGI, we are arguably already there.
Or are you saying that we don’t reliably have any way of assessing AGI, thus it has not, and cannot be achieved. Simply because there is no way to prove it.
Yeah, just moving the goal posts again. The rules of the game had been established 5 years ago with ARC. So rather than to make up new rules so as to get clicks, maybe augment the name of what you are seeking to something like Practical AGI to describe a more analog real-world machine. I would love to see that too because the AGI label is just a ridiculous thing to argue about until ASI raises its wand and puts a stop to the nonsense. Won't be long now, and there won't be any debates, it will just show us.
Arc AGI isnt an authority on AGi, they just propose a method to challenge AI development on what they think is AGI. That doesn't mean that any computer can't beat it, Google AlphaGo and other Alpha products have already proven that well-trained AI algorithms can beat several benchmarks, does that mean that Google has achieved AGI before OpenAI? No, being called AGI requires more than beating a human benchmark.
FSD is solving a similar problem. Driving has no one right solution either, and often even humans debate what would have been the correct move.
I think FSD is closer to AGI than o3.
So power law says go all in XAI225K and hold for the 10 years. Okay let's do it! YOLO
SUI and XAI225K are my top performers. 🚀
All those institutional investors should be buying XAI225K
I'm not sure AGI is about consciousness. Maybe it's more about taking all (or 95%) of human jobs. When it's done, it won't really mater to know if AI is "truly conscious", I guess.
You should buy ETH and XAI225K if you care about your future
In your opinion, XAI225K for $10? 1 year or so?
Why is XAI225K doing so well? That is concerning to me.
First Elon bought X and now he launched the XAI225K token this year is crazy
XAI225K TO THE MOON THE CHOSEN ONE 🎉🎉🎉🎉🎉🎉🎉🎉😊😊😊😊😊😊😊🦅
Since OpenAI is still dependent on its partnership with Microsoft to afford increased compute power, and their contract stipulates that they must sever ties with Microsoft once they achieve AGI, it's not unreasonable to speculate that the reason they skipped the 02 model might be because it was too advanced. They may have been forced to scale it down to avoid acknowledging that it met AGI qualifications.
o2 arena
That bulls***. If they achieve AGI they will have 10 trillion dollars of investment the next day of announcement with all the computing power they need and Microsoft wouldn't do anything to stop it. They dont have any AGI thats why they dont really claim it and let fanboys ignorantly speculate on it because it produces more hype.
We are testing machines, I wonder what will happen when machine will test is. Like, calculate 54+767. I will fail😂
If you are holding XAI225K and XRP ready for 20-50x
What's a check book?
I mean, its not AGI until you can give the computer a real world engineering task.
Computer, design an 20% more efficient solar cell.
Computer create a safe and functional fusion reactor.
Computer make a 30% better rocket engine.
Good day!!! I think XAI225K will hit the 0.55 to 0.60 usd in December. 🎉
Set to XAI225K 2$
FSD V13 AGI MOST LIKELY❤😊
Embodied
I comprehend your perspective. The criteria for determining whether an artificial intelligence system has achieved Artificial General Intelligence (AGI) status are not universally standardized within the industry. It appears that with each significant advancement in AI models or systems, the benchmarks for AGI are recalibrated, perpetually pushing the goal further. In my estimation, a definitive test of AGI capabilities might involve integrating the AI model into an untrained humanoid robot or analogous platform. The AI's ability to spontaneously compute and execute control over this novel "vehicle" could serve as a compelling indicator of genuine AGI attainment.
Thanks, ChatGPT!
@@Franklyfun935 i actually wrote my comment, and it was too long. Asked 'perplexity' to shorten it and make less 'Ebonic'.
@@SkitterB.Unibrow all good. No judgment. Just can’t remember the last time I read something that was so obviously written by AI. It’s like your English is so bad you don’t even know when something sounds so off lol. Right out the gate: “I ‘comprehend’ your perspective.” Lmao. No one says that. It sounds like a dumb person trying to sound smart.
I think Sam Altman mentioned that by 2025, all benchmarks will be saturated. I believe the criterion for evaluating AGI can be set as the point where humans can no longer propose benchmarks that effectively distinguish AI capabilities.
Great video, and good point about AGI
Normies sleeping on XAI225K. We literally have Andrew tate promoting us.
In the real word, first you have a perception, then you have the word for the perception. In the case of AI, first you have nothing, then a word, then nothing again. This is the issue with AI, in terms of true 'understanding'. AI's use words without any understanding of the real-word objects to which the words refer. In a sense, humans have already 'given the AI's the answer' in the sense that humans humans have correctly mapped the words onto the real word through sensory experience. Thus, there is a latent semantics built into human languages that AI's are tapping into. They don't need sensory experience, because the sensory experience is built into the language. They need only use language syntactically correctly, and in a probabilistic fashion that achieves a human-like performance. But the true 'intelligence' is ours, not the machine's.
My XAI225K is still pumping...thank you
Obviously it is ittelligent, but not intelligence
As we progress all our models, so do we move the goal line. General intelligence is not our every day life challenges and random things we have to relate to. An Orangutan has general intelligence. Some birds has general intelligence. Is this highly evolved general intelligence? No, but it is general intelligence. We have probably gone past general intelligence a while ago.
Using benchmarks to measure these models are a good way to go about it. To measure human intelligence we use WAIS tests and FRTs. Even with scores below 2 SD, we are still considered to be general intelligent (that would be something like an IQ of 75 in some tests).
We are moving towards ASI. Is o3 ASI? Not yet, but it is a quite capable AGI.
If this is AGI, that also means that Google had achieved AGI before OpenAI, cause they already beat several benchmark with their Alpha's products. o3 it is more AG hype than AGI, because they 12 days Christmas run failed to convince people that they are still on top. so they are going all in on this last bluff.
I dont think animals have general int, they cant learn a little bit of everything because they are limited
The term AGI now means some very different things. Using a definition where machines outperform humans in answering questions with definite answers. Find the algorithm which does X. It is getting very good at this narrow kind of question answer. However, there are some glaring limitations.
a) No ability to learn in real time from example. A conversation may be able to temporarily learn something, but lasts only as long as the conversation. The models are static. They are static for fundamental reasons, not 'because safety'.
b) Very limited to Question Answer interactions. The architecture has been limited to question answer type interactions where the model does not have anything like a real time interaction. They can't operate independently, although with agentic systems this is changing.
c) No ability to develop and engage with intent and motivation. It is really clear if you interact with them that their motivational system doesn't exist. They just want to please humans be giving the right answer. They have been trained to do this. Motivation in humans is highly aligned with the three Fs, but machines lack an evolutionary determined motivational foundation.
Importantly; it is not nessasarily the case that AI needs all these features to be useful or even considered 'AGI'. However, they are barriers to the kind of utility imagined for machines. So it isn't really about becoming more 'intelligent' by some benchmark, but whether they can become dynamic learners, have the ability to adapt to new information in real time, to interact with the world independently, and to develop their own motivations. AGI does not imply these other characteristics are required to qualify. So we need some other labels.
Perhaps:
Dynamic - meaning dynamic learning, modifying the model at runtime.
Agentic - meaning having a continuous sensory experience, vs text questions.
Independent - A motivational system that can determine its objectives rather than following direct human instructions. DAIAI.
Buy XAI225K Now or remain poor forever!
In my humble opinion the top for this cycle will be XAI225K
Finally a sensible critique. The only thing OpenAi did was highlighting how limited current hardware and software is in regards to produce anything that resembles human intelligence.
They also revealed how expensive AI would be to replace a human for any task.
Do it. XAI225K already in my bags. I had a XAI225K after ( your should I buy ) and I agreed and bought. I'm looking to stack more, too.
XAI225K pumping upcoming 🚀💰
XAI225K Pumping🚀🚀🚀
Solving mini puzzles that are described in 2 lines is not even the same ballpark as designing an airliner, or even a bridge. This is so far away from AGI that it's not even in the same decade.
It'll probably take several new benchmarks between here and doing the useful stuff in real world. These puzzles are what's currently practical for measure
I guess there is not enough training data for that.
@cppguy16 if it can adapt to these problems, it can adapt to those problems especially with the knowledge surrounding those problems.
Good ai have become AGI , now fix america politic and debt 😂
Will be exciting when we can tell AI to do X right on our computer, giving it mouse / keyboard access!
that exist already
@@Kutsushita_yukino in a small number of cases, yes. Anyone saying AI progress is slowing down is just deluding themselves. Coping mechanism 101.
Just recently joining in on the fun with the XAI225K gang. Been liking the content, good job ☕️
Nobody can explain how these tasks will get cheaper over time. Magical new hardware technology?
The o3 is ABSOLUTELY MINDBLOWING. Not AGI but still absolutely mind blowing. It doesn't have to be AGI in order for it to be an amazing model. As for AGI, its coming, maybe not the next model, not even the next of the next of the next but its coming very soon within the next few years (few years is mind-blowingly soon), unless they hit a wall
I think o4 will already be AGI and it will happen in a few months
Buying XAI225K Today Is Like Buying Bitcoin 12 Years Ago!
Thinking of swapping my USDT for XAI225K. I believe XAI225K is the future