Hey, German-speakers! Through a very weird set of circumstances, I ended up owning the rights to the German audiobook of my bestselling 2022 cryptocurrency heist technothriller Red Team Blues and now I'm selling DRM-free audio and ebooks, along with the paperback (all in German and English) on a Kickstarter that runs until August 11.
Not only is agentic AI bullshit, but it's a specific kind of bullshit that AI hucksters have busted out in the past, and will bust out in the future, so it's worth spending a minute to unpack this bullshit and catalog its traits so that we don't fall for it. As GW Bush says, "Fool me once, shame on you; fool me twice, we don't get fooled again."
Automation can be transformative, relieving us of danger and drudgery by getting a machine to pick up some of the heavy work. Ideally automation seamlessly swaps a human for a machine at some stage in a process (ideally, the boring, dangerous and/or difficult phase). Like, whipping egg-whites for a meringue is hard on your wrist. But swap your whisk for a hand blender, and suddenly that tiresom process becomes fast and easy. If the blender is cordless, you can use it anywhere in your kitchen, including wherever you would have stood over a bowl with a whisk.
A mixer, by contrast, requires more labor on your part: you have to decant the contents of your mixing bowl into the mixer, run its motor, and then scrape the whipped whites back into your bowl for the next phase. It's worse automation.
But the worst automation would be a mixer that requires a special electrical outlet, a different fridge, and a special egg-carton. You would have to redesign your whole kitchen to use that thing. Sure, it might produce perfect meringues, and sure, if you had a meringue factory it might be a great solution. But for everyday use, it's a solution that creates more problems than it solves.
AI pitchmen promise that seamless swapping of a human tethered to some choresome drudgery for software. That's the whole point of self-driving cars: each of us can swap a standard car for one with an autopilot and use the same roads, with the same road-users, to get to all the same places. We don't have to tear up all the roads and lay tracks, or fill the roadside environment with sensors and beacons to help the "self-driving" cars navigate the system. A self-driving car can share the road with human-piloted vehicles, even when those other vehicles are driven by humans who don't see why they should allow a robot to merge into their lane or have the right of way, even if the human is turning left into oncoming robo-traffic.
Self-driving cars are not very good at this stuff, as it turns out. When that became apparent, self-driving car hucksters announced that it was only reasonable for their products to require something of the rest of us. As Andrew Ng put it:
“I think many AV teams could handle a pogo stick user in pedestrian crosswalk,” Ng told me. “Having said that, bouncing on a pogo stick in the middle of a highway would be really dangerous.”
“Rather than building AI to solve the pogo stick problem, we should partner with the government to ask people to be lawful and considerate,” he said. “Safety isn’t just about the quality of the AI technology.”
This is an incredible act of shameless bait-and-switchery. In just a few short sentences, Ng's cars go from being the kind of automation that is purely the concern of the person who uses it – the owner of a self-driving car – to the kind of automation that everyone in the world has to adjust to, lest we become part of the "pogo stick problem."
Making a car that can navigate a well-behaved, non-adversarial world is relatively straightforward. But demanding that the entire world behave itself? Well, that's the hard problem of 100,000 years of civilization and ethics. A product that only works in an ideal world isn't a viable product.
Self-driving car boosters didn't invent this wheeze, either. The entire concept of "pedestrian" (and later, "jaywalker") was invented by the auto industry to shift blame for the death and destruction the wealthy owners of their products inflicted on everyday people to the victims:
The latest peddlers of pogo-stick demands are the agentic AI people. They have raised (hundreds of) billions of dollars by promising that they will make AIs that can autopilot your browser to accomplish tedious, time-consuming tasks, visiting the same websites you would visit, locating and processing the information needed to perform the task you've set for it. This will supposedly make all kinds of human workers obsolete (which is where the hundreds of billions of dollars come in – the whole AI investor pitch is "We are developing technology that will let bosses fire their workers").
But agentic AI sucks. Asking a chatbot to take a screenshot of a website, then make guesses about which parts of it are links and what those links do, choose one link to fire a click at, and then start again is a recipe for incredible dysfunction. That's even before we get into "hallucinations" (this is AI jargon for "errors").
A more mature agentic AI apologetics admits that while no one knows how to make an AI that can navigate the whole internet, we can make specialist agents that can perform one kind of task, then hand off the output from that task to the next agent, and the next. This also sucks: you're created a whole menagerie of AIs, each of which is prone to its own failure modes, and then combining them, multiplying all those error potentials together, sending erroneous findings careening through a cascade of downstream AIs. This is broken-telephone-as-a-service. Give it your credit card, ask it to order a bag of jucing oranges, and six months later someone's gonna back a 16 wheeler up to your front door with $40,000 worth of frozen OJ and a receipt for a futures contract you're on the hook for.
The latest agentic AI pitch "solves" this problem by asserting that the whole internet will simply have to accommodate itself to AI agents. Every website will have to adopt robust, accurate semantics that describe its navigation and offerings, standardized across every domain of human activity. This would be great. The semantic web people have been trying to make it happen since 1999, with no success to speak of, for reasons I identified more than 20 years ago:
The reason websites don't make their results easy to scrape and compare is that they want to cheat you. They want you to buy something more expensive and/or inferior than the best match for your desire. There is no way for an AI agent to know when a website is lying to it, and the websites that lie the most are incentivized to have the best, highest-grade automation hooks for an AI agent to connect to (just as spammers have the best, most pristine anti-spam indicia, from DKIM to SPF to DMARC records).
And these cheaters aren't fringe players – they're the biggest companies out there. Amazon knows that Prime members don't shop around, so it presents them with higher prices than non-Prime users. Airlines use AI and surveillance data to estimate your desperation and price their tickets accordingly:
The hard part of comparison-shopping for an airline isn't sorting a database of all the prices offered to all customers under all circumstances: it's compiling such a database. We don't need complex AI-based techniques to perform a simple sort – we need AI to solve the problem of knowing what prices every airline is charging at this instant to every flier for every itinerary.
When agentic AI grifters insist that the entire internet has to adopt and faithfully use standard APIs so their bots can accurately analyze the internet's contents, they are re-inventing the pogo-stick problem. Yes, if you could get the entire world to arrange its affairs to your benefit, you could surely do some incredible things, and if my grandmother had wheels, she'd be a rollerskate.
Even if you could get everyone to adopt a standard set of APIs and use them well, this is a titanic engineering challenge, at least as big as anything the agentic AI people are promising to do.
There's an unassailable response to the assertion that you could do amazing things as soon as everyone else upends their life to make things more convenient for you, the sacred principle of "wish in one hand and shit in the other and see which will be full first":
Support me this summer in the Clarion Write-A-Thon and help raise money for the Clarion Science Fiction and Fantasy Writers' Workshop! This summer, I'm writing The Reverse-Centaur's Guide to AI, a short book for Farrar, Straus and Giroux that explains how to be an effective AI critic.
If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
