The Twilight of the Chatbots
The article argues that chatbot technology, particularly large language models, is hitting fundamental limitations. Despite impressive early capabilities, these systems struggle with reliability, reasoning, and truly understanding context, leading to a "twilight" period where the initial hype fades and their practical shortcomings become more apparent.
Background
- This piece is by Ethan Mollick, a Wharton professor who researches AI and startups, and writes the popular "One Useful Thing" Substack about practical AI impacts.
- It argues that chatbot-style interfaces (a text box where you type a question and get a text answer) are giving way to "agentic" systems — AI that takes action on its own (e.g., booking a flight, writing and running code, managing a workflow) rather than just chatting.
- Key context: Since late 2022 (ChatGPT launch), the public has mostly experienced AI as a conversational bot. But behind the scenes, companies like Anthropic (Claude), OpenAI (GPT, Operator), and Google (Gemini, Project Mariner) are racing to build "agents" that can use tools, browse the web, and complete multi-step tasks autonomously.
- The "twilight" refers to the idea that the chatbot era is ending not because the tech is failing, but because it's being subsumed into more powerful, less visibly chatty AI systems. Mollick also flags risks: agentic AI is less predictable and harder to control than simple chatbots.