Frozen weights or weightless learning. When to use which.
Yes, a GPT-class model can be put inside a humanoid body. The question is what it can do once it is there. Its knowledge is fixed in its weights on the day training ends. In the field it cannot change its tactics between attempts, it cannot tell you why an attempt failed, and it can state things that are not true. If the job needs the machine to adapt and get better outcome after outcome, that is the wrong tool. If the job is your email, it is the right one — and Perslis is overkill.
The short answer
What frozen weights means
A neural model learns during training. Everything it knows is spread across billions of numbers, and those numbers stop changing when training ends. The model on day 100 of a deployment is the model from day 1 unless someone takes it back, gathers new data and trains it again, somewhere else, for days.
Three things follow for a machine in the field:
- It cannot change strategy between attempts. If an approach fails, the next attempt comes from the same weights. It may phrase a new plan, but nothing it learned from the failure is kept.
- It cannot show why it failed. There is no rule to point at, only numbers. It can write an explanation, but the explanation is generated text, not the cause.
- It can make things up. It generates its answer, so it can generate a confident wrong one — in an email that costs a correction; in a machine that acts, it costs the outcome.
What weightless learning means
Peel keeps what it knows as rows and named rules that a person can read, not as weights. When an attempt fails, the failure is recorded with the situation it happened in. One change is tested against it in paired matches and kept only if it wins clearly; otherwise it is thrown away. The next attempt runs on the new rules — no retraining, no data centre.
Because every decision is picked from the moves the orders allow, from facts it read, with a written reason, a failure can be traced to the rule that made it. Nothing is generated, so nothing is made up. And the learning sits inside a floor it cannot widen: it gets better at the mission, never past its orders.
What it is not: the learning happens between attempts, not inside a single fight, and in the tank arena the menu of changes the loop could try was written by engineers. The loop chose which ones to keep.
The evidence, from one arena
In our BZFlag tank arena, five tanks fought each other: BZFlag’s built-in AI, our rule pilot, and three language models driving through the same controls.
- Decision rate. In one ten-minute match the rule pilot made 60,140 decisions (about 100 a second). DeepSeek made 415, llama3.2 213, Claude 58 — at a median of 7.6 seconds each through its command-line tool.
- Knowledge without time pressure. On an eight-question written exam of tank situations: rule pilot 8/8, Claude 6/8, DeepSeek 5/8, llama3.2 4/8.
- Getting better between fights. The learning loop took the hand-written pilot from −1.17 to −0.26 net kills per tank-minute against BZFlag’s AI (z = 2.93) by keeping two changes and throwing the rest away. The models’ weights did not change between matches.
- What it did not do. It still loses to BZFlag’s own AI: 45 kills to 74 in the same time. Its lower death rate (1.53 against 1.73 a minute) is not statistically significant. The claim is that it improved and can show why, not that it wins.
They work together
This is not a choice between neural networks and rules for the whole machine. A trained model is the best tool we have for seeing and reading: it turns camera frames and speech into named facts. Peel takes those facts and decides, inside the commander’s rules, and learns from the outcome. The model is the eye. It does not command.
If all you need is to answer email, write summaries or search documents, use a model on its own — Perslis adds nothing you need there. Perslis is for the machine whose outcome has to get better every time it fails, and whose every decision has to stand up at the after-action review.