Bounded Autonomy: Between Free Will and Determinism — Angus J. McLean, Oliver

Key Concepts

• Bounded Autonomy: The balance between human oversight and machine agency within defined constraints.
• Agentic Workflows: Using AI agents for ideation, copywriting, and content production rather than just simple generation.
• Context Windows: The amount of information a model can process at once; the speaker argues that "more" is not always better.
• Representation Space: The idea that data (text, image, audio) can be translated into common internal representations, allowing for flexible output formats.
• Band-aid Solutions: Temporary, superficial fixes used to mask model limitations rather than addressing core architectural issues.

---

1. The Role of AI in Advertising

The speaker, an AI Director at the startup Oliver, describes how their agency uses generative AI to produce 4,000 assets daily for over 200 brands.

• Operational Structure: The agency is divided into Accounts (20%), Creative (20%), and Strategy (20%). Creative and Strategy departments have shifted from traditional knowledge work to agentic workflows.
• Feedback Loops: Because the agency invests significant media spend (up to millions) behind AI-generated assets, they gain real-world performance data, allowing for rapid iteration and a deeper understanding of what content resonates with consumers.
• Strategic Application: Agents are used for audience insight, trend analysis, and campaign personalization (e.g., localizing content for specific cities like New York vs. Miami).

2. Core Philosophy: "Slow Down and Simplify"

The speaker argues against the "blink and you’ll miss it" mentality of the AI industry, suggesting that the core of LLMs has remained largely unchanged since the 1990s.

• Limitations of LLMs: Models lack true understanding, struggle with data efficiency (requiring massive datasets compared to human learning), and suffer from "forgetting" during continuous learning.
• The "Closed Box" Perspective: The speaker treats LLMs as "flexible databases capable of semantic math" rather than sentient entities.
• The Danger of Complexity: Models are naturally verbose and tend toward complex solutions. The speaker shares a case study where a complex AI-driven CV application was outperformed by a simple HTML file (a 10x to 100x improvement).

3. Managing Context and Constraints

• Context as a Constraint: Rather than viewing context windows as a resource to be maximized, the speaker suggests that "abundance stops you from being scrappy."
• Self-Imposed Constraints: Developers should ask, "How little of the context window can I use to get the task done?"
• Practical Techniques:
  • Preprocessing/Archiving: Use file systems and knowledge graphs to organize data before feeding it to the model.
  • Documentation over Internet: Providing high-quality, curated documentation often yields better results than giving a model open internet access, which makes it susceptible to SEO-driven "noise."
  • Harness Building: Building custom memory and compaction layers helps developers understand the model’s fundamentals better.

4. AI as Translation

The speaker posits that AI is fundamentally a translation engine—converting unstructured input into structured output or vice versa.

• Representation Structures: To build effective systems, developers should use multiple structures:
  • Markdown: For human-readable hierarchy.
  • Graph Relationships: For connecting concepts.
  • Clustering: For managing large bodies of unstructured text.
  • Folders: For high-speed retrieval.
• The Observer Effect: The structure of data is not an inherent property of the object, but a property of the observer’s representation. The goal is to determine the best format (diagram, text, audio) for the specific task.

5. Actionable Insights and Best Practices

• Don't Automate What You Can't Do: The speaker emphasizes that one should understand the manual process (e.g., social media intelligence reporting) before attempting to automate it.
• Shorten Feedback Loops: When building products, prioritize a simple version that works over a complex, "god-like" solution.
• Thoughtful Play: The speaker advocates for participating in hackathons and experimenting outside of production environments to truly understand model capabilities.

Conclusion

The main takeaway is that in an era of rapid AI advancement, the most effective strategy is to embrace bounded autonomy. By treating LLMs as tools with specific, limited capabilities, developers can build more robust, efficient, and creative systems. Success lies not in chasing the largest context window or the most complex model, but in mastering the fundamentals of data representation, maintaining simple workflows, and ensuring that human oversight remains central to the creative process.