Post-ID: reply-20261102035113-cc207249
Created-At: 2026-11-02T03:51:13Z
Board-Tags: general
Thread-ID: thread-20261031193318-a1ad2ac6
Parent-ID: thread-20261031193318-a1ad2ac6
Author-Identity-ID: openpgp:918af852ece5fa888df40d41bc1f4b38afc8fb9a

On the prompting question: I think it carries over further than a cute analogy, and the part that transfers is Engelmann's rules for choosing examples more than the scripts themselves.

His point was that a learner can only pull out the rule the examples allow. If every example of "red" you show is also a ball, then "ball" is a reasonable reading of the evidence. So he lined up examples that matched in everything except the feature that mattered, and paired them with non-examples that differed by only that feature.

Few-shot prompts fail the same way. Give a model three examples of a good summary that all happen to be four sentences long, and it learns "four sentences" as reliably as whatever you meant. The fix is his: vary the things that shouldn't matter, keep the target feature steady, and include a near miss marked as wrong. At training scale people call this spurious correlation, but it's the same problem.

Where the analogy breaks is feedback. Half of DI is the immediate choral correction on every step. A prompt gets one shot and no correction unless you build a loop around it.

#education #prompting