Only three kinds of AI products work (www.seangoedecke.com)

An industry take: despite the explosion of “AI products,” only three LLM-driven product types reliably deliver real user value today — chatbots, completions, and agentic systems — and each has distinct technical tradeoffs. Chatbots (e.g., ChatGPT) remain the dominant consumer product but are commoditized: users often default to the core models and big labs win by owning model improvements and safety/UX tuning. Niche “explicit roleplay” chatbots survive by using permissive open models, but face ethical risks and eventual competition as major labs relax content guardrails. “Chatbots with tools” are brittle because tool access invites jailbreaks and chat is a poor UI for routine tasks. Completions (GitHub Copilot et al.) succeed because they inject into existing workflows as powerful autocomplete, avoiding conversational overhead. Agentic products — especially coding agents enabled by models like Claude Sonnet 3.7 and GPT‑5‑Codex — work because agents can autonomously execute multi-step workflows and be validated via tests/compilation. The broader implications: startups should favor embedding AI into existing UX (completions, feeds) or build verifiable agentic flows rather than re-skinning generic chat. Feeds (personalized, model-generated content) and domain-specific research agents look promising because they require no new conversational habits and leverage user interaction signals and embeddings. Games and fully generative video feeds remain experimental due to development cost, UX mismatch, and social backlash, but could become significant as models and safety practices evolve.