
Just keep going until you hit Vegas.
If you believe tech industry marketing, here are the things you might reasonably expect computers to do today:
- Drive my kids to soccer practice
- Land my plane
- Remove my infected appendix
- Lay out a defense strategy for my white-collar crimes
- Make me a CRM for my florist shop
Here at Aboard, we’re building with AI, so sometimes people come to us with the expectation that we can deliver things like the examples above. We have to help them see through some of the marketing.
The fantasy we’re sold is a “Built to Order” (BTO) universe, where little robot chauffeurs and butlers attend to us and do our bidding. The nascent AI industry is more guilty than most of pushing this fantasy. In the unmitigated enthusiasm of AI innovation and solutions™, the promise is that AI will do the whole thing: Reliably get the kids to soccer practice via self-driving cars. Or land the plane. Or remove your appendix. Or remove your appendix while landing the plane.
But alas, we can’t. At least not yet. There are three big reasons:
Reason 1: No Context
The prompts above are essentially end results—in other words, goals. AI does not work back from the end goal. It tends to forget the goal. Imagine a soccer field where, mid-game, the players just stop and get a cup of tea. Not because they’re lazy; they just forgot they were supposed to win. That’s what we’re wrestling with here.
To manage this problem, we now have AI that “reasons”—it tries to produce a set of logical steps towards a goal so it can improve its results. Every step in its thinking has to make assumptions to go to the next step. Sometimes it works, but often it doesn’t—and the result is an artificial context that reads to us like bad judgment after the fact.
Reason 2: No Flow
One of the most under-reported aspects of human intelligence is our ability to take in a constant flow of data and adjust our reaction in real-time. In fact, we don’t even know we’re doing it. Psychologists call it “procedural memory.” From Wikipedia:
Procedural memory guides the processes we perform, and most frequently resides below the level of conscious awareness. When needed, procedural memories are automatically retrieved and utilized for execution of the integrated procedures involved in both cognitive and motor skills, from tying shoes, to reading, to flying an airplane. Procedural memories are accessed and used without the need for conscious control or attention.
Not only are we good at course correcting based on real-world contexts, we don’t even know we’re doing it. It’s orders of magnitude more complex than how LLMs work, because we don’t assemble words together to figure out what to do next. Instead, we’re processing inputs in real time, using our own human LLM to make adjustments constantly.
AI is still in batch mode. The problem with BTO prompts is that we can’t step in and make adjustments as the AI is doing its thinking. Instead, it’s on its own while we sit back and wait for the final result. Reasoning helps, but it’s still a far cry from the raw, real-time algorithmic power of human procedural memory.
Reason 3: No Judgment
AI absorbs public content without judgment. LLM-makers determine what’s in and what’s out, and for now they seem to be focusing on quantity. Their products “learn” from everything—good and bad. They absorb bad answers and bad code, sloppy images and dumb takes from message boards last updated in 2006.
Can AI do brain surgery? It can index and save enormous amounts of neurosurgery info and digest millions of research papers and charts. But the highly specialized skills that a neurosurgeon spends years building are, well, incredibly specific and specialized. They’re a combination of #1 and #2 above—they’re working with context and information in the flow, and that information goes back years, even decades.
That combination is what makes skilled humans extra-amazing, and that particular “learning” takes is deep and narrow. I’m not saying that AI will never be able to do brain surgery! I don’t know what’s coming next. But right now, Sam Altman wouldn’t let AI open his skull, and neither would you. You want knowledge, flow, and judgment.
So What Works?
Am I down on AI, then? Not at all! I love it as a tool that supplements human knowledge, flow, and judgment. But there’s a better way to leverage what we have today, and it involves putting us in the driver’s seat rather than alongside AI as a passenger.
People using AI tools are seeing enormous improvements in productivity, especially coders. But it’s because they keep their hands on the wheel. Harper Reed recently shared his LLM codegen workflow. It’s worth reading because it illustrates how an expert engineer applies a sort of slow-motion version of procedural memory to step through the thinking and work behind building an app. It’s many, many steps, and it directly attacks a lot of the pitfalls that come from just trusting AI to ship you a built to order app.
For now, that’s the way to work with these tools: Not issuing commands and stomping around, but more like driving a strange, powerful vehicle, constantly making little adjustments and tweaks and keeping your destination front-and-center. AI cannot keep its eyes on the road (Waymo cars aside). Built-to-order is still a ways off, but accelerated task completion is here today, and it’s great.
As people discuss AI and how it’s going to gobble up jobs, there’s a bit of irony in revealing that AI really badly needs your human brain to get it across the line. It’s comforting, but also a challenge to all of us to level up. Maybe one day an AI robot will guide another to carefully navigate these steps, but until then, let’s keep our hands on the wheel and go.