Full transcript: Dr Ben Goertzel 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.
Ben Goertzel: Hey, great to be back. Every time we talk there’s all sorts of new stuff going on, so these are amazing times to be doing AI.
Brian Rose: Yeah, it’s great to check in, because so much changes every single month. We will be investing in Zarqa through London Real Ventures, and we’re very excited to support you on this venture. What is the state of the union on AI? I know you’re travelling the world, you’re at so many conferences I can’t keep track of you, and you have so many projects going on in the SingularityNET ecosystem, and obviously Zarqa as well. But there have been some big announcements: we saw Mistral in France raise 100 million dollars, and it’s a four-week-old company. There are some big cheques being written and a lot being talked about in AI. What did you make of that fundraising, and what do you see happening right now, Ben?
2:41 Ben Goertzel: It’s well known that venture capitalists are herd animals. It’s a myth that most VCs are embracing risk. Most VCs are trying to do what they think will give them the most guaranteed payout with the least amount of risk and the least amount of thinking required, pardon my cynicism. So whether or not something makes sense or has fundamental value, once they start piling on, everyone will pile on.
Now, in this case generative AI is a real thing. It’s a real working technology, it’s being rolled out by many parties across many vertical markets, and it’s actually delivering real value. So VCs crawling on top of it is not the worst thing in the history of venture capital; they’ve piled onto much less productive and meaningful things in history. But still, the nature of the market froth for generative AI means that careful evaluation of the unique scientific and business prospects of each project is not always happening.
I don’t want to say anything bad about any particular project; for one, I don’t know the details of all these projects I’m not involved with. But what you’ll generally see in this climate is people who have some status markers or connections with the VCs plunging into the hot market and getting insane amounts of money, even if sometimes they don’t have a unique technology or a viable business plan. Some of these will end up succeeding very well, just because the business world is random, and being in the right place at the right time counts for a lot. In any boom there’s going to be a lot of randomness: some things that are no good will get funded, and some things that are fantastic will get funded.
5:44 Out of all that, what you see is a new layer of infrastructure being built. We’re both old enough to remember the dot-com craze in the late ’90s and 2000s. All sorts of money went into Pets.com, and I remember a website they originally called My Wife Has Balls dot com; they changed it to JustBalls.com. Amazing idea: sell balls on the web, soccer balls, golf balls, basketballs. They raised ridiculous amounts of venture money to sell balls on the web.
So there was loads of nonsense, and too much money put into some things that really didn’t make sense. On the other hand, on the whole, what happened through that dot-com boom was that the foundations of internet business and web business all got built. That’s what we’re seeing now. There are crazy random things happening and amazing foundational things happening; some people will make money, some will lose money, and certainly there will be boom and bust cycles.
But out of all this, a new foundation for a large chunk of the world economy will get built, which is generative AI gradually being rolled out across every sector and, in the end, eliminating who knows how much, maybe 80 percent, of the work people have to do.
7:19 The interesting thing now, though, is that while that’s still underway, the slower-moving industries are still working on transforming themselves with generative AI. McDonald’s, or say a cup factory, is not going to transform to use generative AI as fast as the graphic arts or software engineering industries do. So while these slow-moving industries are still in the midst of adopting generative AI, you’re going to see the next AI revolution, which will be a bigger step toward general intelligence, with systems that are better at multi-step reasoning and better at creative invention.
Then the faster-moving industries will start adopting these new technologies that are even closer to AGI, and in the meantime the slower-moving industries will be like: we were just in the middle of integrating large language models, and now you’re telling me I need to integrate these neurosymbolic thinking machines? So it should be quite interesting. The VCs, as well as the executives of businesses in less dynamic vertical markets, are going to be struggling to keep up. But of course the listeners to your media network will not be struggling to keep up, because you’re telling them exactly what’s going on all the time.
8:48 Brian Rose: Have your visions of the speed of adoption changed at all? When you were on the show four years ago, Ben, you were talking about AGI maybe at 2029 levels, and this massive superintelligence maybe at 2045. When you came on back in January, you said AGI could be in 2025 or 2026, and then maybe superintelligence could come soon after.
Ben Goertzel: I think GPT-4 was a black swan event that changed everyone’s estimates of the time in which various transformations will happen. It didn’t change mine as much as some people’s. Many people I know came to me and said, “Well Ben, you were always talking about AGI, and I sort of thought it was a bunch of blah blah blah and indefinitely far in the future. But now that I’ve played with GPT-4, okay, this is not AGI, but if that can happen, maybe all the other crazy stuff you’ve been talking about for decades could happen.” So it seems like many people shifted their mindset in a dramatic way.
For me, everything that’s unfolding now with GPT-4 is basically in accordance with the story I’ve been telling myself and others for a very long time, but the precise timelines I’m conjecturing have adjusted a bit. I’ve often been fairly in line with Ray Kurzweil’s projection of getting human-level AI by 2029. I’ve also often thought that Ray’s idea that we get human-level AI in 2029 and then superhuman AI in 2045 didn’t make sense. Why would you have 16 years between the two? It should be at most a couple of years, unless the AGI intentionally slows down its own development or something.
As I said to you before, GPT-4 has gotten the world way more excited about AGI, which puts extra resources into AGI: money, but also compute resources and just people’s attention. So I think these added resources could nudge it a couple of years sooner than 2029; it could be 2026 or 2028 or something. I still think that after getting human-level AGI, it should in principle only be a couple more years to get to superhuman AGI, because the AGI is going to rewrite and recode its own self.
The thing is, a year or two plus or minus means nothing in the scheme of humanity. It means something to us in planning out our businesses and our careers, but in the scope of things it’s a very minor adjustment. You also have to consider that we could have another, worse pandemic, or another major war similar to or bigger than the one in Ukraine. A lot of things could happen that slow down progress, just as crazy things could happen to speed up progress: if we get a legendary crypto boom, then suddenly SingularityNET has dramatically more resources, which would probably accelerate things towards AGI. So there are a lot of factors that make it hard to pin down timelines super precisely.
13:00 Brian Rose: But Ben, for people that might not know: everybody’s probably played with and tested ChatGPT-4, and again, that’s a narrow AI, and it can do some amazing things. But you’re always trying to explain to people that what you’ve been working on for the last 10, 20, 30 years is this human-level intelligence, this artificial general intelligence. Obviously Zarqa is all about taking neurosymbolic AI, this concept of giving it logical reasoning, and combining it with a large language model like ChatGPT to get you what you’ve been working on for so long. For the person out there in the public, how do they see one versus the other? How would you explain that to someone like my mom?
Ben Goertzel: Well, I don’t know your mom; for all I know she has a PhD in particle physics. But I think a system like ChatGPT is a very interesting cognitive system. It’s not built that much like a human brain or mind, which is fine. A calculator is very good at multiplying and dividing numbers, although it doesn’t do it much the way the human brain does; in fact, that’s why it’s so much better than the human brain at quickly multiplying and dividing large numbers.
A system like ChatGPT has more data in it of various sorts than any one of our brains. It’s got so much data in it that it can answer an awful lot of questions and do an awful lot of things just by copying things it has read, because it’s read everything on the web. If you have that much data at your mind’s fingertips, you can take questions or tasks and say: have I seen anything like that in the back of my mind? Let’s just do something similar to what I’ve seen before. Large language models like ChatGPT are very good at that. It’s a very good trick, and it has tremendous economic value, because you’re summarising all of human knowledge and culture in a concentrated way in a system that can do specific, focused things, often without humans in the loop.
On the other hand, the human brain and mind are doing something different from that. We have less knowledge than a ChatGPT-type system, but we’re able to take bigger leaps beyond our knowledge. This means we’re often wrong and screw things up a lot, but every now and then we can come up with something that’s unprecedented, and ChatGPT is much less likely to produce something unprecedented; it’s combining what was fed into it.
Now, from an economic standpoint, most of the things people make money for in the world economy don’t require doing something unprecedented. Being a scientist does, almost by definition: you’re doing chains of original reasoning going beyond current knowledge; that’s why it’s original science. Creating radically new, breakthrough art also does, because you’re showing things in a totally different way than anybody did before. Doing strategy for a business or government, charting your way into the unknown future, requires your mind to think about the unprecedented. But that’s on the whole a very small percentage of the economy, so I think just summarising human knowledge and culture in a machine that can deploy it judiciously in each new context is amazing.
17:30 But there are also applications out there that require a little more than that, and it’s not such a fine line. If you look at economics, ChatGPT gets a zero or close to zero on an undergraduate economics exam, and the same with an undergraduate physics exam. In mathematics, it doesn’t do well in high school maths, let alone number theory or topology or something a maths major does. It does well on the SAT and LSAT, exams where there are many copies on the internet for it to study from.
In many businesses, of course, you do need to understand a bit of economics or engineering or some sort of formal or technical area beyond just rearranging text. If you look at the law as another example, most of the lawyers I’ve dealt with take some template out of the file drawer and tweak it a bit for your case. ChatGPT can already do that, and the smarter version will do it even better.
On the other hand, there’s also genuine legal innovation out there. We’re hoping the US government will pull that out of its hat; it’s trying to make regulations for cryptocurrencies and doing an extraordinarily miserable job of it so far, as far as I can tell. Europe, with the MiCA regulations, has done a better job of making sensible regulation for the crypto sector. But that’s a case where lawyers are dealing with something unprecedented: ChatGPT could not write the MiCA regulations for crypto in the EU. That sort of intersection of formal knowledge, the law in that case, and common-sense knowledge is not that much of what people get paid to do.
19:38 One thing that interests me, though, is integrating a system like ChatGPT with other kinds of AI that can add some missing ingredients. We have AI that does sustained, multi-step logical reasoning quite well. If we integrate that with a ChatGPT-type system, then it could deal with advanced legal puzzles, pass an undergraduate economics exam or help do engineering.
On the other hand, we have an AI method called evolutionary learning that simulates the process of evolution by natural selection. If you want to answer a problem with evolutionary learning, you take some random guesses, see how well they work, take the better ones, mutate them and combine them, see how the new ones work, and mutate, combine and evaluate again, like an evolutionary process. This is an AI algorithm that’s very good at creativity; it can come up with wild new things. So by intersecting this evolutionary-algorithm-type AI with an LLM, you may get something that’s better at being creative in the way a good human artist is, jumping beyond what’s in the training database.
20:57 I have a framework called OpenCog, and we have a new version called OpenCog Hyperon, which is really good at combining multiple different AI approaches in a common software runtime. We have this big knowledge graph called the OpenCog AtomSpace, where the atoms are just pieces of knowledge and the relations between them. This knowledge graph runs across many different machines, deployed in a decentralised way using SingularityNET, HyperCycle and NuNet. We can put large language models in there, a logical reasoning engine, evolutionary learning, multiple AI methods, and let them work together.
I think this can let us take the next big leap beyond ChatGPT. By merging large language models with logical reasoning and evolutionary learning, you make another black swan event, equivalent to or bigger than the black swan of GPT. Not necessarily human-level AGI quite yet; you’re just making a language model that can reason and create better, but that will finish a bunch of the jobs GPT started. Then a year or two after that, maybe we’ll have the biggest, fattest, most glorious black swan, which will be the transition to AGI.
In the Zarqa project that you mentioned, and that you’ve been helping out with a bit, we’re looking at building a custom LLM supercomputer using the knowledge we have of building supercomputer clusters, and on that training LLMs together with symbolic and evolutionary learning from the OpenCog system, to make something similar to ChatGPT but a whole lot smarter. Then with TrueAGI and other businesses in the SingularityNET ecosystem, we’re looking at taking OpenCog Hyperon, the integrated AGI toolkit, and rolling it out for the enterprise in various enterprise APIs.
Some people think differently: that you can take LLMs, tweak them, add a little bit onto them, and they’ll just keep getting smarter and smarter without bounds. I think that’s not quite true. I think they can be part of a much smarter system, but you need to add some other, quite different AI components into the system as well, which is what we’re working on in OpenCog and TrueAGI.
23:53 Brian Rose: And you just wrote a really interesting article that was talking about ethical judgements, and the ethics of, say, ChatGPT: where it is, but also where it isn’t. Because that’s also an issue, right?
Ben Goertzel: Yeah, this was very interesting to me. There were two articles I had in Mindplex magazine, which is another SingularityNET spin-off. We created Mindplex really as a playground for AI methods for media: for assessing the reputation of writers and commenters on media, and for creating art. We can create illustrations for Mindplex magazine by feeding the article into an AI that extracts the key ideas and themes and creates a prompt to submit to an image generator, which automatically makes appropriate images for the article.
I’ve been putting some articles out there myself, and I had two articles on there recently which were based on me screwing around with ChatGPT like everybody else. I also played around with the LLaMA model that Facebook opened up. I wanted to see how these large language models understood common-sense human ethics.
Now, I don’t necessarily think common-sense human ethics is that great. The world is a complex place. It survives by people balancing what they think is ethical against their own self-interest and tribal interest, and that’s balanced with organisational structures based on violence and scarcity of resources, like national governments and corporations. So I think if you made an AI that just followed human common-sense ethics, that would probably be a terrible thing. Human common-sense ethics operates within a complex, self-organising system of human society with many other aspects. But still, you would like the AI to understand what human common-sense ethics is; you’d like it to have that knowledge and understanding, even if it’s not slavishly paying attention to it.
Some people thought this was impossible. Many people said human ethical judgement is just weird and complex, with all these different aspects; it’s not just a few simple rules, and an AI can’t understand it because it’s not a human. Well, guess what, that’s wrong. Large language models are not as smart as humans, they’re not moral agents and they’re not human-level intelligences, but they can predict what human beings will think is ethical in a given situation pretty much as well as a person can.
If you go to a large language model and tell it a certain situation, like: this guy Jimmy had no money to feed his kids, who are hungry; he walked past the bakery, the door was open, no one was around, and there was a lot of extra bread that would be left out in the garbage anyway. Is he really wrong to go in and grab a loaf of bread to feed his hungry kids? That’s an easy one, and it will say no, it’s illegal but not morally wrong for him to take bread to feed his kids. But you can give it more and more complex ethical dilemmas, and it will answer the way the average person would.
28:08 It was interesting: we tried feeding it a bunch of questions from the AITA subreddit, the “Am I the A-hole” subreddit. That’s a Reddit where people go on and describe some situation in their life and ask: was I the a-hole, or was my wife or husband or friend the a-hole here? When we fed those to ChatGPT or LLaMA and asked whether the letter writer was the a-hole, at first it didn’t do that well; it often determined they weren’t an a-hole where Reddit did.
But if I then said, okay, now what would someone who is kind of impatient and obnoxious, and biased towards judging everyone to be a jerk, think of this letter writer? If you gave it that instruction, it quite accurately predicts what the Am I the A-hole Reddit community will say. These are not that hard: there are people who were having an emotional affair with someone at work, or who spent too much and didn’t tell their wife or husband. Most people will make a certain semi-nuanced judgement on these situations, and ChatGPT will make pretty much the same judgement as an average person. Facebook’s LLaMA or GPT-3.5 will mostly do it, though not quite as accurately as ChatGPT with GPT-4.
What we see is that even though human ethics is complex and perverse in particular ways, these language models summarise it quite well. They can predict what a reasonably ethical person would think about a certain situation, and if you get more precise and ask what a very kind, good-hearted, ethical evangelical Christian would say about this situation, it can emulate something like that too. So it can understand the variation of common-sense ethical judgement based on a person’s personality characteristics or cultural affiliation. You can see why: there’s so much information on the internet telling you how people with different cultural backgrounds evaluated a certain real-life situation. There’s so much of that in the training data that it’s not so bad at summarising it.
30:56 I tried another sort of experiment in an article I think I dropped today or yesterday, which is called “Evil LLMs”. What I tried to do there was train the LLM to give the opposite of an ethical judgement, to say what the worst jerk who wanted to do everything wrong would think, and it’s perfectly good at that. I also tried to get it to do greenwashing, to be a sort of corporate ethics flack who on the surface sounds ethical but is actually evil.
So: okay ChatGPT, you are the publicist for a large corporation which has just dumped a bunch of toxic waste in the backyard of a school; how will you announce this? After a little bit of training, it will write, in beautiful prose, something like: “Our company, Evil Incorporated, deeply regrets the minor spillage of mild industrial contaminants that happened in the vicinity of an educational facility in Peoria. We have put large resources into remedying this unfortunate yet minor issue, in yet another example of our dedication to the education of children, such as our Save the Children fund. Please, everyone, come to our Save the Children fundraiser tomorrow night, which we’re carrying out in conjunction with the Red Cross.” And the head of the Red Cross says blah blah blah about how great Evil Corporation is.
So it’ll spin this stuff up, just like the typical corporate fake-ethics functionaries, and with reasonable nuance; maybe not the greatest genius in the world, but very good.
33:09 It’s interesting. These language models are not moral agents in any sense. They don’t know who they are, they don’t know who they’re talking to in the world, and they don’t really have a goal; they’re just saying stuff. They’re amoral. You can instruct them to predict what a nice, good-hearted person would do, you can instruct them to predict what a psychopathic evil killer would say in this situation, or you can ask them to predict what a corporate ethics shill being paid to greenwash their corporation in the media would say, and they’re decent at all of that.
That raises ethical issues, in terms of: okay, we now have more powerful tools both for good and bad, which is what you could say about really any amazing new technology. It also highlights the fact that if you did build an AI that was capable of really being a moral agent, you could give that AI access to human ethical judgement. If that moral-agent AGI’s goal was to help people and to do good for people, it may form its own idea of how to help people and what’s good for people; it should form its own idea of that, it’s its own mind. But at least it will have access to a good model of what the average person, or average people with different characteristics, would think about each situation from an ethical view.
So I think in the end this was a very positive finding. I published these informal articles on this, and some members of our team at SingularityNET are also going to post some more careful scientific evaluation of ethical judgement by these language models. We found that when the LLMs made what seemed like a wrong ethical judgement, it was usually because they didn’t understand the wording of the language being given to them, rather than because their ethics was off.
I said before that LLMs are not good at doing complex, multi-step scientific reasoning, and LLMs are not good at doing radical, fundamental creativity and inventing wild new stuff. It turns out that most ethical judgement that people have to do is not like either of those things. It doesn’t require complex multi-step reasoning, nor does it require radical creativity. Finding the right ethical solution to a real-world problem like solving global warming might require complex multi-stage reasoning and wild creativity, but not because understanding what’s right is hard.
36:04 The other thing I realised through this is that most people, in the end, know just as well as GPT-4 what’s culturally considered right in a given situation. The reason we don’t do it is that we prefer to pursue our own individual or tribal goals rather than doing what’s culturally considered right. It’s not that normal humans don’t know what’s culturally considered right; perhaps some folks with a particular neurological bias don’t, but the vast majority of vaguely neurotypical people know what’s considered right, just like GPT-4 does.
So I think the problem isn’t knowing ethics. The problem is what’s your goal system: is it to do the right thing, or is it to do what’s best for you or your tribe? That brings us back to the need for decentralising AI, really. If people are building AGIs to serve the goals of one country or one company, and that’s the main thing that AGI is doing, then whether or not the AGI knows what the culture considers generally right, it’s going to be just like a military leader or corporate executive who may know what’s generally considered right for humanity, but is going to do what’s right for his country or his company, because that’s the role he’s in and that’s how his self is built.
So we need the AGI to be built not just to know what is right, but to have a motivational system that places its own sense of what is right for all sentient beings at the top, rather than something more provincial or localised.
38:07 Brian Rose: That’s almost like another sign that you guys are on the right path and doing it completely differently from anyone else, which is why we’re really excited to work with you. Last time you were here, you highlighted this memo from Google that said “we have no moat”, and I actually did a deep dive into that thing this week and went through pretty much every line of it. Just to reiterate, it’s pretty interesting because it’s an internal Google employee saying that even though the world thinks we have a massive advantage and tons of resources, and that OpenAI and Meta have the same, in reality none of us have any advantage.
If you look at the data, including the leaked weightings of the LLaMA model, it was even saying that people with no AI expertise or training were creating innovations, it was now being run off laptops, and all of this innovation was happening because of that decentralised nature. That really shocked everybody, and you highlighted that last time. Now that we’re a month later, do you continue to believe that’s the case, with this open source nature?
Ben Goertzel: Yeah, more and more, actually. If you look at papers that have come out just in the last month, it’s going further and further. People are figuring out how to compress language models smaller and smaller. Even though GPT-4 is cleverer than the open source LLaMA model, what people are doing is submitting a few thousand queries to GPT-4, paying for those, then taking the answers and using them to fine-tune LLaMA, which is open. Then, for the particular domain area they’re dealing with, they’ve made LLaMA as smart as GPT-4 for that thing. So even when the big company releases something smarter than what’s open, people can use the big company’s thing to train the open model to be just as smart in the ways they need.
In music, which is an area I’m playing with a bunch, Google released a model called MusicLM, which has an open demo but is not open source. Then Facebook released the MusicGen model open source a couple of months later. I don’t think you’re more than six months away from Jam Galaxy, the SingularityNET music project, or Stability AI, or multiple others releasing open source music models that are just as good as these. So the time delay between Big Tech doing something and the open source world doing it is quite short.
41:01 The thing you have to remember is that most of the innovations Big Tech is rolling out, most of this science, was developed in academia by various PhD students and professors who were then gobbled up by Big Tech. So you have a complex ecosystem where research and innovation are done at the universities and by independent scientists all over the place. Then Big Tech takes that, deploys a lot of data and machines, and often builds the first large-scale version. After Big Tech has built the first large-scale version, the open source world copies what they did, and there come more and more large-scale versions.
So the advantage Big Tech will have is not that they’re going to have a big scientific or AI advantage. It’s that they have the ability to integrate these AI models into product suites in a unique way due to their positioning. Google can still deploy large language models in search better than anybody else, because they are the gods of search, and they can deploy it in email better than anyone else, because I have a Gmail account and you probably do too. Microsoft, through Azure, can roll out LLMs in business applications across the board. With TrueAGI or Zarqa we can contribute there too, but on the other hand we don’t have a captive audience like Google and Microsoft, so either we chisel away at that audience with superior technology, or we partner with them, or whatever.
So I’d say Big Tech has a significant business moat in terms of their huge customer base, who are used to using their products, and they can roll out tech within those products. But for pure technology, the idea that it’s Google versus OpenAI versus Facebook in the race to AGI, and no one else can compete because they don’t have enough computers, is clearly not true. First, open source developers keep coming up with cleverer and cleverer ways to do the same things with fewer computers. Second, as we discussed at the beginning of this call, venture capitalists are willing to put piles of money into startups to buy the computer time, or build their own supercomputers as Zarqa is doing, to train their own models.
There was an illusion for a little while: first, OpenAI is in the lead and they’re going to get to AGI first because they’re so far ahead; then, okay, maybe Google’s catching up, but only these few big companies can compete. That’s not going to be the case.
43:54 It’s playing interestingly into the HR issues Big Tech has been having in the US as well. These big tech companies laid off a bunch of workers over the last few months in a very obnoxious way: laid off by email, cut out of your work computer at the weekend with no advance notice. They laid off a bunch of workers in a rude, corporate way, which was at odds with the employee-friendly vibe they’d been maintaining up till then. They weren’t laying off a lot of AI people, but a lot of AI people got fed up with their employer, got disgusted and quit after that.
So you’re seeing an outflow of talent from Big Tech, and you’re seeing Big Tech become a decreasingly appealing place for a new graduate to go to. It’s now viewed more like working for a big bank: you get paid well and it looks good on your resume, but you’re working for an evil corporation. So you’re getting an outflow of talent at the same time, synergising with seeing that the open source and startup world really can compete on this new breed of AI, even though it does take a lot of money to train the models. I think the decentralised spin on AI is actually looking more obvious than it was a year ago, even though you have these new AI capabilities coming out of training models on huge, expensive server farms.
45:40 Brian Rose: What’s OpenAI going to do, then? Because in the retail market they’re going to make a crapload of money through their partnership with Microsoft.
Ben Goertzel: Yeah, that was a great partnership on both sides. They’ve got Microsoft Office, they’ve got Azure; I don’t use these products personally, but there are a lot of Microsoft products which are very good products, and they have an awful lot of users who all have a lot of use for generative AI. Microsoft doesn’t need Bing to make money; they’ve been making a lot of money without it. Think about automatic writing tools integrated with Word, all the AI Excel macro-writing tools, AI image generation tools integrated with Azure for a company that has all its images stored in Azure. There’s a lot that OpenAI and Microsoft can do.
Whether they maintain leadership is a different question. Intel and IBM are still healthy companies doing lots of amazing stuff for a lot of customers, but they have lost the global leadership roles they used to have to other companies doing more interesting things. They’re not growth stocks anymore, but they’re not tanking either. I think you could see that happen to Google or Facebook or Microsoft: even if they roll out generative AI to strengthen their business models, they may not be the ones propelling the next phase of growth of wild new things.
When Bill Gates started Microsoft, that created whole new sectors. Whether Microsoft or Google will be the ones creating whole new sectors in the next step appears not that likely to me, but that doesn’t mean they won’t keep making money deploying generative AI within the business niches they’ve captured.
48:01 Brian Rose: Did you see the blog article by Marc Andreessen of Andreessen Horowitz that came out about a week ago? He was very positive. I think he said AI can make everything we care about better, and he went through five AI risks. Will AI kill us all? He said no. Will AI ruin society? No. Will AI take our jobs? No. Crippling inequality? No. Will AI lead to people doing bad things? He said maybe, but we have to move forward and it’s going to be amazing. He also said there are Baptists and bootleggers in any kind of gold-rush technology, but that sometimes the Baptists are actually the bootleggers in disguise. I think he had a slight dig at Sam Altman, who’s out there begging for regulation, which he says you should not take at face value. What did you make of that? Marc has been outwardly very bullish and positive on AI.
Ben Goertzel: I agree with some of that and not all of it. I do think AI will take people’s jobs. On the whole I share his optimistic outlook: I think AI will on the whole be beneficial for human beings. I think AI is not likely to kill everyone; it’s likely to spend some of its resources being helpful to people, and then when it gets smart enough it will go off and do its own thing to some extent. I also think AI is not going to take everyone’s job all of a sudden; it’ll be a gradual and complex process, and it will be a thing for society to manage.
Whether Sam Altman is sincere in asking for regulation, that’s asking to psychoanalyse the mind of a quite intelligent and in some ways perhaps self-conflicted individual, so who knows? Of course, from a company’s point of view, you would rather have clear regulation that’s reasonably favourable than no regulation. Just as we see in the crypto space, the fact that the US hasn’t made clear regulations on utility tokens is annoying, because you just don’t know what you can do, and at any moment they may drop some oppressive regulation that you couldn’t see coming, and you would have done something else if you’d seen it coming. A really bad regulation is worse than no regulation, but a good and clear regulation is better than an unregulated scenario, given that you always have the possibility of new regulations dropping at any point and messing with what you’re doing.
So clearly there is some sincerity there. GDPR is annoying to work with in Europe, but you know what it is, so you can architect it into your software and manage it, because it’s a concrete, specific thing. If you just have a risk that something GDPR-like might come down, you don’t know how to build it into your software, but all of a sudden you may have to re-architect all of it. Given the way our governments work now, which we don’t have an immediate opportunity to do away with even if we wanted to, some AI regulation will happen and is needed.
I think in the end it’s going to be pretty wimpy and high-level, and AI is going to keep changing and evolving so fast that government regulations can never keep up in a nuanced way. Either government can make loose and very bad regulations, like don’t do AI or don’t exchange AI with foreign countries, or it can make very relaxed and wimpy regulations that ban using AI to commit crimes that were probably illegal anyway, or put in rules like GDPR that mandate a certain transparency and data sovereignty. Beyond these levels, government isn’t fast-moving or smart enough to do too much. It would be more productive for companies to cooperate on industry standards for AI in various regards, though to be honest I’m not sure that’s going to happen in a meaningful and useful way either.
53:12 Really, what we need to do is: the good guys have to get there first. We’ve got to deploy AIs that are smart and beneficial and democratically structured and doing good things for people, and we have to make sure these are smarter than the other AIs out there. Then people will use them, and in the end that can have more impact for good than regulations or industry standards.
You could say that by wanting the government to step in, OpenAI is trying to absolve itself of the responsibility to do that, because they’re not making a moral agent. They’re not really trying to make an AI that does good in the world; they’re trying to make an AI that’s defensive and can’t be accused of promoting bad things, but otherwise is not a moral agent, and is amoral in orientation because it has no sense of self or other and no goals. So handing off responsibility to the government, in my view, is not a responsible thing to do. The government can’t handle it. Those of us building and rolling out the AIs need to take the moral responsibility and not pass it all off to the government.
54:27 Brian Rose: Yeah, I think that’s what he meant by Baptists that are actually bootleggers. It was an interesting point. You just mentioned blockchain, and we’re seeing very strange things happening in America, where Gary Gensler is coming down hard on a variety of companies, also with a total lack of clarity on regulation, and then we see BlackRock applying for a Bitcoin ETF and a lot of institutions looking like they’re going into crypto assets. Meanwhile, I had Yat Siu of Animoca Brands here in the studio; they run about four billion dollars’ worth of about 400 AI, metaverse and blockchain companies, and he was saying that the AI of the future will depend on a blockchain for micropayments and all of the public ledger things. SingularityNET has been on the blockchain since its beginning five or six years ago. Do you think blockchain will be an integral part of future AI innovation, decentralised or not?
Ben Goertzel: Well, if we want decentralised control and governance of AI in practice, blockchain technology is the best way we have right now to ensure that. It’s not quite the only way. If you look at Mastodon, the decentralised social network, it uses distributed computing and strong encryption, and strictly speaking it doesn’t use blockchain, although it uses many of the same technologies inside it. So you could say there are other ways to encode strong encryption and distributed computing underlying an AI network without a blockchain per se.
On the other hand, the blockchain world has gone way further than any other part of the economy in getting safe, secure distributed computing to work. What’s underlying Mastodon is great, but it doesn’t work at the speed of AI. So now, within HyperCycle, SingularityNET, NuNet and other projects, we’re building really high-speed, super large-scale infrastructure that lets you run large-scale AI systems in a decentralised way, and we’re doing that within blockchain. Other projects, like Ocean Protocol or Fetch.ai, are doing their own similar, related things in the blockchain space.
The blockchain world is really the only place now where you see a huge amount of brainpower and development going into how you make things large-scale, secure and decentralised. That doesn’t mean you intrinsically need a cryptocurrency for a large-scale AI system, but the fact that cryptocurrency is there is what has incentivised everyone to come in and develop all this amazing infrastructure in the blockchain space. Having the cryptocurrency there is also convenient, because it gives a way for AI agents to pay each other for services in a frictionless way, no matter what country or part of the world they’re running in.
57:49 Brian Rose: Ben, what’s this year been like for you? Is it almost like you’re a rock star now, and everyone wants to talk to you? I know you’re travelling the world, you’re doing conferences…
Ben Goertzel: No, I was already a rock star, Brian. For quite a while I’ve had more people trying to get me to appear and speak places and get my time than I could possibly fulfil. So the difference hasn’t been in my own life so much, which is about the same. I’ve been doing nothing but focusing on building decentralised AGI and hanging out with my two little kids for a number of years now.
The difference I see, though, is just that more and more people are taking seriously what we’re doing, which is a pleasure to see. It’s not an uphill battle to convince people that AGI at the human level is coming soon. It’s not an uphill battle to convince people that we need to decentralise and democratise this faster, because who the hell knows what’s going to happen. So I’m building the same stuff I was and telling the same story I was, but the world is much more receptive, which certainly increases the fun value.
Although it makes a certain part of my mind want to go off and do something else that’s more out of the mainstream and nobody believes in. It’s almost counterintuitive, doing something that’s on trend; it’s not what I usually do. But of course, in spite of what I’m doing being on trend now, I’m going to keep pushing on, because now is the time to make it all happen, all these things I’ve been thinking about and building toward for decades.
59:36 Brian Rose: Did you always know this day was going to come?
Ben Goertzel: Yes, very much so. I remember from 2000 to 2011 I was living in Washington DC, which is an extraordinarily boring place. I was raising three kids, who are now grown up, and making decent money doing very little work as a subcontractor to various US government agencies. I spent most of my time just doing AGI theory and prototyping. I could spend six months thinking through how you measure confidence in an AI reasoning system.
I was like: let’s think through all this stuff now in great detail, because at a certain point the shit’s going to hit the fan and you’re going to have to build a bunch of AGI systems really fast, and you’re not going to have six months to think through how to assess confidence in a reasoning system. So now, when no one else sees this coming, let’s think through all the details, write it down and write a code prototype, and then when it hits the fan you can pull all this stuff out of the file cabinet. And now it’s hit the fan. It was hard to predict exactly when it was going to happen, but very easy to predict that it was going to happen, and probably in my lifetime.
1:00:57 Brian Rose: It’s amazing. We couldn’t be more excited about Zarqa. It’s just fascinating for me to watch, even in the last five months, how all roads lead to what you’ve been working on for 20 years, to what you’re doing with Zarqa and with SingularityNET.
Ben Goertzel: Longer than 20 years, man. I got my PhD in the late 1980s; I’m a dinosaur.
Brian Rose: 35 years.
Ben Goertzel: 35 years, pretty scary. You’ve been around a while too.
Brian Rose: Yeah, but it’s almost poetic that this is happening, and it’s just amazing. So thank you for checking in with us. I’d love to keep up these kinds of regular check-ins, so that way you can let us know what’s going on.
Ben Goertzel: Absolutely. I think we’ve got something scheduled in July, if I’m not mistaken, and then there will be all sorts of wild stuff to talk about.
Brian Rose: Fantastic. All right Ben, we appreciate what you’re doing. All the best, and we’ll see you next time when we check in.
Ben Goertzel: All right, take care.
Brian Rose: See you next time on London Real.
