LLMs Explode the "Intelligence Explosion"
Large language models like ChatGPT represent an advance. Why does superintelligence seem as far away as ever?
Hi everyone,
This has been a paid post, but as I’m getting considerable feedback/questions on it, I’m making it a “public service announcement” post available to everyone. As always, please consider a paid subscription if you’re finding value in my writing. And thank you all again. Here we go…
I’ve been candid about Large Language Models, repeatedly describing them as innovations in the field of AI and an advance in the power and scope of our natural language processing (NLP) capabilities. Here’s what nags at me: why does superintelligence seem as far away as ever? Let me do a super quick recap of what’s happened in the field leading up to and after the release of ChatGPT in November 2022.
By the late 2010s, the language model phenomenon had arrived. Prior to ChatGPT, Language models like BERT and RoBERTa were already showing improved performance on a range of NLP tasks, like text summarization, word sense disambiguation, co-reference resolution, entity recognition, and sentiment analysis.
By early 2023 OpenAI’s ChatGPT had a hundred million subscribers. Developers (and managers) were noticing it could generate Python and other code snippets from simple prompts, solving actual coding problems. And that’s not all….
An entirely new Question/Answering capability appeared seemingly overnight. Suddenly the smartest guy in the room was in your computer. Folks started talking about AGI. Folks started declaring the Turing Test passed.
Hundreds of millions of users noticed that the new Super AI would occasionally start yammering out bullshit, like your uncle at the Christmas party. Code would have obvious, even moronic mistakes in it, solving for instance the wrong problem entirely. Medical responses would cheerfully recommend that imaginary patients should commit suicide. Flummoxing and tripping up the new Super AI became online sport. The models seemed simultaneously to be smart and knowledgeable and on mescaline.
Superintelligence and AGI enthusiasts popped their heads out early on, and then weirdly lurked more in the background. Odd. AI finally acts like an intelligent thing—even given its slip ups—and the futurists go mum. It’s like it never happened.
I expected more aggressive predictions (to be sure, some enthusiasts have stayed the course). But the mood of futurists, by my lights, has been decidedly ambivalent and even cool on our latest advances. How do we make sense of this? Are LLMs really not much of an advance? Or is something else going on?
Here’s an idea: LLMs expose the folly of mechanical “intelligence” even more starkly. I think it’s this. In what follows I’ll try to unpack why LLMs are actually working against the futurists. We’re seeing what a powerful AI looks like, and we’re also seeing how…. mindless it is.
A special message to all readers of Colligo: I write for the love of it, but I also write for a living. Please consider a paid subscription to help me continue to produce quality content and share my ideas with the world.
But Nick Bostrom is Still Bostrom
Nick Bostrom, the Oxford philosopher who coined “superintelligence”—replacing I.J. Good’s 1960s term, “ultraintelligence,” I guess words matter—and wrote a book with that title in 2014 has been a go-to voice for prognostications about the future of AI for decades. Following other luminaries like Ray Kurzweil, Bostrom sees a coming intelligence explosion once we achieve AGI, leading to superintelligence quickly. It’s the exponential growth idea, the “hockey stick” graph that crawls along and then rockets toward infinity.
In the 2010s, Bostrom seemed to be prescient. Deep Neural Networks were, well, deep. They could recognize images like faces and recognize speech. They improved autonomous navigation on the ground and especially the air, with drones. Image generation started working, as did deep fakes.
Recommendation systems for movie, music and other products improved. Drug discovery and genomics showed promise (still a long way off). Champions at difficult board games like Go were vanquished by AI systems. On and on, DNNs moved the needle in AI—in data-driven—AI.
In the next decade the world screeched to a halt with Covid-19, and Reddit groups started grumbling that self-driving cars weren’t really happening. Even Elon Musk admitted so, after some notoriously bad press. When LLMs burst on the scene in November 2022, it seemed all the grumbling would—must—finally stop. I expected to hear Nick Bostrom on the nightly news. Futurist extraordinaire Ray Kurzweil on Good Morning America. Bostrom would no doubt go on a multi-country speaking tour, warning captive audiences we had only a few more years of human culture (maybe he’s right, in a different way?).
But the story of the 2020s so far is more of the same—AI is a decade or two away, as always. LLMs don’t seem to have mattered. As far as I can tell, they’ve pushed the mythical superintelligence question even further into the future. Huh?
A Glimpse of Mechanical Intelligence—Naah.
Famed neuroscientist Kristof Koch weighed in on the superintelligence question for IEEE Spectrum last year. He starts the important-sounding interview with typical comments that intelligence is “ill defined,” and basically we don’t know what we’re talking about enough to make predictions. It’s a recent interview with an acknowledged brain guru, but it feels like it could have happened ten years ago. Twenty.
Author and podcaster Sam Harris interviewed venture capitalist and LLM booster Marc Andreessen on his Making Sense podcast last year as well. Andreessen spent most of the interview joyfully prodding a recalcitrant and sceptical Harris with examples of ChatGPT already demonstrating something like AGI, as with its sophisticated takes on moral dilemmas. It was obvious Harris wanted to save the idea for some other technology still in the future. A mythical one? Seems so.
And that’s the rub. The future of AI will never arrive. It’s not supposed to arrive; that’s the point. AI futurism represents a powerful myth, and it loses its grip if the big moment is coming tomorrow or next month (egg-in-the-face time). It loses its grip if it’s a hundred years a way as well. The ideal time to keep the discussion going forever seems to be about one to two decades. It’s amazing to me that this “one to two decades” has been a moving target since essentially the inception of the field.
These are shopworn examples, but certainly make the point: In 1965, Herb Simon, one of the original scientists at the Dartmouth Conference, where AI as a field of study was inaugurated to much fanfare, declared we’d have AGI in two decades. Marvin Minsky—also at Dartmouth—claimed two years later that “"within a generation … the problem of creating artificial intelligence will substantially be solved." (Define “solved.” Better: define “substantially.”) Kurzweil, to his credit, has been settled on 2029 for decades (and 2045 for the Singularity). It’s five years away now, and given our ongoing problems with moronic responses and hallucinations, five years seems—wait for it!—too close. What innovation will make that happen? What’s the science of any of this? Welcome to the future, same as the past.
The Truth and Nothing But…
So what’s really going on here? Mythology. That’s what’s going on. In his quite readable 2016 book Rise of the Machines: A Cybernetic History, Thomas Rid explains the myth of true AI in the context of our love of myth generally. “[M]yths overcome the limits of fact, experience, and technology at a given moment in time. The rise of the machines was always projected into the future, not into the present or the past. Evidence was always in short supply.” He continues, “…mythologies are remarkable not for their content, but for their form. The basis of the myth appears as fully experienced, innocent, and indisputable reality: computers are becoming ever faster; machines are ever more networked; encryption is getting stronger.”
But the form is also emotional. That’s the key. They transform the present and so provide us with emotional material about the future, once inchoate and uninhabited, now full of extensions of our boring world today into the world of make-believe. “Mythical narratives form a path between the past and the future to keep a community’s shared experiences in living memory,” says Rid. In many non-technological myths, the stable emotional core is in the past, behind us. Wars, religious events, and political moments form our myths in history books and ceremonies. In AI and technology generally, the emotional core is—must be—always in the future. In a very real sense, it can never arrive.
What’s always approaching but never arrives? Tomorrow. - Mission Impossible: Dead Reckoning.
And here’s the paydirt with our world full of marvelous and hallucinating LLMs. Of course they’ll be fixed—in the future. Clearly they’ll be sentient—in the future. And other technologies will no doubt appear. Those future technologies will be super smart and sentient.
In other words, it’s not the set of facts, the science, the evidence that we’re even talking about. We’re talking about the emotions of technological myth that are of necessity set sometime in our future. Not so far that we don’t care. And not so close that we can be proved wrong.
I’ll make a final point here that I’ve already broached to some degree but not really fleshed out completely. It’s a bit like the “uncanny valley” in image generation AI.
The closer we get to a truly human looking image, the weirder it seems. To be “almost human” is worse than to be a cartoon. Our minds can make sense of cartoons as artistic creations. They have a tougher go of it with high-powered generative AI getting us to the uncanny valley. Text generation systems like LLMs are a bit like this as well. As the largest models—like GPT-4—get quite close to generating consistently human textual responses, the occasional and inevitable confabulation and bullshitting and hallucination seems even weirder, more uncanny if you will. No wonder the futurists are still of different minds and babbling about the future like business as usual. Getting to the destination isn’t really the point.
LLMs today remind me of the old saw about getting to the moon by climbing a tall tree (this too is a shopworn example, but well-worth repeating). Yes, you’re making progress. No, you won’t be arriving. Such is life in AI as well. What we can be sure of, come hell or high water, is that the mythology of AI will continue as long as humans keep being human.
Erik J. Larson
> Bostrom would no doubt go on a multi-country speaking tour, warning captive audiences we had only a few more years of human culture (**maybe he’s right, in a different way?**).
Got the chills when reading the "in a different way?" part. It's as if LLMs amplify what's been wrong with us humans for quite some time. In a way, it might be a wake-up call, a reminder that our obsession with tech kinda let us forget that we need to obsess with our humanity in the first place.
Erik, thank you for this simple and brilliant analysis of the mechanics of a myth, and the specific features of the technological myths, well extracted. I will be attending a conference in Prague on 24.4 April - so pls let me know if it is ok, if I reference these points in my panel talk, of course, quoting you.
https://qubitconference.com/