The first investment bubble in “AI” happened in the 1980s. As I mentioned before, one of the things which kicked it off was Japanese investment in the fifth generation computing project. Go look at Blade Runner for an idea of how people thought of Japan back then: everyone figured they were the country of the future. It was probably true in a sense, but nobody was looking at demographic collapse back then: they just figured the Japanese were so successful they’d continue their successes, making AI brain in a can computard as easily as they dominated cars, semiconductors and everything else, ultimately owning everything. We even had dipshits back in the 1980s talking about how AI was going to end th…
The first investment bubble in “AI” happened in the 1980s. As I mentioned before, one of the things which kicked it off was Japanese investment in the fifth generation computing project. Go look at Blade Runner for an idea of how people thought of Japan back then: everyone figured they were the country of the future. It was probably true in a sense, but nobody was looking at demographic collapse back then: they just figured the Japanese were so successful they’d continue their successes, making AI brain in a can computard as easily as they dominated cars, semiconductors and everything else, ultimately owning everything. We even had dipshits back in the 1980s talking about how AI was going to end the job market back then, and bring about some kind of singularity where everything would be different and white collar workers needed to look for something else to do.
just like 1984
Back then, the special sauce was mostly expert systems shells, as opposed to LLMs. Same basic idea drove the hype; both technologies worked with a sort of verbal user interface. Expert systems shells were considerably less automated, but still give nice verbal answers which make you think like you’re talking to a canned brain. With an expert systems shell, it could tell you stuff like why it came to certain conclusions: they were more deterministic. They were of course trained on much more limited data sets, which were at least relevant to problems people in businesses had. There was other stuff going on too though; from autonomous vehicles to voice recognition. There were also a few projects designed around neural systems. But it was mostly the expert system wordcel which drove the craze; same reason for the current hype, people anthropomorphize and like talking to brain in a can, even when it’s a glorified if statement.
Companies like MCC, Intellicorp, IIM, Syntelligence, Verbex, GO Corporation, Teknowledge, Alvey, Gold Hill Computers (remember Golden Common Lisp? Me neither), Envos, Applied Expert Systems, Kendall Square Research, BBN, Brattle Research, AICorp, Neuron Data, Carnegie Group, Prologia, Votan, Nestor, Mind Path, Palladian, Logicware, Airus, IntelliGenetics, Aion, Apex, Silogic, Lucid, Symbolics, Cognitive Systems, Thinking Machines, LMI, Inference Corporation: most of them don’t even have wiki pages and may as well have never existed other than employing notable historical personalities, and appearing in the history books on the era. Amusingly if you search on some of the names, such as Thinking Machines or Verbex (there are more), even the goddamned names have been recycled for the current year bubble. These companies had thousands of employees; tens of thousands total in a time when software engineering was a much smaller field.
The hype of the era was as insane as it is now. Celebrity scientists like Feynman got jobs at these Potemkin companies. Ed Feigenbaum kind of invented the expert systems shell, and played a role somewhat like current year Yann LeCun. Mountebank CEOs of these companies were treated like movie stars; one guy flying around in his personal B-25 bomber, cavorting with models and so on to the adoration of the media. Everyone was on the bandwagon or had a plan to get on the AI bandwagon. There are even funny stories about google-like benefits for the engineers in these places. Catered food (fromLegal Seafoods no less), massage therapists, trips to Disneyworld, foosball, stock options, limousines (limos were a big deal in the 80s); all the excesses of the first dot com eras happened back in the 1980s at these companies and probably originated with them (assuming it wasn’t taken from Wall Street). The business press hype about huge investments in “AI” was present. Also the same the lack of curiosity about actual use cases which generate economic benefits for the customers of these tools. Amusingly many of these companies had confidentiality agreements with their customers which didn’t allow any comment on the efficacy or utility of their projects.
Special purpose hardware was also created in this bubble; Lisp chips were made by TI (the NVIDIA of its time), Xerox and Symbolics among others. Apple, largely considered a has-been company being run by Soda Pop guy, got considerable nice PR by shipping a Mac-II with a TI Lisp chip in it: such innovation. NPC Soda Pop CEO = do what everyone else is doing. There were pilot projects galore: everyone from Campbell Soup to Arthur Little to Travellers Insurance to General Motors invested in various initiatives designed to automate away important jobs using the magic talking “brain in a can.”
Just like today, there were plenty of large established businesses going long this stuff. Texas Instruments, Xerox, Fujitsu, Toshiba: they all ultimately did OK because they had positive cash flows from other lines of business. Though you could make the argument that TI for example would have been better off developing RISC chips or even 32 bit processors like the 68000 rather than dead-end lisp machines. All of these companies promised the same nebulous crap as current year AI hype babies: brain in a can. Using expert system shells rather than neural nets. Basically because everyone else was doing it.
There are essentially only two survivors of the first AI bubble: Wolfram and Maplesoft. They wrote computer algebra systems which solved a real, albeit small problem and were founded towards the end of the run. Both companies used expert systems shells for solving mathematical problems. Essentially they each built a programming language which had internal expert system shells for solving differential equations, integrals and intricate mathematical relations. Neither company to my knowledge ever claimed to have a General AI solution: just a nice tool for solving mathematical problems, saving students and scientists a lot of paperwork attempting to solve their math homework. At the time, there were large classes of problem they couldn’t solve (Greens functions for example: stuff involving the Calculus of Residues basically), but I would imagine they’re pretty decent at this by now. There are a couple of other less hyped companies who survived more or less by continuing to service DARPA contracts; I assume Cyc fits the description, BBN certainly does.
There are differences between the bubble of the past and the bubble of the now. Back then, the computer industry was much smaller as a fraction of the economy. This bubble was also greatly subsidized by government R&D projects in the defense spending extravaganza under the Reagan Administration. There were more things to invest in back then. Lots of stuff in tech which wasn’t AI was worth dumping money into: databases, microcomputers, microchips, operating systems, disk drives, middleware, CAD, text processing, accounting software, video games, networking technology, financial data software, anti-virus software. There was also plenty of novelty in the industrial economy which looked like a good bet: cable TV, budget airlines, consumer electronics, financial and agricultural conglomerates, oil infrastructure, banks, payment cards, drug companies. Since technological progress has slowed since the 1980s, there are fewer likely lads out there to put your money to work in. Now a days the alternatives are things like online gambling or raising the rent on poor people. Worse though, at present there is a lot of capital looking to be put to work. The structural problems with the boomer retirement, offshoring, ridiculous deficit spending and outsourcing require the money the US government keeps printing to keep the circus turning over must go some place. Finally, it was a lot easier to float an IPO in the 1980s than it is now: quite a few of these companies had public listings. IPOs are healthy: they make companies do stuff like earn revenues and file honest financial reports. The present reverse SPAC situation only enriches underwriters while deferring responsibility. Public companies going bankrupt is good for dispelling woo.
If we use our trusty AR(1) model for figuring out how the present bubble might shake out, the outcome is pretty straightforward. Once things start exploding (OpenAI is most obviously over the cliff) most of this horse shit will go away. Established companies which went long this crap will suffer, but persist. The actual hyped technology, LLMs, will find a couple of niche uses helping schoolkids do their homework (expert systems shells are still used beyond stuff like Maple; insurance companies use them to understand their own policies for example). Also probably as a front end to a search engine, the way Brave does it, by including source material. Pieces of the technology that fed the bubble may also be useful: Sun Workstations and database engines came out of the AI bubble as a sort of byproduct. Maybe people will find GPU chips, giant databases that feed the LLM and vector databases economically useful for something else. Also the other technologies developed along with LLM horse shit will find their use cases: lots of cool machine learning stuff from the 80s is pretty normal now: Kernel Regression, Hidden Markov, 3-layer neural nets with SGD: that sort of thing. Maybe emacs will sprout something as useful as ^X doctor which uses neural nets.
Fun source material: https://dspace.mit.edu/bitstream/handle/1721.1/80558/43557450-MIT.pdf;sequence=2