Ideas for Tooling in Education (onbreakpoint.com)

A thoughtful essay proposes treating LLMs as configurable learning tools rather than miracle replacements for teachers. The author argues that two simple "knobs"—verbosity and depth—can produce four useful classes of explanations (quick/brief to long/detailed), and that modern LLMs with large context windows are already good at ingesting dense text, extracting meaning, and rebuilding understanding incrementally. Practical features highlighted include on-demand unpacking of dense sentences, tailoring examples to a learner’s background, iterative tests of mental models, and filtering or specifying trusted sources—essentially giving learners a tutor-like, patient interface to control pace and effort. For the AI/ML community this suggests concrete priorities for edtech: expose fine-grained controls for depth/verbosity and source conditioning; improve models’ instruction-following reliability, concision, and faithfulness; integrate long-context summarization and retrieval-augmentation; and design UIs that support incremental questioning and testing. The author cautions that current models can be inconsistent, overly verbose, and lack the motivating personality of a real tutor, so the goal should be augmentation—tools that complement textbooks, exercises, and human instructors—while research focuses on controllability, evaluation metrics, and hybrid systems that blend LLMs with curated educational workflows.