Conference Presentation, Keynote, Other
What's next for AI agentic workflows ft. Andrew Ng of AI Fund
- AI capabilities are expected to expand dramatically in the current year driven by agentic workflows, representing a "natural evolution" from single-agent to multi-agent systems that mimic expert coder and reviewer roles.
- Human workflows will shift to delegating tasks to agents and waiting minutes or hours for responses, as the speed of token generation in agentic loops is critical for enabling rapid iteration that outperforms human reading speeds.
- Four specific design patterns are planned for categorization and adoption to achieve a quick productivity boost: reflection, planning, multi-agent collaboration, and tool use.
- Reflection patterns are considered robust and can generate improved code versions (e.g., v2 or v3) through iterative conversations, whereas planning and multi-agent patterns remain "emerging," "finicky," and not always reliable at this moment.
- Multi-agent systems, such as those in ChatDev, are anticipated to improve in generating complex programs like "GoMoki" through extended iterative conversations, though outcomes depend on the base model (e.g., GPT-3.5 vs. GPT-4) and specific implementation.
- Agentic workflows may yield better results than slower, higher-quality single runs by using slightly lower quality LLMs for faster token generation, allowing for more iterations within the loop.
- Early tool-use capabilities originated in computer vision, with large language models previously being "blind to images" until developments like GPT-4V and Lava; upcoming models including Cloud 4 (Claude 4), GPT 5, and Gemini 2.0 are expected to enhance these building capabilities.
- Running applications on early models with agent reasoning may achieve performance levels approaching theoretical "GPT-50 shot" benchmarks on specific applications, although results are not guaranteed and depend on the underlying model quality.
- The path to AGI is characterized as a long journey where agent workflows take a "small step forward," with research agents already being utilized for tasks that take a few minutes, acknowledging a success rate that varies.
- While agentic workflows outperform standard zero-shot prompting on benchmarks like Human Eval, the term "agents" has been overused in abstract reports, prompting a focus on concrete design patterns rather than vague predictions.