# Google’s AGI Plan Just Got Clearer (Demis Hassabis Explains)

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=j0Gnn6KdLFk
- **Дата:** 01.03.2026
- **Длительность:** 16:39
- **Просмотры:** 23,845

## Описание

🎓 Learn AI In 10 Minutes A Day - https://www.skool.com/theaigridacademy
Get your Free AGI Preparedness Guide - https://theaigrid.kit.com/agi
🐤 Follow Me on Twitter https://twitter.com/TheAiGrid
🌐 Learn AI Business For Free AI https://www.youtube.com/@TheAIGRIDAcademy


Links From Todays Video:
https://x.com/ianmiles/status/2025345151053496422?s=20

Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.

Was there anything i missed?

(For Business Enquiries)  contact@theaigrid.com

Music Used

LEMMiNO - Cipher
https://www.youtube.com/watch?v=b0q5PR1xpA0
CC BY-SA 4.0
LEMMiNO - Encounters
https://www.youtube.com/watch?v=xdwWCl_5x2s

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

## Содержание

### [0:00](https://www.youtube.com/watch?v=j0Gnn6KdLFk) Segment 1 (00:00 - 05:00)

So Deisabes has posed a new test for AGI. Let's talk about it. So essentially Deisabes came on an interview and he was talking about AGI and future development and essentially it came onto the conversation of talking about genuine artificial intelligence. So what is the true test for AGI? Because when you look at benchmarks, reasoning, when you look at robots, there isn't one broadly accepted definition for AGI. Sure, people say general intelligence, but what does that really mean? So, Demos puts forward this kind of test that is genuinely innovative. And he says that the kind of test that he would be looking for is training an AI system with a knowledge cutoff. Let's say for example 1911 and then seeing if it could come up with general relativity like Einstein did in 1915. And that is the kind of test that he thinks is a true test of whether you have a full AGI system. — We uh look my definition of Agi has never changed. So which I can tell you is we've always defined it and I've always defined it since I started working on this 20 30 years ago as a system that can exhibit all the cognitive capabilities uh humans can. Now why is that important? First of all, because the brain is the only existence proof we have that we know of. um maybe in the universe uh of a general intelligence. That's also partly why I just studied neuroscience because I wanted to understand the only data point that we have that this is possible, right? And understand that better. And um and so that's the definition I use. It's quite a high bar because it means uh if you wanted to test the system against that, it would have to be capable of all the things humans can do with this brain architecture which is incredibly flexible. It's clear today's systems uh although they're very impressive and they're improving, they don't do a lot of those things. So true creativity, um continual learning, uh long-term planning, they're not good at those things. And another thing that is missing is general consistency across the board at capabilities. You know, of course, in some circumstances, they can do get gold medals in international maths olympiad questions like we did last summer with our systems, but they can still fall over on relatively simple math problems if you pose it in a certain way. So, you that shouldn't happen with uh a true general intelligence. It shouldn't be sort of a jagged intelligence like that. So, there's still quite a lot of things missing. I think the kind of test I would be looking for is um maybe training an AI system with a knowledge cutoff of say 1911 and then seeing if it could come up with general relativity like Einstein did in now I think this is a genuinely interesting benchmark for AGI and I think this highlights something important the difference between pattern matching on existing knowledge versus genuine specific scientific reasoning from first principles. Now I think the challenge of this test though is that it is pretty hard to do because it is going to have to reach a conclusion without the tools a human would. For example, Einstein didn't just have knowledge cutoff of 1911. He had years of obsessive focus, physical intuition built through experiments and access to Lorent Maxwell's equations and other things. So the AI would need all of that context and the creative leap. Now, what makes this a compelling test is that it rules out the main criticism of current AI, and it's basically that we're just sophisticated retrievers. If a model could genuinely derive something that wasn't in its training data through pure reasoning alone, that's a qualitatively different kind of intelligence. Now, of course, the counterargument here is that this might be achievable without without, you know, true AGI. You could imagine a very capable reasoning system that you can extrapolate physical laws without having broader general intelligence. Solving one hard scientific problem is not as equal to generalized agency. And so I want to show you guys this short clip from Moonshots where they ask Ray Dalio if he's concerned about the moving goalpost of AGI. One of the things that I've started to realize in this AI space is that individuals will say, "Oh, AI can't do something. It can't pass this fundamental test. " And then when it does actually pass said test, the goalposts are moved to, "Okay, but it cannot pass the test in this way or that way. " And so if goalposts keep moving, how do you determine when you've hit AGI at all? Ray, are you at all concerned about goalposts getting moved yet again as we see happening over and over again with definitions of AGI and otherwise that we will pass your definition of the singularity? But nonetheless, most commentators will be arguing with each

### [5:00](https://www.youtube.com/watch?v=j0Gnn6KdLFk&t=300s) Segment 2 (05:00 - 10:00)

other for a long time after that whether the singularity has actually happened. — Well, mine is actually pretty strict. I mean to pass my definition of AGI uh you have to be an expert in thousands of different areas which is actually more strict than most definitions of AGI. So I I think I have a suitably strict definition of it. But what else has Deis Habis said when it comes to achieving AGI? We take a look at this video. He actually talks about the fact that there are still probably two or three more breakthroughs needed for AGI. some of which Google are currently working on. Take a listen because this is of course very true when you step outside of the ALM bubble and we look at true reasoning. There are a key few issues that we genuinely need to solve if we're going to get to the AGI that he does talk about. — And I'm definitely a subscriber to the idea that maybe we need one or two more big breakthroughs before we'll get to AGI. And I think they're along the lines of things like continual learning, better memory, longer context windows or perhaps more efficient context windows would be the right way to say it. So don't store everything, just store the important things. That would be a lot more efficient. That's what the brain does. Um and better long-term reasoning and planning. Now, it remains to be seen whether just sort of scaling up existing ideas and technologies will be enough to do that. Uh or we need one or two more uh really big insightful innovations. I'm probably if you were to push me I would be in the latter camp. Um but I think um no matter what camp you're in, we're going to need large foundation models as the key component of the final AGI systems of that I'm sure. So I don't I'm not subscriber to someone like Yan Lun who thinks you know that they're just sort of some kind of dead end. I think the only debate in my mind is are they a key component or the only component? — Now in that clip you can see he talks about a few things. you know, you're going to need two more breakthroughs, which is of course two more giant step functions like continual learning, better memory, longer context windows. But something he also mentions is, of course, the very infamous Yanlen. Now, the reason I'm talking about Yan Lakun is cuz we're going to go full circle here and I'm going to show you guys why everyone can be correct at the same time. And he talks about the fact that, you know, Yanlen basically believes that LLMs are a complete dead end to AGI or whether or not they are the full solution to AGI. And I think he's kind of right in the sense that LLMs are probably not the fullest solution to AGI, but a wider part of what AGI actually is. Now, if you're unfamiliar with Yanukan statements, absolutely no way. Um, and whatever you can hear from some of my uh more adventurous colleagues, uh it's not going to happen within the next two years. There's absolutely no way in hell to, you know, pardon my French. um the you know the idea that we're going to have you know a country of genius in a data center that's there's complete BS right there's absolutely no way what we're going to have maybe is systems that are trained on sufficiently large amounts of data that any question that any reasonable person may ask will find an answer through those systems and it would feel like you have you know a PhD sitting next to you but it's not a PhD you have next to you it's you know a system with a gigantic uh memory and retrieval ability not a system that can invent solutions to new problems. Um, which is really what a PhD is. Okay, this is actually — it's you know connected to this post that uh Tom Wolf made that uh um inventing new things you know requires uh the a type of skill and abilities that uh you're not going to get from from LMS. So, so now of course the thing is that Yanukan talks about LLMs not being leading us to AGI and it's a super fascinating question because remember he says that LLMs are basically a dead end to AGI but if we look at the ARC AGI leaderboard we can see that with every iteration of new large language models you can see that they've been increasing in capabilities. Now, the ARC AGI benchmark is one of those that is notoriously difficult because the way how the test is designed is essentially to basically prove that only a human could get above the human baseline, which is around 80 to 90%. But in just a few short space of months, maybe even just over a year, we've seen models go from maybe 5 to 10% all the way up to 80% like Gemini 3 Deep Think is doing. Which begs the question, are these things smart or are they just pattern matches? It is very hard to wonder what is happening. Now, here's the problem. Okay, benchmarks are unfortunately being gamed. I'm not accusing any of these AI labs from doing so, but benchmarks, I don't think, are that much of a useful target anymore when it comes to figuring out if we're close to AGI. So, let's

### [10:00](https://www.youtube.com/watch?v=j0Gnn6KdLFk&t=600s) Segment 3 (10:00 - 15:00)

take a look at what exactly I mean by that. So, I was doing some research and I came across this very, you know, there wasn't that much information about this Twitter thread, but I did do a little bit, you know, of a deep dive and I saw someone, okay, post about the fact that Arc AGI might not be as good as you think. So, Milani Mitchell did some research and she said, "Of course, this is what got me excited about Ark AGI in the first place. I worry that this goal has been lost amid the rush to achieve higher and higher accuracy on ARC. At least some ARC tasks can be solved using shortcuts sparious correlations in the task data. For example, our group found that the numbers representing the colors in the inputs can be used by LMS to find arithmetic patterns that can lead to accidental correct solutions. And we found that if we change the encoding from the numbers to the other kinds of symbols, the accuracy goes down and the results are going to be published soon. And they've also identified other possible shortcuts. And so essentially there was like a you know a complete Substack post on this but the main point from it all is that high benchmark scores don't necessarily mean that the AI actually understands the task. It may be getting the right answers for completely wrong reasons. And the best, you know, example from the article they talk about is that when AI models solve ARC AGI correctly, they can only explain the right reasoning about 70% of the time. Humans explain it correctly 90% of the time. So roughly one in three correct AI answers is basically a fluke. And they talk about the clever hands hook. And there was a horse essentially in 1904 that could apparently do math, read clocks, identify playing cards, and everyone thought it was a genius. But it turns out that it was just reading tiny unconscious facial expressions from the questioner to know when to stop tapping. Brilliant at something, just not what anyone thought it was. And apparently that's probably what's happening with these models on benchmarks. So the thing is that accuracy alone can tell you almost nothing. Why the model actually got it right matters just as much as whether it got it right. And that is genuinely a key point here. If someone just blurts out the answer to your question, if they don't truly understand why they were able to derive that answer, they don't really understand. And of course, that isn't really intelligence. I will of course be making an update when they do come out with the results to be published soon because of course there's probably going to be some criticism, some push back on ARGI, so it will be interesting and I haven't seen anyone talk about it. And so one of the things I continue to think about is the fact that AGI is going to be multimodal. Most people talk about LLMs, learning this and that, but you have to understand that general intelligence requires many, many more things more than text. Okay? You've got vision, you've got audio, you've got, you know, touch, look at all the senses you have, you know, the ability to reason in the physical world. There are many things that are going to be needed for true general intelligence. And so I came across this tweet from Chris who says, "Wouldn't it be hilarious if Figure Robotics accidentally solves AGI before the pure LM labs because their vision language action model is forced to build a hyper accurate predictive world model to navigate physical reality like Jeppa, which is Yanl's system predicting the next physical state. If a model can perfectly predict the physical physics of the real world, it would be funny that if that spatial and temporal reasoning became the actual foundation of AGI, it would be fascinating to see if their Helix model is able to do something like this because they have the best humanoid robots in the world. And they could also, if they crack AGI, they'd be worth 25 trillion before 2035. And you can see that Brett Acock, the CEO here, says the AI chatbots I use today from the Frontier Labs still feel pretty dumb. They're a research tool for advanced internet searches. AGI will be multimodal. It will listen to you. It will talk to you, see the world, have nearperfect memory, deeply personalized, and able to interact with the world. And so that point about AGI being completely multimodal. I wholeheartedly agree. When you think about if you compare a pure LLM system to a multimosal, you know, VLA like the one that Figure is using, like their new Helix system, I think maybe that could lead to AGI more in a way because of what they're actually doing and the true definition of AGI. Now, if you're wondering about AGI and what actually could define AGI, I think Yosha Benjjo puts it very clearly here. He says that AGI isn't going to be some one defining moment. It's actually a spectrum of capabilities and as these capabilities increase, you just get better and better models. — It's not a moment. Um the reason is simple. Intelligence isn't just like one number. We have people who are very smart on some things and stupid on other things. And it's the same with AI. We currently have AI systems that are even much stronger than humans in some ways

### [15:00](https://www.youtube.com/watch?v=j0Gnn6KdLFk&t=900s) Segment 4 (15:00 - 16:00)

in their knowledge, in their abilities with like so many languages and so on. And in other ways, they're stupid. They're like a child. And yes, uh, progress will move on all fronts probably, but it's not it's unlikely we'll end up with the same capabilities as humans across the board at any moment, which means that we shouldn't be thinking of like an AGI moment. We should think of particular skills that AIs are you know becoming better at. Track those skills and for each of these we should ask the question you know how useful or beneficial it can be for what purposes and also how it could be misused or if we do get loss of control how any could use it against us. — Whether or not we get AGI in the next few years people a lot of people like are starting to not really like care about that question. they still expect the next 25 years or the next 50 years to play out kind of like the last 25 years or the last 50 years where you know there there's a lot of technological change between 2000 in 2025. Um but it's like a moderate amount of change and they kind of expect that in 2050 there will be a similar amount of change as there was between 2000 and 2025 even if they think that we're going to get AGI in 2030. They think AGI is just like what's going to drive that sort of continued mild improvement. Um whereas I think that there's a pretty good chance that by 2050, you know, the world will look as different from today as today does from like the hunter gatherer era. You know, like it's like 10,000 years of progress rather than 25 years of progress driven by AI automating all intellectual activity. — So let me know what you guys think about AGI and I will see you guys in the next one. It's been Andrew Black. You've been watching the AI grid. Hopefully you guys have a wonderful

---
*Источник: https://ekstraktznaniy.ru/video/11426*