Three ways I learn with AI (interjectedfuture.com)

You can learn from LLMs — but only if you use them as disciplined tools, not passive sources. The author outlines three productive patterns: (1) a patient teaching assistant for reading academic papers (upload PDFs, ask for plain‑language explanations, concrete examples or domain metaphors, then spot‑check claims against the original passage, web sources or canonical docs); (2) a rapid ramp‑up for new languages, frameworks or ecosystems (have the model generate a working first draft for a real project to absorb syntax, conventions and "administrivia" quickly — code is often self‑verifying via compilers/type checkers); and (3) a conversational partner for software architecture (use iterative prompts so the model generates options, pros/cons, counterarguments and rebuttals, then critique and refine). Technically and practically, this approach requires active verification and prompt engineering: instruct models to push back (avoid sycophancy), ask for counterarguments, and record failure cases as system eval traces for later error analysis. Be aware of common failure modes — hallucinations, median/over‑engineered designs, and shallow generalizations — even in advanced models (the author cites GPT‑5 and Claude Opus). The takeaway: LLMs lower activation energy and accelerate iteration, but they’re assistants not replacements; maintain the human-in-the-loop for critical thinking, verification, and building lasting expertise.