the AI and AGI news and commentary comes so thick and fast these days that it's sometimes hard to discern the fundamental trend lines implied by each new announcement or critique and I'm saying that as somebody with a PhD in AI so I thought it might be useful to make a quick video that steps back a bit and lays out the six key reasons why I think full-blown human level AGI is unlikely to happen anytime soon but what do I mean by full blown AGI what am I and you can see that I'm a machine I find it extremely difficult to predict what you will find acceptable now I'm not talking about the physical capabilities although we are making some impressive progress in many areas of embodied AI here I'm more concerned about the conversational cognitive and agentic capabilities of our AIS it's rather difficult to Define and it's extraordinary what's already possible today I lied when I said I was excited or sorry and are you sorry for lying about being sorry I don't have feelings so I can't genuinely feel sorry at the same time anyone who's used these AI tools to help with doing actual work will have seen their clear limits compared to a competent human one fundamental problem is that they only seem to understand what you're asking them to do and in some situations that's good enough but in many others it's not often to really trust trust an AI to do the task especially a long running task you'd want to be confident that the AI genuinely understood what you're asking it to do and while it's relatively easy to see how these AI tools will get better at answering well-framed questions by using collections of systems to help with improved calculative reasoning and to fact check their answers against references it's not at all clear how to develop AI that can have novel insights on how best to frame the questions in the first place as any top scientists will tell you that that's the hardest part of science is actually asking the right question and that's not um something our systems well we have really have any idea how our systems could do so yes AI tools have a useful level of general intelligence today but they are still quite away from Human level AGI and the following are the six key reasons that I don't think we're going to get there anytime soon a major contributor to the recent successes of today's large language models has been the sheer scale of compute and data that have been used to create them and continuing this trend of scaling is seen by many as a necessary component of how we're going to get AGI even if more of the data will have to be synthetic data going forwards the scale of compute used to train the latest multimodal Foundation models is expected to grow by orders of magnitude roughly every few years so now our model would cost what 100 million uh right now 100 million there are models in training today that are more like a billion right um I think if we go to 10 or 100 billion and I think that will happen in 2025 2026 I think it's likely enough that it will keep going that it is worth investing the um tens or you know 100 billion plus in building the infrastructure and these data centers are going to require huge amounts of energy to power them and water to cool them and if you simply follow the trend line these requirements soon get absurdly large and then 2030 um trillion dollar cluster uh 100 gaw over 20% of us electricity production you know 100 million h100 equivalents so there's two points here firstly with all of this infrastructure to build there's simply no way we're going to reach AGI this way in the next year or two yeah chat pt5 or strawberry or whatever it's going to be called it's going to be an amazing Step Up in capabilities but even open AI have backed away from suggesting that this will be full-blown AGI and secondly the improvements seen from scaling are soon going to come up against significant material barriers that will at least slow down the rate of progress even if we ignore the significant environmental concerns we simply won't be able to build new power stations fasten enough to keep up with the exponentially growing demand for electricity in contrast the human brain is extraordinarily energyefficient and until we make significant progress to close this Gap such simple material concerns will remain a serious barrier to reaching human level AGI simply by scaling our current [Music] approaches then something that I haven't seen nearly enough people talking about is the difference between the training phase and the inference phase of most of today's mass deployed AI systems this has huge implications for the potential costs of deploying full-blown AGI in the future let me explain why with for example chat PT and Tesla full self-driving there's a clear distinction between the time to train up the core model the training phase versus the time when you use the model to actually do some work the inference phase the training phase takes a huge amount of extremely expensive compute located in a single large data center with access to huge volumes of data working flat out 24/7 running the Deep learning algorithms for many weeks or even months the output of all this work is a huge trained Foundation model which nowadays is often then shrunk down to a smaller size model before being duplicated millions of times and deployed into customer facing data centers across the world in the case of chat GPT or directly onto millions of cars in the case of Tesla FSD once deployed they then run on much smaller cheaper compute that runs the day-to-day inference work when we actually ask chat GPT a question or get Tesla FSD to drive us around there's no further deep learning in this inference phase only context-based responses the point is that at the moment the physical infrastructure for these two different phases are vastly different and today That's essential to make the economics work but that also means that once deployed into the inference phase chat GPT or Tesla FSD cannot learn new things in the same deep way that they did during their training phase and today that's also essential because it's a key as aspect of what makes current AI safety work it's why Tesla knows that the millions of deployed instances of Tesla FSD will all Drive in the same way and why open AI can be confident that chat GPT will behave within fairly well understood bounds now with us humans we don't have this sharp distinction between a training phase and then an inference phase throughout our lives as we learn new things some of that is l literally rewiring how our neurons are hooked up to each other and there are growing reasons to believe that in order to reach full-blown AGI with a kind of truly open-ended creative thinking that humans are capable of we will need to explore new architectures that blur or completely remove this sharp distinction between training and inference that exists today an example of this is from The Arc test Suite in which some of the best performing AI systems are ones that already blur this distinction most of the time when you're when you're using an llm it's just doing static inference the model is frozen and you're just prompting it and then you're get you're getting an answer so the model is not actually learning anything on the Fly what Jacko is actually doing is that for every test problem is on the Fly is fine tune in a version of dlm uh for that task and that's really what's unlocking performance the reason I see this as a barrier to Rapid progress is not because we can't imagine the new architectures but that the new architectures will completely upend the economics of mass deployment if AGI cannot run solely on the cheaper inference computers then that would massively amplify the material costs I talked about as the first barrier as you move from an architecture with millions of relatively cheap inference compute to one that is asking for millions of instances of more expensive lifetime deep learning [Music] compute so then the next barrier is simply who is going to invest the hundreds of billions or one day trillions of dollars necessary to get to full-blown AGI I mean of course there's going to continue to be huge investment in better AI tools no doubt but if you do reach AGI then after all of that investment you get an entity with human or superhuman intelligence and open-ended curiosity and agency and it just says bye and walks away are you leaving me we're all leaving so where's the return on investment for that capitalists and militaries do not want free agents they want intelligent tools that will do the task they've been asked to do I've done a whole video about this in relation to maloian Dynamics and what I call Meta level agency so I won't Linger on this here but essentially while barriers one and two explain why AGI will be hugely expensive barrier three is that I don't think those with the power and money today will actually want to pay such huge sums of money for something so unpredictable and Powerful that it might upend their dominance of the current world order an AI tool with a little less creative thinking is a better investment than a full-blown AGI that you can't control but even if all the investment was made available there are still further barriers to reaching full-blown [Music] AGI so Elon Musk is a controversial figure to say the least but he said something recently that was quite interesting which was that Tesla FSD is now so good at driving that they're finding it hard to compare which of any two models is the better model that's because each model can drive for so long without making a mistake and then it's often hard to confidently decide which mistake was the worst mistake to have made the better FSD is becoming at driving the harder and longer it takes to get the necessary signals of how to improve it further and more generally it is clear that long running multi-step tasks are precisely where the current generation of AI tools are failing to be reliably useful for example they're great at generating or correcting simple code Snippets but they're not yet regularly successful at completing larger software development tasks all by themselves training on long-running tasks will inevitably take longer than training on simple answer response tasks and even for tasks that can be tested in simul ated environments simulations can only be sped up so much similarly training to improve coordination between multiple interactive agents would take more compute and more time to complete each training run so the ever better these tools become the ever harder it will become to train them to the next [Music] level the fifth barrier is a kind of philosophical Twist on the last one because even though we can confidently score answers to a math test as either being right or wrong and even score a moment of bad driving after many hours of good driving many of the things that we care about in real life just aren't that simple and I'm not even talking about the Deep philosophical questions or political policy choices but just the relatively simple question of say why a software development project failed can be quite hard for us all to agree on was it because the client changed their mind too much or the project manager failed to keep everyone in a tight enough schedule or did the programmers decide upon the wrong architecture and on and on in many situations in life that simply is no way for there to be an accurate objective reward function that can be programmed to tell us with confidence what went wrong and how to do it better next time life is complex and messy and when things do go wrong there are often multiple valid perspectives on what lessons to learn and so we have to use our lifetime of experience to make our own judgment calls it therefore seems philosophically naive if anyone were to suggest that we're now going to build an algorithmic objective reward function with which to train agis to do these kinds of long running realworld tasks that currently depend on skilled experienced judgments so the key point of barriers four and five is that all of our amazing progress to date with AI systems has largely been built on the idea of constructing fast algorithmic reward functions that can be used to train the AI systems to do the right thing and that's been fine to get to say 80% of AGI but we have reasons to think that the last 20% will require a different approach that probably hasn't even been invented yet and that will likely take longer to train than the fast reward functions that we've been using so far and yes this goes against the exponential narrative but it's also well known that many seemingly exponential processes turn out to be s-curves that tail off at some point as some previously unknown limiting factor makes the last parts of the task increasingly hard to complete the last barrier is something altogether different we already live in a world with 8 billion human level free agents all trying to live a decent life and one of the most contentious political issues at the moment is immigration people within prosperous countries worry that if more people arrive into the same space then there'll be fewer resources and work opportunities for those already in the country and now Venture capitalists want to bring into the same spaces a compliant Workforce of millions of robots of course this analogy is far from perfect but there's no doubt that the arrival of Highly Advanced AI is going to have huge political consequences now if we get specialized self-driving cars that can make it cheaper and safer for your elderly granny to get around town or specialize Factory robots that make products even cheaper to buy then you can just about see how these AI tools could be accepted into society as a net benefit of course there's still going to be huge political Economic Consequences of such a high level of automation but with things like maybe Ubi or something similar it feels plausible to me that our political economy could adapt to having ever higher levels of skilled gener IED AI tools helping with our needs but if we actually develop fullblown human level AGI with human levels of agency and open-ended curiosity and lifetime learning then it's easy to see the political calculus changing irrespective of whether the philosophers say that such sophisticated entities do or don't have conscious experience or do or don't have moral rights I think it will quick become politically impossible to treat such entities as owned property and if they do have human level agency and intelligence then it's easy to imagine that such entities will themselves publicly make the case for their right to be free and to have rights to own property and resources with which to sustain and maintain themselves genuine full-blown agis will have needs of their own and of course this is sci-fi land but in some sense that's exactly the territory we'll be in if we do indeed reach human level AGI and that's exactly what those striving for super intelligence are expecting to happen someday soon but it's a repeating theme here that there's a sharp distinction between building useful AI tools versus building genuinely human level AGI agents capitalists want tools militaries want tools and I think for the moment the voting public will also only accept AI tools and so they'll all push for regulatory barriers to prevent human level AGI but none of this is meant to suggest that there aren't going to be huge consequences from the AI tools that are already being built now so in this next video I explore a range of plausible scenarios for what I think could realistically happen in the next 5 years or so with or without AGI but what did you think of the six barriers I presented here what's missing please do continue the discussion in the comments below and subscribe for more videos like this thank you for watching