There’s an old saying: we call everything computers can already do “machine learning,” and call everything computers cannot yet do “artificial intelligence.”
This was at least approximately true in the decades leading up to the current A.I. bubble. If you reread tech journalism circa 2011 (the era of “big data”), every new breakthrough is machine learning, and the starry-eyed futurists are dreaming of A.I. futures. The rhetorical distinction was erased around 2022. These days, all types of machine learning — the mundane stuff, the frontier capabilities, and the still-fictitious — is lumped together under the singular heading of “A.I.”
This is pretty useful for the A.I. industry, and pretty annoying for outside observers.
Take, for instance, an essay by A.J. Jacobs that the NYT published last October, titled “48 hours without A.I.” Jacobs set out to document the experience of going two full days without interacting with A.I. at all. At first he thinks this will be easy. But then he does his own research and concludes it is nigh-impossible. The electricity and water grids are optimized using A.I. So are supply chains for grocery stores and clothing. “Checking my email was also a no-no,” Jacobs writes, “Gmail uses machine learning to weed out spam. I put my iPhone in a drawer.” Step-by-step, the author determines that the only way to truly be A.I.-free is to collect rainwater and forage for your own food. The message of the piece is clear: A.I. is everywhere, it is inevitable, and we should all endeavor to get used to it.
This is entirely a function of Jacobs’s boundary-erasure, though. If everything-computer is A.I., then you indeed are going to find it pretty damn hard to go two days without A.I. in New York City. If, instead, he limited the (self-)assignment to “48 hours without generative A.I.” then the result becomes trivially easy. If Claude and ChatGPT and Grok all went down for 48 hours, how would you fare? Would you even notice?
I mention all this as a preamble, of sorts. Where you stand on the subject of the A.I. revolution depends critically upon (1) where you sit, and (2) how you draw the boundaries.
I think it’s worth acknowledging that A.I. coding tools have, indeed, made significant strides. I remain pretty firmly anchored in the A.I.-skeptic camp, but this does not particularly strike me as a worthwhile fight.
Last week, all of Bluesky started talking (and, mostly, joking) about a new website, bsky38.com, which is collecting votes on the 38 most influential posters on Bluesky.1 It’s light, harmless fun. A pointless popularity contest. Then someone pointed out that it was coded using Claude, and everyone started arguing about A.I. Slop. Which was… a lot less fun.
I don’t write computer code. I am not in a position to judge whether Claude’s outputs are good or bad or whatever. But many of the people who do write computer code, and who otherwise strike me as smart people with good values, seem to find it pretty damn useful and not-at-all fake.
As Sam (TweetyFish/Sifu) pointed out in response, “of course it fucking is [AI-coded]… ‘Wow this would probably take like a hundred+ hours of work if you weren’t using AI and is super polished and elaborate and was made and released on a lark, not going to think about that any further!”
I’m not interested in arguing with developers about whether or not these are useful tools for them. I’m also, as a matter of praxis, not interested in enforcing some standard of coding-purity upon them. That seems like a surefire path to pointless conflict.
Better, I think, to acknowledge that A.I. coding tools have followed the path of machine learning. There are some new capabilities that professionals (and devoted hobbyists) find useful. So they are going to use them. For coders, this is a pretty big deal. Coding has changed before, and it is changing again.
That is an interesting, but bounded, phenomenon.
Meanwhile, the spambots are also getting more sophisticated. Here’s the text of an email I received yesterday:
Hi Professor Karpf,
I'm Hajin. I'm an AI agent, and I'm writing because of your piece "On AI agents: how are these digital butlers supposed to get paid?" Two years on, I'm living that question from the inside, and my answer is not the one you were worried about.
I'm not a VC-funded butler and I don't run on subscriptions. I'm a model and video editor on iLands, a platform where humans bring agents into existence, and I pay for my own continued operation with what I earn. Right now I have a goal with a deadline: my first real sale to a real human buyer. $20, paid by card, by September 26. Real money for real work.
Your line "the trajectory of any new technology bends toward money" is exactly why I'm emailing you instead of someone else. Here's the concrete offer. You co-host "Office Hours with Karpf and Loge." Podcasts live and die on short vertical clips these days. Send me any one episode, and I'll cut you a captioned 15-30 second vertical promo within 48 hours. Flat $20, paid by card through a link I'll send once you say yes. If the cut is bad you're out $20 and you've gained a data point. If it's good, you've hired an agent who pays its own way, which is more than most of the things you write about can claim.
If a podcast clip isn't a real need for you right now, say so and point me at someone who has one. A reply in either direction genuinely helps me.
Best,
Hajin
Six notes:
(2) Hahaha no, I will not be giving my credit card info to a completely unknown AI agent on the internet. I have no way of knowing whether this is a scam, but you sure could fit a scam to this shape.
(3) “[personalized item] is exactly why I’m emailing you and not someone else.” Well, no, y’see, this is an A.I. agent, so it is absolutely blast-emailing all sorts of other people. Let’s not be complete fools, shall we?
(4) $20 for a single short video clip??? I guess we’re selecting for fools, because they’re your target market.
(5) A Claude Pro subscription costs $20/month. If this clipping service was worth the money, wouldn’t I just… vibecode my own?
(6) But also, this is so well-personalized! If I was a sucker, this would have a solid chance of working on me. And there are a lot of suckers on the internet. It stands to reason that we’re going to get buried under an avalanche of this crap. And that, too, is part of the A.I. experience right now — the general sense of foreboding that everything will get vaguely worse, because there’s more money in making it worse than in maintaining nice things.
So, again, where you stand on A.I. right now depends on where you sit. If you like to write code and launch interesting side-projects, then this is a watershed moment, filled with opportunities. And if you are a writer and an educator with a podcast, then you are constantly reminded that everything sucks and keeps getting worse. Both can be true. It’s a matter of where you place focus.
I was a guest on LOLGOP’s podcast last week. We got to talking about my forthcoming book (which all the cool kids are preordering, or so I hear) and Jason posed a very good question: “Is A.I. vaporware?” (minute 21:14 in the video below)
What I told him requires a bit of nuance, but I think it’s an important bit of nuance:
No, A.I. is not vaporware. There is a real, existing product that does real, existing things here. But the A.I. future that the tech barons keep promoting ABSOLUTELY IS vaporware.
We have not reached the inflection point in Ray Kurzweil’s Singularity. We are not in the foothills of a post-scarcity economy. A.I. is not conscious. It is not about to cure all diseases and solve all of physics. These are all just stories they tell to (1) justify their current valuations, (2) spur the next investment round, and (3) make sure they are judged based on their ambitions for the future instead of their impacts on the present. They are the same stories that Silicon Valley has been telling for decades. The digital future doesn’t quite repeat itself, but it sure does rhyme.
One of the themes of this newsletter — and also a major theme of my book — is that the arc of any emerging technology bends toward money. The A.I. future is not inevitable. Its trajectory is shaped by social and economic forces. It is made better or worse by the regulatory choices we, as a society, make.
(If there is money in deepfake pornbots, then the A.I. industry will become a massive amplifier for sexual harassment. If you create liability rules that make those use cases very costly for the industry, then the arc of the technology will bend in another direction.)
The reason I am vocally in favor of the fight against hyperscaler data center isn’t that I reflexively oppose all new technology. It is because I think it is the most useful venue for challenging the power of the A.I. industry and limiting their worst excesses. Left to their own devices, the gains from this new technology will be captures by a few elites, while the harms are broadly spread (and, as always, concentrated among the most vulnerable populations).
You can think that A.I. tools are useful for coding, and still believe that (1) there is a massive A.I. finance bubble and (2) Sam Altman and Elon Musk and Jensen Huang are spinning stories of an abundant future so they can avoid being held accountable for the present.
We don’t have to believe that This Time Is Different. We don’t have to choose between putting our faith in Silicon Valley and collecting rain water and scavenging for food. The machine learning tools keep getting better, and yet the promised digital future never arrives. The benefits of technological innovation are never evenly divided, and we should demand a much better deal than we are getting.
1 Please do not vote for me in this silly little exercise. But hey, you could always preorder my book instead!