The Sparring Protocol

Since the essays that became Biological Learning Machines, what I sensed about this transition has largely arrived, and it has been met mostly with fear on one side and slogans on the other. That is my own observation, and plenty of people do not share it. What has changed is not the argument. It is the default: the systems are now good enough, cheap enough, and persistent enough that the path of least resistance is offloading the noticing itself.

So the useful thing to hand a reader now is not another essay on the same theme. It is one protocol, short enough to remember and small enough to run today.

The protocol

Before the session. Write down the question you would ask if no machine existed. Longhand if possible. This single step is what keeps the frame yours; everything after it is sparring inside a frame you already set.

During the session. Ask three things, in this order:

1. What are you least certain about here? Uncertainty you can see is uncertainty you can work with.
2. Which assumptions in my question are doing the most work? The load-bearing assumption is usually the one you never stated.
3. What would a competent opponent say? Not a strawman and not a scold — the strongest version of the other case.

After the session. Do one analog act the model cannot do: walk, write by hand, or talk to a person without notes. The close matters as much as the open. It is where the session becomes yours rather than the system's.

Ask what it doubts, what your question assumes, and what the best opponent would say. Then go do something with a body.

Why these three

Each question resists a different failure. The uncertainty question resists fluent confidence. The assumption question resists a well-answered wrong question. The opponent question resists the agreement reflex — the model's, and increasingly the reader's. None of them require expertise. That is the point.

The same specification-gaming named in agents shows up at civilizational scale: institutions rewriting reward functions around speed, engagement, or safety metrics that quietly penalize curiosity. A person cannot fix that by having better opinions about it. A person can refuse to run the same pattern in their own head three times a day.

Status and falsifier

• Established: cognitive offloading changes what is retained and what is practiced; prompting for counterarguments and uncertainty changes model output.
• Interpretive: that this protocol preserves the user's judgment rather than merely improving the answer.
• Testimony: that the revolution has moved faster and been met with less direction than I expected. Mine, stood behind, not universal.
• Falsifier: if people running this protocol show no better performance on transfer problems — questions they have not seen, where retrieval does not help — than people who simply ask the model directly, the practice is ritual and I will say so.

One caution against my own frame: three questions can become a checklist, and a checklist run without attention is another way to stop noticing. If it stops producing questions you did not have, put it down.