AGI Within 7 MONTHS , New Q-STAR Paper , Strict AI Regulations and More

So, with a crazy week in artificial intelligence, today's video is going to cover around six different stories that are pretty intense, starting with AGI in seven months. Now, this has been something that is pretty crazy, but this one right here predicts AGI by November 2024. Now, that isn't long away at all. We've only got a couple months until November. And this is Allen's conservative countdown plots. Predicts AGI by November 2024. So if you aren't familiar with Allen's conservative countdown to AGI, essentially it's a method, essentially it's a tracking metric developed by Dr. Alan D. Thom, which aims to estimate the progress towards the development of artificial general intelligence. Now, AGI is defined, of course, as a routine that performs at the level of an average median human across a wide range of cognitive tasks. And the countdown uses a percentage as a scale to represent the progress towards achieving AGI, with 100% representing AGI. Now, of course, you can see right here, this is his conservative countdown to AGI. And numerous announcements have made this thing tick forward. Now, I'm not surprised at this at all because it is very, very fascinating to see the recent developments that we've had. Now, the countdown does include milestones such as the elimination of hallucinations in large language models and the physical embodiment of AI in robots, and of course, the ability to pass Steve Wi test of AGI. Now, one of the things that we've seen here is pretty crazy. He says a round score of 80%. This is where an AI in a robot plays, passes Steve Wo's next test of AGI. And this is the test where a robot can walk into a strange house, navigate available tools, and make a cup of coffee from scratch. So you can see right here that this is, of course, 100% and this is in November. But he's stating that we are very, very close to this being done. Now, I think this is a very interesting benchmark because for an robot to be able to go inside a house and to be able to look around, find the door, open the door, find the coffee materials, be able to get those, make a cup of coffee from scratch, I think that that is something that an advanced AGI system is going to be able to do, because a lot of the times there are adversarial disturbances and there are a lot of things in the environment that the robot is going to struggle with. Now, with Figure's recent demo, if you haven't seen it, it's pretty crazy. You can watch it. It's Literally insane. But they've combined the actual LLMs with the robots and now there is the physical embodiment of an AI system. And this is of course one step closer to AGI. So of course people are stating that for hold up, it needs to be at least 80 by the beginning of June and 90 by the beginning of September. Meaning for this graph prediction to hold up, what we would need to see for this to hold up is we would need to see an AI robot walking into an unseen environment and making something like a cup of coffee. Now is that possible considering that we only have, I think around three to four months left until June? I don't think that that is impossible considering this company is moving at lightning speed. And with a few breakthroughs from the echo chambers that I've heard, it seems that robots are about to get a huge jump up with the embodiment of these AI systems. I think it was one of the integral parts that these AI systems were missing. So within the next three to four months we're going to be able to see how quickly things move up. And of course, you know, someone made a very valid point. They said, to be honest, it seems like a 50, 50 chance to me. Sora, Seba, Figure 1, Claude Opus all happened in a mere three weeks from each other and the pace doesn't seem to be slowing down. So this seems like it could very much so be a reality. So do you think Allen's AGI countdown to AGI within seven months, do you think that is RA conservative? Do you think it is something that is too, I guess you could say, eager. I mean, either way it's going to be able to see. And this is someone extracting the data. By the way, he doesn't say that AGI is going to be achieved at this date. This is someone that's extracting the data based on the previous polynomial projection. So this is based on all the previous data points. So yeah, it looks like we are on that S curve. So I mean, if that date does happen, I personally don't think it's going to happen. But then again, I do remember looking at these two and this is something that I did miss, something that I forgot to include in the video and at the start of the year actually suggests that people who are working on their entrepreneurship startups build with the mindset that GPT5 and AGI will be achieved relatively soon and that most GPT4 limitations will get fixed in GPT5. So that is crazy. So of course Sam Altman stating that 2024 and 2025 are going to be big years for AI. That does mean that potentially this kind of chart doesn't look as crazy as it does now, and that AGI being at 70% doesn't seem that crazy. Now, of course, with AGI, there are some doubters now, of course. Christopher Manning, this is someone at the Stanford AI Lab. He says that I do not believe that human level artificial intelligence or the commonest sense of AGI is close at hands. AI has made breakthroughs, but the claim of AGI by 2030 is as laughable as the claims of AGI by 1980 are in retrospect. So what he's doing here is he's looking back on when we used to claim the AGI was around the corner. But this is from 1980, and technology has since rapidly evolved since then. So let's take a look at this statement. We can see that he says, in from three to eight years, we will have a machine with general intelligence of an average human being. I mean, a machine that will be able to read Shakespeare, grease a car, play off its politics, tell a joke, have a fight, and at that point the machine will begin to educate itself with fantastic speed. And in a few months it will be at a genius level. And in a few months after that, its powers will be incalculable. In retrospect, this is a pretty crazy statement. However, looking into the future, things are on an exponential, meaning that the future is much harder to predict, that if we look back into the past. So now, of course, there are some more doubters. But take a look at this first from Elon Musk. He says AI will probably be smarter than any single human by next year. By 2029, it is probably smarter than all humans combined. So, extrapolating this information out, I'm guessing he's stating that maybe we achieve some level of virtuoso AGI by 2029, and by next year, I'm guessing that we probably have, across the benchmarks, probably 98 to 99% across all of the top reasoning benchmarks, meaning that AI is going to be smarter than any single human next year. Now, of course, this claim was, you know, in response to Ray Kurzweil stating that we will achieve human level intelligence by 2029. And of course, he states that we're not there, but when we will be there, it will be 2029, and it's going to be matching any person. Now remember, this isn't that far away. It's 2024, and in five years we're going to see some major AI developments. One thing that you have to understand that whilst 2029 could be the singularity before then, we're still going to see some massive shakeups in the economy and pretty much everything else. Because it's not like, boom, we get AGI and then everything just falls off a cliff. We're going to be seeing, you know, LLMs increase their capabilities, vision systems go crazy, video systems go crazy. Remember, we're going to have four years of AI development before an AGI system, you know, conservatively is released. So that is going to be another prediction there. But then again, of course, some people are predicting AGI in seven months. But then again, it's just an extrapolation of data going out. Now of course, some people on the website have predicted the GPT5 date and Alan expects his date to be August 2024, which does kind of line up with the OpenAI dev day which is scheduled for November. So that could be kind of a funny, interesting time to see what they release at that day. Now, like I said as well, back to the doubters of Elon Musk. This is of course Yann Lacan and I tweeted about this earlier. I tweeted the fact that the top AI researchers could not agree on some of the really, really big issues shows us that the future is going to be near impossible to predict. Because on one side you have people like Yann La that says no, we're not going to have AI smarter than us by 2029. If that were the case, we would have AI systems that could teach themselves how to drive a car in 20 hours of practice like any 17 year old. But we still don't have fully autonomous self reliable driving even though we, you have billions of hours of label training data. So on one side we have Yad Lecard, which is a very, very respected AI scientist who is at the top of his field that says that we are not going to have this by 2029. And of course we do have, you know, this guy right here in retrospect stating that we were pretty crazy to think we would have something. And then of course on the other side we like Elon Musk, Ray Kurzweil and Sam Altman stating that, you know, GPT5 will achieve AGI, not GPT5 will achieve AGI, but GPT5 and AGI will be achieved very, very, very soon. But what do you guys think about that? Where are you on the spectrum? Do you think that Sam Altman and Elon Musk, do you think they're wrong. Do you think they're too hopeful or do you think that, you know, Yann Lecan is just delusional or naive? I mean, either way, I think that, you know, the only way we're going to know is to be in the future. And then whenever the AGI system is developed, then we can look back and say, okay, this person was right and that person was wrong. Because either way, these people are the some of the most respected people in their fields and them disagreeing. It's kind of a concern because it means the future is really, really uncertain. Now, in addition to uncertainty, we do have the laws surrounding AI. Now, essentially what we have here is EU's AI Act. So the EU AI act has a few things that I wouldn't say are cause for concern, but the vagueness of them gives me some headache because essentially one of the things they said here was that some applications will be banned outright because they threaten citizens rights, including emotion recognition systems and schools and Workpl place, amongst others. Now, I'm sure that potentially this is going to be in more detail probably on the article, but the point here is that I think the rate at which AI is developing poses a really hard problem to legislators because the rate at which it's developing means that they're going to have to dynamically update these laws because there are going to be new and newer problems that they just simply haven't predicted. There might be even emergent behaviors or properties that make the system behave in a way that we didn't predict, which means that certain laws, it's going to be a bit vague on how we kind of govern those AI systems if they have certain abilities or capabilities. So that's why I think this kind of stuff is important to pay attention to because, you know, there were some. And a crazy thing as well that you can see here is that it says, you know, a complete set of regulations, including rules governing chatbots will be in effect mid-2026 according to the European Parliament, which noted each EU country will establish its own AI watchdog agency. So that's the thing, okay? They're developing these laws and they're stating that, you know, a new rules governing chatbots are going to be there by mid-2026. How much do you think chatbots are going to change in the next two years? In the next two years we could literally have artificial general intelligence systems or systems that have such advanced reasoning capabilities that trying to implement rules on these is going to be very, very difficult. So I mean, it's going to be so hard to do that because it seems like whatever rules they implement, they're going to be pretty much outdated by 2026. Because literally last week we saw Claude 3 Opus Figure 1, all of these insane updates. And I mean it's only been like, I think it's only been 450 days since ChatGPT was initially released and from GPT 3.5 to where we are now is absolutely incredible. And with that being said, I mean we've already seen how the economy's responded to that. And by them stating that another two years, literally, we really don't even know what the AI space is going to look like. So good luck to you if you are on that. But then again, of course, as long as we pay attention, I think it's going to be good because I do think we do need some laws surrounding that now. Now the EU wasn't the only place implementing some laws. Of course the us, they were doing some stuff. So essentially they were publishing a document warning about the dangers of AGI and they were saying that it would introduce catastrophic risks. And I don't want to gloss over the catastrophic risks. I do think that whilst those are true, I think the main thing here is that they said that AGI is the primary driver of catastrophic risk and that OpenAI, Google, DeepMind, Anthropic and Nvidia have all publicly stated that it could be reached by 2028. Now, whilst like I said, 2028 is a pretty much good year. 2028, 2020, I think all of those three years we would likely see an AGI system. We have to understand that this thing is developing super rapidly and something as well that they also did state that most people didn't pick up on. They stated I think this might be llama3, I'm not too sure, but they stated that one individual at a well known AI lab expressed the view that if a specific next generation AI model were ever released as open access, this would be horribly bad. So I have a controversial viewpoint here. Okay, and hear me out because I'm going to reference the video in the future. If I am, I'm true. I personally don't believe that artificial general intelligence is ever going to be open source and I really am wondering if they're ever going to allow the standard person to access an AGI system. And I know that might sound a bit insane, but there are two points, okay? Number one, if they open source this, of course there are some second order consequences that could be quite horrific because AGI, it's going to cause wide scale problems in terms of if it's open source, people can use it for nefarious means, which is of course quite bad. Now think about it like this, okay? I'm not going to say that OpenAI is going to be the only car, not the only company to achieve AGI, but it might be harder than we think because remember, okay, all these other companies, it took them quite a while to catch up to just where GPT4 is. The problem is with that is that AGI seems like it's going to be a huge step above that. And if it is, it means that these other companies, how are they going to catch up if they do have an AGI that's working for them in those systems? So I don't think that once they do get it, they're going to open source it because it just wouldn't kind of make sense. Now I think them deploying it, I also do think that there's going to be compute issues around that because it's going to be a capable model. I mean, it could be highly efficient. We don't know. But yeah, it's going to be fascinating to see how that AGI thing goes down because some people are stating that open source AGI is going to be awful. But then again, you do have to think about it like open source AGI, it kind of levels the playing field in terms of everyone getting access to it. So I mean, Geoffrey Hinton also did say that if AI is more dangerous than nukes, why aren't we just giving everyone nukes? It doesn't make sense to open source AGI if it's that much of a capable system. And I think if everyone gets that mentality, I don't think we ever see open source AGI because it might be seen as reckless to do so. So I think in future, once we actually see what the AGI is capable of, then I think the conversation changes because right now people say AGI this, AGI that. But once you see an AGI system and you're really understanding what it's truly capable of, then you're going to be like, okay, now we can make a decision. So with that being said, there's also the framework that they did talk about, okay. And in that framework they spoke about three strategies that I'm going to include here. I'm not sure why they haven't publicized it because I had to go through so many links to find it. But yeah, I'm going to show you guys that and then kind of come.

Back to another research paper, 2022 ChatGPT, came out. The US government commissioned an assessment of the risks to national security from advanced AI on the path to human level AI, or artificial general intelligence, as it's sometimes called. Now, that assessment was also to include an action plan to address those risks. Now, Gladstone has just finished performing that assessment. Our team worked for over a year and talked to more than 200 people, from top executives at Frontier AI labs to cybersecurity researchers, WMD experts, and some of the most senior national security officials in the United States and around the world. We also met informally with some of the most knowledgeable individuals inside the world's most advanced AI labs, some of whom spoke to us on condition of anonymity. Now, this is the first assessment of its kind, and it's informed by an unprecedented level of access, not just to Frontier Labs, their researchers, and even their CEOs and other executives, but also to an unparalleled range of US national security officials, from the White House to working level experts at the top of their fields. And along the way, we learned some sobering things behind the scenes. The safety and security situation in advanced AI seems pretty inadequate relative to the national security risks that AI may introduce fairly soon. Frontier AI Labs are publicly saying that they think they might be able to develop artificial General Intelligence, or AGI, within the next five years. And privately, many researchers are telling us they see much shorter timelines as plausible. You can check out our explainer on the new era of AI for more context. But the upshot here is that these claims are rooted in hard data and are surprisingly plausible. And that's great in so many ways, right? I mean, it could allow us to make amazing economic, scientific and medical breakthroughs that would have been the stuff of science fiction. But on the other hand, the AI systems we might have in the next few years could have capabilities that could be weaponized to cause devastating damage. At the same time, Frontier researchers are also privately worried that beyond a certain level of capability, they may not be able to maintain control of the AI systems they're creating. And quite a bit of early evidence is pointing in that direction. You can check out our explainers on AI weaponization and loss of control. But the upshot is that although we can't know for sure, there's compelling data to suggest that AI may introduce weapon of mass destruction, or WMD like risks sometime in the relatively near future. That data seems to be on solid enough scientific ground that these risks ought to be taken seriously. But this assessment isn't just about the Risk. It's also a blueprint for what we can do about it. It culminated in an action plan, a set of coordinated whole of government policy proposals designed to mitigate the catastrophic risks that evidence suggests could come with future AI progress. All while positioning us to reap the incredible present and future benefits of AI as much as possible. Plan starts with LOE 1, which outlines possible strategies the executive branch could use to buy down catastrophic AI risk in the near term, while setting the conditions for successful long term AI safeguards. Now, any safeguards we put in place for AI need to be carefully scoped so we can keep realizing the incredible benefits of the technology. But at the same time, AI systems can already be weaponized in concerning ways, and AI capabilities are accelerating almost inconceivably fast. Some could plausibly have WMD like or WMD enabling capabilities in the next one to three years. To understand why, you can check out our explainer on weaponization and alignment risks. So if we want to buy down that risk in the short term while positioning ourselves for the long haul, we can do three things. First, the US government needs better situational awareness of AI threats. It could achieve that in part by creating an AI observatory, a hub for AI threat evaluation, analysis and information sharing with horizon scanning and emergency preparedness functions. Second, as frontier AI models are being scaled and improved, they're developing new weaponizable capabilities at an unprecedented rate. So there's an urgent need to establish rules for responsible development and adoption of advanced AI systems. The US government could set up an interagency AI Safety Task force to establish and coordinate the implementation of those rules at our most advanced AI labs. And finally, advanced AI safeguards in the US won't achieve much if other countries don't take similar measures. Fortunately, the US can use its unparalleled influence over the AI supply chain to do things like establishing a licensing framework for developers who want to use US cloud services services and taking steps to put in place monitoring systems on us designed AI hardware. Now, LOE 2 outlines actions that the US government could take to increase its preparedness for rapidly responding to incidents related to advanced AI and AGI development and deployment. Now, the accelerating power of AI is amazingly good in so many ways, but it also increases the destructive footprint of malicious actors who use it. Over time, that could raise the prospect of new kinds of cyber warfare, make it easier to plan and execute biological and chemical weapons attacks, and even increase the risk that we might lose control of the systems we're developing. So if we want the ability to prepare for and respond to AI incidents quickly and in a technically informed way. We can do four things. First, the US Government can keep coordinating on national security risks from advanced AI and centralize its efforts in that area by setting up working groups and overall ownership of the problem set set at the executive branch level. Second, if you want informed planning and informed responses, you need an informed workforce. And that means setting up education and training programs in AI across the US Government to increase our preparedness and response capacity. Now. Third, future AI incidents could have WMD scale impacts and unfold very quickly. So the US Government could consider developing a framework of indicators and warnings for advanced AI and AGI threatening threats to national security. That would help build out capacity to anticipate and detect emerging catastrophic threats from weaponization or loss of control, both domestically and abroad. And finally, while a warning system is crucial, what do we actually do if we detect signs that a serious threat is about to materialize? Well, here's where contingency planning comes in. The US Government could consider directing a contingency planning process to develop response options for various threat scenarios. LOE 3 outlines actions the US government can take to strengthen domestic technical capacity and advanced AI safety and security. AGI alignment and other technical AI safeguards. Now, frontier AI labs are locked in a race to build more and more powerful AI systems. And that race is driving them to invest more and more in AI capabilities and less and less into technical safeguards and procedures against catastrophic risk risks. It's already an open secret that many labs are worried they'll be able to develop extremely powerful AI systems before they can figure out how to reliably control those systems. You can check out our explainer on alignment risks. But long story short, controlling very capable AI systems is an unsolved technical problem. Now, if we want to close that technical gap as quickly as possible, we can do two things. First, the US government can directly fund fundamental research in AI safety and security, including in technical alignment approaches that are intended to scale up to AGI level systems. It could establish AI centers for both open and classified research and establish protocols and infrastructure for secure information sharing. And second, the US government can continue to support the development of standards for responsible AI development and adoption to give frontier labs benchmarks for safe and secure practices. But there's an important caveat here. These standards shouldn't rely too heavily on our attempts at evaluating what AI models can and can't do using today's approaches. See, the best AI evaluation techniques today can't tell us for sure if a model does or doesn't have A dangerous capability that we care about, or even which kinds of dangerous capabilities we should be looking for in the first place. So safety and security standards should encourage the development of new techniques built on a fundamental scientific understanding of the behavior of AI systems.

So now that you've seen that, there was also another research paper called quietstar, and this is one that is really fascinating, essentially that it's about having an internal monologue. Now, essentially, I think it's either 50% of people do, 50% of people don't. Now it's essentially an internal monologue. Now, some people might have an internal monologue that helps them with things. And it's not like another person, like, talking to you, it's just yourself speaking to yourself internally. It's very hard to describe, but it is a concept that some people have is. It's kind of like theory of mind in a sense that, you know, you can think about something and kind of, you know, reason with yourself. It's kind of like an extra reasoning engine, I guess you could say, but. But I mean, it's kind of weird because they kind of use this in LLMs and LLMs were able to double their performance. And it's something that I did a video on the second channel and I will include that. And the reason I'm going to include that in the video is because it was something that was widely discussed on Twitter because some people don't have internal monologues. And Yann Lecun was also debating internal monologues, stating that he doesn't have one. So some people were stating that maybe.

That'S why he doesn't believe AGI is.

There, because he doesn't realize how internal monologues work. It's just an entire debate, but I will include that.

When humans write or speak, we often pause to think and reason about what we want to say next. This reasoning is usually implicit and unstated happening between the lines of the text or conversation. Recent AI research has shown that large language models, LLMs, AI systems trained on vast amounts of text data, can be prompted to show their work and output explicit chains of reasoning to answer questions or complete tasks. This allows them to solve more challenging problems. However, this has required carefully curated datasets of question answer pairs to train the models. A new approach called QuietStar aims to allow language models to learn implicit reasoning from arbitrary text without the need for specialized datasets. The key idea is to train the model to generate useful thoughts between tokens of text that help it predict the next parts of the text more accurately. In this way, it learns general reasoning capabilities that are embedded throughout language. How Quiet Star Works Quiet Star operates in three main steps that form a learning loop. 1. Think as the model processes each token of text, it generates some thoughts or reasoning statements relevant to predicting what comes next. Multiple possible thoughts are sampled at each point.

2.

The model then makes two next token predictions, one based solely on the original text and one that incorporates the thoughts it generated. These predictions are combined based on a learned weighting learn. The model is updated based on which thoughts led to better predictions, receiving a reward signal that encourages it to generate more useful thoughts in the future. Thoughts that improve prediction accuracy are reinforced. This process allows the model to incrementally learn to generate reasoning that improves its language understanding and generation without explicit supervision. The researchers used some key techniques to make this computationally efficient and stable, such as parallelizing the thought generation and using special tokens to control the reasoning process. Results and significance the researchers tested Quietstar on two challenging question answering benchmarks, Common Sense QA and GSM 8K math word problems. They found that pre training a model with quietstar on a large web corpus led to strong improvements on both tasks compared to a regular LM without needing any task specific fine tuning. On common sense QA accuracy improved from 36.3% to 47.2% and on GSM8K it improved from 5.9% to 10.9%. Importantly, performance scaled with the length of the reasoning chains generated during pre training, confirming that the models were learning useful multi step reasoning. Qualitatively, the thoughts generated by quietstar during the fine tuning often related to retrieving relevant facts and performing logical steps pertinent to the text even if those facts and reasoning steps were not explicitly stated in the original text. This work demonstrates an important new paradigm for imbuing language models with reasoning capabilities through self supervised learning on broad language data. Rather than relying on narrow datasets and explicit chain of thought prompting, models can learn to reason quietly as they process text to improve their predictions. The fact that this leads to strong transfer performance on reasoning tasks is a promising sign that the models are learning generalizable reasoning abilities. This opens up exciting avenues for further scaling up language model reasoning in more open ended ways. Conclusion the Quiet Star technique shows that language models can learn general reasoning capabilities through self supervised learning on text text without explicit reasoning supervision. By learning to generate chains of thought that improve its predictions, a model develops reasoning skills that transfer to challenging question answering tasks. This is an important step towards imbuing language models with more human like reasoning that happens between the lines as we process language. With further scaling and refinement, this approach could lead to language models that more closely reflect the flexible reasoning and generation capabilities of human intelligence.

Now that you've seen the entire video, let me know what you guys think about AGI in seven months. Do you think this is a ridiculous claim? Do you think this is going to happen due to the exponential nature of AI? Do you think that Sam Altman stating that GPT5 and AGI is going to be here makes the future building much more difficult? Or do you think it makes it much more easier? And do you think super intelligence by 2029? Like Elon Musk said, all humans being outsmarted by next year is going to be true? But yeah, let me know what you think about that. And I would say for those of you who are super interested in AI, don't forget to check out the video coming in 3 hours because it is absolutely insane. Just as long as the information isn't fabricated because there is some a recent announcement. It isn't being covered by everyone, but it's pretty incredible. So that being said, I'll see you guys the next one.