Sam Altman (OpenAI) says helpful agents are AI's killer app. What will the path there look like?

Some potential options: - Scale the present: Take the current LLM architecture and scale it up to even more human data (text, image, video, audio, and more). - Embedded minds: Attach LLM and other architectures and connect them to the real world; let them learn directly from the world without being constrained to human input/output. - Social interfaces: Smarter chatbots that start answering questions, then start automating requests, and build around human value. What are others?