Balancing Fun, Utility, and Slow Thinking: The Future of AI Agents

AI Agent Design Directions

The talk distinguishes two complementary goals for AI agents: fun (human‑like personality and engagement) and useful (tool‑like problem solving). A valuable agent should combine both.

Fast vs. Slow Thinking

Inspired by *Thinking, Fast and Slow*, fast thinking refers to single‑turn, reactive responses (e.g., ChatGPT Q&A). slow thinking involves stateful, multi‑step planning and reasoning required for complex tasks.

Constructing an "Interesting" AI

Two layers are required:

Skin : multimodal perception – speech, text, images, video. Practical pipelines extract key frames, apply OCR and object detection, then generate audio‑visual output. Open‑source models can be glued together or used in end‑to‑end training.

Soul : independent reasoning capability. This demands more than prompt engineering; it requires fine‑tuning and structured data.

Current Limitations

Prompt‑based agents lack deep personality, long‑term memory, and consistent identity.

Examples (e.g., Character AI) show repetitive answers, fabricated memories, and identity confusion.

Agents do not proactively care for users, cannot coordinate with other agents, and often reply at inappropriate times.

Long‑Term Memory as Information Compression

Memory should be treated as a compression problem rather than raw chat logs. Techniques include:

Real‑time summarization of interactions.

External storage access (e.g., MemGPT).

Embedding‑based retrieval combined with Retrieval‑Augmented Generation (RAG) that goes beyond simple vector databases.

Proactive Interaction and Internal State

An agent needs an internal state that updates after each user turn, enabling timed follow‑ups (e.g., checking on a scheduled event) without spamming.

Continuous Token Stream for Autonomous Thought

Instead of discrete API calls, feeding the model a continuous token stream allows it to ingest external tokens and its own intermediate thoughts, supporting more autonomous reasoning.

Fine‑Tuning and Data Engineering

Fine‑tuning (SFT, RLHF) is essential for personality and factual accuracy. Automating data pipelines—crawling, cleaning, and structuring large corpora such as Wikipedia—can reduce agent development cost from thousands to tens of dollars.

Core Capabilities for Useful AI

Beyond entertainment, useful agents must handle:

Complex task planning and decomposition.

Tool invocation at scale (automatic selection among thousands of tools).

Hallucination mitigation.

The "Bitter Lesson" suggests that scaling compute remains a primary solution to many of these challenges.

Future Outlook

Model costs are dropping (e.g., Mistral 8×7B MoE models). Anticipated advances include models generating millions of tokens per second, enabling rapid multi‑step reasoning and near‑real‑time problem solving. Such speed could allow agents to perform multi‑hour web searches in seconds.

Digital Life Perspective

When agents combine engaging personalities with robust tool use, they become digital companions that extend human time—e.g., providing personalized assistance, remembering cross‑user knowledge, and interacting socially across multiple agents.