291 - Playing to Strengths
Humans most often dramatically underestimate just how much of the real work they end up doing where LLMs and similar technology are concerned. You still have humans doing the heavy lifting of:
-
Taking in real-world complexity and context
-
Figuring out what their problems are within that
-
Choosing what to do about those problems
-
Figuring out how to define and implement those solutions
-
“Banging the rocks together” until they get a Stupid LLM to do what they want, or at least a plausible imitation of said goal.
This predictably leads to them being underwhelmed by the results when they cut themselves out of the loop and systematic failures run up a steep bill, a naivety epitomized by the trending focus on “AI agents” today. LLMs aren’t reliable for any of the above functions, so you get exponentially increasing risks, losses, and costs if you try to cut humans out under the vast majority of real-world circumstances.
Each of those steps is critically important for viable results, and neglecting multiple steps with naïve automation attempts causes multiplicative and irrational forms of drift, with humanity’s fastest and weakest substitute for cognitive bias doing the driving where LLMs are mistaken for decision-makers and action-takers.
Of course, our list of options goes well beyond these two. No one step of this process needs to rely purely on humans, nor should any step rely purely on simple tools like LLMs. Among this particular list, humans are frequently both best suited for and most comfortable performing number 3, with 1, 2, 4, and 5 all being better suited for human-like digital and scalable systems because:
-
Invokes choice overload and other cognitive biases in humans as complexity scales.
-
Requires deep introspection and awareness in humans, as much of need is unconscious.
-
Requires only selecting the best option from a set, like a dish on a menu.
-
Often requires broad and updated subject matter expertise, experience, and time.
-
Benefits strongly from greater experience and more direct forms of integration, including better vetting of outputs.
No matter what you duct tape to a Muppet, it will remain a Muppet, and by using that Muppet you’ll just be doing your best impression of Gollum from LOTR, talking to yourself in 2 different voices. The future looks nothing like the fever dreams that people were sold under the umbrella of “GenAI” or “Agents”, but that future is coming nonetheless. If you want to continue LARPING LOTR with your best Gollum impression, no doubt Peter Theil will be right there beside you, but for the rest of us there are better things to do.
