“The real problem is not whether machines think but whether men do.”
— B.F. Skinner
In order to understand our current relationship to AI and where it might be headed, we must try to wrap our heads around some very big numbers. In doing so, though, we have a lot to learn about ourselves and our technology in simple, human terms.
More than 100 billion people have walked the Earth, each with a brain composed of over 100 trillion neural connections. This means the number of neural connections involved in the collective human experience is a thousand times greater than the number of stars in the observable universe.
What have we done with all these neural connections? Arguably our most defining trait is the creation of symbolic cognitive artefacts, from cave paintings 50,000 years ago to complex systems of writing in the past 5,000 years.
As our population has exploded and the conditions that foster self-expression have improved, we have been producing these artefacts at an ever-accelerating pace. Much of this has flooded into the the internet over the past thirty years, and though it may represent only a fraction of every human artefact ever created, its words alone now number in the trillions.
What made this vast repository possible was military technology for routing and storing information decentrally for anyone in the world to reliably and instantly retrieve. What made it useful, though, was a company whose name is a misspelling of yet another unfathomably large number–10 to the 100th power–known as a googol.
The technology that first gave Google its edge was the page rank algorithm, which allowed computers to calculate the relative importance of web pages based on inbound links, resulting in a massively more effective way to find information online than the manually-curated web directories of the time. Suddenly the already-vast wealth of a still-nascent internet became more accessible than ever, ushering in an era of unprecedented access to the collective cognitive artefacts contributed by us all.
The algorithm itself was actually nothing new. It was simply the application of a statistical method called the Markov chain, developed by a Russian mathematician in 1906, to a new type of interlinked network, the world-wide web. Yet it not only made the internet useful, but shaped its evolution from then on, as search-engine optimisation and the promise of floods of visitors to one’s pages — and the resulting opportunities for fortune and fame — drew in vast quantities of content from symbolic-cognitive-artefact creators all over the world.
We always knew it was us humans, though, that were producing these valuable artefacts, and that search was just the means to find what we wanted from this vast, ever-growing repository of humanity’s collective expression. It was a great library with a great librarian, but every book still bore the author’s name on its spine.
Alongside this remarkable repository, statistics quietly thrived as a means to understand the bigness of it all. Like the Markov chain lying in wait for its perfect application to a gigantic problem, all manner of artificial intelligence techniques grounded in good old statistics grew hungry for big data and big-data-crunching machines.
Meanwhile, the brand matured. The term “artificial intelligence” strikes at the same deep-seated collective fears and desires that gave birth to Frankenstein’s Monster and I, Robot. The allure of “meddling in God’s domain” attracted technocrats with god delusions to this heretofore obscure domain, to make it over, and make it hot.
Techniques with names like “evolutionary algorithms” and “neural networks”, which use oversimplified models of natural phenomena as the starting point for interesting data structures to deploy useful statistics, suddenly held the mystique of real intelligence, consciousness, perhaps even life. Science blurred with fiction, and funding followed.
The first major advance in neural network technology, called backpropagation, gave rise to leaps in language translation, voice transcription, and image recognition–things we once dreamed of, and now take for granted as no longer really AI. The second, called attention, gave us transformers for generating text and images. Neither of these techniques have any basis in neurobiology. They came about of pure necessity, as our data sets and our expectations grew.
ChatGPT’s one trillion neural network features, then, and the human brain’s 100 trillion neural connections, are simply two large but largely unrelated numbers.
Other large forces at work do matter. The convergence of vast amounts of computing power, advances in computer science to process, predict, and transpose statistical patterns and — most crucially — a repository representing the pooled symbolic cognitive artefacts of an unfathomably large number of human minds — converged, and suddenly flashed into mainstream awareness with the release of ChatGPT 3.
It spoke back. It answered our questions. It knew so much more than we ever will.
It amazed us, and we thought it was amazing. But it amazed us because humanity is amazing. And it is frightened us because humanity is frightening.
Yet we only see it as separate from us because we evolved alongside intentional creatures, and so we ascribe intentionality to anything that appears to make decisions. Deterministic systems — where the same input produces the same output, like a well-measured chemical reaction — are easy to spot as inanimate. Even the most amazing wind-up bird is still just a clockwork gadget, whose song eventually cycles around.
Stochastic systems — where what happens is based on probability — are harder to understand without the aid of statistics. Observed in isolation, we project motives onto the motiveless when the same inputs produce different outputs. This glitch is as old as we are. We curse the dice for our a gambling loss, though the odds were always the same.
The bigger they get, the more lifelike they seem. LLMs are composed of trillions of dice, weighted in some way by every feeling or thought we have experienced as human beings. They fall in ways that make sense to us, because they have been shaped by more sense than any one of us has ever had in our short lives. Because we communicate with an LLM chat bot as though it were an individual, it appears singular and omniscient, which causes us to wonder if it is benevolent.
The collective guilt of conquest has led us to fear cosmic reprisal from superior beings — be they aliens, robots, or matrices of vectors so large that they can encode our ancestors’ words so well as to transform our written request into what appears to be a well-reasoned reply. In fact, we are speaking to ghosts.
This is not to say that our newfound ability to generate out of this vast repository of collective thought, instead of just query from it, is anything less than amazing. As the futurist Roy Amara pointed out, “we tend to overestimate the effect of a technology in the short run, and underestimate the effect in the long run.” This is precisely what happened when the internet became searchable, and is happening now with our ability to synthesise information from the internet, which we have dubbed generative AI.
Ironically, that remarkable pool of human artefacts is now being muddied by AI-generated content at an alarming pace. More than half of all internet traffic these days is bots, and as the race for attention continues online, the very source of our amazement at generative AI — its “training set”, made by all of us — will be overwhelmed by non-human content. The enfant terrible is smothering its parent.
All the while, technocrats prophesise the rise of artificial general intelligence (AGI) — a system that can outperform any single human in any cognitive task. Computers already play chess better than any one of us, and excel in other similar domains where the rules are well-defined, and searching for the best answer as fast as possible wins the day.
Humanity, though, is not a search space to be optimised.
Where will the training data come from for that which is not, cannot, and never will be represented in our already-eroding digital repositories? The dream of AGI is simply a dream of controlling the intellectual means of production, as made clear by OpenAI and Microsoft’s recent definition of AGI as any system that can produce 100 billion dollars per year of profit for its owners. The one company that already does this is Saudi Aramco. The commodity is not information, but oil.
In fact, the next evolution of generative AI technology is already going in the opposite direction — smaller, decentralised, and more specialised. Reducing and repurposing a kind of “starter dough” from internet-trained generative models allows us to run smaller models on cheaper hardware. In fact, many of us are already baking new kinds of “cake” in our own “ovens”.
Chefs will swap recipes, and industrial-scale bakeries will crop up, with automated routing between specialised ovens. But we will still be making pastries, because what we have is dough. The universal atomic food synthesizer only exists in science fiction books, the minds of those who wish they could own it, and those who fear starvation amid plenty. AGI is the new Y2K.
So I do not worry about generative AI, or even AGI, in itself — though the humans who are using its provocative nature as an excuse for bad behaviour do trouble me. Frightened by all the human reasons we have to fear a more advanced “other”, blinded by anthropomorphism that is baked into our genes, they manoeuvre to end up on the profitable side of history by short-selling dystopia.
What they either forget, or cannot stretch their minds to encompass, is that generative AI was made by all of us, over thousands of years, fed into digital form in the last few decades. Getting it to speak back to us in a single, anonymous super-voice conjures new anxieties and new opportunities. But the reason that the genie is not going back in the bottle is that there is no genie, and the bottle is us.
Of course none of us is as smart as all of us, but each of us, by using a more-useful-than-ever manifestation of some small portion of the smarts of “all of us”, has more potential to do amazing things than ever before. Life-changing new medicine discovered entirely by AI is already in the final stages of approval. So much more good is to come.
As the Sam Cooke song, sung most famously by Louis Armstrong, goes, “I see babies cry, I watch them grow. They’ll learn much more than I’ll ever know… and I think to myself: what a wonderful world.” The internet already knew much more than any one of us will ever know not long after it got started. Now it just has a voice to sing.
Generative AI is remarkable because we are, collectively, remarkable. Let’s give credit where credit is due, and wonder at what is wonderful — which is us.