Holy hell has it been a wild week in AI discourse. And (...*checks non-existent paper calendar on desk*…) it’s somehow still just Tuesday, friends. But even with almost every other conceivable angle and hot take exhausting itself, we haven’t really seen anyone offering what we’d recognize as the “practical futurist’s” position. So here we go.
1) Remember that the worst possible outcome is never the only possible outcome, and it’s rarely the most likely outcome. So when you see Jacob Coxon tweet that “people building AI earnestly believe that it could kill us all by the end of the decade,” well… sure. That’s probably pretty close to a worst possible outcome, which means that it exists at one extreme of a spectrum where the opposite end is occupied by fully automated luxury communism and Peter Diamandis’s post-scarcity techno utopia where the biggest question is simply what kind of art and literature to create with all of our abundant free time. Almost every other possible outcome literally exists somewhere between these two extremes. Many of those alternate outcomes are more probable, and taken as a whole, the collected messy, syntopian middle of the spectrum of possibilities is much more probable than the isolated extremes.
2) Related: Remember that probabilities are supposed to involve real numbers that actually come from somewhere and real math that, you know, actually adds up. So when Evan Hubinger tweets that “We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,” you should be wondering what the basis of that figure is and whether it’s ultimately just a hype-baiting WAG (Wild Ass Guess).
Remember also that there’s a loooong tradition of hype-baiting WAGs in the historical AI discourse and that many of these WAGs have been strategically deployed not only to stoke hype but also to draw attention away from issues that the big AI companies would prefer to ignore. The philosopher and information ethicist Luciano Floridi called this out over the weekend and pointed to a new paper appropriately titled “Not Even Wrong 2: An Audit of Public AGI Prediction, 1950–2026.”
3) All of that being said, it remains important to not dismiss the whole overheated discourse as bullsh*t – despite all the bullsh*t. As Floridi argues (along with Amy Webb, Timnit Gebru, and many others), there are PLENTY of very significant near-term risks and even present harms associated with generative AI / LLMs that aren’t waiting on a superintelligent paperclip maximizer to actualize. But the companies developing, promoting, and betting the house (and more!) on frontier models aren’t being held accountable for the risks and harms created by their products in the way that companies in other fields routinely are.
Imagine for just a moment a pharmaceutical company developing a product that they claimed might one day cure cancer or might result in the death of every human on the planet. (This is a claim made by Anthropic, OpenAI, and any number of AI boosters over the years about the systems they’ve been working to develop.) Now, ask yourself whether that pharmaceutical company would be allowed to bring that product to market.
If you want to see something done to hold AI companies accountable for their products, do something to help elect serious people with a demonstrated record of addressing real problems and serving the public interest.
4) In the meantime, make a habit of differentiating between irreducible uncertainties and reducible uncertainties and focusing your mental energy on the latter. This means turning down the volume on some of those underspecified conversations around AGI / ASI / machine consciousness / meaning in a hypothetical post-scarcity, post-work world and instead, digging into the deep but not unfathomable AI-related uncertainties where you can actually explore and experiment your way toward applicable new knowledge that can benefit the communities you belong to and the organizations you lead.
If you’re looking for an eminently practical set of questions to explore around AI futures, software, and org transformation, jump right into Ben Evans’s most recent essay. And if you’re looking for thoughtful (i.e., less hot, more nuanced) takes on AI tools and their implications, follow folks like the practically-minded Ethan Mollick and the LLM-skeptical Gary Marcus.
And if you’re looking for guidance on how to better explore and experiment and reduce uncertainty while building clarity on the scope of possible futures, good news: You’re already in the right place. Also: check out Pascal’s new book.
5) Breathe.
Sorry! I had to do it.
My biggest fear is that all of this agitprop, potentially being deployed or co-deployed by a political enemy for their own benefit, will slow down if not stop valuable progress.
THAT is the existential threat re: AI that worries me.