PostHog: not every task needs an LLM
PostHog's Jina Yoon writes that decision models can be up to 444.6x cheaper and 193.6x faster than frontier LLMs, returning typed answers with explicit probabilities. In its Replay Vision product, two yes/no questions screen session summaries for personal data and only scores above 0.4 trigger a Gemini rewrite; in production 98.7% of requests skip the LLM call entirely.