WARNING: Summaries are generated by a large language model and may be inaccurate. We suggest that you use the synopsis, short and long summaries only as a loose guide to the topics discussed.
The model may attribute to the speaker or other participants views that they do not in fact hold. It may also attribute to the speaker views expressed by other participants, or vice versa. The raw transcript (see the bottom of the page) is likely to be a more accurate representation of the seminar content, except
for errors at the level of individual words during transcription.
AI is revolutionizing the way humans do things. With the introduction of chatbot technologies, people can automate mundane tasks in minutes, leading to increased productivity. Major tech companies are competing in an AI arms race, taking different approaches to the technology. Privacy concerns are a key factor in the development of AI technology, and software is unique in its ability to be adopted quickly and easily. AI has immense potential to revolutionise actions undertaken by humans, and its rapid adoption could lead to economic rents and startup opportunities.
John Carmack believes AGI is possible with current computing power and that he has a small chance of producing the key insight. He has set up Keen Technologies to pursue old ideas that have been missed, and is unlikely to have a monopoly on whatever is found. Attention, which underlies Transformers, could have been developed 20 years ago, but technological advancements often require a surplus of resources to be discovered and pursued. Carmack's example and the potential rewards of AGI research may make it difficult for governments to control the race dynamics. Commander Keen was a game made by the progenitors of ID Software before Doom and Quake.
Chat GPT has become the fastest growing consumer application in two months, reaching 100 million users. This rapid growth has highlighted the potential of AGI technology and has sparked a discussion about the potential existential risk posed by AI. Google and Microsoft have invested heavily in research institutes and Open AI respectively, showing the importance of this technology. The iPhone was also a pivotal moment, as it gave people a view of a new kind of future, and this is a similar turning point. Governments, universities and schools are now developing policies to respond to the potential disruption of AI.
The introduction of the iPhone and App Store in 2007 revolutionised smartphone technology in a similar way to the Wright Brothers' first powered flight in 1903, which changed conversation, planning and investment. Chatbot Technologies have been adopted much faster than expected, surprising even the developers. It is possible that Google and DeepMind knew the capabilities of the technology, but chose not to release it due to potential risks. The Innovators Dilemma has been a key factor in the adoption of chatbot technologies, showing that the potential of the technology was underestimated.
Google has launched a chatbot technology in their search process, but the current models have flaws and could lead to a competitive period between tech companies for market dominance in AI applications. There is a risk of errors that could cause a drop in stock price, and companies have been embarrassed by their bots' racism and biases, leading to hard lessons and a need for caution. This has created an arms race between companies disregarding risks in pursuit of the prize.
Major tech companies are competing in an AI arms race, taking different approaches to the technology. Apple is investing in privacy and local computing, while Amazon attempted a bot that lived in homes but was unsuccessful. People are highly sensitive to their privacy rights, making local computing a strong foundation for AI technology. This could be a distinguishing factor between companies, with those investing in data center computing and those making use of local computing. Privacy concerns will be a key factor in the development of AI technology.
People are giving up their privacy to access increasingly indispensable software and internet services, even though some of these services are used for illicit activities. With the rise of AI services, people may give up all their privacy and personal data to access them. Corporations and countries are using AI models, raising questions of sovereignty as control over these systems may be outside of the borders. Open source alternatives to the development of Transformer and diffusion based image generating models have been discussed, although the training requirements for language models are costly and difficult to compete with the Giants. Distributed super computing initiatives such as SETI and Folding at Home have been discussed as potential open source alternatives, but it is uncertain whether they will provide a useful democratizing function.
Software as a Service (SaaS) is a rapidly adoptable technology that can harness the computational power of people's home computers. However, the gap in performance between the latest semiconductors and what people have in their home machines makes distributed training of cutting-edge models difficult. Chat GPT requires insider knowledge and manual labor, making it difficult to democratize. Software is unique in its ability to be adopted quickly and easily, leading to rapid transformation of AI. Traditional patterns of technology adoption do not necessarily apply to software.
Chat GPT is an automated system used to quickly and efficiently perform mundane tasks such as writing essays and emails. There is a meme where Chat GPT is used to demonstrate how manners can be used to convey a sentiment without actually saying anything. Evolutionary psychology suggests that manners form an important social function in most cultures. Software has enabled automation of many cognitive tasks, such as those related to greetings and rituals, resulting in disintermediation and the removal of physical equipment. Chat GPT is an example of this disruption, providing a time-saving and cost-effective solution.
Computing is a general purpose technology that has enabled mechanization and disintermediation of existing specialized technologies, as well as new value creation. It is unique in its ability to transform activities and has the potential to revolutionize productivity. AI is a narrow technology that has a profound impact, while GPT and software have improved with the speed of computing. AI is still in its early stages, but has the potential to transform the way we do things.
AI has immense potential to revolutionise actions undertaken by humans and should not be dismissed as science fiction. Experts have not fully grasped the potential of these technologies, leading to the public not understanding their impacts or how quickly they will be adopted. With a few downloads, individuals can automate 75% of their job, and the technology is only becoming more powerful. Get-rich-quick schemes have been used to spread awareness, but more effort should be put into explaining why these beliefs are reasonable.
Rapid technology shifts create opportunities for professionals to scale their caseloads, leading to the emergence of startups built on chatbot and other technologies. To understand the push and pull incentives that drive adoption, a research team is modelling disruptions and an epidemiological model. Adoption follows an epidemiological model, where individuals act as carriers of the new knowledge, but there may be economic rents to be extracted from those yet to become aware. Looking forward, the question is how much faster adoption can get than 100 million people in two months.
John Carmack expressed a belief that AGI could be possible with the current computing power available and that he himself had a small chance of producing the key insight. Hopfield networks, a simple system with simple interactions, were influential in neuroscience and the modern Transformer is similar to this. Attention, which underlies Transformers, could have been developed 20 years ago. However, it did not happen that way.
Technology advancements often require a surplus of resources to be discovered and pursued. This is due to the inertia of the existing system which makes it difficult to explore new ways of doing things. However, those who can identify these opportunities before they become available can gain a competitive advantage. This idea can be explored further, as it could lead to the development of new business models based on technology.
Carmack's Keen Technologies is an example of how difficult it is to coordinate in the race for AGI. He has the resources and freedom to look for old ideas that have been missed, but it is unlikely that he or his company will have a monopoly on whatever they find. It is highly unusual for someone to have the resources and freedom to pursue this, but there is a non-zero chance that Keen Technologies will be up there with DeepMind and OpenAI. Governments will be unable to effectively control the race dynamics due to Carmack's example and the potential rewards of AGI research. Commander Keen was a game made by the progenitors of ID Software before Doom and Quake.
Chat GPT has become the fastest growing consumer application in two months, reaching 100 million users. This is an unprecedented rate of growth and highlights the potential of AGI technology. It is difficult to predict how governments may react to this technological advancement, however, given the potential risks posed by AGI, it is important to consider all options.
With the advent of chat GPT, governments, universities and schools worldwide have been developing policies to respond to the potential disruption of Artificial Intelligence (AI). Google has invested 300 million dollars in Anthropic, an AI research institute, and Microsoft has invested 10 billion dollars in Open AI. This has sparked a discussion about the potential existential risk from AI. This is a turning point, as the capabilities of chat GPT have been latent for 10 months, and the awareness and anticipation of new capabilities is increasing. The iPhone was also a pivotal moment, as it gave people a view of a new kind of future.
In 2007, the introduction of the iPhone and the App Store revolutionised smartphone technology, changing the way we interact with and use our phones. This was comparable to the Wright Brothers' first powered flight in 1903, which was a major shift in conversation, planning and investment. There were no precursors to the Wright Brothers' event, as they were sceptical it could be achieved so soon. Within 15 years, Aviation had become an integral part of warfare and changed the course of World War One. The moment of powered flight was primitive, yet it pointed to much more powerful Aviation capabilities emerging quickly.
Chatbot Technologies have been adopted much quicker than expected, surprising even the developers. There was a systemic industry-wide underestimation of the value that these technologies would bring to users. It is possible that Google and DeepMind knew the capabilities of the technology, but chose not to release it due to potential risks to their business models. It is also possible that they simply did not realize the full potential of the technology. The Innovators Dilemma has been a key factor in the adoption of chatbot technologies.
Google's introduction of chatbot technology into their search process carries extreme risks due to the potential for errors to be highlighted and cause a drop in stock price. Microsoft and Open AI can take these risks more easily due to their different positions, but Google has been forced to move quickly and launch their product in search. This could be a mistake as chatbot technology is not yet ready to be integrated in this way and the current models have many flaws. This could lead to a competitive period where a prize is to be claimed for the successful implementation of this technology.
There is intense competition between tech companies for market dominance in AI applications, with the potential for immense profits. This has led to an arms race, with companies disregarding risks in pursuit of the prize. The previous status quo of a clear leader and gap between runners-up allowed caution and risk acknowledgement to take place. However, some companies have been embarrassed by their bots' racism and biases, leading to hard lessons and a need for caution.
Google and DeepMind's attempts to be ethical and risk-averse have been overshadowed by the competitive landscape, making it difficult for upstarts to coordinate with major players. This has led to an arms race in AI technology, with the tech giants taking different approaches. Apple is conspicuously absent from the discussion, but is likely to emphasize local compute and privacy. Amazon attempted a bot that lived in homes, but it was unsuccessful. As technology develops, Apple has made large bets on privacy, investing in health and other areas.
People are highly sensitive to their privacy rights, making local computing a strong foundation for AI technology. This could be a distinguishing factor between companies such as Apple, Open AI, Microsoft and Google, who are investing in data center computing, and those who are making use of local computing. There are various levels of sensitivity to privacy concerns, from those who take action to protect their privacy, to those who are unaware of the issue. This will be a key factor in the development of AI technology.
People are unaware of the implications of giving up their privacy to access software and internet services, which are becoming increasingly indispensable in daily life. Although some of these services are used for illicit activities, people do not seem to hesitate in giving up their privacy. It is likely that when AI services become as indispensable as expected, people will give up all their privacy and personal data to access them, knowingly and unknowingly.
AI models are increasingly being used by corporations and countries, which raises questions of sovereignty as control over these systems may be outside of the borders. This could lead to a situation where a company's lines of communication are embedded with AI playing crucial roles and if access to the API is cut off, the company may not know how to communicate. Australia may not have the supercomputing capabilities to have one of these models inside its borders, making it a key sovereignty risk.
There has been discussion about open source alternatives to the development of Transformer and diffusion based image generating models. While there has been more movement on the image generating front, the training requirements for language models are so costly that it is difficult to compete with the Giants that have massive cloud computing resources. Recently, distributed super computing initiatives such as SETI and Folding at Home have been discussed as potential open source alternatives for training large language models. However, it is uncertain whether these collective efforts can provide a useful democratizing function in the arms race or if they are just a nothing burger.
SoftWare as a Service (SaaS) is a rapidly adoptable technology that can be used to harness the computational power of people's home computers. However, the gap in performance between the latest semiconductors and what people have in their home machines is too large for distributed training of cutting-edge models. Additionally, the complex systems required to produce chat GPT require insider knowledge and manual labor, making it difficult to democratize. Therefore, SaaS remains a rapidly adoptable technology, but democratization of cutting-edge AI models may not be possible.
Software is unique in its ability to be adopted quickly and easily. Chat GBT reached 100 million users in two months, a speed of disruption that is much faster than the typical 10-15 years for other disruptions. This has implications for AI, where software can be quickly and easily adopted, leading to rapid transformation. This suggests that the traditional patterns of technology adoption do not necessarily apply to software.
Manners are often used as a performative function to demonstrate respect and politeness, even when there is no real content to be conveyed. Chat GPT was adopted rapidly due to the latent demand for automating mundane tasks such as writing essays and emails. There is a meme where a person inputs 'say thanks' and Chat GPT outputs a two-paragraph professional sounding email, which is then summarized by Chat GPT as 'thanks'. This is a demonstration of how manners can be used to convey a sentiment without actually saying anything. Evolutionary psychology suggests that manners form an important social function in most cultures.
Software has enabled automation of many cognitive tasks, such as those related to greetings and rituals. This automation is seen as a form of disruption, as it is time-consuming and costly. Disintermediation is a key part of this, as software has removed the need for physical equipment and replaced it with digital services. This has enabled a wide range of tasks to be automated, as the water level rises, and the chat GPT moment is an example of this.
Computing is seen as an enabling technology for disintermediation and displacement of existing specialized technologies. It has a broad sweep of capabilities and can facilitate new value creation. Other examples of technology with transformative potential are language and writing, and electricity combined with motors, which enabled mechanization. Electricity can be transformed to and from kinetic energy, allowing it to capture and release energy in a way that was not possible before.
Computing is an example of a general purpose technology that has allowed for the mechanization of a wide range of activities. This is a phenomenon that is unique to computing, and possibly also intelligence. It takes time for people to shift activities to the new technology, and this is why AI has not yet had a major impact on productivity. GPT is a great example of this, as improvements to the technology automatically make any use of it better. Software is another example, as the speed of computing has improved all office software. AI is its own category, but it is still narrow and has a profound feature.
Artificial Intelligence has immense potential to be embedded in and factor through virtually any action humans undertake. This could lead to almost unbounded transformative potential, yet its potential is not widely appreciated. The public, media, reporting, and entertainment have long dismissed AI as science fiction, viewing the internet as a glorified phone book before it became a reality. AI could lead to countless new possibilities and should be taken seriously.
Experts involved in the development of new technologies appear to have not fully grasped their potential and how profoundly useful they will be. This has led to the public not fully understanding the impacts of these technologies, and how quickly they will be adopted and become disruptive. With only a few downloads, individuals can automate 75% of their job, and the technology is only growing more powerful with each generation. Get-rich-quick schemes seem to be the main vector for spreading awareness of the capabilities of these systems, but more effort should be put into explaining why these beliefs are reasonable.
Rapid technology shifts create opportunities for professionals to scale their caseloads by 10x, but this won't last long. To raise awareness, people are building startups on the back of chatgpt and making money, but there is a question of whether it is worth trying to convince people of the potential of these technologies. A research team is working on modelling disruptions and an epidemiological model to better understand the push and pull incentives that drive adoption of new technologies.
Adoption of new technologies, tools and ideas follows an epidemiological model, where individuals act as carriers of the new knowledge and infect others. However, this model may break down when it comes to the details of how and why infection occurs. During the inter-regnum period, there may be economic rents to be extracted from laggards who have yet to become aware of the technology. Looking forward, one question is how much faster adoption can get than 100 million people in two months, as it is possible to imagine technologies that would be adopted with astonishing swiftness by everyone who is conscious.
GPT 5 has recently seen a rapid adoption, with 500 million users in one week. However, there are some who have been unimpressed with the technology and have disregarded it. It is difficult to determine the percentage of people who have had this experience, but it could be significant. Computers may seem like a magical technology to some, but it is necessary to understand why it is impressive. The models have evolved significantly and become much better, but people may not be aware of chatbots or other similar technologies.
Adoption of OpenAI's chatbot is driven by particular use cases, with only a small number of people having figured out how to make use of it. Adoption is likely to vary across different populations, with younger people more likely to adopt it than older people. It is important to consider the value the technology creates, as this will determine which population is most likely to adopt it.
so I I actually I read a um an interview with John Carmack um I think you must I saw that interview yep yep and um uh I mean you know I think there's a he seems to be a fairly grounded person I don't know how you know how crazy or egotistical or anything he is but he expressed it seemed fairly um barely reasoned uh and not hyperbolic um he was he was I I thought you know he didn't really come it didn't really present as unhinged to me let's say but there was an extraordinary claim that he was making and the claim he was making was that there was a small but non-zero chance and he was saying one or two percent chance that he himself as an individual could uh produce the Insight produce the insights or perhaps a key Insight that made AGI possible and I think he I think he suspects that there is a large enough Hardware overhang right now that in fact I'm quite sure he believes that um there's a large enough Hardware overhang right now that AGI would already be possible with the compute we have available and it is uh it is just a a software um yeah and and sort of research challenge at this point it's no longer a a situation this is his claim is that it's no longer a situation where we we are lacking the the Computing horsepower to make something like a system like this work I thought those were pretty extraordinary claims but he's an extraordinary individual so yeah uh I don't know what you made of that sounds very reasonable to me uh so you could say that the the underlying algorithms of the Transformer are very similar to Old ideas um so the very one of the very first pieces of neural network research how should I say this like kind of a mathematical physics take on neural networks that had an impact in Neuroscience was hopfield's work so the hopfield network was a basically a mathematical physicist took some of the simplest statistical mechanics models something called the icing model modified it a bit and used it to show how simple ingredients like a a simple system with uh simple interactions could give rise to a an approximate memory and this was very influential in neuroscience and the modern Transformer is in many ways very similar to a hopfield network uh quite a few people have made a made hay out of this particular grass so the attention idea that underlies Transformers could have been developed 20 years ago it was and that isn't really quite how it happened so help field networks aren't really I think what inspired the authors of the Transformer paper
but that's an example of an idea that was latent and its true power wasn't realized until somebody uh pointed it at enough resources and enough data and that could I think CarMax point is that that could be out there again and it may be enough to get us all the way to AGI you know some old paper of schmidhubers might actually contain all that we need if you just really do the engineering properly on it and I think suitskiver has said similar things on Twitter so that's interesting it makes me think that they're and and I believe Owen's with us and so I uh this is something for my team our team gets to think about it and look through the the historical cases of disruption the pattern there would be uh maybe not necessarily just disruption I'm sorry but technological advancement in general but the pattern would be something like a a new way of doing things is revealed when enough Surplus capacity of some sort becomes available and the the an enabling Surplus uh that facilitates exploration and Discovery and innovation then in hindsight shows that the same the same advancements or the same capabilities might have been achievable with fewer resources in retrospect but they they worked discoverable or pursuable in absence of a you know sort of a condition of abundance or a condition of surplus and that may be a that may be a general feature for the inertia emergence for example of a new business model based on a technology or something like that where yeah the the the the you have to reach a certain stage of development in the progression of capability whatever that might be in order to discover a new way of doing things even if in principle doing things in that new way would have been possible feasible perhaps even economically or financially viable and competitive uh with fewer resources or less capability that's that strikes me as a general phenomenon we might want to explore and it would be of interest because you could then imagine a competing landscape or circumstance circumstance environment maybe is the way to say it a competitive environment where folks are looking for examples of that it could be that that rather than wait for a surplus to reveal these new opportunities through the normal historical mechanism of Discovery amidst plenty you might you might accrue a competitive Advantage by trying to find those opportunities before they are you know that the um or sooner maybe that's already already part of the competitive Dynamics anyway it's an interesting I think so
what I'm saying is that this may be a maybe a more General phenomenon than is simply specific to Computing Technologies yeah I think that's that's CarMax point isn't it so he said in that interview that uh maybe that doesn't happen really so you might imagine that given the potential rewards there would be people out there trawling through the old ideas to look for things that were missed uh but somehow were much more of a bandwagon hopping species than that so the vast majority of the resources is spent in a very small number of directions for reasons to do with how we coordinate our activities and incentives and careers and so on and it's actually highly unusual for someone like Carmack to have those resources and be sufficiently free of all those structures to just think hey you know maybe it's out there and I'll go take a look um so yeah I actually think he's got a pretty good shot at I mean it would be very difficult to keep it secret so I I very much doubt that he or his company will um you know have a monopoly on whatever it is they find but he certainly seems like there's a non-zero chance that Keen Technologies ends up being up there with deep mind and open AI right like if they just get lucky they could do it so from the point of views from the point of views yeah from the point of view of safety uh I think it's probably indicative of how difficult it will be to coordinate in the race dynamics that have already emerging when people like comic uh can just go do their thing and have a non-zero chance of making it work it's part of the reason why uh this isn't controllable right the the idea that you have a government even if it starts looking like it's going to work as in like a year ago um the idea that governments will be able to come in and uh lead coordination among the leading labs in order to slow down or impose more safety constraints or whatever that isn't that might might be possible in some kinds of Technology like nuclear weapons for example but what Carmack is a walking counter example to the idea that you'll be able to do anything effective in that direction with AGI I think the name of the company is Keen Technologies as in Commander Keen you're probably too young to know Commander Keane but us ancient people may have played um Commander Keane it was one of the games that was made before doom and Quake and all the famous ID software games I think it may have even been made by the apogee games when the progenitors of ID software worked there may not be an ID
software again yeah I can imagine I mean certainly a society certainly in in Western Democratic uh you know patterned societies that you know where there's where it would be implausible to expect that you know government could just snatch some individual person who posed a risk of you know developing AGI in their basement off the street I mean we did it you know sort of X-Files and and and uh you know Cloak and Dagger kind of uh Hollywood depictions of of the US governments and its agencies notwithstanding I think this is not a reasonable expectation that something like this would happen I mean it is if if somebody were I suppose engaging in potential terrorism like bioterrorism or something like that that you could imagine governments taking action and snatching up individuals or small teams and putting them in the proverbial Guantanamo Bay kind of uh environment or Worse um but I I I I I I think there are some places in the world where it's not unreasonable to expect that individuals could be um you know they could be targeted the question is whether or not we could know or any government could know in advance and monitor that activity and take action soon enough so there's sort of two questions there there's one the question of uh is is is it knowable is is the information is the knowledge available such that you could identify sources of risk uh before you know key breakthroughs were made and then it's a separate question of could you could you act or could you expect a government or an agency to act on that information and um yeah I don't know what the answers to those questions are and I would normally find it very distasteful to even contemplate that except that we are potentially in a situation not not without parallels to for example the development of nuclear weapons right that I think most of us agree that there is at least some potential that AGI could be as impactful and potentially as dangerous as that um so I suppose nothing is off the table at least in principle given the stakes and the risks but what we what we might reasonably expect uh yeah it's it's difficult it's difficult to say yeah this is touching on our earlier conversation around timelines and maybe it's worth updating a little uh given recent events on that discussion before we move on to the the topic that um that you identified for today Adam so the just to recap what's happened in the last week uh so chat GPT has crossed 100 million users in two months and is the fastest growing consumer application in
history uh universities worldwide schools worldwide are developing policies to respond to chat GPT in terms of assignments and so on um I think there are very few academics who haven't been preparing for that as they return to the next semester uh many governments people have given speeches MPS in the UK and Australia an MP a federal MP from Victoria where I am gave a speech in Parliament and actually mentioned existential risk from AI directly which is the first time I think that's been mentioned in that body I think there's been similar speeches in the US and the UK probably elsewhere so it's the the idea of AI progress reaching some kind of disruptive level and potentially dangerous that's still on the fringes I don't think that most people for most people it's not about uh existential risk or anything like that but this discussion is definitely undergone a phase change in the last several weeks the proximate cause being chat GPT which is it's kind of strange because the capabilities of chat GPT have been latent in the API essentially for SM Altman asset uh 10 months at least so this is well not surprising a delayed reaction but there definitely has been some updating of timelines I would say over the last couple of weeks Google stock price is down by nine percent today as a result of being from Microsoft launching with chat gpt-like capabilities Google has just invested 300 million dollars in anthropic an AGI Research Institute oriented around safety that was got a lot of investment or putative investment from the FTX Foundation when that collapsed maybe it looked like it was on Shaky Ground but now they've got plenty of money to train bigger models and there's the 10 billion dollar investment of Microsoft into open AI Microsoft is rolling out chat GPT like Tech across their entire office suite and thereby to millions and millions of businesses around the world um so that's like in the last couple of weeks um yeah it's been an astonishing time and it does qualitatively feel like a sea change in in terms of awareness and anticipation of new capabilities this this strikes me as a turning point a pivotal pivotal moment and we don't see very many of these the iPhone and the smartphone was an example it really did give many people a view of a new kind of future how the future was likely to differ substantially from what came before the iPhone was in a more limited domain perhaps but still impactful made very big waves very quickly and it wasn't just the idea that
you would have a very capable smartphone with a beautiful interface and and so forth it was also perhaps it's easy to forget but the the idea of the App Store was also a an integral component of the iPhones um evolution of smartphone technology and it was really the the um the app concept taking off that that had such a profound impact on how we use these tools we use our phones in related to um to the devices and interacted or interacted and began to uh shape Our lives our behaviors and our work around these devices so that was there were moments in 2007 when major tectonic shifts in conversation and planning and Investments were occurring as a result of what was now suddenly a very visible new terrain ahead and I'm getting I'm very much I'm trying to think of other uh analogous situations throughout history I'm sure there have been some the the flight by the first powered flight Wright brothers 1903 must have been similar it's perhaps slower and it's because of the you know speed of communications and expectations of things actually that's probably been historical with that example as well I mean this moment we're in with chat GPT is of course just a preview of the real uh Wright brothers moment when when an AGI is unveiled so is this I don't know if there's really an analogy where there was some event before the Wright brothers first flight that made it pretty obvious to people who were paying attention that this would happen I I don't remember the history well enough um I don't think so because the I I know I saw I saw the anecdote quick recently as a matter of fact I believe the Wright brothers themselves were skeptical that that powered flight would be achieved as soon as it was and so I think that it came as a surprise I don't think it was something that was latently expected um it so it I don't believe it had precursors in the same way that chat GPT and other narrow AI are are at least potentially precursors to much more powerful systems but I don't know if that is exactly the right metaphor or an or analog because the right flyer the actual device was very you know primitive and perhaps in some ways underwhelming but it did point to much more powerful Aviation capability emerging quite quickly which of course it did I mean within 15 years Aviation wasn't was an integral part of warfare and World War one uh would have proceeded very differently without Aviation so uh and yeah the the it could be the the moment of flight uh a powered flight with the Wright
brothers was you know a little bit like a maybe the iPhone like Steve Jobs presentation of the iPhone and then it you know took a while afterwards before the real capabilities got into people's hands and started emerging in this case chappy GPT it really is you know it's a preview it's been I think it surprised a lot of people how capable it actually is I think that's that's perhaps one of the one of the extraordinary features of this particular example of techno technological adoption is that the the uh I think there were women I I maybe I misreading this situation you you guys may know better than I do but my read of the situation is that uh there was a systemic industry-wide underestimation and under appreciation of how much value of the public and and general users users would obtain out of these chatbot Technologies it's it it I think that this has come as a surprise even to the developers of these Technologies how much value people are getting out of them and how much interest therefore there has been in them and how and of course how rapid the adoption has been as a result of all of that it I could be wrong about that it could be that there was you know Google and deepmind knew exactly and it didn't none of this came as any surprise at all and they cynically didn't release the technology in the way that open AI did when they could have considerably earlier because it would pose a threat to their own you know uh their own business models elsewhere across the Google you know the alphabet ecosystem search being the key you know big Target in the crosshairs there but uh I tend to not to be cynical and the easier explanation is that they just didn't notice didn't realize it and I've seen quite a few uh inside insiders suggest these sorts of things like going on T wasn't new we've been able to do this for you know a year or two now um and I think that I seem quite respectable quite quite imminent people say this people from Facebook and other major research is and it's a little and so it does beg the question well if you knew how powerful and how amazing this was how come you didn't do it and it it uh so it isn't that the capability well I think there was an underappreciation of what that capability the value that that capability represented represents to the broader public well I think probably the inside is just they're very aware of the risks and Google is a case in point of the innovators dilemma over the last few days right so as a result of in their demo of their upcoming
chatbot which nobody can interact with directly yet in the demo it got some fact wrong about the James Webb Space Telescope and journalists made a big deal out of this and this is being cited as a reason behind a nine percent drop in their stock price so for a behemoth like Google who Stakes its entire income on the relevance of its search results the risks from introducing these Technologies into that process are so extreme that it it really must be I mean whatever they did over the last few weeks to get this to a position of launching it I mean it really must have taken the CEO saying personally don't worry if it screws up and a journalist points out an error in our demo and we lose billions and billions of dollars in market value it's not your fault right that's what it must have taken to to get them to move on this and they would not have done it I mean they're waiting for the next generation of tech right they they know that the current models have many flaws and really aren't in a position it's not really ready to be integrated into search in this way Microsoft can do it because nobody uses Bing open AI can do it because they've got nothing to lose and it's kind of a marketing thing for them basically right and for them to attract investment for the other things they're doing Google's in a very different position and they've had their hand forced and it may have been a mistake because well I mean I wouldn't be using these the the strong point of these Bots currently or the technology is not in search it's not in trying to get factual answers to things I mean Microsoft claims to have this new tech that wraps around chat GPT but I don't know I'm not using it for this purpose the way I interact with chat GPT is for very different things right so they they somehow should have acted earlier and launched a product in a different segment than search but somehow they then got pressured moved quickly launched it in Search and actually I'm not optimistic about Google being able to make this work for them it'll be definitely interesting to see how the situation evolves we are we're and the Innovative The innovator's Dilemma is a good observation it's perfect of course and one of the things that goes along with that is we are likely now to see a the emergence of a new uh intensely competitive period where there's a rush of startups and well not startups but upstarts is a better way to say it um where there is very clearly now a prize in view to be claimed to be to be
seized mainly in the form of market share but but you know fortunes made and and power obtained adjacent those things as well and there will be intense competition amongst uh uh players who had settled into a you know a fairly um what is the right way to say it some sort of homeostasis some some stable condition at least for for a time that's not to say that that first and second and third and fourth place holders don't exchange those positions periodically they do um but you know the relationship like as measured by a marketer the relationship between Google and uh you know Bing as search engines and so on they were those weren't they weren't jockeying and intensely competing for for their uh where they were standing on the podium uh for first second third place so and so forth I see that now having shifted dramatically uh first place is up for the up for grabs again in a way that it hasn't been for a while and we upstarts are now also part of the picture and I think we are going to unfortunately enter or have entered Armor's race conditions as well where the prize on offer is so great and I'm not even talking about AGI like we were at the beginning of this conversation I'm just talking about the the enormous potential uh revenues that could be obtained with market dominance in these narrow narrower AI applications still talking about the absolutely immense amounts of of income for the companies that dominate uh it they have every incentive now to disregard any and all risk uh because there's there's so much to gain from being the winner and you know as ever this these arms race conditions uh you know pose a real challenge for society or civilization at large and they're very great cause for concern so the status quo that we had before with a fairly clear leader and and a fairly large gap between the runners-up and so forth that created the circumstances in which some caution and some acknowledgment of risk to take place and I honestly do suspect that Google in deepmind in particular were being cautious and were being mindful of the potential negative impacts and with some hard lessons knocking some sense into them in that respect along the way um I mean we earlier chat Bots were quite famously became embarrassments for you know having their racism in their and so forth their biases revealed and Mills did burn the companies I believe that was Microsoft it burned the companies that did you know uh tentatively put them expose them to the world and it's and all the scrutiny that goes with
that but uh to the extent that Google and deep mind could be ethical and could be uh risk averse and try to be responsible and safe that is now out the window I think the the the competitive landscape has changed such that that no longer seems like a viable option unless there is massive coordination between the major players and then and then on top of that the um the the incentives are not there for upstarts to participate in any coordination there unless they're so then we were you know then we're uh left to Hope for regulation to step in and perform and enforce that that function but when you've got you know the stable diffusions and the mid-journeys of the world and all of the other you know dozens of others now uh in the open source Community aiming very expressly to to explore the riskiest aspects of these new technologies gleefully in some cases uh yeah there's there is no avoiding an arms race now and all the risks that go along with that at least that's that's to my mind I'm I hope you guys can convince me otherwise that's where I've landed on this we're well into it now so on the topic that you let's pivot a little bit and talk about the the tension between centralization and decentralization which is what I one of the themes I read into the statement you made about software disruptions being uh much more rapid than other kinds of disruptions and they're being uh a open question about whether at what rate we can roll out the infrastructure for all these capabilities to be scaled up to the level of say Google search from where they are now uh and that scaling brings with it so one of the differences between the tech Giants how they're going to approach this will be to do with what happens on your device versus what happens in the data center and that will play to the different strengths of the the companies involved so Apple's name is kind of conspicuously absent from a lot of this discussion around Advanced AI capabilities but they're likely given their position in the market to to try and emphasize local compute and privacy and so on and that may work so it's Amazon's attempts with what's called Alexa to build a a bot that lives in your home and you talk to flopped that division sort of been axed or on the way to being axed probably just because it doesn't work very well not because people are concerned about privacy but as this technology develops I mean Apple's made very large bets on the Privacy angle and being trusted because they've been investing in health and areas where
people are much more sensitive than they are in other parts of their lives to the Privacy rights so they have a very strong foundation for moving forward with local compute for these kinds of Technologies so if you're integrating an AI tutor or an AI assistant into your life in a very thorough going way questions around where is the computation happening what's being sent where um they will again be topics of discussion I think in general maybe people don't care very much I care I think more people should care but I think this is likely to be something that will be a distinguishing factor and it's an open question whether you can do this locally and to what extent and what needs to be processed in a data center versus whether your phone is powerful enough to do some of the inference so Apple has been working very hard to make technology like stable diffusion run very quickly on iOS devices for example and I would imagine that there's a large team inside Apple working to make Transformer models large ones or rather sort of compressed versions quantized versions run on your devices and building dedicated Hardware to run them anticipating how important this Tech will be in the future whereas I guess open Ai and Microsoft and Google and so on will by the nature of their Investments be leaning towards a future where most of the compute happens in in the data center yeah I suppose it is an interesting technical question of how do you distribute the computing uh footprint whatever the requirement is and the where where controls where control is seated or granted of that resource um I I think I'm with you Dan I suspect that these new capabilities will be so profound and they will represent so much value they'll be so alluring and very quickly so indispensable that people will will just sacrifice their privacy uh both both knowingly and unknowingly and so there are I I would I would I think there are sort of you could draw you could say there are perhaps three or four sets that overlap to a greater or lesser degree of people of users in their relation to to privacy concerns they're sort of they're sort of or maybe it's maybe it's simpler than that but at any rate there are individuals and organizations I suppose that are both aware and very sensitive about private their privacy concerns and take action to protect the Privacy very aggressively then there are those who are aware that privacy is an issue but don't care and then there are those who are not aware that privacy privacy is an issue and
therefore act as if they don't care now whether or not they do or don't care we don't know for that group because they're not they they haven't made that choice they're simply unaware and and I don't know how the the total population of internet and software users is distributed across those categories and maybe I'm missing a category or two if I'm not being complete there but in my mind the the second two categories people who know and don't care and people who don't know and therefore also don't care those are the overwhelming majority of users it leads to my I don't have data but I I strongly suspect that that's the case and these are for services that are valuable and to some degree indispensable for work and personal life now compute just access to Computing in general and the internet in general you can't live without it basically you really can't function without it I don't think that's an exaggeration there are some people who still manage it but it's extremely difficult I think the new AI tools will be very much in that same category of rapidly becoming indispensable and I have no reason to expect that users will differ in their in their uh stance as far as privacy and and security goes towards these new technologies any more than they do today um and there are things that people do online now that that are you know they're illegal in some instances copying and downloading things that are that are you know they're genuinely illicit activity that people undertake and then there are you know socially uh what's the right word things around which behaviors around which there is social stigma I'm thinking pornography for the most part here uh on the internet where people uh consume that you know that kind of content and they give up their privacy whether they know it or not in in accessing the providers of that sort of content and people don't seem to hesitate that seems to be you know a large and absolutely you know thriving uh portion of the internet so again I just I see no reason to expect no reason at all to expect that these AI uh services and capabilities if they become as indispensable as I believe they will that they that people will not basically give up all of their privacy and personal data and much else to have access to them and we'll redraw boundaries around their lives both knowingly and unknowingly in order to have access to those capabilities I have every expectation that that's the case yeah and it's one thing at the level of individuals that's
a discussion we're used to having it's privacy and technology is yeah that's a debate that's been active for a long time and especially in the context of the internet but what I think is new is this also applies now at the level of um institutions corporations and countries so what we're discussing scaled up to those that level is actually a question of sovereignty so I think it's legitimate to consider this phenomenon that we're just describing where we seed our data but also effectively some part of our decision making and the uh sense making capability that we embed into our institutions we're going to seed that to the AIS because for convenience but also partly for competitive reasons and if those AI models are centralized in say the United States then a country like Australia will lose what little sovereignty it has remaining in that relationship what I'm thinking about is suppose a large number of corporations in Australia use the kind of technology that we've been discussing here right build chat Bots to look inside your institution and understand what's going on and interpret it for you raise awareness of risks that are latent in all the documentation you have and all the data you have interfacing between people within your organization all of that you know there's a flood of startups building that Microsoft is going to be heavily investing in that it'll suck to start with right and it'll have all sorts of problems but it will work and when it does work if that if the control over that system is outside of your borders do you really actually control your cooperation it it continues at the the whim of Microsoft or some other company and that's already true with operating systems right the maybe your company will just fall over if you're Microsoft subscription stops working uh but this is this is a kind of different thing right it's at the level of if open AI turns off access to its API to our company suddenly we don't know how to talk to each other anymore and uh the the lines of communication within our company we don't really understand what they are because they've sort of been embedded uh with the AIS playing crucial roles so I I do I do think it's appropriate it seems a bit silly now but soon it won't I think to to think of this as actually a question of sovereignty and I'm actually genuinely concerned that Australia doesn't have the supercompute capabilities to to have one of these models inside our borders I think this is a key sovereignty risk
that is being just completely ignored um actually it's an interesting question this is an aside a little bit but the I've seen some discussion I don't know what to make of it about open source alternatives to the development of these models uh I I don't so I'm mixing up the Transformer based language model technology here a little bit with the diffusion based um uh image generated generative uh capabilities those have both been spectacular uh developments in the last year these capabilities but they're they're they're not the same thing as is to my understanding and there's been I have seen more movement on the on the image generating front you know the Dali uh two Alternatives that are open source then on the on the um the language models but my understanding is that there's you know there's intense interest already in open source development of these large language models but then of course there are you know the training uh requirements super computer basically the compute requirements are so formidable and so costly that this is very difficult to do competitively with the Giants that have the massive you know cloud computing and server Farm sort of resources but then but then um one thing that I was reminded of recently and I completely forgotten about is the um distributed super Computing that's done by the the at home uh project or network I it used to be seti where you could help assess the this was going back 30 years now or more um to be able to process radio telescope signals to search for signs of extraterrestrial intelligence so that's seti the search for extraterrestrial intelligence that the that initiative broadened out into other Collective efforts to provide Computing resource so for protein folding it was folding at home that project uh and I believe there are those efforts still continue are you does anybody do you guys know of any efforts to have sort of collective open source alternatives for training you know specifically large language models using uh a cloud of volunteer user compute capacity as opposed to privately owned or publicly you know government-owned super computer capacity and what is the relevance of that is there any hope that that could provide a uh or serve some useful democratizing function in this arms race or is it just a a nothing Burger yeah it's an open question I don't have any idea at all I think unfortunately it's a nothing Burger yeah there's quite a lot of projects out there um a lot of them are cryptocurrency
related right where you would get some sort of token in in return for putting up your computational power um that worked for seti because there was kind of this Hardware overhang where people's computers were relatively powerful enough to be able to do some of that processing and there's a Fair number of gpus sitting in people's homes but they're just nothing compared to a100s or h100s so the the progress in the chips and the the the Gap in performance between the latest uh semiconductors that are baked into nvidia's products for example and what people have in their home machines it's just too big I it may it may be actually economic for like lagging Tech so like two generations old right you probably can't have a distributed system that would train gpt2 um and democratize that so I think there's a place in the ecosystem for distributed training to be coming along and making very cheap things that are a couple of generations Behind The Cutting Edge uh but yeah so that may make a difference I mean I guess stable diffusion is an example of you know that's certainly at The Cutting Edge uh but it's not only the data and the training so that's the place where I think this democratizing aspect falls apart a little is that these systems will become increasingly complex and difficult to produce it's not it's already you know you need some Insider knowledge to be able to retrain gpt3 or something but many groups have done that but chat GPT is is not just that right there's this whole layer of reinforcement learning from Human feedback on top which involves going and collecting human data interacting with the systems there's like a whole manual schlep involved in producing chat GPT which anthropic is reproduced right in order to train their Claude agent and that's part of the reason why they're getting this investment right there's like this extra stuff on top of the somewhat well understood process now of training gpt3 that you have to do to get it to the position of being a useful agent um and at some point I don't think that becomes that's the secret Source now right uh so at some point I'm not sure that is I'm not optimistic about the democratizing efforts I'm not sure that stable diffusion is really a strong indicator of how this is going to go um well maybe we can we can pivot there into another aspect of the discussion uh topic that we talked about for today and that is how soft where in particular as a technology is a a type of service that can be adopted very swiftly
in a way that that it is more difficult for other goods and services to be adopted so swiftly software is sort of it's in an extraordinary category of its own there you can you can download and apply and update a software update and begin adopting its new capabilities and realizing the new utility of it with so little effort so swiftly at such low cost that it the potential for for a new software capability a new software a new piece of software let's say to capture market share in a given you know in a given within a given Market to capture the share of the user base or or the time of people's uh participation or utilization um uh in a given domain the potential for that to happen so quickly is really uh it's it's sort of [Music] yeah it it doesn't obey necessarily the same rules doesn't certainly doesn't follow the same patterns that I think we see for other disruptions that occur where availing yourself of the new capabilities requires adopting new equipment going up a steep learning curve right reorienting or reorganizing yourself or your your uh your Enterprise whatever it might be your Institution uh it it this these these things that we've seen historically at any rate examples of historical examples of disruption have typically taken years even in the swiftest even in the very very very fastest cases we don't see physical products disrupting their respective markets or Industries in you know a matter of months or weeks or days but in software it's different in software there is the potential for adoption unmasked to occur astoundingly fast and of course chat gbt reaching 100 million users in two months is an example of that that is just so much faster than the normal pattern of disruption disruptions historically have taken I mean really at the low end seven or eight years and typically more like 15 13 to 15. seldom much longer than that if they take much much longer we stop calling them disruptions for one thing but we uh our our typical uh pattern that we see is sort of 10 to 15 years for most disruptions but um you know a hundred million people adopting a new technology in two months is just an astonishingly fast uh transformation and so I thought it would be useful to see you know what what uh other aspects of disruption of the patterns that we see historically from other precedents of technology technology adoption what what of those even hold for new software Technologies and in particular these these ones around AI where you know we're talking about
an order of magnitude for more rapidly tours of magnitude more rapidly than other disruptions this is this is just you know it's it's it's a different game it seems different bowl games yeah when he was speaking what I was thinking about was the the role of it's like a latent reservoir of jobs to be done or something that is relevant here so when you think about why was chat GPT adopted so rapidly well some people are doing genuinely interesting things with it but I have to think that the reason it became a social phenomenon is that lots of people have [ __ ] essays to write performance reviews to write [ __ ] emails to write that actually have no content and nobody really reads them but you need to write a nice email to the person and you know you get the tool to do it they don't even read it they just skim it and they're like okay yeah they get like a sentiment from the email right so they read the email it's four lines long they skim it and they get the sentiment thanks but if you just said thanks then it would seem rude okay so there's like this this latent jobs to be done of just actually there's no information there but it's a task that humans do essay writing is a lot like that most essays just are a sentence really but you're trained in high school to write these very you know content free multi-page things to say almost nothing um this is a thing that is a culture we like to do and there's this huge latent demand for just automating that process from both ends you know I saw this this meme where um I think it actually yeah I don't know who it was but on Twitter where somebody has an email to write what they put into chat GPT is say thanks what comes out of Chachi PT is this two-paragraph nice sounding professional email the recipient puts the email through chat GPT to summarize it and gets thanks and everybody's happy right [Laughter] I mean in some ways all across maybe I don't want to be hyperbolic here in most cultures it may be a genuine human Universal although those are rare uh manners are performative in exactly that way so there is a performative function and there's an expectation of that performance that we call having good manners and it is it it performs the function um of demonstrating and at least my understanding I've read a little bit about this in the past um in some evolutionary psychology texts in in my undergrads this is very old information but my I'm left with the impression after all of these years that manners form uh perform important social functions
you know Advanced cohesion they show respect they they um are a uh a demonstration of the willingness to be to cooperate as opposed to computer conflict and so forth and so they perform these important functions but they are time consuming you know they are performative you have to do this song and dance and they serve an important function um and we have we have grandfathered in the the the the baggage of that Legacy into modern corporate culture bureaucratic culture institutional culture and it it definitely does feel wasteful in some some sort of objective sense right um and that's separate that's a separate from the rest of bureaucratic uh what's the right word well again I'm I think it's fairly performative but bureaucratic fluff and and um [ __ ] basically so endless form filling and so forth is a different manifestation and it perhaps a properly different phenomena but certainly the performative aspect of good manners and the ritualistic greetings and and so forth these are these are uh deeply embedded rituals and they are costly there's no doubt about it yeah so regarding disruption I wonder if there's a general phenomenon there of I guess there must be which is that uh having that latent supply of things that I mean yeah the risky and tautology but a latent supply of things to be automated uh kind of like low level low-lying tasks and when the water level rises by one inch suddenly it spreads very far inland right because there's this kind of low-lying territory um that seems to be I mean that's that's actually a very kind of General mathematical principle that's that's a key to why phase transitions happen that that kind of configuration so I guess it's typical of disruptions in this case you know we will discover stage by stage uh the latent supply of cognitive tasks like that that are ripe for being automated as this proceeds and the chat GPT moment is arguably to do with a particular class of such tasks um but yeah software in general seems well the actual question I wanted to ask in connection with that was the role of disintermediation and disruption so um software disintermediated a lot of tasks right which used to be I mean you used to have to sit at a desk that had a particular piece of equipment or whatever right so a particular like the calculator the calculator was a thing you went and sat at a desk and then you used it the computer disintermediated a lot of that and took the the physical environment out by just providing all those Services through a single screen
and that that facilitated the second I mean there was maybe A disruption where you first put all those tools inside one box but then that played an important role in the the later disruptions of software where people moved to software as a service and lots of things started running in the cloud and so on right the the unification of all those tasks behind a single interface was then crucial for the next step and the next step so I wonder if there's a general how General that is in disruptions where disintermediation is a kind of enabling role I'm not sure there are I'm not sure there are any examples as powerful as computing performing that function the disintermediating is a good way of of framing it I think I'm going to have to think about that length here the way I've thought of Computing in the past is as a a general enabling technology and disintermediation and displacement of existing specialized Technologies as a part of that but as a general enabling technology Computing also makes fundamentally new capabilities possible that weren't in the past so it performs you know these multiple extraordinary functions and it is Computing really is quite in a class of its own in that respect I think that we have only a handful of other historical examples of technology that have so much reach and so much transformative potential because they they embed Within such a broad sweep of capabilities so language and writing is another example perhaps because it facilitates so much new value creation and so much new capability um energy probably less so energy is energy does facilitate it is an enabler but well maybe yeah I think that's a good example like maybe electricity combined with motors is a way to think about that because that facilitated mechanization in a general way that wasn't possible before and mechanisms had to be powered in a sort of bespoke fashion or anything else that had to be powered had to be done in a bespoke fashion before either with human or animal labor and then to a much much lesser extent some clever mechanical uh means of storing and then releasing energy winding up clocks or suspending weights uh uh or or pumping or carrying water up to a higher elevation for release later and so forth there were a few a handful of ways but electricity was transformative in that respect in combination with electric motors which and generators so so both to capture and then release that transformed electricity to and from kinetic energy basically and that
allowed the mechanism the mechanization of a huge sweep of of activities but yeah I think yeah I'm sorry I'm thinking out loud here I think my answer to your question is that it I I'm not aware of this being a general phenomenon I think it is something that is quite extraordinarily um if not unique that certainly a stands out far more prominently in the case of computing than any other technology I can think of and it may be that intelligence is another perhaps even more spectacular example of that um yeah I think that's a good point the gender maybe that's a defining factor in general purpose technologies that they they cause this to happen where lots of tasks and activity becomes sort of factored through the technology so uh once everything [ __ ] I mean there was the Industrial Revolution took a long time to adopt electricity right so but once things were sort of factored through electricity then you could improve everything by improving the electrical grid and by making just general the general purpose technology improve so everything kind of got to lift up on that rising sea um and but it sort of I think that's part of the reason why people haven't seen productivity growth from um AI yet right this has been a mystery for at least 10 years people have been like Bruno Olson and others have been talking about how there would be this huge disruption from AI but it never quite showed up in the productivity figures uh and that seems to probably be because it takes a long time actually for people to shift activity to lie behind the new general purpose technology so that it sort of benefits from automatic improvements in that technology but you can see that with chat GPT right GPT has had a few updates making it better in various directions and it just automatically makes anything that makes use of it better uh GPT is like that I mean these Bots here at meta uni if they drop the prices on GPT they'll just become 10 times cheaper right and they'll just get better and I won't have to do anything so as everything starts to move behind this general purpose technology uh I guess you you get these automatic benefits and that's um certainly true of software right speed of computing just automatically made all your office software better yeah the um in the case of in the case of AI and not General AI of that that is its own category but even broadly capable broadly more capable but nevertheless still narrow in artificial intelligence is I think I think that the single most profound feature
of of that technology and it is is is precisely its potential to be embedded in or as you said Dan for for other things to be factored through virtually everything that virtually everything um just just so many different aspects of Our Lives of the work that we do uh of the activities that we engage in and almost by definition any action that humans undertake will have some intelligence behind it somewhere and uh I guess there are I guess there are categories of unthinking action where there's where a decision is made long ago and it's you know the the system is is continuing on on that basis without any new intelligence in it and you could be cynical and say well you know there are systems that are that have spun out of our control and are really dumb and and counterproductive now and so forth you can certainly argue that but I'm thinking on a smaller scale um one could imagine embedding intelligence just ubiquitously around human beings and all of the all of the things that we engage with every action and all virtually every object and and and so the scope for for artificial intelligence to create value and for Value to be realized through the utilization of this new tool is is I mean it's it's as close to unbounded I think as anything we've ever seen um it is is I mean again I struggle to wrap my head around how profound the potential the transformative potential is here and this is again without going to fully General artificial intelligence and then certainly without going to the Sapient uh version of fully General artificial intelligence just within the narrow sense so yeah I I and several things continue to surprise me here the first is that this is just not widely appreciated I don't know why it it I mean it just it boggles my mind frankly that uh not only the public don't see the potential of this and dismiss it as you know science fiction in the same way that you know people dismiss the the potential of the internet people just I mean the public did not see it for a very long time for a very long time the public the media the reporting the depiction and film and entertainment the internet before it became a real thing outside of a pretty narrow sphere of futurist and science fiction you know thinking um that's very much a subculture it's a small minority of people the internet was just basically going to be a glorified phone book a little bit interactive you know that was it and that was that was what we people imagined the internet would be that was
the extent of it um and and this is just mind-numbingly short-sighted okay um well fast forward here I think you know that you know fooled me once we shouldn't be expected that the same thing would happen again with the general public okay and in general the public is perhaps not grasping how profound these impacts are or how swiftly they will happen but what surprises me even more what I'm truly dumbfounded by is that as we talked about earlier in this conversation even the industry itself didn't realize what they were sitting on not necessarily that that some people knowledgeable experts people involved in the development of the Technologies themselves didn't fully appreciate how profoundly useful they would be and therefore what you know an extreme interest that would be manifest and how quickly they would therefore be adopted and how disruptive they would therefore be uh you know the in in we will intimately involved in the development of these Technologies appear to not have fully grasped its potential this shocks me I just I I don't get that and uh so we're still in this situation where you know we are literally only a few downloads away from life-changing transformations in the capability uh of individual people you know to get certain things done in your life that you need to get done it's they're really only a download away um you know to automate 75 of your job a download away I mean it's this is this is you know just I'm flailing around here uh failing to articulate this ineffable uh impression I have of how useful these technologies will be and therefore how impactful you know their deployment and adoption is going to be and again as I said this is all well before you know any full-on science fiction Singularity sort of stuff with Sapient artificial general intelligence I'm just talking about you know chat GPT and the Next Generation or two of that technology it is going to be absolutely transformative We've Only Just Begun to see the big the first inklings of it yeah it's interesting that I I see on YouTube that it seems like the main Vector for spreading public awareness of the capability of these systems is get rich quick schemes so uh the the way in which you know I was just thinking through why why don't the people who have these beliefs spend more time and effort trying to uh explain why they're reasonable and why why it's not just mere speculation or boosterism um and it there is a kind of lack of something in the middle right where uh what's the what's the expression
um something something lack all conviction the um is that Tennyson uh the the good lack all conviction so when you when you have an environment like this with a rapid technology shift coming well there's lots of money to be made right I saw someone point out that for the next few years uh professionals say lawyers I mean if you're a capable lawyer you can probably scale your caseload by 10x um I mean I don't I don't know if that's accurate but that's uh maybe it's not true for lawyers in particular but it's true for some professionals I'm sure where you have you could hire a lot of people hire a lot of AIS maybe we're not quite there yet but maybe gpd4 will be enough I don't know at some point you'll be able to do that but it won't last very long right so there'll somehow be an interregnum where you can make a lot of money by being the first to adopt and scale these tools you know you could say that you're a Kind of Traitor to the human race or something but uh okay so so what's the incentive for convincing everybody that that's true I mean if you believe it's true just go make your money um so I think a lot of that is going to happen where you know you can see all the people building startups on the back of chatgpt uh well what's the point in trying to convince the public that the even more profound things are coming they'll just try and sell what they're up to and indirectly that's going to actually raise awareness right so the the people pay attention to what makes money the potential of the internet reached everybody's awareness when they were thinking about well I should get my kid to study computer science because then they'll make a lot of money and so that's how this will happen but uh yeah is is there I guess the question is is there really any point in trying to convince people or make the case for this uh it seems to me that it's important to do it but um the uh yeah I'm just not sure what the uh if this ever is the way that people really become aware of disruptions that someone tells them about it and makes the case and convinces them to pay attention foreign yeah there are there are push and pull incentives in place certainly right for the the um the sort of contagion that the sort of social contagion that at least in part drives adoption of anything new uh one of the things that that Owen uh our our newest member of my research team is is working on is modeling disruptions and um the epidemiological uh model it's conceptual more than than formal
but the epidemiological principle for the adoption of any new technology or any new tool or any new idea is is a model of contagion that individual carriers of that uh new knowledge or new tool infect others and the the place where an epidemiological model breaks down is in the details about how and why infection it might occur and what you pointed to Dan of course are reasons why individuals would choose not to be a vector for uh that could contagion for why individuals would choose to take actions such that they don't infect others with this new idea they don't give people the the knowledge or the um awareness that there's a new tool to be utilized and that they should adopt it and that's interesting we haven't don't know that that's something we would necessarily model as part of our work on other disruptions but I can certainly see it being relevant in this particular uh case and it may be that it's relevant up to a point and then and then it stops being effective if you know past a certain Tipping Point or threshold General awareness is simply too high to worry about about that but certainly in this like you said in in this inter-regnum period their is there will be opportunities to extract what economic rents of some sort I suppose from the uh from from the laggards the lagging markets the lagging Industries and so on that that haven't yet become aware of the technology and it's scope and its implications it's yeah it's it's uh it's interesting uh one question that I have for us looking forward is how much faster can adoption get than 100 million people in two months what is the limit I mean what would and maybe we should I mean we could indulge in in you know a thought experiments that that take us to the Limit and then back off from there like I mean you one could easily imagine um you know properly magical technology that that virtually every single last human being would adopt instantly if they could you know if if aliens arrived and said you know swallow this pill and it will you know X Y or Z and whatever it is make you immortal and you know transform you into the perfect physical specimen by your own standards or um uh uh whatever you could one could imagine you know magic wand style Technologies or that that would be adopted with astonishing swiftness basically more or less immediately by everyone um who was conscious I was awake it wasn't it wasn't a sleeper or a coma or something like that but if we back off from that extreme you know what is the
what's the Practical limit in the near term is this just we just done a step change and then this is about as fast as it will get and we're not likely to see this again until you know we're perhaps closer to True Singularity kind of territory or you know is this are we going to have a speed up where adoption is it will be 500 million users in one week a year or two from now for the latest version of um you know GPT 5 or whatever it is do we have does anybody do we have any thoughts on this very fast adoption but like and how much of it is people just like hearing reading a new story about it and wanting to try it out themselves right like you can just make an account with like Google um I wonder it's like like what proportion of that represents the kind of real adoption in the sense that it's actually being used for some purpose yeah it's a good point I don't know if that figure is just accounts or like daily active users which might be a more indicative figure hmm yeah I I guess I assumed it was accountable I don't know maybe it's and I should I should add to what Owen is saying here I don't know if you guys have encountered this or how often but I've certainly seen in various online chatter online forums and Twitter and social media and all that sort of stuff people saying oh yeah I tried GPT and it was useless yeah I tried it for 10 minutes and it's and it's dumb don't waste your time I mean I've seen a fair bit of that it's not like that is a you know an ultra rare experience I don't know what percentage I would you know wager a guess at it being but um I've seen enough instances of it to imagine that it could be you know it could be something significant it could be one or five or ten percent of the people who've used who tried GPT have been unimpressed with it enough that they disregarded yeah I I just recently showed it to my mother-in-law um she didn't seem particularly impressed um and I wonder like I was sort of thinking about it well I think computers seemed like a bit of a magical technology to her right right it's kind of um like they are already amazing and you you already sort of don't know how it works or like you don't I think you need to have understand a little bit of to see why it's impressive maybe while I come I mean one thing is like you know you look at sort of how the how the models have evolved and how much better they've got that's kind of one thing but but if you know people are not aware of you know chatbots or like it's just kind of you know
what's next yeah yeah I think it's it's the adoption's probably very much tied to particular use cases so out of that 100 million you know 60 million are students using it to write essays and uh 30 million uh people writing emails or performance reviews and they learned about it from some article they saw or some Tick Tock or some video on YouTube and there's like enough of these particular use cases that they've made up some user account but I doubt I think very few of those people will have just loaded up chat.openai.com and been like interacting with it and figured out something useful I think that's a vanishingly small set of people will have just figured out for themselves a way they might make use of it I think it's like a they might like go from one task and try other things and find out that it works but I think the adoption is very much driven by much more narrow tasks and in that sense that's why people are pointing out that well you can already do this with uh you know the open AI playground you didn't need chat GPT but somehow it was um necessary for this step to happen we're gonna have to wrap up in a few minutes but I think this question of uh how quick can adoption get would be a good topic to pick up next week what do you say yeah that sounds that sounds good um maybe some things to think about mull over between now and when we pick it back up are um Owen's comment made me think about who are likely to be adopters because it isn't again deviating from the epidemiological model of an infection well maybe not do maybe actually not maybe not deviating from that but the not all individuals in the population are equally likely to adopt a new technology and so from that we get some sort of distribution of you know uh across the population of when they're likely to adopt younger people probably more much more likely to adopt than older people but you know within limits you're not going to have eight-year-olds adopting this technology you're probably not going to have 98 year olds adopting this technology and then some uh you know some some sort of distribution in the middle and again it depends on the sort of the value where this is created if it's a lot of students using it to write essays uh then you're going to have a lot of school age young adults uh who are being who are the early adopters adopting very rapidly and that population you know is computer savvy anyway so you have has a lot of things going for it one could imagine if certain use Kate so one thing we