Europe 2026
Bounded Autonomy: Between Free Will and Determinism — Angus J. McLean, Oliver
Overview
This talk explores the evolving relationship between humans and large language models (LLMs), particularly in the context of agentic systems. It argues against the hype surrounding rapid AI advancements, emphasizing that core LLM capabilities have not fundamentally changed. Instead, many current tools act as temporary fixes for inherent model limitations like data inefficiency and a lack of continuous learning. The core thesis suggests that understanding and leveraging these limitations through self-imposed constraints and diverse representation structures can lead to more effective and creative AI applications.
Who should watch
- AI engineers and builders experimenting with agents.
- Product Managers and designers evaluating AI capabilities.
- Individuals overwhelmed by the pace of AI development.
- Experts seeking new perspectives on agent design and LLM interaction.
- Those working in fast-paced, high-risk environments where AI output has significant impact.
Key takeaways
- Focus on fundamental LLM limitations such as data inefficiency and the inability to continuously learn without forgetting, rather than chasing superficial advancements.
- Recognize that many current AI tools are "band-aids" – temporary fixes that mask underlying issues rather than solving them.
- Context windows, while increasing, will never be sufficient due to the exponential growth of global knowledge; therefore, managing and filtering context is crucial.
- Self-imposed constraints, such as intentionally using less of a context window or employing simpler models, can foster creativity and lead to more robust solutions.
- AI fundamentally operates as a translation or summarization process, converting data between different representation formats (text, image, audio, etc.).
- Leveraging multiple representation structures like Markdown for hierarchy, graphs for relationships, and clustering for large text bodies can improve AI system design.
- Prioritize building simple, functional versions of AI applications and shortening feedback loops with reality, rather than over-engineering with complex AI solutions.
- Experimentation and play, such as through hackathons, are vital for understanding and developing effective AI applications.
Notable quotes
*Today's large language don't actually understand the data they're presented with.*
*Context windows keep getting larger, but they'll never be enough.*
*Constraints actually create creativity.*
Unofficial community note. Prefer the recording for nuance.