Could a race with 20th century computer technology plausibly develop general-purpose AI?

Question

I have a race of aliens whose technology roughly corresponds to humanity’s in the early 1970s. Wireless communications via microwaves have been developed, but are only available to important groups such as law enforcement; everyone else has to make do with electric telephones. They do not use fossil fuels, (because duh) and instead use water-power to generate electricity, (although some households have solar panels to use during power outages). They produce biodegradable plastics from distilled plant oils, and have a good knowledge of chemistry and a growing knowledge of physics thanks to particle accelerators. Televisions have been around for a while, and personal computers have just arrived on the scene for office work and even some household gaming, (if you can afford them). These aliens have just made their first crewed mission to another world, (albeit their rather inferior and close-orbiting moon) and are looking forward to a bright interplanetary future.

So, naturally, I want to spoil all that by having them suddenly invent artificial general intelligence, capable of learning to perform any task that these beings can; resulting in widespread unemployment and all kinds of other bad stuffs, to delay their expansion by another three decades or so.

But first, reality check here: is it possible that a race with the 1970s-level technology described above could plausibly invent such a powerful AI, given that we in the age of youtube and the humorous panda video have yet to do this?

Answer 1

I'm not going to retread the ground of the other answers, which accurately characterize how pitiful 80s computing technology truly was, and how monumental a task "true" artificial intelligence is.

However, there is a way that you can still make the premise of your story work, nearly as written.

I want to spoil all that by having them suddenly invent artificial general intelligence, capable of learning to perform any task that these beings can; resulting in widespread unemployment and all kinds of other bad stuffs

If your requirement is for a computer-powered entity that puts workers out of work, you don't need to take it out of human hands. You just need to combine cybernetics (no, the other kind) with some historical projects that didn't get their fair shake OTL.

Bureaucracies as artificial intelligences

By itself, even a general-purpose AI does not make much of an impact on anything. It needs to be embedded in a system that allows it to transform inputs into decisions and actions. Every "AI" is really a cybernetic system, with humans servicing a complex algorithm to one degree or another.

And software doesn't have to run on silicon. Any bureaucracy in the information age is different from an AI only by matter of degree:

Markets, bureaucracies, and machines are inventions designed to process information at speeds, in quantities, and with accuracies that surpass human capabilities. In all three systems this information processing is made possible by reducing reality to narrow inputs (e.g., bits, prices, entries on bureaucratic forms) and then detecting patterns and pattern conformance from these inputs. Recognizing this commonality, this paper treats these systems as members of a set of artificial intelligences and uses the experience with markets and bureaucracies to suggest descriptive, predictive, and prescriptive insights about machine intelligence.

But could you make an interesting enough computational bureaucracy to make this a story about AI and not a story about hedge funds?

Yes. In fact, it almost happened!

Socialist Computer Science

In 1971, Chilean socialists had a very cool idea - using a big ol' computer to do the central management of their economy. During the system's limited time in existence, it was extremely potent:

the system's telex machines helped organize the transport of resources into the city with only about 200 trucks driven by strike-breakers, lessening the potential damage caused by the 40,000 striking truck drivers.

All we need is a sprinkle of plausibility to make Cybersyn into an even more outlandish tool from the realm of science fiction. For that, we need to cross the Pacific to borrow from another socialist project: Setun. Unlike the primitive binary computers of the decadent West, this was a ternary computer, which is 1 better. While in our timeline, this research didn't really get much traction, we have enough to propose an alternative timeline.

The Realms of Plausibility

Alongside advanced ternary computing, your aliens have developed management cybernetics beyond our 21st century's applications. Their version of Cybersyn was wildly successful - while no individual employee or computer within the system was too far beyond what 80s computing of our timeline could do, the system itself is configured in such a way that it is able to bring in information and issue commands orders of magnitude faster than human managers, economists, and planners.

Even to the operators themselves, this system might as well be an AGI black box. Its instructions would appear inscrutable, because no individual user can see what it sees. It would eliminate inefficiencies overnight, throwing the world into chaos as jobs vanish and entire companies and industries are made redundant.

Advanced computing throws the world into disorder and panic, just like you wanted. The power of labor and capital have now both been broken by computational logic, and it's going to take a long period of adjustment to put the world back together. It just has a bigger human element than you might have expected.

Answer 2

Impossible

I'm only guessing, but I suspect you were born after 1980. Let's look at some highlights from 1979 (Source).

• CompuServe became the first commercial online service offering a dial-up connection to anyone on September 24, 1979.

• The first commercial version of SQL (Structured Query Language) was introduced in 1979 by Oracle.

• Atari introduced a coin-operated version of Asteroids in 1979.

• Usenet was first started in 1979.

• Robert Williams of Michigan became the first human to be killed by a robot at the Ford Motors company on January 25, 1979. Resulting in a 10 million dollar lawsuit.

• Bjarne Stroustrup, a Danish computer scientist, begins work on the programming language "C with classes" that will later be renamed C++.

• By 1979, the TRS-80 offered users the largest selection of software available for a consumer microcomputer system. (If you have never used a TRS-80, you need to hunt down an antique and use it to really understand the overwhelming nature of this one historical fact.)

• The Intel 8088 was released on June 1, 1979.

• Hayes markets its first modem that becomes the industry standard for modems.

• The Motorola 68000, a 16/32-bit processor is released and is later chosen as the processor for the Apple Macintosh and Amiga computers.

• IBM introduces the first disk drive to feature thin-film inductive heads and an RLL (run-length limited) coding scheme (IBM 3370).

• IBM announces the 4300 processor, featuring multilayer ceramic packaging and 64 Kb memory chips for the densest packaging of memory and logic circuits in intermediate-sized IBM systems. (Compared to commercial-grade CPUs in the 2020s, what IBM did in 1979 was invent a digital version of the abacus.)

First of all, you didn't clearly define what you mean by "Artificial Intelligence." That's a problem because right now "AI" is following the trend of identifying products as "green" or having the letter "X" in their names. It's being slapped on anything and everything whether or not "artificial intelligence" is involved at all. It's the latest "the world is coming to an end!" hot potato that everyone's talking about but nobody actually understands.

Ignoring fiction (and that's mandatory to answer your question), the beginnings of "artificial intelligence" stretch clear back to 1940. But you need to understand what that means. Scientists were thinking about models of intelligent, self-learning activity that would let them physically create it. (Source)

And here it is, 2023, and we have not reached that goal.

We're getting there, but there isn't "artificial intelligence" on the planet yet. Programs like ChatGPT are little more than very efficient data aggregators. They "learn" from the perspective of improving the efficiency of their core goals — but they're not thinking for themselves in any sense of the concept. (I'd give my left big toe to see ChatGPT truly embarrassed by the quality of what it spits out.)

So, what's a "powerful AI?"

I'm going to assume you're thinking of the science-fiction version of "powerful AI" and not the real-life version of "powerful AI" like automation with feedback analysis or a ChatGPT database analyzer. If that's what we're looking for, you can't have in the 1970s what we don't have in the 2020s.

If you scale that back to, for example, automation with feedback analysis, then please note the death of the first human by a robot at Ford. You already had that kind of "artificial intelligence" in the 1970s.

I'm assuming you want the former more than the later, so I continue to assert that it's impossible.

What are you lacking?

1. A massive world-spanning database containing vast amounts of human knowledge. Remember, the first commercial version of SQL was introduced by Oracle in 1979. It could not do what databases do today.

2. Multi-core computational arrays operating with gigahertz clock frequencies. (In the 1970s the max clock frequency was +/- 16 Mhz and the backplane was a LOT slower than that.)

3. Dense high-volume, very fast RAM.

4. Exabytes of storage — possibly Zettabytes of storage — worldwide.

Keep in mind...

• Laserdiscs were introduced to the market in 1978.
• Seagate introduced the first hard drive for home computers in 1980. It was 5 megabytes.
• The first terrabyte hard drive didn't hit the market until 2009. (Source)

5. Finally, a world-spanning reasonably fast network (separate from the database of #1). The Internet as you recognize it fundamentally didn't exist prior to 1983 when TCP/IP was introduced. There was no world-spanning Internet in the 1970s. There were modem-accessed bulletin board systems¹ (BBS, I actually miss some of those early BBS sites), but that's like claiming you have a Ferrari when what you really have is a 2-cycle lawn mower go-kart. And that might have been a very generous comparison.

Thanks for the walk down memory lane, but does it matter?

Having said all that, none of it matters! Authors today are really stuck on the idea that things must be entirely "realistic." Baloney! Just because you can't produce the schematics for the computer that operates your AI or the software that instantiates it using 1970s tech doesn't mean you can't have it.

The idea of machines that could intelligently take over the world began with E.M. Forster's story "The Machine Stops" in 1909. They by no means had the technology to produce the omnipotent machine of the story (we still don't over a century later) — but the story is an absolute sci-fi classic.

So, after having told you that technologically it's impossible, I want you to confidently present with absolute assurance that you can succeed a magnificent vulgar hand gesture and go write that story.

If I wanted fiction based on realism, I'd read the biographies of politicians.

But I do recommend that you take the time to specifically define what you mean by "artificial intelligence." Rationalizing the tech in your world would be a lot easier if we knew exactly what you wanted to achieve. And remember that there are ways to creatively solve the problem without depending on "Real Life" technologies. Go watch the Star Trek TOS episode "Spock's Brain."

---

_{¹ To be fair, the ARPANET existed in the 1970s. It connected whole dozens of computers at different locations in the U.S. using, primarily, phone-based MODEMS. In other words, it existed, but it wasn't on all the time... and dozens of sites don't make the Internet, if you know what I mean. The technology was emerging, but hardly capable of creating Colossus.}

Answer 3

Far-fetched. We know that general purpose AI is possible because we do it in our heads. It was a tempting thing to try on a computer, and people were trying to write self-training network-like algorithms for language and image processing in the 1950s. The recent breakthroughs have come about thanks to the speed and parallelism of GPUs, databases of many billions of images, and computers with quarter of a million cores. Once we have trained a network, it can be simulated on a single good graphics card. That is much more than anything we had in the seventies. I was working on PDP-11s in the early 1980s with clock speeds of 80 KHz and program memory in 4KByte chunks to a maximum of 64K.

Even in 2001, we could not have made HAL.

If they could make an organic neutral net, they might be able to train it to do simple things such as recognise objects the way our current AI tools do, but using the computers of their day. This would give them the processing power they need, but without the debugging tools that have lead to our current understanding. They might then assemble modules of their trained neural net to mimic how a brain works, with its various specialised processing centres, and generalised processing. People tried to do just this, but without any great success. However, it would be a lot more plausible to suggest that someone made an artificial intelligence by growing cells than by using 1970s electronics.

Perhaps your creatures could naturally have some modular brain unit equivalent to a typical 20-layer neutral net. Suppose this could be grown and replicated in the lab, and then trained; rather than the freeform, asynchronous tangle of cells we use for brains, that grows and trains at the same time.

Answer 4

No, they cannot, because their economy would not allow it. Water power just isn't that potent, and looking at the Earth, we have already deployed it pretty much everywhere it makes sense (and even in some places where it doesn't), and it still produces only 15% of the world's electricity and only 6% of the world's total energy. An alien economy operating on water power would need to fit within those much sharper limits, and would as a result be smaller and unable to afford the necessary investment (and it would also have much less need for it, again due to being smaller). I would not be surprised if their technology already had all the advances achievable without the use of fossil fuels.

Answer 5

Absolutely. Ish.

Wait, let me explain.

If you try to build an AGI on top of a general-purpose computing platform then there are a lot of interesting issues that have to be solved. What you'll end up with is something that solves a particular class of problems quite well, like processing a Large Language Model to interpret input and produce consistent outputs, as ChatGPT and other current "AI" projects do. The problem is that these systems require truly ridiculous amounts of processing and storage to do what they do in part because they're running on GP computing platforms.

But GP computing is far from the only option. Nor is digital computing.

Almost all AI today - from the most trivial image recognition ML to the LLMs that are all we seem to hear about these days - are based on simulations of neural networks on GP digital computers: complex algorithms running on massively parallel digital processors (usually banks of GPUs) to process each generation of data, consuming truly stupendous amounts of power. For the last few years there has been a rising acknowledgement that this is terribly wasteful, and that the power/compute bottleneck is a major issue... and that analog neural network hardware implementations are the future of AI development.

So can we build analog neural networks with 1970s technologies? Sure we can. They're not going to be compact, they're not going to be particularly performant due to inherent speed limits of inter-neuron communication, but it could be done. And for your aliens who have no other option, if they're building neurons for an analog neural network then of course they'll use that tech.

The problem is one of scale. Imagine we could build an analog neural network comprising a million analog neurons with all of the associated connections between all of those neurons, perfectly emulating (not simulating) the brain of a cockroach. Using modern technology the whole assembly would be larger than that Amazon warehouse. In 1970s tech the individual neurons would be much larger, and the whole thing would probably be the size of a major city. And it'd consume the entire power output of a medium-sized country.

And there's the rub. Yes it's possible, but it's far from feasible. And a machine the size of a city that consumes massive amounts of power - and produces commensurate amounts of wate heat - that thinks it's a cockroach just isn't particularly useful.

But perhaps we can produce an analog neural network that isn't just a model of an existing brain. Instead we could experiment with these neural networks and see how to use a tiny fraction of the neurons to focus purely on conscious processing without wasting all those neurons on moving the emulated cockroach's non-existent legs and so on.

But is it worth it to create something significantly less intelligent than a 12 year old human child?

Answer 6

There are too many "yes" and too many "no" answers being given.

The fact of the matter is that we still know very little about how to build an AGI, so we can't say whether there's an approach which would work well with relatively primitive computing technology.

The substantial advances of Hinton et al. over the last few years have not been because of some leap in computing power, but because they've worked out how to apply the matrix operations supported by graphics cards to neural network processing.

The whole thing is very likely to turn out to be a "we didn't know how to do it" rather than a "we didn't have enough computing power to do it". As such, we don't at present know "how much" will be "enough" as far as the required algorithms are concerned, which makes any definitive answer to OP's question impossible.

Answer 7

Yes, I could image that, if we would have set different priorities back then.

In the past, we went down the road of digital computing, which allowed our computers to do exact calculations. Before that we used analog computers.

But these days, we re-discover the power of analog computing as it is way faster and way more energy efficient. The main challenge with analog computing is that it hard to be precise. But this is less an issue in machine learning (on which much of AI is built), as it deals with uncertainty anyways. In fact in ML we trade precision for other qualities, like speed (see quantization).

A good place to dig deeper in the high-level interactions between analog and digital is "Geoffrey Hinton - Two Paths to Intelligence". Hinton is a Turing-Award winner (which is the "Nobel Prize" for computer since)

Note, that some areas of AI still profit from digital computing (like linear programming) as well as quantum computing.

I could imagine a realistic past, where we didn't invest only in digital computing, but in a combination of analog, digital and quantum computing. All have their strengths on a path to a general AI.

The obvious counter-argument is that much of our AI-development came from faster hardware. A single A100 graphics card has 6912 CUDA cores and 432 Tensor Cores, each with 1000-1400 MHz Compare that to a Cray-1 from the 70s with 80 MHz in total! And don't get even started with memory and memory speed.

Answer 8

Yes and No: because we actually solved GAI by 1970

All the important math, processes, and the principles needed to develop learning algorithms were all solved and well understood for making a GAI by 1970 but research into AI mostly died out by the early 90s because computers were still millions of times too weak to do anything particularly useful. So on the 1 hand, YES they can invent GAI with 1970s tech, because we did it... but on the other hand, they can't make it without at least 2010s tech because that is how advanced of hardware you need to run the software.

So Side-Step the Problem

Because we are talking about an alien civilization here, there is no guarantee that thier tech tree has progressed at all the same as ours. While humans have become masters of linear calculations, differences in the way an alien thinks could lead them down a whole different path of computer technologies. Perhaps thier computers intrinsically work more like an organic brain. Slow to do a series of tasks, but able to aggregate trillions of datapoints in parallel. In this way thier computers will feel like our garbage 1970s computers if you want to calculate a ballistic trajectory or sequence a genome or something like that, but when it comes to using its collective knowledge to make a single decision, it can do it in split second just like a brain can.

Answer 9

It depends entirely on how wealthy and how committed they are. All the various computing concepts we have today have existed since the 1970s -- with the exception of quantum computing. The differences are just a matter of size and efficiency. The power of your cell phone today would have taken a couple of rooms worth of equipment back then.

So, as long as there isn't something which renders an electronics-based AGI impossible, then from a technical point of view, they could build pretty much anything we can build today, just bigger, slower, and more power-hungry. All of those things mean "expensive".

To put it in perspective, the lowest estimate I've seen for the computing power of a human brain is around 10 teraFLOPS. Our best supercomputers didn't hit that speed until the early 2000s. Building that kind of power with 1970s computers would probably end up being a city-sized facility, with a city worth of technicians to maintain it, and a handful of nuclear power plants to run it all. And you'll still have to figure out how to program it. And it'll be about as smart as one person. Just maybe without a need for sleep or recreation. So be generous and call it three people.

So the hard part for your story will be coming up with a plausible reason for why they would spend that kind of effort.

Or else somebody would need to have an epiphany in analog computer design. Analog computers are considerably more efficient than digital ones, but far, far harder to design and program. Digital was just easier, so that's largely the way we went. But if some savant had spotted a way to build an easy to build and use, general purpose analog computer that might get the construction requirements down to something more reasonable. But probably still wildly expensive.

Answer 10

As others have noted, scaling is the big problem. One way to overcome this is for your alien AI to be bootstrapped, and take advantage of "computing overhang". This doesn't require a complete "seed" AI; it can grow from "centaur" approaches to research.

The first working prototype of any technology is inevitably crude, inefficient, and barely capable. Refinements and improvements can be made later, once the system/approach is better understood and tested. In the case of AI, we've been taking these clunky initial implementations and throwing them straight at hard, commercially-useful problems like natural language translation, image classification, self-driving cars, etc. so of course we have to give them massive amounts of compute. They also need masses of data about the world, since there's no way to "derive from first-principles" facts about the external world, like English grammar, or what dogs look like, etc.

Your aliens could instead focus all their AI effort on scaling their AIs. For example, instead of making ELIZA which analyses English text, they make a superoptimiser which analyses machine-code looking for speedups; instead of automating chess strategies they automate microchip layout strategies; instead of populating a database with common sense knowledge, they populate a database with theorems regarding algorithm complexity; etc.

Despite their systems being excruciatingly slow, they would occasionally stumble across a significant discovery: shortening an instruction sequence to save a few microseconds; or rerouting wiring to allow more space for cache; or proving a theorem that rules-out a large part of some search space; or tweaking an experiment design to allow more measurements without loss of statistical power; etc. These improvements would compound, with each efficiency gain making it slightly more feasible that others would be reachable, ultimately leading to AGI-level capabilities from shockingly few resources.

An unfortunate by-product of this approach is that the system is so ultra-specialised that its one big "black box": every part is so inter-dependent on the others, and serves so many purposes, that it's impossible to identify any clear "components", let alone understand what their purpose might be. The hardware itself seems to exploit some unknown physical effects (found through blind evolution), but attempts to study or harness this are hampered by the sheer inscrutability of the whole; preventing any offshoots which might help other information technologies to advance. Querying the system itself doesn't help, since it doesn't know: it was optimised for efficiency, not explainability. As the most complex object in the known universe, figuring out why its self-improvements work is beyond even its super-human(oid) capabilities (the same way those second-most complex objects, our brains, still don't understand themselves).

Answer 11

Yes, it is possible - if AGI is possible, which is an actual discussion in our own real world. That's too long a discussion to have here though.

If AGI is possible, then what really makes up an AGI is source code running on some computers. If they manage to come up with the source code, then there it is. Their hardware might seem jurassic compared to ours but distributed computing was already a thing in the 70's (as well as things such as artificial neuron networks).

The AGI will need a really large cluster in order to run with feasible performance - the whole setup might be as big as a building or maybe even a small city.

It will also consume a gargantuous amount of electricity per computation compared to our own world's nowadays computers, because time and technology is not just about making computers faster - it is also about making computations more efficient, including in terms of energy.

Finally, if it turns out that AGI requires quantum computing, that is not a blocker. Quantum science was already s thing much earlier than the 70's and scientists could figure a way to make qubits back then if they had the necessity and proper pressure to do so. Again, compared to nowadays capacity and tech, you will have something that will be bulkier, costlier, and which will require much more power to achieve the same computing performance of a 2020's quantum computer. But that is not technically impossible.