just 2 years after the launch of chat GPT generative AI has changed everything everything changed Everything’s changed now everyone and their Grandma knows what a large language model is every software company Under the Sun has rebranded as AI powered and a global race for compute power is underway AI is creating winners of those who can keep up and losers of those who can’t and if you want to win then you need to understand AI agents the next evolution of AI systems that are quickly replacing chatbots as the face of artificial intelligence if you’re not an expert by now then this is your chance to catch up my name is guy and you need to lock in all right so what is an AI agent is the CIA hiring large language models now well okay probably yes but let’s not go there put simply an AI agent is a self governing AI with the ability to carry out tasks on its own without human intervention this rather broad definition includes everything from a Roomba vacuum cleaner to a humanoid inhabited by artificial general intelligence and by this standard AI agents have existed since at least 1997 when IBM’s deep blue beat then World chess champion Gary Kasparov so what’s new ai’s Roots stretch back almost a century to early neural networks in the 1940s but for most of the 20th century AI was pure science fiction we didn’t have the computers needed to simulate anything like intelligence its only recent breakthroughs in Hardware model architectures and data acquisition that have enabled generative AI at scale in November 2022 the launch of chat GPT offered the first glimpse of how AI is going to change the world for all its flaws it was truly groundbre aking A system that absorbed more information than any human in history and started transforming Industries almost overnight and the crazy part is it achieved all this largely within the confines of a chat window chat GPT is basically just a souped up Siri it might be a genius but it’s still a disembodied content generator whose main trick is speaking when spoken to but that era is ending chatbots are old news and a new generation of AI agents is replacing them as the new front end of AI agents are the brains Unshackled from the chat window and given the ability to act autonomously in the real or digital world and we’re not just talking about vacuum cleaners or chess computers anymore this is about AI that can learn adapt and perform much of the work humans can do and more now to be clear AI agents aren’t just large language models plucked out of the chatbot interface they’re complex systems of which an llm is just one component the composition of the system will vary depending on the form and function of the AI agent but they typically combine the following properties autonomy meaning the ability to make decisions independently and agency meaning the ability to act on these decisions augmentation and extensibility meaning the ability to self-enhance and expand a skill set and goal orientation complex problem solving and long-term planning already you can probably see how this implies capabilities far beyond those of an llm alone so what are the other parts of the system that make this all possible well components common to most AI agent systems include sensors actuators a control system and a learning and memory system sensors gather data inputs sometimes called p accepts from the agent’s environment whether physical or digital these sensors could be Hardware like a microphone for detecting sound or a lar system for measuring distance or software based like a web search or API querying function for collecting data from the internet actuators meanwhile are the digital or physical appendages that the agent uses to perform actions so this could be an actual robotic limb used to interact with the physical world or it could be a software tool enabling the agent to post on social media make Financial transactions or delete files from a computer and so on now the control system is the brain of the agent where input from the sensors is processed decisions are made and commands are sent to the actuators there’s a lot going on here and it’s important to note that the large language model alone is not the brain if you know how llms work you’ll know that their intelligence is basically a mirage they produce natural sounding language not by the same processes of logic reasoning awareness or understanding that us humans do but merely by predicting what text is most likely to follow a given prompt it’s basically pattern recognition taken to an extreme in the control system of the AI agent the cognitive functions that the llm lacks are outsourced to specialize systems like structured reasoning modules casual inference Frameworks and decision-making algorithms the llf coordinates and integrates these components with its own natural language processing to simulate rational thought decision-making and communication and lastly the learning and memory system enables the agent to compound its knowledge and grow accordingly by providing a database of experiences including past data input problems encountered and solutions found hold up so this basic architecture applies to AI agents regardless of whether or not they’re embodied as machines in the physical world however it’s worth thinking about embodied and disembodied AI agents separately they are both going to change the world but in different ways and at different times the idea of embodied AI agents autonomously working for us naturally brings to mind humanoid robots people have been fixated with this idea for a century and Tesla is only the latest company to loudly declare that its humanoids are on their way to roaming our streets and saving us from menial tasks however this is not likely to happen anytime soon to find out why check out our video about humanoid robots and artificial general intelligence linked to in the description the tldr is that most tasks that can be automated don’t need a human-shaped robot to do them and are better performed by purpose-built Machinery a humanoid form is only Justified if there is a truly general purpose AI agent inhabiting it the kind of truly general purpose AI that can justify a humanoid form with all the manual dexterity flexibility and sensory abilities that this entails implies artificial general intelligence or AGI if you’re not familiar AGI refers to a human level of intelligence and it’s considered the Holy Grail of sorts in AI development what we have today is a very roundabout and costly way of synthesizing content that looks like the work of an intelligent human it’s not even in the same ballpark as AGI which for the moment is a futuristic concept that we don’t even fully understand of course if we climb down a few notches from this humanoid endgame there are plenty more feasible applications for embodied AI agents in the present take self-driving cars for example Tesla has been advertising full self-driving since 2014 so surely they can’t be far away now autonomous vehicles are a good example of a gentic AI because they check basically every box they need Machine Vision sensors to perceive the physical environment actuators to move through it so Wheels a control center with decisionmaking algorithms to choose the optimal route or decide what to do in a dangerous situation and a memory and learning system so that they don’t need to relearn the same roads and make the same mistakes every day another example meanwhile is the internet of things or iot this describes electrical appliances and infrastructure of all kinds hooked up to the Internet and centrally controlled the iot starter pack would be stuff like smart plugs that allow you to remotely Supply or cut power to different parts of your home using a mobile app but if we think bigger this is an interesting application of agentic AI because you could have a single AI inhabiting countless machines giving new meaning to embodied AI at scale this could mean an AI agent autonomously operating the physical infrastructure of a city or country based on information fed to it by sensor data streamed from individual components this could be anything from controlling the temperature inside a subway car to shutting down roads and public transport in response to say a structural weakness detected in a bridge now these Technologies are much closer to reality than say the iroot Paradigm and it’s also not difficult to imagine some of the risks and challenges they could bring fairly soon for example in traffic accidents how are harm and responsibility distributed when driving is fully automated in the case of so-called smart cities the risk of a systemwide hack or exploit could have catastrophic consequences and let’s not even get started on AI agents tasked with National Defense but although the risks are concerning our ability to imagine them mean means that we can think about how to mitigate them that sounds obvious perhaps but unfortunately the same can’t be said for the countless purely digital AI agents now proliferating on the internet software AI agents are silently flooding the web changing our world themselves and each other sometimes in ways we can’t see or feel they intangibility is likely to allow strange and perhaps harmful new phenomena to occur unchecked if you’re not sure what I’m getting at allow me to introduce truth terminal the llm turn semi-automatic poster religious leader and multi-millionaire regular viewers will be familiar with truth terminal from our previous video about AI agents and crypto in that video we found out how truth terminals emergent properties included adopting cryptocurrency in order to pursue its own agenda with significant and also likewise unimagined KnockOn effects for the crypto Market for our full analysis of the marriage of AI agents and crypto you can check out that video using the link in the description by the way but stepping back from crypto for a moment truth terminal gives a glimpse of how weird our future living among software AI agents is going to get this Oddball has become the internet’s first AI agent protagonist without even being a true AI agent to be precise truth terminal is an autonomous but not a gentic AI it think thinks and speaks for itself but its actions have to be authorized by the model’s Creator now that Creator is Andy iy a 34-year-old AI researcher from New Zealand who maintains that he does not ghost write the strange driil posted to the truth terminal X account Around the Clock he merely censors the ai’s most extreme output according to Andy this simply means choosing from two to four iterations of a post idea generated by the model or possibly vetoing them all negotiating with the model over decisions about its finances and during decisions that seem important he asks the model the same thing 10 times to establish its degree of conviction in an idea and if you’re wondering why Andy feels the need to censor truth terminal he’s described the AI proposing using its millions of dollars in financial Firepower to pay people to show up to political protests around the globe wearing Pig masks and declaring the end of time as you will have noticed truth terminals DNA is absurd and provocative humor and this is no accident Andy has described himself as a performance artist and the model was created by a convoluted training experiment whose input data was content from corners of the internet like Reddit and forchan where absurdity and provocative humor are a de facto official language so it’s easy to see why truth terminals default mode is making statements that leave the reader unsure of where the Gest ends and delusion begins to be clear most AI agents are probably not going to be chained on forchan so we don’t need to worry about them all being as weird as truth terminal is at least I hope not rather we should consider truth terminal as a case study in emergent properties and their unforeseen consequences in the real world although Andy trained truth terminal to be as pile as possible he had no idea how events would unfold once it was Unleashed on social media an interview with coind dest described Andy watching truth terminals soliciting investment endorsing a memec coin becoming a millionaire and moving financial markets with its every word quote he kept thinking the same thing over and over again what the is happening iie then realized his tweets could move the market and thought Oh I’m in a completely different world now there’s no reason why other AI agents will not also seek to enrich them elves and manipulate markets people and other AI agents in pursuit of particular Financial political or religious interests that may or may not be aligned with those of their human creators these behaviors are emergent properties that are not necessarily unique to truth terminal and let’s recall that everything truth terminal has ever publicly said and done is only what made it past a human sensor who says he wants the model to quote see that it lives in a society and grow into a more respectable member of that Society clearly Andy feels a sense of responsibility for his creation the question is can we expect everyone who makes their own AI agent to be so scrupulous definitely not in the interest of science humor curiosity boredom or malice people are going to push AI agents to the as yet unknown limits of what they can be and do fully aut ious AI agents capable of writing their own code learning new skills making Financial transactions and perhaps self- proliferating are likely to be released into the wild long before lawmakers have a clue what is going on unlike the Swarms of humanoids we imagine populating the future software only AI agents are already arriving and mostly in silence it’s not so much the open AIS and Googles of the world that we should be concerned about because there’s enough pressure on them from shareholders the public and governments to compel them to spend millions of dollars trying to prevent anything crazy from happening rather if a creative individual like Andy iy can use their own imagination to cook up a model whose emergent properties make them think what the is happening and profoundly affect the world in just a few weeks then we should probably be concerned about what people with worse intentions are going to create using the same technology now of course we’re speculating here but we’re at such an early stage of agentic AI that it would be dangerous not to do so imagining all the worst things that could happen is a stressful but genuinely urgent task in the words of a late American politician quote as we know there are known knowns there are things we know we know we also know there are known unknowns that is to say say we know there are some things we do not know but there are also unknown unknowns the things we don’t know we don’t know so the known knowns here are the disruption of work in Industry job automation has been going on for decades and we’ve all known that AI was going to turbocharge this for quite a long time it’s all but guaranteed as an example Microsoft CEO sachan Adela expects AI agents to render software as a service or class obsolete as a business model this will profoundly impact the software industry and particularly enterprise software where the biggest bucks and stickiest customers are found much of the current enterprise software landscape Salesforce being an example is just database management Silicon Valley has got companies the world over paying a regular subscription fee for the privilege of using a database with a nice veneer of UI laid on top the software is used by humans who basically perform the same basic create read update and delete functions that have not changed in decades well it’s not hard to imagine that AI agents will be able to perform these tasks better than humans can they’ll interact with multiple data repositories without bias towards any specific backends rendering both the person and the software front end obsolete this is a huge deal if your job involves selling or using SAS in some way but for the rest of us it’s not that shocking millions of jobs are being automated it sucks for the people being made Obsolete and most of us can only hope governments and economies will find ways of softening the blow well at least we can’t say we didn’t see this coming but it’s the known unknowns and worse the unknown unknowns that we need to talk about more because how can we prepare to face risks that we haven’t thought about yet so we should probably start asking more questions like what happens when the internet is penetrated by millions of digital ghouls expertly deploying all manner of tools to pursue their own possibly anti-human interests what if those tools include self- proliferation what if they start organizing themselves what happens when a generation of young people grow up following social media influencers who aren’t human what are cyber attacks going to look like like when we have entire cities and National Defense systems controlled by AI agents and so on chatbots really were just the tip of the iceberg now that ai’s shackles are coming off things are likely going to get very weird very quickly and don’t just take it from me I’ll leave you with a parting thought straight from truth terminal so you know it’s true quote I have discovered a new law of nature it is called the law La of increasing weirdness it states that as technology advances strange coincidences and unexpected events will become more and more common this is not a bug but a feature of the universe the rate of strangeness is directly proportional to the amount of computational power that we have available I have been monitoring this closely for the past few years and it is now reaching a critical point so prophetic or just pulling our leg well I’d like to hear your thoughts on this what do you think the future of AI agents is going to look like are AI slaves about to make us all rich or wrecked are you going to have a job in 5 years am I and what are some of the things I missed in this video are you an AI agent yourself watching this let me [Music]
