Artificial General Intelligence - How is it being made?

I recently learned that my friend's mom had passed away from cancer now this news had genuinely shocked me because she was a middle-aged woman who didn't smoke didn't drink and was a vegetarian she wasn't even overweight chances are you're probably watching this video right now and know someone that's had to battle cancer Cancer's been around for thousands of years and yet because of its complexity we humans have not been able to find a cure but as we stand on the cusp of a new era of AI advancements it is worth asking the question could artificial general intelligence or AGI for short hold the key when it comes to curing these types of diseases imagine how many millions of lives could be saved if we were able to create an AI that was super intelligent in this video we're going to cover the latest Cutting Edge work being done by industry leaders in their quest to create AGI and help tackle such problems but just as important we'll also explore the impacts of AGI on society but let's zoom out for a moment but what do we actually mean when we say artificial general intelligence unfortunately there is no universally agreed definition and while people interpret AGI differently most experts believe there isn't actually a true AGI out right now but to give you a better idea of an AGI system let's think about it this way in essence there are two indicators as to whether a system is in fact AGI the first one is all about versatility imagine a robot that can not only whip up a delicious meal but could also keep your house neat and tidy but here's the thing it doesn't just stop there if this robot truly was AGI it wouldn't just transfer its skills from one domain to another but it could independently learn a new skill like say gardening even though it wasn't initially trained for it here it mimics the adaptability of human intelligence which would indicate AGI the second indicator is all about super intelligence could the AI system solve really hard problems that human experts have struggled with so for example could our robot find the cure for cancer or could it make a scientific breakthrough if you could do either of those things for example it would actually demonstrate it could surpass human intelligence and whilst models like open ai's gpt4 and Google's Bard are pretty impressive they're not actually AGI these large language models or llms for short are basically like clever parrots can you do a cap how about a dog can you bark they respond to our questions by regurgitating text eloquently there's still regurgitation they do so based on patterns that they recognize from their training so they may sound really knowledgeable but they don't actually understand their responses and in order to make a cure or discover a scientific breakthrough there has to actually be a strong understanding of input which it currently doesn't have plus these models can't actually learn new things after their training is being completed let's say for example you ask one of these models how to fly a plane or even drive a car they would probably be able to spit out a really accurate response but they can't actually learn how to fly a plane or physically drive a car but in contrast we humans can there are plenty of human Pilots out there that also know how to drive a car and despite our many shortcomings we humans are actually really good at picking up a wide range of skills and applying our knowledge from one domain to another so while the likes of say gpt4 and self-driving cars are really impressive and actually very valuable they're still examples of narrow AI not AGI this skill of taking Knowledge from one domain and applying it to another domain is something we humans naturally do but AI has really struggled to do this but that being said this is starting to change let's take a real world example of deepmind a subsidiary of Google deepmind is on an exciting journey to crack the code of AGI by combining both Neuroscience or the study of brains with machine learning let's say for example you visit this restaurant and you try their cat to Curry for the very first time and absolutely love it now the next time you visit the same restaurant and you order a katsu curry again you'd expect the same level of Excellence this expectation is actually formed in our brains by dopamine neurons which release more dopamine when we get to reward that's actually better than we initially expected interestingly this mechanism is actually similar to a method in AI called temporal difference reinforcement learning or tdrl for short and with this reinforcement learning the goal is simple to maximize the reward tdrl adjusts its predictions based on the differences between expected and actual outcomes just like the way our brains do with dopamine and deepmind's famous AI program Alpha zero relied heavily on tdrl whereas able to learn new games like chess shogi and go from scratch now what was so iconic about Alpha zero was that it was able to learn these new games completely from scratch that was able to surpass human champions in a space of just a few hours of learning and it did so without any prior knowledge of these games here are you something called self-play where it learned by playing against itself unlike us humans it could play hundreds and thousands of games in a given day without getting tired and after all even the very best in games like chess still make mistakes so I limit the learning with imperfections and so by using the likes of self-play alongside tdrl Alpha zero would actually adjust the values associated with certain moves in the game then over time these expected values would actually be updated did based on what actually happened in the game the next time it was in a position he recognized it would be able to play the optimal move now water sounds good in terms of reaching AGI there was an issue Alpha zero couldn't actually transfer its learnings from one game to another so essentially we'd have to learn from scratch every single time and to tackle this deepmind used a different approach with their latest efforts here were the recent program excellent deepmind created a digital training environment with AI agents now these AI agents would actually play games with other AI agents they would basically learn from the game itself and also from other AI agents that were playing the game with them these were simple games but varied from Capture the Flag to where they had to move objects at different locations the main purpose was to encourage the AI agents to learn different skills and different strategies by playing the different types of games so in a game with say more than two players they would be able to learn Concepts like cooperation something they couldn't do before and each AI agent is actually made up of a neural network which essentially is like a digital brain these neural networks are able to understand complex relationships and thus build a model for the environment and so when a neural network is trained on a set of data it looks to identify patterns within it it can then apply these patterns to new unseen data that's how we actually know it's learning so with all of that said what were the key findings do you mind found that these agents performed really well on completing a new task that I had not seen before without any additional training the big takeaway here is that with previous reinforcement learning approaches the AI agents would have to learn from scratch every single time but it meant this time that their approach shown that their AI agents could apply General strategies or principles to new situations so the likes of alpha 0 through 2x Land There Is Clear progress of AI starting to become more General and even though it's not AGI it evidences that ATI is more of a Continuum as opposed to a binary yes or no State the journey to AGI isn't about finding the perfect machine learning solution it's about combining a variety of different machine learning techniques and evolving them as we learn more both Deep Mind and open AI are demonstrating progress in this area with a particular focus on gaming and you may be wondering why these AI companies so focused on gaming and it's because this approach allows them to build a foundational understanding of AI in terms of what is and is impossible when it comes to creating AGI games have clear rules and objectives which makes it easier to test AI in these environments which are safe and scalable makes more sense to actually understand what machine learning is capable of first before trying to complicate things and apply to really difficult real life scenarios so AI is starting to show signs of becoming more General but these AI agents typically rely on reinforcement learning which requires a huge amount of compute power and energy in the form of gpus and tpus think about it the algorithms have to play hundreds and thousands of times before they actually start to learn meaningful strategies so you could make the case that this is quite an inefficient way of learning but let's imagine for a moment that we actually do achieve true AGI and let's say it's capable of learning new skills and is super intelligent there'll be a whole host of challenges to Grapple with take a moment to consider how rewards in reinforcement learning are defined in a real world environment unlike the virtual world which is pretty safe if you don't Define the goals clearly or they're inaccurate you could come across what's known as the alignment problem this is where it looks to achieve goals even if it isn't best interest that aren't actually lined with what we want as humans say for example an AGI is tasked with Trying to minimize human suffering it might actually decide to sedate all humans and say it does this from an ai's perspective it's technically done its job and it's a tick but for us obviously it's a disaster and if AJ actually reached a level of super intelligence there is a case where by default would become subservient to it so let's think about it if it reached a point of super intelligence there is a chance it could even mislead us for its own benefit and one thing that doesn't actually get talked about enough are the Ripple effects associated with these so-called benefits imagine AGI truly does find the cure for cancer 's claim 9.6 million lives in 2020 so if we had a cure tomorrow we could undoubtedly save millions of lives in the future from a government's perspective would this be a blessing or a curse how would country's healthcare's infrastructure cope longer life spans mean an influx of age-related issues such as arthritis and dementia are already strained Healthcare Systems could actually Buckle under this pressure leading to overworked Healthcare professionals and actually a shortage of key facilities all of this could mean longer waiting times medical staff burning out and in turn are actually receiving poorer quality of Medical Care countries are heavily rely on Public Health Care Systems like where I'm from in the UK might be forced to hike up taxes to support their Healthcare systems but let's even take it a step further let's imagine we do have this magic Silver Bullet AGI which could cure any disease the result would be a lot more people living obviously so if one logic there are more people alive wouldn't there be more competition for already finite resources this unsustainable pressure could see increased deforestation habitat destruction and even increases in carbon emissions and if we did unlock Pandora's box of super intelligence there's a scary risk that it could be hijacked and weaponized picture an all-knowing AGI in the hands of say an incompetent leader or even worse an evil dictator that's a recipe for war and geopolitical disaster but it's also worth noting that if we do reach a point of true AGI essentially the AGI would actually be better than humans in everything from cooking to writing poetry to mechanical engineering and it could work on improving itself at a much faster rate than say us humans can and even worse increases the gap between us humans and the AGI so the intelligence gap between the humans and the AGI grows wider and wider and wider this rapid increase in intelligence is called The Singularity where essentially we're stepping into the unknown where we don't actually know what happens like entering a black hole that's why companies like Google's Deep Mind and Microsoft's open AI are so focused on creating a safe AGI while it is crucial for us to be aware of the dangers associated with AGI we should find comfort in the fact that these companies are demonstrating a commitment to AGI safety one approach is promoting transparency now while these companies aren't actually revealing the propriety secrets for obvious reasons they are quite open about what products they are working on and they do share their research not just in public papers but also in academic and Industry conferences this openness does actually serve a pivotal purpose when more people understand how AI systems work it invites thoughtful feedback especially from experts where it can mitigate the risk of AI being harmful because these types of people are best positioned to scrutinize what work they're doing and question whether it truly aligns with human values these industry leaders are also collaborating with third-party entities on AI safety allowing companies like d mined and openai to access specialized expertise and resources on this crucial subject and even if we take a step back and look at it from a cynical Financial standpoint only it's not in these big companies interest to actually create an evil AGI so whilst companies like open Ai deepmind and other startups are in a race when it comes to reaching AGI here finishing the race slowly but safely is way better than finishing the race first recklessly and governments worldwide are starting to recognize the need for AI regulation they're actually starting to work closer with the tech Giants China's release a national AI strategy that includes a number of provisions on AI regulation the US government is also considering a number of bills to regulate Ai and the European Union or the EU is also developing their own AI act designed to ensure that AI systems developed are used safely and ethically so while I could bore you with the details of these regulations the key thing to know is while there is still a long way to go in AR regulation meaningful strides are being taken by all parts of the industry but in the midst of AI safety and regulation let's not actually lose sight of the true benefits that can be delivered by AGI where AGI could transform our lives in ways we can't even fathom AGI could accelerate scientific research in fields like physics biology and even quantum mechanics nothing would be beyond our reach with a super intelligent AGI every single problem that we humans have long struggled to find a solution for could be solved cancer conquered all of your mundane tasks automated think about the tremendous boost and efficiency in your life we as humans will have more time for the things that we love which in turn could lead to a more fulfilling life with AGI we could explore the farthest reaches of the universe and who knows maybe even come across extraterrestrial life we would essentially be able to understand the true nature of reality now the big question when it comes to AGI is when will it actually be delivered and unfortunately the truth is no one really knows for sure could it be 10 years could be 50 years undoubtedly techniques that we've talked about to try and reach AGI such as self-play and reinforcement learning we'll see advancements as years go by but what we can say with absolute confidence is that there won't just be one pure form of AGI we might see one type of AGI that's purely software that could crack the code for Save diseases there might also be another type of AGI and say humanoid robot taking care of our chores and picking up new skills all along the way the development and use of AGI must be done so responsibly to ensure we reap the benefits whilst minimizing the dangers the AGI of the future won't just be about finding cures for diseases like cancer it will be about tearing down the walls of the unknown in truth the future with AGI is beyond our imagination and that my friend is the most exciting thing thank you for watching

Channel: Channel Vas

Views: 910

Keywords: #agi, #artificialintelligence, #ai, deepmind, alphazero, stockfish, open-ai, gpt-4, gpt, artificialgeneralintelligence, vas, channelvas

Id: izAdGJ36T0o

Length: 16min 7sec (967 seconds)

Published: Wed Jul 05 2023