I wrote this article, and made this Web page, shortly after Webmind Inc. closed its doors on April 2, 2001
For info on the current Webmind Inc. situation, see www.goertzel.org/webmind.htm
For a personal essay on the state of Webmind Inc. as of April 2001, see below…
Waking Up from the Economy of Dreams
April 9, 2001
Rationally speaking, bankruptcy was a clear and present danger right from the start. The seed money we got was only enough to last us a few months, and we had no idea where the next round was coming from. But at that point, we never really considered we might run out of money. It was early 1998, the economy was booming, the Internet was all the rage. There were only a handful of us then – six company co-founders, and I was the ringleader. I’d recruited the others to help me out with the not inconsiderable task that I had set for myself: creating and commercializing the world’s first truly intelligent computer program.
Of course, in the present state of the financial markets, the idea of starting a company with a goal like this would get you laughed out of any conversation with any serious businessperson. Creating a thinking machine, and then commercializing it? Well, fine, but how are you going to make money while you’re creating the damn thing?
But in 1998, all problems were solvable. We had the answer to that. Sure, the thinking machine might not get finished in the 3 months that our seed money would last us. But, in the meantime, to tide us over, we’d solve a simpler problem: we’d use some of the bits and pieces of our unfinished AI engine to predict the financial markets. The technical co-founders and I had been working on the first version of the AI engine for many months, by the time the seed funding came in. A healthy amount of software code existed (although the code itself wasn’t entirely healthy).
The seed money had come from the family of one of the six co-founders, Lisa Pazer. We six were a motley crew indeed! One of the co-founders, Ken Silverman, was an old friend of mine from Simon’s Rock College, an institution that offers college education to high school age students. We’d met in our freshman year, when I was 15 and he was 16. Onar Aam, living in Norway, was a philosopher/composer/hacker whom I’d known through the Net for yeas; Paul Baclace, living in California, was an experienced Silicon Valley software dude who had contacted me and asked to become involved in the project. Jeff Pressing, living in Melbourne was a psychology professor, an expert in physics, music, financial prediction and a whole host of other things. He was another quasi-prodigy, having entered CalTech at 15. We had a lot of collective brainpower, but Lisa was the only real businessperson among us. She was our first CEO, and she was also the inspiration for turning our AI tools to the problem of market prediction. Many times during her years as a market analyst, she’d been frustrated by watching the quantitative market analysts sit in a room by themselves, analyzing charts and crunching numbers, while the qualitative analysts like her self sat in another room analyzing news events, trying to limn the market’s psychology. What if one could use the computer to read the news and predict the markets based on what it read? What if it could combine this information with the standard number-crunching formulas? Quantitative and qualitative analysis would be fused into single prediction! She asked me if my software could do this. I told her, sure, compared to true human-level intelligence and fluent conversation, that’s easy. I knew I had Jeff’s prior work on nonlinear-dynamics-based market prediction to fall back on, and the partially-finished AI Engine code as well – the puzzle pieces were there, they just had to be fit together….
Lisa and I had met over the Internet, in late 1996. I was living in Perth, working as a Research Fellow at the University of Western Australia. She had recently quit her job as a Wall Street market analyst to stay home with her kids, and devote herself to fiction writing. She found some of my crazy half-finished fiction on the Web, but she was particularly intrigued by an article describing my vision of the global brain, the emerging network of transhuman intelligence that, I said, was about emerge out of the computing and communication networks all around us.
The global brain wasn’t something any one person could create, not even me; I knew that. It was something that had to emerge. But there were things you could to seed its emergence, encourage its crystallization. One of these things, I believed (and still do believe), was the creation of real AI. By creating a computer program that could read all the information on the Internet, understand much of it, and place new information back on the Internet, enhancing the environment that it lived in, one could cause a phase transition in the development of the Net. One could jolt the Net into a different state of being – effectively causing the emergence of a new organism, a new form of intelligence and life. I wasn’t the only one to have this vision, of course. After the idea became crystallized in my mind, I learned about my predecessors thinking in this direction, including Valentin Turchin, a brilliant Russian computer scientist whom I know personally now, and who published a book reporting such ideas way back in the early ‘70’s
In early 1997, I moved back to the US, to work with Lisa and another friend, John Pritchard, on creating superintelligent AI software that would achieve exactly this. John had business connections as well as deep software ideas, and he felt we could get a new company off the ground quickly. As it happened, Lisa, John and I were not an effective threesome, and John went off to build his own software in peace. (Actually, his software system, a kind of super-efficient Java-based Web operating system, is still not finished after many years of efforts, and still defies easy verbal description; see www.syntelos.com) Lisa and I plotted world domination on our own, until I began recruiting other friends to join the effort. We struggled to get VC funding for a while, but apparently we didn’t have a good enough story. Fortunately, over a year after I’d returned from the US, her family chipped in some of their hard-earned cash, and we were off. We were a few months away from being broke, which was a lot better than actually being broke, which had been our previous situation.
What saved us from bankruptcy at that point? Well, actually, it was a newspaper article that did the trick. A friend introduced me to Tom Petzinger, a Wall Street Journal columnist (and a remarkably pleasant and intelligent guy), who wrote a piece on us that appeared on the first page of the second section of the Journal. Ba-bing! This didn’t bring us VC money, but it brought us much-needed cash from a host of wealthy WSJ readers. We were alive for another half-year or so.
But we couldn’t keep going on the strength of that article forever. Even a follow-up New York Times article wasn’t enough. Our efforts to sell the Market Predictor software were going much slower than we had expected, even though initial test results were good. And, shock of shocks, creating real machine intelligence was taking longer than I’d initially projected. The design I’d started out with had come out of my academic work, which was fairly theoretical. I’d built some simple prototypes before, but I didn’t have experience with large-scale software engineering. It became clear that, the essential conceptual aspects of the system aside, there were a lot of hard pure software engineering issues that had to be plowed through in order to get the thing implemented in a workable way.
Bankruptcy was very close at that point – yet again A substantial, last-minute loan from the same friend who had introduced me to Petzinger saved the day. It carried us through until we could close a deal with our first VC, a small Chicago firm one of the partners of which was a friend of Lisa’s brother-in-law. This was a time period where, more than once, the money came into the bank account a couple hours before payroll checks had to go out. Our employees at the time never realized quite how close we were to dying. In late 1998, the economy wasn’t quite as good as during the first part of the year – the market was in a downturn (which the Market Predictor had anticipated), and pretty much no one closed VC deals in New York during that fall. But in the first months of 1999, the market got happier, and everyone’s pending VC deals closed. Whew!
Did these first VC’s, or our earlier individual investors, really understand what we were doing -- how powerful our software ultimately could be, how difficult our task really was? A few of the individual investors did. One of the ringleaders of the WSJ-inspired investors was Frank Mosca, a psychologist (and ex-spy) friend of mine with his own wild and crazy theory of the mind, and no small measure of sympathy for my own. Another significant investor at this stage was another visionary psychologist friend, Terry Marks. But these two, and a couple others, were the exception. By and large, these people were putting their money into us because, quite simply, we looked like really smart people doing really cool stuff – and it was the middle of the Internet bubble.
In fact, one of the partners of our first VC firm used to raise money for us like this. He’d tell people: “This is the best cocktail party investment you’ll ever see. Think about it this way. If these guys succeed in building a thinking machine and predicting the financial markets, you’ll be made incredibly rich, and you can say you got in on the ground floor of something remarkable. And if they fail, well, at least you can tell people you were in on a really interesting swindle. Either way, you’ll have something to talk about!”
Today you’d never see this happen – the markets are so tight, no one is about to consider investing in something they don’t 100% believe in. But in those days, companies with no original technology and no business model were seeing tremendous IPO’s. No one could understand why one company wildly succeeded and another one stagnated, or got acquired for a modest sum. So all the rules were thrown out the door. People bet based on pure intuition. They had the intuition that we were really clever, and would do something really cool. As one investor said at the time, “Will the software work? Well, that depends…. What does ‘work’ mean? I’m sure it will do something interesting.” Of course, this faith placed in me and my team by strangers was flattering. But I felt it was largely justified. We really did have a better idea about how to make computers think. We really did know how to predict the markets using the news.
In thinking about this time period, I’m often reminded of Frank Zappa’s analysis of the history of the music industry. During the late 60’s and early 70’s, he said, the music business was run by fat old men wearing suits and smoking cigars. These men had no understanding of rock and roll music at all. To them it all just sounded like noise. But the kids loved it. It made money for the record companies. So the old cigar-smoking record company executives would just sign contracts with any rock band that looked like it was doing something cool. Frank Zappa sold the Mothers of Invention as a “white-boy blues band.” When the record label heard their first recording, their response was: “Hey, that’s not white boy blues.” Well, there were a couple blues tunes on there, but perhaps what the company was perplexed by was the 15 minute track consisting of modern-classical-style quasi-tonal improvisation played over the sounds of conversation and orgasm. But it didn’t really matter. People were buying the stuff. It was all part of the Movement.
But then, a few years later, these old execs were retiring, and the record companies were taken over by younger people, people who had grown up on rock and roll, people who felt they understood youth music. The result was a stultifying conformity. From “anything goes, because we sure don’t know what’s going on”, the pendulum had swung waaaay back to the other side, to “we know exactly what sounds good, and if you’re not playing it, tough luck.” The same exact cycle hit the rock music industry again in the 90’s. The birth of alternative rock in the early 90’s was a time of wild innovation. All sorts of bands became popular under the label of “alternative rock,” including bands like Ween and Mr. Bungle whose music is barely rock at all, because no one knew what alternative rock was, except the kids who were listening to it. Well, before long, record labels hired people who had grown up on Nirvana and Pearl Jam, and alternative rock became conventionalized. The innovative phase had ended.
A lot of our early investors were like Frank Zappa’s cigar-smoking record-company executives. In fact almost everyone was in this position in the middle of the Internet bubble. Everyone knew a technological revolution was happening. You didn’t have to be the sharpest pencil in the pack to figure that one out. It was possible to make some general predictions about the outcome of this revolution, such as “Most goods and services will be sold over the Web” and “A good search engine is important” and “Some kind of global brain will eventually emerge.” But to get from these general predictions to specific judgments as to which start-up companies were going to own a piece of the grand Net future – no one was smart enough to do that. In fact this may have been an unsolvable problem, objectively, not just a consequence of the limitations of human analytical intelligence. Social and economic systems are chaotic, especially at times of transition.
What’s happened in the tech industry today is, in my view, a rather unfortunate combination of no-more-cigar-smoking–fat-guy syndrome with a severe market crash. Internet technology has been around a while now, long enough that a lot of investors figure they’ve got it down pat. Or at least, they’ve hired a few young guys who have been around, and seem to know what they’re talking about. By and large, investors are no longer interested in funding the tech equivalent of a Ween or Mothers of Invention, because there are too many companies that are offbeat but bad, along with the handful of unique and precious gems. But at the same time as this natural pendulum-swing toward conservatism, we’re seeing a terrifying market crash, so what we have is a combination of conceptual overconservatism with necessary fiscal conservatism.
But I digress. I was explaining how a loan from a friend, followed by a VC investment, saved us from bankruptcy in early 1999. At that point, odd as it seems today, we were actually in the position of trying to convince the VC’s to invest less than they wanted to. They told us we needed more money in the bank, just to be safe. We were more optimistic about how soon we’d start bringing in revenue, and how easy it would be to raise more capital. The markets were looking good.
And what a collection of brilliant people we’d accumulated! The diversity of our tech staff is such that there’s no way to do it justice. We had our share of PhD’s in relevant disciplines but that wasn’t the heart of it. It’s easy to have a crew of knowledgeable and competent PhD’s with no imagination. The Webmind vision attracted all kinds of fascinating, brilliant, independent minds, people who had taught themselves various disciplines and had been working for years in their spare time on this or that intriguing project. Anton Kolonin had been sitting in Novosibirsk creating his own project called “Webmind,” a project focused on visualizing the global brain whose key ideas would gradually wend their way into the AI Engine design. Anton’s polymath mad scientist style jarred wonderfully with the calm precision of Pei Wang, our true AI guru, who brought to us his own half-finished thinking machine, many aspects of which were incorporated into our system. Every few months, it seemed, Youlian Troyanov went from believing AI was easy to believing it was impossible. He had his own theory of the mind and universe, but no one could understand it except (presumably) him. Stephan Vladimir Bugaj, who had a few years before accomplished the remarkable feat of becoming a Bell Labs researcher without a bachelor’s degree, came to us from a boring job running and Internet ad agency, and immediately branched out into every aspect of Webmind Inc. operations, from IT to product design to AI theory to endless meetings with prospective clients in various industries, to arguing with the building administration about how many servers could be placed per square foot without breaking the floor. In true start-up mode, Stephan and I veered back and forth from being what he called “Uberjanitors” – dealing with mundane issues like ordering supplies, fixing broken computer networks, returning phone calls from interested businessfolk and assorted crackpots and so forth – to exploring the most ethereal realms of science, staying late into the night arguing about the viability of different ways of importing prefabricated linguistic knowledge into a digital mind without obstructing its ability to learn on its own. And, alongside the AI brains, we were accumulating more and more engineers with serious product development experience, some in Brazil, some in the US. It seemed that we had the capability to create a thinking machine and engineer high-quality products around it, that would enable us to, basically, take over the world….
And, speaking of brilliant minds, while we were in the process of bringing in VC money, Lisa found us another great connection: a guy named Andy Siciliano, a true titan of the financial markets. Andy had made a huge amount of money trading derivatives in the 80’s and early 90’s, and immediately when I met him, he struck me as one of the shrewdest people I’d ever met. I’m sure he has a fine creative mind and a broad-ranging intellect, but against the backdrop of the Webmind Inc. AI crew, these weren’t the qualities of his that stood out right away. Rather, it was his ability to carefully take stock of a situation, think slowly and dispassionately about all aspects, and then make a measured decision. His minimalist communication style was definitely at odds with the compulsive verbalization atmosphere that Lisa and I had fostered in the company so far. But in spite of being opposites in many ways, Andy and I hit it off pretty well. We respected each others’ very different minds. I had the feeling that, having met a lot of businessmen over the past few years, I’d finally met a really good one.
Andy had a suburbs-to-riches story to make anyone jealous. An MIT engineering undergrad, he’d quickly changed his major to finance, figuring money was where the action was. Out of MIT, he’d gotten a job at O’Connor and Associates, a young derivatives trading firm. O’Connor was just plain smarter than the competition. There was an advanced mathematical theory of derivatives and options trading, based on the Black-Scholes formula, well-known in academia but not used on the trading floor. Until O’Connor did it, making many of its staff very rich, including Andy and his friend David Solo. O’Connor was then acquired by Swiss Bank, a much larger firm which was in the process of computerizing and centralizing. After a brief period, the O’Connor management – a young bunch of “blue jeans and polo shirts” guys – had taken over many of the top management positions at Swiss Bank. These guys knew more modern ways of doing things than the suits they displaced. And then Swiss Bank merged with Union Bank of Switzerland (UBS), and the O’Connor guys again wound up at the top.
Then Long-Term Capital Management came along. This was a huge hedge fund, run by a Nobel Prize winning economist, using mathematics similar to that underlying O’connor’s success to trade the international currency market. But the real world failed to live up to the assumptions underlying the mathematics, and when a number of foreign economies tanked at once, LTCM lost billions. Andy was running the Foreign Exchange group at UBS, which possessed a large investment in LTCM from the pre-merger days. When LTCM died, Andy’s group lost many hundreds of millions of dollars – and UBS, out of concern for its public image, encouraged Andy to retire, which he did. When I met him he was 6 months or so into retirement, and bored out of his mind. A man, he explained to me, needed to be involved in doing real things out in the world, or else his testosterone level would get too low. I wasn’t sure about the metaphorical physiology, but I knew what Andy meant. Sitting in his mansion and manipulating his investments wasn’t as stimulating as pushing to dominate the world of high finance, and nor was funding movies, his main hobby at the time. Lisa and I tried to convince him to start a hedge fund using the Market Predictor, but he was more interested in doing something in the Internet space. In early 1999, financial prediction was old hat. The investment bankers who’d made piles of money in the 80’s and early 90’s were jealous of the new wave of Internet entrepreneurs, some of whom had become far richer than the investment bankers. The Net was the place to be. Andy decided to pass on the hedge fund for the moment, but he offered us his services as CEO, together with a substantial amount of investment money.
While all this fundraising, and repeated narrow avoidance of bankruptcy, was going on, the technology work was proceeding steadily in the right direction. In moving from my initial conceptual design for a thinking machine, toward a real working system, we were answering questions about computational mind design that no one else had even gotten far enough to ask.
On the highest level, the starting-point of Webmind Inc.’s AI Development work was a conceptual theory of mind called the “psynet model,” which I wrote about in four academic books, during my years as a math, psychology and computer science professor, before starting the company. In this theory, the mind is viewed as a complex, self-organizing system – a system consisting lot of little agents, each one acting somewhat independently, but interacting with each other constantly. The agents are recognizing patterns in each other, they’re creating new agents, and they’re giving rise to complicated patterns of organization. Of course, it all comes down to what the particular agents are. A lot of my academic career was spent figuring out how to explain all the different things the mind does in these terms -- perception, action, cognition, memory, learning, and so forth. In doing this figuring-out I relied heavily on past work in artificial intelligence, neurobiology, psychology, and loads of other disciplines. For years I struggled to pull my ideas together into a fully rigorous mathematical theory of the mind, but I failed. Either I, or modern mathematics, just wasn’t up to it. Anyway, what I ultimately developed in my academic research was a conceptual framework, worked out more fully in some areas than in others.
This conceptual framework, the psynet model, didn’t tell you exactly how to build a thinking machine – it just told you how not to do it, and pointed you in the right general direction. I built a prototype psynet-model-based AI in 1994, in an obscure, beautiful and fantastically inefficient programming language called Haskell, but just about all it did was make the computer run out of memory. Then I discovered the Internet – clusters of powerful machines linked tightly together … millions of less powerful machines with spare CPU cycles linked loosely together…. This, it seemed to me, was the computational brain to underlie the world’s first digital mind! Each computer on the Internet could be a neuron in the global brain! More powerful computers could be like whole groups of neurons, or whole brain regions! The view of the mind as a distributed network of heterogeneous agents, giving rise to emergent patterns, fit perfectly with the Internet, a distributed network of heterogenous machines. It all made perfect sense. There were just a few details to fill in….
In 1997, when I moved back to the US with the bright-eyed idea of starting a company to manifest my vision of Internet AI as the seed of the global brain, I made the next real real attempt to create a psynet-model-based AI engine. Unlike before, I was thinking about practicality as well as the simplicity and beauty of the underlying software code. I was determined to exploit the power of the Internet. I defined a set of particular “mind agents” and their capabilities and interactions. I did the programming in Java, a Net-friendly language that seemed to strike a middle way between elegance and practicality. I started coding the thing in April, and in September, Ken Silverman, one of the other Webmind Inc. co-founders – and my friend since we started university together at age 15, way back when -- took over most of the actual programming from me.
When we got our seed funding in early 1998, progress picked up steam considerably, and we hired experts to help us create agents with particular capabilities. I already mentioned Anton (nicknamed by himself the Siberian Madmind), and Pei Wang, an expert on logical reasoning, with his own detailed theory of how to create AI, which has contributed hugely to the Webmind AI Engine. Also crucial was Karin Verspoor, a computational linguist who had been working at Microsoft Research in Australia. And company co-founder Jeff Pressing, who had been a relatively minor company co-founder in the early, pre-funding days, grew into a bigger and bigger role. Jeff was a psychology professor with a physics PhD who had been spending several years trading the Australian financial markets in his spare time. He also had studied drumming in West Africa, was an accomplished composer and had been the dean of the music school at LaTrobe University in Melbourne. Jeff contributed a lot of ideas about how to make mind agents dealing with numerical data, as well as a large amount of general cognitive science knowledge.
Before long we had found substantial success using this early-version system to recognize correlations between trends in the news and movements in the financial markets. But practical experimentation soon led us to two significant concerns. First, making the system work effectively across a whole bunch of different machines was going to be a very big job in itself. The problem was solvable – and it had to be solved if the whole “AI seeds the global brain” vision was going to work – but it was more tedious than we’d initially thought. Secondly, using just a small assemblage of agents was not going to work – to make a really smart system we’d need billions of agents all working together. Which is not terribly surprising given that the brain has hundreds of billions of neurons in it.
There was a bit of a change in approach at this point, as we moved into 1999. It wasn’t a break from the psynet model “mind as a population of interacting agents” view, but it was a significant change in emphasis. We were focusing more and more, not on the framework by which all the mind-agents would interact with each other, but on the mind-agents themselves. In retrospect, this shift of emphasis was only natural. A human brain is born with a lot of specialized “wiring” for various types of intelligence (linguistic, visual, temporal, and so forth), and similarly, we found ourselves “wiring” our digital brain in various specialized ways by creating specialized agents. Yet we found we were able to do this without sacrificing the complete adaptability of the system and its potential for self-organizing emergent general intelligence.
With this shift in focus, the project became more integrative in nature. We drew in more and more bits and pieces from other peoples’ AI theories. We began to make extensive use of evolutionary programming, a computer science discipline based on simulation of evolution by natural selection. We integrated specialized techniques for numerical data analysis: prediction, association and causation finding, trend analysis, etc., developed by Jeff Pressing, myself, and others. Most critically, we took a big bunch of ideas from Pei Wang, our first paid employee, whose Non-Axiomatic Reasoning System (NARS) system is my second favorite AI system. The earliest version of the system had relied on some “neural net” like algorithms, simulating the spread of electricity in the brain; now that we knew what we were doing better, we moved further and further away from the brain as a concrete design inspiration.
We divided the system into modules, each module containing agents dealing with a particular aspect of intelligence: a reason module, a language module, an evolutionary learning module, a psyche module (dealing with feelings, goals and motivations). As we saw the different modules growing and beginning to work together, we thought we were almost there. Real intelligence was just around the corner! We had an unprecedentedly detailed understanding of how the different types of mind-agents should interact. Neural-net-like agents were used to find loose associations. These loose associations were used to guide reasoning. New concepts were formed using evolution-like methods, then evaluated by reasoning. All the parts of the digital mind fit together, like pieces in a huge multidimensional puzzle. We AI folk were talking so enthusiastically, even the businesspeople in the company were starting to get excited. This AI engine that had been absorbing so much time and money, now it was about to bear fruit and burst forth upon the world!
Well… things never work out quite that nicely, not in reality. As 1999 progressed, we found the increasing complexity of the various agents in the system was stressing the codebase. After a lot of difficult debate, we decided to grit our teeth and rewrite the core of the system from scratch. And, in what was a very difficult decision in terms of internal company politics, I decided to let our Brazilian team do the redesign.
How did we end up with a Brazilian team? Well, about the same way as we ended up with our New Zealand and Australia teams, and our assorted programmers and scientists scattered among South Africa, Romania, Siberia, Georgia (the Republic, not the US state), Arizona, and so forth. I decided early on that the global brain should be seeded by a global team. Why not recruit the best people I could find, anywhere on Earth (I would have gone off Earth if we had an interstellar Internet connection)? If I had it all to do over again, I wouldn’t globalize the team quite so free-flowingly – I’d hire fewer isolated people and focus more on building small localized teams -- but at the time it seemed like a good enough idea. Certainly we got much better talent than we could have obtained in New York City – the New York job market has a lot of depth in the finance and media areas, but not in advanced computer science.
So as part of this “let’s ignore the constraints of physical space” philosophy, toward the end of summer 1998 I hired a guy named Cassio Pennachin, who responded to an online job ad I’d placed on an obscure software jobs website. Cassio was a 21 year old university student in Belo Horizonte Brazil. He replied to the ad and offered to work for free for a month to prove how good he was. It turned out he was pretty damn good. The first job I gave him was to fix up some code I’d written for evolving new mind-agents by simulating evolution by natural selection. This was the beginning of what was soon to become a company tradition: Brazilian programmers receiving American code by e-mail and responding very politely with comments like “Excuse me, but would you be terribly offended if I made a few changes to this code?” Of course, you say yes, and a few days later you receive a completely new version of the software, containing exactly three lines from your original code, but much better designed and also more efficient. Well, Cassio proved to be an outstanding manager as well as an excellent software engineer and designer, and we let him accumulate assistants until, at one point, we had 60 people there out of a total company staff of 130. When the need to rebuild the system became clear, it was only natural to throw the task down to Brazil, and during fall 1999 and early 2000, Cassio and his Brazilian software gurus worked with me and the other American old-timers to create a new AI Engine, similar to the old but better-engineered.
But all this re-engineering was spending money, month by month. We were quickly realizing what should have been obvious from the get-go – that getting our thinking machine to work could well be a multi-year pursuit. Each month we were more and more confident that we knew what we were doing. We had the right overall system architecture, we had the right set of specialized agents – we’d moved, between 1997 and 1999, from a conceptual picture and a crude software design to a really nice, thorough and professional software design embodying the same conceptual picture. And we had all these bits of working code. But none of it was bringing in any revenue, that was for sure.
In the meantime, our work on forecasting the daily financial markets continued very successfully. We were using the AI Engine for part of the market prediction (the text analysis part) and a separate piece of software for the rest (the purely numerical analysis). After some thought and some experimentation, we realized that we could build a simpler system embodying the key AI processes needed for text-based nonlinear market prediction, without the overhead (or full emergent intelligence) of the full AI Engine. And so the Webmind Market Predictor product was born.
But one thing that we learned from Andy – a realization that had been creeping up on us for a while -- was this: a better way to predict the markets is an almost impossible thing to sell. Naively, we’d been thinking that if you build a better market predictor the world will beat a path to your door. People should be begging you for the chance to trade their money using your system, and give you a percentage of the profits. Heh heh. Yeah, right. More likely, the world will beat you over the head. It turns out to be close to impossible to convince anyone that your market prediction system actually works. There are too many frauds out there, and too many really smart people who think they have the golden glowing solution to market prediction but are just fooling themselves. Even with Andy Siciliano, a bona fide Titan of the Financial Markets, on board, the sell was still difficult. One meeting after another, in which people nod their heads and look fascinated but don’t end up writing you a check. Now, through Andy’s reputation and his connections, we’ve finally got a small fund set up based on the Market Predictor technology. But it took one hell of a long time.
Given the trouble with selling Market Predictor, and our new appreciation for the magnitude of the task of finishing the AI Engine, we knew we had to move in a different business direction. We had to find something else that our technology could do well, easier to sell than Market Predictor and easier to finish than real AI.
With this in mind, in early 1999, flush with new funding from Andy and the Chicago VC’s, we began playing around with financial message boards, doing categorization of texts drawn from financial message boards using AI Engine technology. At first we figured we could let the Market Predictor pick the best quality messages out of the large mass of drivel. Just show the user the messages that the Market Predictor uses to predict the future of the markets. As it turned out, this worked pretty badly. The problem was, if there were 100 messages saying “Microsoft sucks” on a given day, this may be predictive of market movements, but the messages still don’t make interesting reading. The patterns the Market Predictor were picking up weren’t necessarily the ones that a human wanted to see.
But the problem of filtering out good messages from bad, and more generally of dividing texts into categories, seemed like an interesting one, with plenty of different business applications. We reckoned we had enough AI technologies in our bag of tricks that we could solve this problem in a completely non-Market-Predictor-ish way. Furthermore, we had on staff a couple guys from Waikato University in New Zealand – where I’d taught computer science in 1994 -- who knew a heck of a lot about automatic categorization. So we began working toward the current Webmind Classification System product, which, like the Market Predictor, doesn’t rely on the AI Engine directly, it just uses some bits and pieces taken out of the AI Engine codebase, together with some other things.
In terms of fundamental AI Engine development, the biggest thorn in our side at this point was natural language processing. We’d been trying to get the system to learn language just by extracting patterns from texts, but this didn’t work very well. So we turned to special databases created by linguists, containing explicit “rules of language.” This worked OK for some purposes, but of course, this we knew this was only a partial solution. If you’re building a real AI, “wired in” knowledge of this kind is only acceptable if it comes along with a way for the system to adapt this knowledge based on its own learning. We realized that we didn’t merely have to design a system capable of thinking, we had to design a system capable of growing from a baby mind into a mature mind, accumulating more and more knowledge and intelligence along the way. We knew we had a framework capable of supporting this, but we had to go back and build some new types of mind-agents.
Thinking in this direction, we began to focus more and more on what we came to call “experiential interactive learning.” Language learning had to be integrated with the learning of cultural patterns of cognition, and this learning had to proceed through interaction with other minds in a shared perceptual/manipulable environment. We realized more and more vividly that, even with all the different modules containing billions of agents working together and spawning intricate emergent patterns, all we were building was a baby. A very capable baby, or so we hoped, but a baby nonetheless. We began to think more and more seriously about the huge task of teaching our baby. We created a mechanism by which Baby Webmind could interact with us in a simple simulated world, in which it could participate with us in various interactions with files, directories, financial data series, and other digital objects. Along with this came a new focus on action as well as perception, something that we’d neglected in the past. We developed what we call the “schema” framework, a kind of programming language for Webmind actions, implemented in terms of mind-agents. We worked out how these little mind-programs could be learned by a combination of evolutionary programming and inference. Anton Kolonin played a huge role here; he proved to have an unparalleled intuition for this aspect of the mind. Cate Hartley, a young cognitive science graduate from Stanford, rose to prominence in the company because of her natural intuition for the Baby Webmind approach. We had enough depth of scientific and engineering talent that, whatever direction the work turned, we had the mindpower to handle it.
Equipped with schema governing its actions in its digital world, Baby Webmind could then ground its linguistic knowledge in non-linguistic social interactions, just as a human child does when learning language. When you talked to it about “moving to a new house,” it would understand this by relating it to its own experiences moving files from one directory to another, or moving data packets from one machine to another. When you talked to it about eating, it would build an analogy to what happens when it reads a text and breaks it into parts. The system would never understand human reality like a human, but it would understand human reality by analogy to its own reality. Or so we conjectured – for reasons I’ll tell you in a moment, we haven’t actually gotten this aspect of the system working yet. Maybe in a few months.
The classic litmus test for AI is the “Turing Test,” which states that a computer program is intelligent when it can flawlessly imitate a human in conversation. A well-taught Baby Webmind wouldn’t necessarily pass the Turing test, because there’s no reason to believe it would ever be able to discourse like a human about such topics as how sex feels, or the smell of the air on a spring day. But this doesn’t matter. I’d rather have a computer program that knows it’s a computer and discourses about its computer-ness intelligently, than one that can successfully pull off a pathological-liar act and fool us into thinking it’s human.
We were doing deep, difficult and fascinating conceptual work, side by side with intense nitty-gritty software engineering. But eventually, this phase of the work ended. The new ideas stopped coming. Not because our minds had burned out, but because they stopped being necessary. The urge to add new kinds of agents to the system disappeared, instead we were decreasing the number of agent types. As the year 2000 progressed, our confidence in the finality of our AI design grew significantly. The work on experiential learning and schema had been the last piece of the mind-puzzle. Before this last piece, we had designd a complete system capable of thinking – but not a system capable of growing up to an adult mind out of an infant mind. Now we had licked this problem too – at least in theory. For the first time, we could review any textbook on cognitive science or human psychology, run through every aspect of mind mentioned there, and explain in detail how we accounted for that. We had designed a complete mind system, with the diverse specialization of the human brain, as well as the creative self-organizing flexibility. And, as the end of 2000 approached, we had written nearly all of the code needed to support this. Again, we were almost there. We were basically done engineering. We were ready to start teaching the thing. Hallelujah!
At the start of 2001, we completed what we called “Webmind AI Engine Version 0.5” – for the first time, an AI Engine incorporating all the modules, working together sensibly in system running across many different machines. Millions of nodes, billions of links, dozens of types of cognitive processing. Tremendous!
Well, there were only two small problems. The first was: Hundreds of parameters, complexly interacting, making the system very difficult to tune. Reasoning had parameters determining how speculative the system’s inferences were. Neural-net-like association finding was governed by a host of different numbers, which controlled the flavor of the system’s associations. Simulated evolution involved a half dozen numbers. The interactions between different machines were quantified by numbers that had to be tuned to the right values for the system to work well. Of course, the human brain is similarly complex, and is governed by the concentrations of various chemicals, the thicknesses of different parts of the neuron, and so forth. It’s taken evolution millions of years to tune these parameters.
And second problem was: The performance of the system still wasn’t anywhere near what we wanted it to be. The thing was way too slow, hundreds of times too slow to even seriously teach it. The system needed to be drastically sped up, and its memory usage significantly reduced. This compounded the too-many-parameters problem, because if it was slow to experiment with the system, then it was impossible to run all the experiments needed to find the right sets of parameter values.
For a moment we thought we might have to rebuild the whole thing again, but as it turned out, the Brazilians had designed the system sufficiently well that this was not the case. Instead, we believe it’ll be possible to do some intensive efficiency-oriented rearchitecture without altering the basic object structure and conceptual framework of the system. We’ve conceived at a number of simple yet radical design changes that, according to our experiments, should improve the speed of the system by several orders of magnitude.
To give a feeling for how radical these changes are, consider that the Webmind system we had a month ago consisted of over 750,000 lines of Java code. One of the Brazilian engineers, inspired by conversations with myself and other tech staff, has just now re-coded all the most critical parts of the AI Engine’s thought process – what we call the “cognitive core” in about 5000 lines of C code. Of course, it’s only due to our years of research and development experience that we figured out exactly what these 5000 lines should be. Most of the 750,000 lines of Java is still useful – it covers issues like communicating with other software processes, balancing processing among different machines, reading parameters from files, and so on and so forth. Necessary infrastructure. But the C version of the cognitive core is lean and mean and fast and easy to tune and play with. It’s given me a whole new burst of optimism as to the rapidity of our ongoing development.
But what do we have to show, at this moment, for all this work on the AI Engine? Aside from a wonderful ¾ finished thinking machine, what has this R&D effort produced that’s of any practical use? Well, unfortunately, the AI Engine 0.5 is not adequate performance-wise to be used inside most kinds of products. It’s too slow, and requires too many machines. This problem may be solved over the next few months through appropriate usage of the C-language cognitive core. But, even apart from that, we’ve found a way to work around this problem – a way that the AI Engine’s intelligence can be used to enhance the performance of software products, even with the system as slow as it has been. The trick is a technique we call “rule exportation.”
One thing the AI Engine is very good at is identifying relationships between various words, and various concepts. It is able to export rules describing the relationships between various words and concepts – in general, or in a particular domain characterized by a particular set of documents. Think of the exported rules as a kind of superintelligent, optimized thesaurus. The exported rules can be used within lightweight products to produce indices for documents, and these indices can be used for applications like text categorization, search and market prediction. Not as cool as real AI, as a program that talks to you like a person -- but still, not a bad thing. The world can use better search engines, better text categorization systems, better predictions of time series. And the code is being sped up, bit by bit, as we speak. Experiments are being run, getting all the parameters set to appropriate values. Digital mind inches ever closer – not as fast as we would have liked, but methodically and definitively, coming along at its own pace, with the expected heady mixture of triumphs and disappointments.
During the right parts of 1998 and 1999 – and probably for a couple years before that, though I wasn’t involved in the scene then -- one could raise substantial funding on the strength of a strong personality and a well-articulated dream. Such an atmosphere doesn’t arise very often. As Andy likes to say, it was a once-in-a-generation period when capital flowed like water.
Well, a lot of this capital that flowed like water was invested in absolute nonsense. To take a well-worn example, how many companies selling pet food online do there need to be? And who buys their pet food online, anyway? Most of us buy our pet food at the supermarket along with all the rest of the food. To go to a special website just to buy pet food is extra hassle, not a time-saver, unless one happens to have a pet with special nutritional requirements. And what possible sense does it make for hundreds of millions of dollars to go into making life a little easier for people with finicky or unwell pets? It’s hardly a surprise, that as the Internet bubble burst, the various online pet food companies went bust or were consolidated into other online enterprises with more general scope.
Several of my friends worked for a company called MacroView, that made a website called sixdegrees.com. Inspired by the film “Six Degrees of Separation,” this personal networking website allowed you to sign up and then enter in a number of your friends as contacts. Your friends would then be e-mailed and asked to sign up. Et cetera. The idea was that any two people were statistically likely to be connected by a chain of no more than six acquaintances. (How do I connect to the Prime Minister of Japan? Maybe my dentist’s sister’s brother-in-law’s masseur’s colleague’s aunt shines his shoes for him.) Anyway, sixdegrees.com amassed a huge database of members, and regularly sent e-mails to all their members. The technical challenges involved in setting up this huge website with millions of subscribers were immense, and my friends on the tech side of sixdegrees.com had a great time and learned a lot. They built some great systems. But what good was it? How was it going to make money? Well, they put some advertising on the website, and in the e-mails they sent out. But Web advertising never lived up to its promise. My tech friends at the firm were always trying to encourage their colleagues on the business side to think about how to use the site to actually provide some useful service. But providing a useful service just wasn’t relevant, at that point in the evolution of the economy. What was important was getting more and more hits on your website. Because the Web was the future, and the more hits you had, the more of that future you were likely to own. Or so the story went….
Another friend was involved with a site called JustBalls.com. Supposedly the original proposed name for the site was MyWifeHasBalls.com. JustBalls sells balls over the internet – tennis balls, soccer balls, golf balls … any kind of balls you want, they’ve got ‘em. Who wants to buy balls online? Surely there’s not enough demand from individual ball-hungry consumers. Well, as it turned out, they created a reasonably healthy business selling balls to schools and other youth organizations. The previous distribution mechanism for balls was inefficient, involving sales representatives from ball manufacturers visiting school and organizations a couple of times a year. Coaches would run out to retail sporting goods stores in the interval, paying prices that were badly inflated compared to the discounts available from the manufacturers on the rare occasions of their visits. Well, JustBalls.com was able to offer manufacturers’ discounts year-round. There was a substantial niche market that hadn’t been identified in advance, but that emerged over time as the site got more and more users. This is a small example of what happens in the business world when money is cheap and experimentation is plentiful. A lot of things get tried, and some succeed, and some fail. The things that succeed get to grow and transform themselves into different forms, and combine with each other through mergers and acquisitions.
It’s really nothing more – or less -- than evolution. There’s a theory of evolution called “punctuated equilibrium,” which states that evolution proceeds in bursts. You have a period of rapid change, when new forms flourish. Then the variety is winnowed down by harsh reality, and you have a period of relative stasis. This may happen on the individual species level, where a species undergoes rapid change and turns into a new species over as little as tens of thousands of years – and then remains basically constant over millions of years. And it can happen on the ecosystem level. The Cambrian Explosion is a historical example. Over a period of just 40 million years, between 565 and 525 million years ago, the ancestors of all living and many extinct phyla of multicellular creatures evolve, out of simple, mostly unicellular ancestors. In evolutionary time, that’s amazingly fast. But then there was a huge extinction. 95% of the species that had been created were wiped out. Innovation still occurred, of course, but never again at such a level.
The Internet Bubble was a kind of Cambrian Explosion. All sorts of things got brought into existence. Including huge new businesses of a type never seen before – Amazon, Yahoo, E-bay. Including huge wastes of money and engineering brilliance like multiple online pet food sites and sixdegrees.com. And including all kinds of other fascinating things that never grabbed the public eye – such as my own company. How wonderful to have obtained the opportunity to hire a team of brilliant people and teach them my ideas about digital mind, and work with them for a few years to bring some of these ideas into reality.
But after the explosion comes the extinction. Now comes the point where you have to fight for survival. The kinds of short-term cash-flow problems that put a company like Webmind Inc. near bankruptcy in 98 and 99 – today are enough to kill a company. The environment shows no mercy.
After Andy came on board in mid-99, we spent a lot of time trying to sell our financial message board categorization system, which we called NewsCruncher. We had a lot of interest, but for a combination of reasons, this never quite happened. Instead we decided that financial message board categorization was too specialized an area, and developed the Webmind Classification System, a general toolkit for categorization of documents. During 2000, we gradually learned how to sell this, and accumulated a few customers, including the US office of a major Swiss bank, and some dot-coms. For instance, one customer, NetCurrents, uses the system to automatically create online reports summarizing the positive and negative messages found on the net every day about particular companies. Companies pay them for access to these reports. We also spent a fair bit of time figuring out how to use the AI-engine-exported rules to enhance search performance. We built what we believe is the world’s best search engine. We developed some other simple products, including a nice little thing called the Recommendation System that allows users of e-commerce websites to submit fuzzy requests (“I want a computer that has a very big monitor, is pretty cheap, and has a moderately fast processor.”).
The pieces seemed to be coming together. The AI Engine project was rolling along – proceeding at a breakneck pace from the AI engineers’ point of view, but at a terrifyingly sluggish pace from a business point of view. The marketing of the products seemed to be picking up steam, and we figured it would be possible to build up a solid customer base, who would then be the first round of beta users for the “real AI” technology when it finally came out of the lab. But just as everything was starting to coalesce, the market went through a series of horrifying collapses. Boom, boom, boom. Before you knew it, Yahoo was trading at a tenth of its historical maximum. Disney was folding its website, Go.com. The carnage was getting worse every day. Most of the dot-com companies who had expressed an interest in buying our stuff, were no longer in a position to buy anything – they were laying off their employees. And most of the big companies we were talking to were stalling their purchasing decisions until the economic outlook became clearer. Tech spending was going way, way down, and a lot of tech companies were going down with it.
We’d been paying the bills for some time with money from Andy and his various friends and associates, including a major Asian financier. We always knew that eventually this source of cash would run dry. But we weren’t much worried about this, because in the summer of 2000, we’d secured a deal with an overseas media company. They wanted to be a strategic investor – put in a bunch of money, integrate our software with their own IT systems, and help us market our software in their country. It was a beautiful story, and we felt awfully clever for having secured funds from overseas while our own country’s economy was in chaos. We were also pursing funds from other sources, including some plain vanilla US VC’s, but in retrospect, perhaps not quite as avidly as we should have. Because suddenly, at the last minute, after 8 months of tedious, painstaking and frustrating negotiations about trivial points of the deal – negotiations that seemed impossible to speed up in spite of our best efforts – the media company that had promised us funding pulled out. They were no longer interested. Period. We offered to concede on every possible negotiating point, and even to substantially change our business approach to suit them. It didn’t matter.
Exactly what went wrong with that deal will never be known, but ultimately, I have to think that they were reacting to the poor condition of the US economy for tech companies as a whole. We were having trouble selling things because our target markets had largely stopped buying things. We had ways to get around this problem, adjustments to our sales approach that I think would have worked – and may still work – but the problem was definitely real.
Some people said we were dead right then. I suppose they were right. But pigheaded as I am, I didn’t want to admit it. Instead, we lined up more money, from Andy’s familiar cast of friends. We hired a bunch of crackerjack salespeople from a company that had recently gone bankrupt. These new sales guys really knew how to sell software. Everyone was hyped. It looked like we were going to scrape by, just barely.
But then a major part of the new investment round fell through, because a minor scandal had broken out in the news, regarding one of the investors. The media had saved us from bankruptcy in the early days, in the form of Tom Petzinger’s Wall Street Journal article. And now the media had killed us. We wanted to go on – we knew we had wonderful technology that was ¾ complete; we knew we had good products that would provide real customer value; we knew we had an amazing international engineering team, with few counterparts in any company or institution. We were full of optimism and good ideas for making the business work better. But there were large debts and small revenues. And the bigger the debts got, the harder it got to bring in new investment money. Who wants to invest in a company that’s going to use a substantial percentage of their investment money to pay off debts?
Finally, we were there, exactly where we didn’t want to be: bankruptcy. We’d been teetering on the brink for quite some time – on and off and off and on again. Always, when things seemed hopeless, something would come up to save us, and we’d keep on rolling. But when the market got too bad, all our would-be saviors found themselves either wracked with skepticism or bogged down with their own personal financial difficulties.
We’d been laying people off for a month or so. It was a particularly painful process because so many of the staff were personal friends of mine, or of the other founders, or of the early key hires. We – me and the other technical executives -- would spend two days agonizing over which of our friends to lay off. Then we’d deliver the news, sometimes losing or irreparably harming friendships in the process. Depressed people stalked around the office. People had to fire their roommates. Copious quantities of alcohol were consumed. One depressed, laid-off roommate of a tech exec wound up in jail, on the charge (false, he says) of soliciting prostitution. Despite the tense atmosphere, no one wanted to leave the company, which was just about the best workplace any of the staff had ever seen. The people who were laid off were dejected, and the ones who remained were guilty. And then, a week or two later, we’d need to lay off more people. The same process all over again. Each time, hours and hours are spent debating the merits of individual people who are “on the brink” of being laid off. And then in the next round of layoffs, all those people who were previously on the brink are gone anyway, so all the agonizing turns out to have been worthless in retrospect. During this period, I began to truly miss my academic job. It was hard to focus on technical work, especially in cases where you knew some of the people you were talking to were going to be laid off soon, but you weren’t supposed to tell them yet. “Don’t worry,” I’d want to say. “You don’t need to know the answer to that question about the software code, because tomorrow your butt is getting kicked straight outta here.”
It had been wonderful to be the big shot, to hire all my friends, to work with them together on amazing projects. Now the big shot felt like being shot. And was paranoidly wondering if he was perhaps at risk of being shot by one of his previous friends or colleagues .. for instance, one of the many people who had been brought over on H1-B visas from overseas, and was now at serious risk of being deported, sent back to Novisibirsk or Tbilisi or wherever…. The office computers were being sold off at low prices. Everyone’s personal possessions had to go home. The office would be re-rented to some other company, presumably one that was actually making money. Definitely not one that was making a thinking machine.
Perhaps the saddest part of it all was the Brazilian office. The New York staff basically knew what was going on. They saw the business execs furtively meeting in secret, coming out of the conference room with sad and sour looks on their faced. The Brazilian staff, on the other hand, were happily coding away, building the thinking machine and the search engine, testing various parts of the AI Engine codebase to see if they’d improve the categorization product – working long hours, thinking hard and enjoying it. The tenuous nature of the funding situation somehow hadn’t worked its way down to Brazil, except in vague and diluted form. And then came a shocking e-mail from New York – a couple short paragraphs saying basically “Sorry, but the day we’ve all been trying to avoid has arrived. This will be everyone’s last day of work at Webmind.”
So what did we do wrong? Frankly, all kinds of things. So many things. It’s basically impossible to do anything interesting without making a lot of mistakes. And of course, the assessment of exactly what the big mistakes were is a highly subjective matter. My own take on it is not going to be the same as anyone else’s, even though many of us were present for the same events.
There was definitely a lot of bad luck involved. If that one key financial backer hadn’t pulled out at the last minute, the company would be alive and well today – we’d be busy restructuring ourselves and reorienting ourselves to ensure our survival in the new market conditions. Of course, that one chunk of cash wouldn’t have assured our long-term success. But it would have given us the “run room” to launch ourselves into the air, as the saying goes. We would have had a fighting chance.
But it’s a cop-out to blame it all on luck. Louis Pasteur said, “Chance favors the prepared.” In the midst of the bankruptcy chaos, I coined a corollary: “Chance f**ks the unprepared.” We were hit by bad events, but we were unprepared for them. We were not tough enough to survive the extinction phase of the evolutionary cycle.
It would be easy to blame the company’s failure on my and my AI team’s failure to create a truly thinking machine as quickly as we planned to. No doubt, if we’d created a program capable of holding a human-level intelligent conversation after a year of work, in early 1999, the business would be pretty successful right now. At very least, even if we’d botched the productization of the thinking machine, we could have sold the company to a major software firm for a tidy sum.
But of course, there wasn’t really any way to speed up the progress of the thinking machine development. In hindsight -- of course, knowing what we know now, if we had to do it all over again, we could do it pretty bloody fast. But learning is what R&D is all about. Many other teams have worked on AI for many years, and, in my view, they’ve all run into dead ends. Our willingness to re-orient and redirect ourselves when we ran into obstacles may have sometimes been bad from a business perspective, because it meant that we didn’t quickly enough produce a sufficient number of productizable components. On the other hand, I don’t believe it was very frequently bad from an AI R&D perspective. There were many dead ends on the path to where we are now – dead ends that we ran partway into, and then reversed course and ran out. If we do get to the end goal, and I believe we will, it will be because we were willing to back out of dead ends, to continually take stock of our current status and progress and analyze whether we had to undo or redo some things to proceed optimally toward the end goal.
The bottom line is, creative research is inherently inefficient – even if you have a basic plan for what you’re doing, and you know what sorts of things you have to try, there still has to be room for a lot of experimentation. You can’t run creative research like a business – in a business you need to be sure that every single action anyone takes is contributing to the bottom line. But in creative research you can hardly ever be sure about anything. You’re always playing, testing, experimenting, fiddling around. We definitely should have constructed a better process for extracting productizable components from the AI codebase, which even now has dozens of capabilities that aren’t accessed at all in our current products. But even a perfect process of this sort won’t eliminate the necessity for iterative and parallel experimentation in creative research.
In retrospect, it would have been really nice if we could have gotten start-up funding for, say, 5 years. Companies building new kinds of hardware can get this kind of set-up sometimes. It just takes years to design all the details of a new chip, for example, even beginning from a well-understood and solid overall chip design. Chip companies are funded to go away and do R&D for a few years, and then come back with a chip to be sold. If I could go back in time to 1998, knowing what I know now, I could probably get 5 years worth of funding from somewhere. It was, after all, the peak of the Internet bubble. As it actually happened though, knowing as little as I did about business back then, I was lucky to get the small amount of seed funding I did.
The slowness of the path to real AI, however, didn’t have to spell the death of the company. An example that shows this is an effort somewhat similar to mine, is the Cyc project, currently running within a company called Cycorp. This project has been going on since 1984, and began with the goal of creating real AI. Their approach was different than mine: rather than focusing on integrating different aspects of intelligence within a self-organizing agent system, they’ve focused on the encoding of knowledge. The goal is to get everything a ten-year-old kid knows encoded in a formal mathematical language, so a computer program can understand it. They’ve been at it an awfully long time, and don’t have a heck of a lot to show for it. Not a thinking computer program, not even any popular commercial products. Just a very big database of knowledge. Some obscure products using their knowledge database exist now, and others will be coming out soon. I’m not sure exactly how they’ve remained funded all these years, to tell you the truth.
One thing that hurt us pretty badly was our failure to come up with a sufficiently sexy spin-off product, leveraging some of the abilities of the AI Engine codebase in a widely sellable, easily marketable way. There’s so much wonderful, intelligent code there, and we on the AI side never did a very good job of making it clear to the businessfolks exactly what this code could do. The Market Predictor was hard to sell due to its market niche, and the Classification System works far better than any other product in its category – giving 98% accuracy where other products give 40-60% -- but apparently this markedly superior performance wasn’t enough to close us a lot of deals. Probably the biggest lesson I’ll take from this is “it’s the marketing, stupid.” There are companies making a lot of money from technology that searches and categorizes documents, that works remarkably poorly. There are other companies making far better technology that does the same thing – not quite as good as ours, but far getter than the best-selling stuff – and they’re foundering. The global brain, the path to Real AI, all this stuff was potentially a marketing person’s dream. But this glorious story had to be synergized appropriately with the actual products we were selling. We needed some kick-ass visionary marketers, and we should have hired them even if it meant getting rid of a few of my super-brilliant mad scientist software engineers.
Another thing that hurt us was our sales approach, with the simple products we’d created. It took us a while to figure out that what we needed was a “solutions approach.” For a long while, we’d sell the Classification System to a customer, and then our engineers and scientists would spend weeks or months playing around with the customer’s data, helping them figure out how best to use the product. Things that seemed obvious to us were not so obvious to them. Finally we realized that we had to get paid for this work – that we could derive just as much revenue from helping the customer understand his business problems and how our cool technology addressed them, and helping him integrate our technology with his other IT systems, than from licensing fees on the technology itself. But we didn’t learn this quite fast enough to bring in substantial revenue from the approach, before the end came.
But of course, marketing and sales themselves aren’t the be-all and end-all solution. Plenty of Internet start-ups poured all their money into marketing, getting $50 million in funding and spending $20 million of it on a single SuperBowl ad, and so forth. They had good salespeople too, but they didn’t have anything of value to sell! The opposite problem we had.
The ultimate answer is that a company is an organism. All the parts have to work together well, and then the whole organism has to perform effectively in the environment. Which currently is a very complex, rapidly shifting environment. We needed better processes for the interaction of AI development and product development. There’s a hell of a lot of useful stuff in the AI codebase, which remains untapped from a practical point of view. And we needed better interaction between product development and marketing, so that our understanding of customer needs would feed back into our product development processes, and ultimately into our AI development. All the ingredients were there – and, frankly, after a lot of experiences, we had figured out how to make them all work together. With our new hires in the last few months, we were finally a whole company, a viable organism, with good salespeople, good people to manage product development, an AI group that understood the balance between long-term “real AI” work and short-term product-focused AI work. But we made ourselves whole and healthy too late. When our funding source pulled out at the last minute, in the middle of a terrifyingly bad market, there simply wasn’t time to bring any of the other interested funding sources up to the point of actually writing a check. This is normally a 90 day plus process, but we needed to pay people tomorrow. In the end, we discovered that in business as in music, timing is everything.
It’s depressing as hell to see one’s company go bankrupt. The only comparable thing I’ve been through personally was a near-divorce. Bankruptcy in a close-knit company like Webmind is sort of like a divorce involving a whole lot of people. But even given all the torment, if I take a step back, I feel lucky. Awfully lucky. The out-of-control bull market gave me the opportunity to create a womb for AI R&D. Unfortunately the womb didn’t last quite long enough, and the baby has been cast out of the womb prematurely. But it’s not so premature it can’t live. It’s just sufficiently premature that it’s going to have to struggle to survive – every breath of air will be hard work, until it gets a little bigger.
I’ve noticed a rather extreme difference in reaction to the current situation, between people closely involved in engineering the AI Engine and others. Some folks not too close to the AI Engine are rather annoyed about all the time, money and effort spent trying to build real AI. But the people working on the AI Engine have a different feeling. “It’s ridiculous that a project like this should be killed at this point,” as one of them said. They can feel the success, the momentum, the power in the work that’s been done. They know there’s a lot of work that remains, but they understand the nature of the problems to come, they have a sense for how to solve them.
Imagine you had a team of 45 engineers building a rocket. The world’s first rocket. No one has ever been into space before. This team of brilliant engineers built all the components of the rocket. They rebuilt some of them. Then they started piecing the components together, occasionally modifying one of the components to make it fit better than the others. They built the rocket from the bottom up, until they were about to put on the nosecone. They realized at the same time that they’d have to rework some of the ignition system to avoid explosions, but this seems like no big deal. They’ve been trying to make money by mass-producing some of the cleverer components they created while building the rocket, and some other related inventions some of the engineers have come up with. But this hasn’t worked quite as well as they’ve hoped. And then, one day, they run out of money.
“Sorry,” the money people say. “You’ve spent long enough on this thing. We’ve lost interest in rocketry. Why don’t you build refrigerators instead?”
“But wait,” say the scientists and engineers. “Can’t you see, we just have to put on the nosecone?”
“And then what? Then you’ll have to rebuild something else.”
“Maybe. But not much. The thing is almost done.”
“It doesn’t fly.”
“No, of course it doesn’t fly. It isn’t finished yet. But almost….”
“Almost. Heh. Well, then.”
“And we have a lot of great products, formed out of the components of the rocket. We have a lot of people interested in buying them.”
“Well, how come you haven’t sold very many of them yet, then.”
“We just recently learned the right way to sell them. They’re new sorts of things, and we found it wasn’t enough just to sell them to people, you have to work closely with the customers to help them understand how to use them.”
“What? Well, I have a meeting in half an hour. I’ll call you later…..”
So what next? Well, companies are, in a way, more complicated than individual living organisms. If a human dies, its arms and legs can’t get up and walk away and find new sources of sustenance. On the other hand, if a company dies, it is quite possible for its parts to live on – to find ways to sustain themselves (thus becoming whole organisms), or to graft onto other organisms that are better able to support them.
The New Zealand office is striking out on its own, aiming to rebuild their own version of the Classification System and sell it themselves. The Market Predictor hedge fund is its own animal too, with its own set of funders. Some of the business staff are talking about starting a software solutions company, focusing on integrating AI software with customers’ IT systems. A few individual staff have already run off to start their own consulting or software firms.
One of the tech staff is going back home to Bulgaria to sit in his apartment by himself and build his own thinking machine – he says it will take him three months. In his spare time, maybe he’ll finish the mathematical proof of the existence of God he’s been working on since I met him. Another guy is going to build his own thinking machine while wandering around Europe – though he reckons it will take him years rather than months. They both find the Webmind AI Engine too complicated and want to try something simpler. I wish them luck. I’m working to make the AI Engine simpler too, but with Einstein’s dictum “As simple as possible, but no simpler” in mind. The brain is not a simple thing either.
And the Webmind AI Engine project? It’s become basically a Brazilian pursuit. A handful of Americans and a few dozen Brazilians are staying on the project, in spite of the absence of financial compensation. Some will work for free full time and live off their spouses or their savings; others will work half-time on the AI Engine and half-time at paying jobs; others will work full-time at paying jobs and work on the AI Engine evenings and weekends. I’m seriously considering moving down to Brazil, where all the action is at this point – I happen to be a Brazilian citizen, since I was born there to American parents, although my family moved back to the US when I was one and I’ve only been back for brief visits since.
I reckon that, at this point, I’m at serious risk of becoming the Charles Babbage of AI. Babbage designed the first computer – a purely mechanical computer, pre-electronics. But it was just too damn complicated to build using the technology he had at his disposal. He spent all his money and his life on it, and never got it done. Most of his contemporaries died believing his design was a complete bunch of hooey.
But my relegation to Babbagedom is by no means inevitable. We may be out of money, but we’re not out of enthusiasm, or ideas, or drive. The unity of the company as a whole is gone, but in truth the company was somewhat fragmented anyway. What’s happening now is that each group within the company that had its own “natural unity,” is going its own way and trying to become its own company. Time will tell how many of these attempted spin-offs succeed. Evolution’s work is never done.
In the end, there’s a very big difference between a company like Webmind, and the typical bankrupt Internet company. When sixdegrees.com went out of business, there was nothing of value left, really, except a lot of memories on the part of the company staff and the loyal site users. On the other hand, Webmind the company left behind the AI Engine, the Market Predictor, the Classification System, the Recommendation System, and so forth. In this case the beautiful excess of the Internet Bubble created a womb within which wonderful technologies were able to come to life. Now these technologies have emerged into the world, a little bit young and unformed, but clever and adaptable, and growing fast. It’s easy to regret the things we did wrong, that led us to be unprepared for the cruel chances fate dealt us. But it’s hard to regret the overall experience of Webmind Inc. A great group of people gathered together for a while … some great technology engineered … and just possibly, the first ¾ of the first real AI, of the seed of the global brain, created.
Maybe a visionary new investor will drop out of the sky and bestow upon us the money to pay our staff to finish the AI Engine. We’ll be cleverer about marketing and product development this time around, and become fabulously wealthy as well as scientifically successful. Or maybe no such investor will appear, and we’ll finish this digital brain-beast in our spare time, and unleash it upon the Net in a non-commercial way, like a Napster with a mind. Who knows?
Perhaps the most important lesson I’ve learned about bankruptcy is that it’s not necessarily the end. Bankruptcy is a legal category, a way of dealing with debtors, a formal scheme for the manipulation of money. Of course, it’s an important thing – we all need money. I have a wife and three kids I have to feed (and even I need to eat occasionally). But what’s ultimately most important is not money but the intersection of ideas and people who grasp them. This is what always really changes the world – for the better or for the worse. Money may come and money may go, but a team of people with a common idea can do more tremendous things than any amount of cash. And, except in peculiarly adverse socioeconomic conditions, great teams clustered around ideas tend to eventually attract the cash to support themselves, one way or another. Great companies are great because they gather a great group of people around a common idea. Webmind Inc. had its own peculiar kind of greatness, for a while. The ideas generated by the company will live on, and it looks like many of the teams formed within the company will live on too, each one associated with a certain cluster of ideas.
At least, that’s the best consolation I can offer myself, as I sit here at home, spending more time with my kids than I have in months, helping to organize the die-hard engineers still remaining on the project unpaid, making endless phone calls begging for funding, and trying hard to adjust to my new role as yet another unemployed ex-tech-company-founder….