Understanding, not slop, is what's interesting about LLMs (blakewatson.com)

LLMs aren’t just noisy content factories — their most interesting trait is an emergent, practical ability to model user intent. Beyond well-known strengths (summarization, code synthesis, tutoring) the author highlights concrete gains: AI assistants that reliably map natural questions to relevant docs (anecdotally, Cloudflare’s assistant), improved speech-to-text pipelines, and tools that let models act on transcribed commands. Technically this isn’t human understanding, but it’s a qualitatively different level of language processing that lets systems interpret intent, retrieve targeted resources, and perform multi-step tasks — provided integrations and safety controls are good enough. The implications for AI/ML are profound: better intent modeling can massively improve accessibility (voice-driven coding with Talon/Cursorless, Whisper transcription, hands-free messaging) and let people interact with computers in true natural language rather than “speaking code.” At the same time, risks remain — hallucinations, brittle integrations, copyright and scraping harms, environmental cost, and the need for safe delegation when models act on users’ behalf. The takeaway: LLMs’ “understanding” potential could reshape human–computer interaction and accessibility, but realizing that promise requires careful engineering, guardrails, and responsible data/use practices.