Full transcript: AI Wars: Google’s Bard Takes On OpenAI’s ChatGPT on London Real
0:00 Brian Rose: This is London Real. I am Brian Rose. My guest today is Dr Ben Goertzel, the scientist, futurist, author and entrepreneur. You are considered one of the world’s foremost experts in artificial general intelligence, and have submitted over 150 technical papers and published 25 scientific books on the subject. In 2017 you founded SingularityNET with the goal of creating a decentralised, democratic, inclusive and beneficial artificial general intelligence.
Your latest project, Zarqa, is a supercharged, intelligent, neurosymbolic large language model at massive scale, and promises to take on OpenAI’s ChatGPT but go much, much further. You believe that your advanced neurosymbolic AI techniques will revolutionise and disrupt every industry, taking a giant step towards artificial general intelligence, or AGI. Ultimately, your goal is to realise a positive and beneficial Singularity for the good of all humanity, by building the sentient beings of the future to be loving, caring, compassionate, inclusive and democratic. Ben, welcome back to London Real.
Dr Ben Goertzel: Great to be back again. There are always more and more strange realities unfolding to dig into.
Brian Rose: Yeah, there’s so much happening in the space, it’s great to be checking in with you on a regular basis. Before we get started, I just want to say that if anybody’s interested in learning more about your new project Zarqa, they can go to londonreal.tv/zarqa to request more information. We will be investing in Zarqa through London Real Ventures, and everybody here on the team is super excited to support you in this amazing venture.
4:24 Let’s jump into it, Ben. There have been a lot of crazy headlines. There’s a bunch of big ones, but the one that got a lot of people’s attention was the re-release of Google Bard as kind of a reply to ChatGPT, which had moved way ahead of Google, and we thought maybe it was leaving Google in the dust. Google came out with Bard, connected it to the live internet, allowed it to plug right into your Google Docs and email and so on, and a lot of people are saying that Google is back, competing head to head with OpenAI. I was just curious if you had any thoughts on this. I’m also curious how you feel about a large language model like this being plugged into real-time internet data, as opposed to ChatGPT, which was only archived back to 2021.
Dr Ben Goertzel: I think these are all pretty obvious developments. It was clear that ChatGPT not being trained on data up to the minute was just a sort of random artefact of when they chose to release it. There’s no fundamental obstacle to training transformer neural nets on up-to-the-minute data, and there’s no fundamental obstacle to making a transformer that you can update in real time as new data comes in. OpenAI didn’t do it that way at first just because they were basically creating a research project that came out cooler than they thought, so they decided to release it early.
6:09 I think Google catching up to OpenAI in LLMs is almost inevitable. Google invented transformer neural nets within Google Brain, right? And Google DeepMind and Google Brain, either one of those independently is a far stronger AI team than OpenAI in a lot of ways. OpenAI made a different call in terms of narrowly focusing a lot of resources only on training big transformer models, and then putting them out as products even when they were really early-stage prototypes with a lot of weird aspects. Google didn’t make either of those calls.
Even though they invented transformers, they didn’t double and triple down on them like OpenAI did. They had very cool transformers in LaMDA, and they chose not to release them publicly because they were too iffy in terms of giving screwball responses. So Google made a bunch of calls that were questionable on the business side in hindsight, but clearly they’re super strong with AI in general, and on transformers, which they invented, in particular. So it’s pretty obvious that OpenAI’s lead is temporary right now.
That doesn’t mean OpenAI’s lead is meaningless. Microsoft also has a really strong AI R&D division in Microsoft Research, in many ways stronger than OpenAI, other than in this specific domain of transformers. So OpenAI and Microsoft are certainly viable competitors to Google, but Google being in the running is hardly surprising.
8:27 We also saw a memo leaked out of Google not that long ago, stating basically: well, we have no moat, and OpenAI also has no moat. Basically, there’s nothing fundamental stopping open source from eating everybody’s lunch in large language models. I would say it hasn’t quite happened yet, but you can clearly see that avenue unfolding.
Right now Facebook’s LLaMA model was leaked onto the internet with a little help from 4chan. Facebook put the model out there but they didn’t release the weights, which are a key part of what makes the model do stuff. Then that was somehow leaked onto 4chan, then everyone had it, and then Facebook was like, okay, fine, everyone can have it for non-commercial purposes. The amount of work that’s been done on the open source LLaMA model just in the last weeks has been tremendous. Right now GPT-4 seems smarter than LLaMA; LLaMA seems to be around the GPT-3.5 level. But it’s clear that’s not a gap that fundamentally can’t be overcome either.
So the Google engineer who wrote the “there is no moat” memo was just looking a little ahead into the future, and he’s seeing that what the open source world will do with LLMs is going to be cooler and cleverer than what big tech companies do with them, just because there are more people out there in the open source community than in any one company. So the lead of having a model with more parameters at a big tech company is not tremendously sustainable, given that VCs are out there funding companies building open source LLMs. Stability AI is there, for example, and EleutherAI is there.
And Zarqa, which you mentioned, which is a SingularityNET ecosystem project aimed at training neural-symbolic LLMs running in a decentralised fabric, is going to do open source large language models also. So the investment community is funding smaller tech companies with open-source-friendly business models. There’s no fundamental reason OpenAI is going to maintain a lead there, nor that Google is going to maintain a lead there, really. What we’re seeing is pretty much the inevitable catch-up of everyone else to OpenAI.
11:37 The dialogues within OpenAI and Microsoft have got to be quite different from what many people would naively think. If people are thinking, okay, OpenAI owns the Singularity, you can be sure that’s not what they’re thinking internally. What they’re thinking internally is: everybody basically knows how to do this, and everybody else is putting so much money and attention into this now, so how do we maintain a lead, given that we shot our biggest load already?
Because in taking the next step beyond where GPT-4 is now, it’s not clear that OpenAI and Microsoft have an especially big advantage. Sam Altman has already been quite public saying that training bigger and bigger models with more and more parameters may be interesting, but it’s not the magic sauce for incredible future progress. But if the magic sauce for incredible future progress is coming up with a fundamentally altered or improved transformer architecture, or if it’s hybridising a transformer with a knowledge graph and reasoning on that knowledge graph, then it’s not clear that in taking those further steps OpenAI has such a huge advantage.
If it’s coming up with a better variation of the transformer architecture, well, why not Google, who invented transformers in the first place? And if it’s hybridising transformers with logical reasoning engines and other forms of AI, why not, for example, SingularityNET, Zarqa or TrueAGI? We have been hybridising neural nets with logic and evolutionary systems for decades. OpenAI’s lead in the next leg of the AGI race is not especially obvious when you dig in deeply, though they’ve certainly gotten interesting mind share and market share.
13:54 You could try to draw an analogy with Google’s lead in search, but I don’t think it’s a perfect analogy. Google had a sort of positive feedback loop with search and ads. They got people locked in to search, then they got people clicking on ads, and because there were people clicking on their ads, companies used AdSense. So there was this advertising revenue model that had a flywheel effect, and their lead in search boosted this flywheel of people clicking on ads and people placing ads. There was an increasing-returns revenue model associated with ads that tied in with their early lead in search.
We don’t see that OpenAI has that now. They don’t have some amazing new revenue model innovation that they’re boosting with their early lead in the conversational LLM market. Their way of making money is API calls and subscriptions, but anyone can do that. That doesn’t have the flywheel, increasing-returns dynamics that give you more defensibility. In fact, it’s not really clear how Google and Microsoft will make money from these things, which is probably part of the reason Google didn’t rush to launch their own generative AI chatbots.
Apart from them feeling uneasy about the amount of unrealistic, non-factual bullshit they generate, replacing a list of search results with a natural language summary doesn’t have an obvious advertising business model associated with it. Google can try to intersperse paid stuff with the Bard results, but that’s uglier than putting ads to the side of your list of links. But OpenAI also doesn’t have such a defensible business model, because subscriptions are not that defensible against anyone else offering the same thing at a slightly lower subscription price.
So ChatGPT was definitely a black swan event from the economic point of view, and it’s a threshold in terms of practically deployed technology. But it’s really opening up a new era, and things are pretty wide open in this new era. It’s not at all clear if OpenAI will be the Google or the Yahoo of this new era.
17:13 Brian Rose: Yeah, very well said. It’s interesting: when Google plugs in, you realise Google has this massive audience, from search to YouTube to Gmail to the documents. And then you look at OpenAI and you realise it had zero audience before ChatGPT was released. And like you said, how quickly that could be done.
Dr Ben Goertzel: In enterprise, Microsoft has a huge audience. So many businesses use Azure and the Microsoft Office suite, and they’re going to integrate OpenAI’s tools there, which may cement their ownership of the enterprise. Google has Google Docs and Google enterprise tools, but they’re not actually meeting the needs of enterprise as well as the Microsoft Office suite integrated with Azure. It’s not clear generative AI overcomes the factors that have caused Google not to make great enterprise products. It could help, because generative AI can be an interface to enterprise software, but it doesn’t fully solve the problem.
So I think Microsoft will do just fine. They can roll out generative AI to transform all of their enterprise and office applications, and that can be very, very big for them. Whereas for retail end users, Google has a lot of users and a lot of data to play with. I haven’t actually switched from using Google and Brave to Bing, and probably you haven’t entirely either. I did use Bing a few times when the Bing chatbot was out there, especially the completely insane version, which was fun to talk to. I’m sure it’s growing Bing’s market share a little bit, but Google has responded fast enough that the feared draining of Google’s audience to Bing has not happened.
19:32 Brian Rose: Yeah, it hasn’t happened. And funny you said that about OpenAI, because I heard Sam Altman say on stage specifically that they are not training or building GPT-5 right now. He said, “We’re not training that right now.” I thought that was interesting.
Dr Ben Goertzel: Well, I take that with many grains of salt. Maybe they’re training GPT-4.5. They went from three to 3.5. Maybe they’re training a GPT-4 Turbo.
Brian Rose: I just took that to mean he doesn’t want to scare people.
Dr Ben Goertzel: Could be. I will guarantee you they’re proceeding full speed ahead on whatever R&D avenue they think is most productive at this point in time, and assigning names and dot-version labels to these things is a marketing action, not a technology action.
20:28 Brian Rose: Okay. Demis Hassabis, founder of DeepMind, which was purchased by Google many years ago for around 500 million dollars, I think. They merged it with Google Brain, and now he’s in charge of the whole Google AI initiative.
Dr Ben Goertzel: Yeah, and Jeff Dean, who was running Google Brain, moved to some loftier executive position, I guess. Jeff Dean has bounced around to a few things. To me that’s quite significant in terms of the global approach towards AGI. Not that Jeff Dean was anti-AGI, but he’s not an AGI guy. He did a bunch of really cool, edgy things. He led DeepMath, which was Google’s foray into automated theorem proving, as well as transformer neural nets. Google Brain did amazing stuff. But while Jeff Dean is amazing, he’s not from the AGI world so much.
Whereas Demis started DeepMind to build artificial general intelligence. And Shane Legg, who was one of his two co-founders, worked for me from something like 1998 to 2001. I taught him a lot about AGI. He also did a PhD thesis on the mathematics of AGI, and within DeepMind Shane has been leading a team focused just on artificial general intelligence, while Demis has been ruling over all the different projects that DeepMind does. So what you now have is hardcore AGI guys running Google’s AI team.
This has a lot of complex aspects. On the one hand, it means resources for AGI will not be hard to come by within Google. On the other hand, it may mean that Demis gets distracted from AGI, because his job is now running all the AI for all of Google, which has to focus on making Google money. I don’t think Shane will be distracted, to the extent his mandate is to run a focused AGI team. He may have had all the resources he wanted for that AGI team already, but it’ll certainly be easier for him to pull Google Brain people into his AGI development module now.
So I think it’s a really powerful mainstreaming of AGI. I don’t think it’s just because Demis is a cool guy, although he’s a very capable guy. I think that’s the Google founders deciding: we need to move faster, we need to move more ambitiously, or we’re going to lose the AGI race. So they’re deciding to double down on winning the AGI race by putting their AGI guys in charge. That’s quite an interesting move. It would be analogous to Microsoft putting all of their AI research teams out of Microsoft Research and under the management of OpenAI, which Microsoft is not doing at this point.
24:17 Brian Rose: Okay. Based on all of this, how does it make you feel about Zarqa? Because two weeks ago, when you mentioned that moat article about Google, you just kept saying all fingers are pointing to open source; decentralisation is the fastest way to improve these technologies. That’s been the bedrock of everything you’ve been doing at SingularityNET, which of course then funnels right into everything you’re going to do with Zarqa. So how are you feeling about your new venture, and how is it going to be able to compete with these massive trillion-dollar tech companies?
Dr Ben Goertzel: I think clearly the AGI race is on, and that’s exciting. Some of us have been waiting for it for a long time. I remember in the period from, say, 2005 through 2011 or 2012, I was working on this two-volume book called Engineering General Intelligence, which had a bunch on deep neural nets and language processing and logical reasoning, mental simulation modelling, long short-term memory, attention. And one thing I was thinking while I was working on these books, together with a couple of co-authors, was: it’s good now you’re taking time to think very deeply through all these different aspects of what it takes to make a thinking machine.
Because at some point things are going to start unfolding like bang, bang, bang. There’s going to be a lot of money, there’s going to be a lot of competition, there’s going to be a lot of pressure, and you’re not going to have six months to think about the best way to interleave uncertainty with deductive logic. You’re not going to have six months to think about how simulation modelling really ties in with hierarchical perception and with cognitive control of perception.
So as I was carefully thinking through all these things over a period of years, playing with prototype code and writing books and papers, it was in my mind that we’re stockpiling knowledge and understanding about how to do things. We can’t do all of them yet, because computers are too slow and we don’t have the money to hire people to write all the code. We can write prototype code, we can write maths, we can think it through. And then when push comes to shove, we’re going to have all this thinking done in advance. We’re going to be able to take all this stuff out of the research closet and deploy it using the tools that are available at that time.
And now I feel like that time has come. The fact that we’ve already thought through how to solve so many of the problems that you confront on the path to making an AGI comes to serve as an advantage for us. Because indeed, in my life right now, I don’t have time to sit in a room and think for six months, walking out in the woods every day, musing undistracted about how you interleave higher-order uncertainty with probabilistic deduction. I’m leading too many crazy projects, trying to deploy impactful stuff too fast. I try to carve out hours or a few days here and there for deep thinking, but I’m not able to carve out months for deep thinking right now. But we’ve got all this stuff that we can bring to bear.
28:01 I think that is one of the two advantages that we have over Big Tech in the AGI race. Compared to the teams working on AGI within big tech companies, not just me but the rest of our team in the SingularityNET and OpenCog ecosystem, some of whom are within Zarqa, some within TrueAGI, some in the Singularity Foundation, some within the OpenCog open source community: I think this group of AGI researchers and developers has actually been thinking about, prototyping and working on AGI-related tech longer than anyone in OpenAI, and longer and with more breadth and depth than most people in Google, for that matter.
We don’t have as many engineers as they do, but we’ve got a lot of length and depth of thinking in our team. We’ve currently got dozens of AGI researchers on our side, versus much larger teams within these big tech companies. But there’s a lot of history in the software world of a modest-size team of the right people, with the purpose and the right mindset and knowledge, upending big companies over and over again.
Now that LLMs have shown what they can do, there’s a lot of thinking about: okay, how do we connect LLMs with knowledge graphs to solve their issues with factuality? How do we connect them with, say, Wolfram Alpha or other more structured reasoning systems, so they can do multi-step reasoning better? But we’ve been working on connecting neural nets with knowledge graphs and logic systems for decades. We haven’t had the compute firepower to do what GPT-4 can do now, but in terms of the next step being connecting LLMs with knowledge graphs and structured reasoning, we have far more depth of knowledge and far more detailed plans of how to do that than Google or Microsoft does. So this is one advantage I think we have: just better knowledge and more in-depth, worked-out plans about what the next step is beyond these large language models that are having so much success.
30:46 The other advantage that we have is shared by the rest of the open source world, and this just goes back to the “there is no moat” memo. I think in the end open will defeat closed, because the breadth of knowledge and the breadth of skill, and just the human power in the decentralised ecosystem of open source developers, is just way, way more. I think that’s a substantial advantage, which we can leverage. EleutherAI can leverage it, Stability AI can leverage it, but Google and Microsoft can only partially leverage it, just due to the nature of their own tech stacks, let alone their business models.
When you get a job working for Google, you spend six months learning their two to three billion lines of proprietary code. They’ve just got really efficient tech stacks of proprietary code; they’re great at what they do. But since they don’t want to open all that, because it contains so much commercially valuable stuff to them, they can’t really operate in close coordination with the open source world. Whereas we can build things in SingularityNET and Zarqa and TrueAGI, in our ecosystem, from the get-go with maximum back-and-forth with the open source world in mind.
That doesn’t mean every single bit of code built within Zarqa or TrueAGI has to be open source, or every model trained for whatever purpose has to be open. SingularityNET will be that way because of the sort of entity it is; these commercial spin-offs don’t have to do only open source things. But they can build their architectures in a way that’s designed to get the maximum out of the open source world, which is not what these big companies have done.
33:03 This gets into technical stuff, but if you look at how these big companies are leveraging their massive GPU server farms, they’re using an algorithm called MapReduce, and everything is MapReduced. Now, MapReduce is great for deep neural nets, it’s great for computer graphics. It’s not great for reasoning on knowledge graphs. For reasoning on knowledge graphs, you need smarter ways to map your algorithms onto GPUs and parallel processors.
Now, we have those smarter ways. We’re borrowing them from a project called RChain. We’re working with a guy named Greg Meredith, who founded RChain. He developed a formalism called the rho calculus, which goes far beyond MapReduce in its ability to map complex graph algorithms onto GPUs. He had that out in the open source world inside his Rholang language. So we’re able to connect our OpenCog Hyperon system to his open source system, and then we can use his post-MapReduce method for parallelising stuff within OpenCog to provide the symbolic part of a neural-symbolic LLM, which we can then roll out within Zarqa and TrueAGI and SingularityNET.
So there are two interesting things here. One is, we have a path to neural-symbolic large language models by connecting OpenCog, which already does symbolic reasoning, to neural large language models that will be deployed on Zarqa’s server farms. Big tech companies don’t have that path, because they don’t have the knowledge of OpenCog’s reasoning engine for sophisticated reasoning on knowledge graphs. But the other thing is, the way we’re going to make this scalable is by connecting to Greg Meredith’s open source rho calculus code from the RChain project. And that’s an avenue that Google, Amazon and Microsoft don’t have, even though Greg worked for Microsoft for many years, just because their tech stacks are big, hairy and proprietary. They’re not built for interleaving with other people’s open source code.
So there are a lot of advantages Big Tech has: they’ve got customers, they’ve got money, they’ve got hardware, they’ve got big teams. On the other hand, we have better algorithms for the next phase, and we also have the ability to interoperate very flexibly with the open source community, which is a very, very powerful force. So I think we’ve also got our own big advantages in the race to AGI, which should make this next phase very, very interesting.
36:04 Getting back to some of the broader political, geopolitical and social aspects we’ve discussed in some of our previous conversations: if the next big leap beyond ChatGPT is launched in the decentralised world, then suddenly you’ve got a substantial portion of the world economy running on blockchain. One thing that Toufi Saliba, who’s running the HyperCycle project, which did a successful token sale just a couple of weeks ago, is doing: this is a ledgerless layer one blockchain, or as Toufi calls it, a layer zero++ blockchain, because at one level it’s like a secure communication protocol, almost like a TCP. On the other hand, you could use it as a side chain on top of other blockchains. But let’s call it layer one for simplicity.
One thing he’s doing there is reaching out to people running Bitcoin mining farms around the world, and Ethereum mining farms, and saying: hey, you know what, you’ll make more money using some of your infrastructure to do AI instead of crypto mining. That’s especially true post the Ethereum Merge, because Ethereum mining is not really a thing anymore; it’s proof of stake. Bitcoin mining is still a thing, but these folks with large server farms that were used for crypto mining, if they upgrade the GPUs in their machines, they can run cutting-edge AI algorithms, then they can run HyperCycle nodes, and then they can make money from AI instead.
So you have a concrete path there to the next generation of AI running on a decentralised hardware network, including server farms in Paraguay, Kazakhstan, South Korea, the Middle East, all over the globe. What’s interesting to me there is the idea of having a global compute infrastructure. SingularityNET has always allowed that, but this business angle of reaching out to crypto mining farm operators to run AI clusters within our decentralised AI network makes that very concrete in a very specific way, because these crypto mining farms really are all over the world, and in all sorts of weird jurisdictions.
38:52 So then when you go to something like Zarqa, the SingularityNET spin-off oriented specifically to training large language models for retail use, you’re basically making a neural-symbolic, decentralised-infrastructure ChatGPT killer. We are going to construct our own server farm for Zarqa, because our tech lead there, Sergey Shalyapin, has a lot of experience building GPU supercomputers specifically for LLMs. But we can also make use of compute power on HyperCycle’s decentralised compute network, spanning people’s home computers and all these crypto mining farms.
So we can use a mix of our own server farm and this decentralised network of machines running on HyperCycle. It’s more flexibility in getting the hardware behind the AI than Big Tech has, which can mean lower cost of operation, which can then mean lower cost of use for the consumer. That would mean Zarqa’s ChatGPT-killer language model software can not only be smarter, because we’re integrating OpenCog’s reasoning, it can be cheaper, because HyperCycle is letting it source some of its compute power from this decentralised network.
40:31 Paraguay has this dam, something like 14 gigawatts of electricity shared between Paraguay and Brazil, and Paraguay is selling a bunch of these gigawatts to Brazil at a low price because they don’t need all that electricity. But we’re talking to Paraguay, who are building huge data centres next to that dam, next to the Brazil border, to use these gigawatts of electricity. Informally, and not yet fully committed, they seem to be down with running a whole bunch of HyperCycle nodes there, in these huge server farms next to their dam. So if we run a bunch of Zarqa’s LLM there through HyperCycle’s decentralised compute network, then you’ve just got cheaper electricity than Google or Microsoft.
41:25 Brian Rose: Can I ask you a question about revenue? Because you did mention that ChatGPT might struggle to make money. I just saw a tweet by Mark Yusko, who runs Morgan Creek Capital, one of the big blockchain investors, and he just said: I haven’t seen the revenue models yet for AI or even AGI. I’m curious, where do you see the money coming in, and how do you think that will play out with Zarqa, for example? Because ultimately these things do have to make money.
Dr Ben Goertzel: I think the money is coming in from everywhere, in the end. That’s a little bit like asking where the money comes in from integrated circuits or something. I think LLMs are a fundamental technology rather than an end-user product, and I think AGI is also a fundamental technology rather than an end-user product. So I think OpenAI is doing the right thing by rolling out the plug-in architecture they did. Plug-ins that use GPT-4 will have their own user interfaces, and they will be serving users in particular niches.
I think a retail chat interface like ChatGPT is worthwhile. It’s interesting, it plays a role; many people have questions they want answers to. It’s not a bad thing, but I don’t think that’s the primary revenue engine. I think the primary revenue engine will be licensing the API calls to plug-ins that then serve particular vertical markets. And I think for Zarqa it will be the same way. We will make a ChatGPT-type interface for web or for mobile, or a voice interface you can use through your Pixel Watch or whatever. We will make that, and people will use it, and it will be cool. On the other hand, we will also make a plug-in architecture which will allow people to build whatever application they want, accessing Zarqa’s neurosymbolic language model with APIs, and the revenue models for these plug-ins will be as numerous as there are vertical markets.
44:11 But our plug-in architecture can be much, much better than Google’s or Microsoft’s, because we have SingularityNET, we have HyperCycle. We’ve put a lot of thought into how you connect different software components together in a way that doesn’t give anyone dictatorial control over the whole thing. If you’re building plug-ins for ChatGPT, in the end you’re the minion of ChatGPT. The plug-in architecture is made in a way that makes the plug-ins subordinate to ChatGPT.
In SingularityNET, we’ve designed things for more peer-to-peer type operations, so people can make software systems that interact with the Zarqa LLM but can be deployed in a whole variety of different ways. Then we have the AI-DSL, which is a language for different components to talk to each other. Another AI plug-in can use this AI description language associated with SingularityNET to describe what it does to the Zarqa LLM, and to describe what it does to another component, so they can figure out how to work together. SingularityNET has been all about how you plug together multiple AI components on the blockchain so they can cooperate to solve problems. I think this will lead Zarqa to have a much more sophisticated plug-in architecture and ecosystem than OpenAI will have with their more centralised and dictatorial model. And I think that in the end is where the real money is: in the API calls being sold to the people making application-specific plug-ins.
46:13 Probably even chat will go that way. I don’t think you really need one particular chatbot to serve all things. I think you’ll have a chat assistant on your phone, and if you’re a researcher you can have another chat assistant who’s a research assistant, helping you talk about and read the research papers. And if you’re lonely, you can have another chatbot that’s your virtual girlfriend or boyfriend, which doesn’t have to be the same as your research assistant, or as your phone, which has a dialogue system whose goal is to help you use your phone and get stuff done better.
So I think the all-purpose retail chat interface we see in ChatGPT is a phase, and where the reality will be is in the plug-ins. That will include some plug-ins that are more specialised chat dialogue systems with their own personalities and functions, and some plug-ins that don’t look like chat systems at all. They just look like something that’s playing your music, or something that’s telling you what to buy, but they use a neurosymbolic LLM on the back end, paid for by API calls. So the bottom line is, the revenue model is the API calls. There may not be a great revenue model for ChatGPT as a product, but that doesn’t matter, because I think that’s only of ephemeral existence, because we don’t have the whole plug-in ecosystem built yet.
Brian Rose: In the end, you’re talking about the whole economy.
Dr Ben Goertzel: Right, just like integrated circuits or the internet.
48:03 Brian Rose: Last question, man, because I know you’ve got to go. Regulation: thoughts? Sam was in front of the Senate this week, and they were talking about how maybe something like an IAEA, like the atomic energy regulator, would be the right thing.
Dr Ben Goertzel: Anything anyone from a big company says about regulation, one has to assume is very loosely related to their true thoughts, sentiments and feelings. Nothing against Sam Altman, who’s a great guy and actually pretty much a straight shooter for a corporate guy. But still, there are a lot of business interests there; these are very high-stakes things. So I think what governments say and what corporate executives say is all very well-scripted theatre, I would imagine. Elon Musk is a bit less like that. He merges into well-scripted theatre now and then, and then just goes back to being a whacked-out tech bro.
49:21 But I think some form of AI regulation will happen. Europe probably will make more stringent regulations than anywhere else in the free world; that’s what happened with GDPR.
Brian Rose: Sam just told the European Union he might just turn it off if they did that.
Dr Ben Goertzel: He didn’t really quite say that. If you look at the quote, the media took that and ran with it. What he really said is: we’ll see how the regulations unfold, we will try our best to comply, and if it turns out not to be practical to comply, then okay, we’ll turn it off. To me that was corporate speak with no semantic value, but the media wants to spin it into “he said we’ll turn it off”.
Brian Rose: To be honest, that’s a bit of a threat. It’s a well-said threat, I think.
Dr Ben Goertzel: But it’s trivially true of everyone, right? Everyone will turn their product off if they can’t find a way to comply. Since it’s not yet clear what the regulations are, it’s not yet clear that there will be a way to comply. I guess there will be. If you look at crypto, the EU has actually been more reasonable than the US, UK or anywhere else; MiCA is quite reasonable. In the end GDPR is perfectly reasonable. It’s expensive and annoying to comply with, but you can comply with it. In the end, you can make a way for someone to remove their data from an LLM’s knowledge base. It’s just more expensive, but the money is there to do it.
51:04 But I think in the medium term a decentralised approach is far, far better in terms of compliance, because it’s not monolithic. You can have different approaches in different parts of the world, and if what you’re doing is decentralised anyway, you just have many fewer obstacles to doing it. Even if it’s the same company, like Zarqa, operating in every part of the world, you can run different servers in different parts of the world, they can run different variations of the software, and they can all talk to each other.
The decentralised infrastructure we’re building, which underlies Zarqa, TrueAGI and everything in the SingularityNET ecosystem, is built from the get-go for the different local sub-networks to operate in slightly different ways and yet still interoperate. It’s not that OpenAI or Google couldn’t build something that way. In China, Microsoft and Google have somewhat done that; they’ve done local stuff in China based on Chinese regulations. It’s just that they have to do a bunch of special stuff to make local sub-networks operate differently, whereas that’s how a decentralised network works from the get-go. So I think we can more cheaply, and in a more agile way, adapt the local functioning of sub-networks of our networks to local regulations, as well as to cultural differences and infrastructure differences. So from a competitive advantage standpoint, Zarqa and the whole SingularityNET ecosystem are generically better off than Big Tech as regulations unfold.
53:07 That said, my hope is that we’re not over-regulated. I have to say it was funny: when our SingularityNET media guys reached out to various media outlets to publish my recent thoughts on the AI pause and regulation of AI, the place they found was Fox News. I thought this was pretty hilarious, because I’m an anarcho-socialist. I’ve never voted for a Republican candidate in my entire life. But on these particular issues, I don’t want to see government over-regulate crypto, I don’t want to see government over-regulate AI. On those particular issues, I am more aligned with Fox News than with the more left-wing media that I agree with on most points.
So it’s going to be a political issue, and it’s going to be interesting to see how it all unfolds. But I think whatever the regulators do, a decentralised network is pretty much better equipped to cope with it than a centralised one, because we can adapt and do different things in different places.
Brian Rose: All roads lead to you, Ben, and all the great work you’re doing at Zarqa and SingularityNET.
Dr Ben Goertzel: All roads lead to the Singularity. All of us humans are just pawns.
54:55 Brian Rose: Look, congratulations on all the great things you guys are doing over there. We’re super excited about Zarqa. Final thoughts from you, Ben, before you go?
Dr Ben Goertzel: I think we’ve said it all, Brian. The AGI race is on, and Big Tech has an apparent advantage. But for the reasons stated in the “there is no moat” memo, and for a lot of other, even deeper reasons that we’ve recounted, that advantage is most likely temporary, and appears stronger than it is. I think the open, decentralised ecosystem has a lot of fundamental advantages, which are going to make things like Bard and OpenAI seem more Yahoo-like than Google-like in hindsight, a few years from now. So the race is on, and I think we’ve got a good chance of winning. And the thing is, we’re not looking to win for ourselves. We’re looking to win for the whole of humanity, which is the beauty of decentralised models.
Brian Rose: Yeah, very much so. All right, Ben, I learn so much every time I talk to you. I really appreciate these conversations, and thanks for continually coming on and keeping us updated. I think we’ve got you back here in another few weeks, so all the best.
Dr Ben Goertzel: Should be good. Looking forward to it.
Brian Rose: All right, great. Thank you so much, Ben. We’ll see you next time, and thanks, everybody, for watching.
Dr Ben Goertzel: All right, good. Take care.
