When President Trump announced a few weeks ago that he intended to change the name of artificial intelligence to “super intelligence,” there were predictable eye-rolls at the man who has elsewhere tried to change the name of the Gulf of Mexico, Lake Ontario, the Department of Defense, the Kennedy Center, Mount Denali, and (I’m sure) plenty of other landmarks and institutions that have slipped my mind. With respect to AI, though, I see what he’s going for. Zooming out, it seems obvious that if we designate the next several generations of GPU-powered intelligence as “artificial,” it will do more to confuse our relationship to the technology than clarify it.
Something that’s “artificial” is cheap, unnatural, temporary, and generally understood as qualitatively inferior to organic creations of either humans or the natural universe. Imagine if we applied this label to all our other technologies. If you were forced to choose between commuting on horseback or embracing the future, then sure, under duress, you might buy a car that’s described as artificial transportation. But you would not feel good about it.
Likewise, without hopping back on my football stadium soapbox from last week, the phrase “artificial turf” calls to mind a massive slab of concrete and a thin layer of plastic grass (and 70,000 maniacs at Veterans Stadium in Philadelphia). No one ever wants anything that’s artificial, and they’re usually right. So yes, to the extent it’s in America’s interest to encourage businesses and individuals to incorporate AI into their daily lives and workflows, then the technology’s current branding could plausibly be an obstacle to fostering the kind of trust and enthusiasm that powers a generation of optimism and economic transformation.
More interesting to me, though, is that super intelligence fails too. For one, the idea of “super intelligence” floating around in various clouds is actually less comforting to me than a version of knowledge output that’s branded as synthetic. I’m not saying I agree with Ron DeSantis here, but I’m not surprised by this sort of reaction and think it would be fairly common:
But equally important, let’s stick with the fundamentals: “Super intelligence” is not quite what this technology is, and not how it should be understood by generations to come.
On this point—and noting for the record that Trump has now decreed by executive order that the technology is called Super Intelligence, while also correcting reporters who ask him about AI, telling them it’s now SI—I’m reminded of a tweet I saw a few weeks ago. Someone asked the latest OpenAI model to compare its political views to those of famous political commentators, and the model came back with comparisons to three well-regarded, milquetoast neoliberal commentators. Which, in the model’s defense, is probably an accurate self-appraisal!
Models are trained to be evidenced-based and risk averse, and they will rigorously reflect contemporary facts, concerns and values as elite cultural institutions understand them today, not as society may understand them tomorrow. If you were to go back through the past 10 years, though, it’s not hard to think of all the different areas in which the policy instincts of that same intellectual consensus have been flat wrong or at least myopic—immigration, geopolitics, energy security, you name it—and models would’ve made similar mistakes. (Go back to February 2016, and today’s super intelligent models would have had no idea how to explain Donald Trump’s appeal and imminent political ascent; for that, you’d have been better off with this Bill James blog post.)
Here let me pause to emphasize that I am very much NOT an expert in how LLMs are trained. That said, I will do my best LLM impression and try to BS a little: after ingesting massive amounts of published writing that’s intended to approximate the full spectrum of human knowledge, model output is then graded both by humans and machines, with grades based on how faithfully the models can extrapolate from their training data to produce useful answers. Reproducing familiar arguments with lots of evidentiary support is a reliable path to getting a good grade, whereas offering an original insight that doesn’t align with our intellectual consensus is likely to be graded as a mistake. In light of the penalties for models that offer answers and theories that are flat wrong or unsupported by evidence, the training process inevitably pushes those models to the intellectual middle (I’m purposely skipping over the recent revolution in reinforcement learning that undergirds today’s models, but that also emphasizes the fact that models are best on “knowable” facts and outcomes).
Regardless, all that’s still very useful! AI can be great for brainstorming, or, if you’re looking to learn more about literally any topic in the world, LLMs can instantly perform a week’s worth of shockingly comprehensive work that might otherwise require multiple research assistants. Ask a model to take that research and highlight the most important insights and questions in an executive summary, and again, it will do the job quite well (and generally far better than a human intern would). It’s amazing.
It’s not, however, offering us the kind of supreme knowledge that genuinely challenges humanity or pushes our understanding forward. At least not yet! Among publicly available models trained for mass consumption, there’s none of the foresight or unique insight that one might expect from “super intelligence.” Ask AI to explain most areas of modern life, and what you’ll receive is closer to midwit intelligence. Its answers will sound smart to dumb people, and sound like derivative and boring pablum to any subject matter expert who happens to have mastered the substance at hand.
I’m not writing this to downplay the potential of the technology, but in hopes that we can all be a little clearer about why it’s amazing. AI is great for coding, and in the fullness of time, I think Ben will obviously be correct about the rise of personalized software and agents that transform life for both consumers and businesses. As to knowledge work generally, AI can ingest and analyze massive amounts of information and attack problems with relentless energy. Its gift is stamina, though—not brilliance. That stamina is why math problems that had been unsolved for decades can now be solved in about a week. That’s also why throwing AI’s analytical capacity at the field of medicine could lead to new drug discoveries, advances in DNA mapping, and all kinds of other leaps forward in health and science. That stamina will help in finance, as well, along with probably 1,000 other industries.
But can we all take a beat before jumping to language that overstates the case? Intellectual stamina is of course an important component of intelligence, but it’s not synonymous with the sort of omniscience and omnipotence that’s implied by super intelligence; nor is stamina the signature trait we typically find in the figures who genuinely change the world with their ideas. In general, many of the most consequential figures in history had specific experiences, talents, obsessions, and theories that developed for decades, at odds with consensus, and they arrived at the right time to reshape their fields, societies, and worlds. They were not afraid to be called crazy or reckless, and where AI models are obsequious and concerned with evidence, history’s great men were often breathtakingly obstinate and fueled by instinct (and in some cases, sure, they were clinically unstable).
Speaking of clinics, the AI labs—which are actually just trillion dollar corporations selling a product, while also working the refs in Washington—are invested (and seeking investment) in the idea that they’re building a super-technology that humans can’t fully control or understand. They claim it could pose existential risks for all of us. Super intelligence branding certainly burnishes that story. But are we sure it’s useful to think about the technology that way? The conspicuous lack of supporting evidence from doomers remains incredible to me, and while AI leaders certainly see themselves as the misunderstood great men of modern times, I don’t buy it.
In any event, our choice of language will matter. Spend the next 20 years selling intelligence that’s “artificial” and these products could have a difficult time building trust and enthusiasm. Call it all super intelligence and we are conferring too much authority on these tools, and likely scaring the crap out of everyone along the way.
So how to describe a tool that can automate the work of humans faster than ever, on an effectively continuous basis, and in a relatively rote and uninteresting way? That sounds like a machine to me, which is of course a label with its own implications. Machines can be controlled, machines can be refined, and when machines malfunction and do lasting harm, machine makers can be held liable. Also, as Microsoft’s Mustafa Suleyman wrote a few weeks ago, machines do not have emotions, values or desires, and Anthropic should really stop training them to approximate those qualities. Finally, machines are great solutions for some problems and less useful for others. Even the best machines are not substitutes for human qualities like courage, foresight, or intellectual curiosity.
Machine intelligence, then. Why not? We already have machine learning — that’s what training is — and machine intelligence is the output of that process. That would be my pick if I were President for a day (assuming that renaming commercial technologies by executive order remains squarely within the presidential remit).
Granted, artificial intelligence branding looks awfully entrenched, super intelligence sounds cooler, and “machine intelligence” is such a boring label that I almost feel the need to apologize at the end of this article—so maybe that’s why not. On the other hand, doesn’t a departure from exotic hyperbole and science fiction visions of the future sound absolutely fantastic?
If not sexy, it would at least be healthy. And of course, even if you think this guerilla re-branding campaign is a long shot, one thing my label has going for it is accuracy.
Sharp Text is extension of the Stratechery Plus podcasts Sharp Tech, Greatest of All Talk, and Sharp China. To subscribe and receive weekly posts via email, click here.