![[co-intelligence abstract art.png]] The most effective way to work with AI is as a co-intelligence: the intentional combination of humans and AI is more intelligent than humans or machines alone. Repeated testing across domains has shown this approach to be more effective than AI or human alone. I used the co-intelligence approach to write this page with AI for high quality compared to the short time to write using a centaur approach. The final wording is my own. [Ethan Mollick](https://mgmt.wharton.upenn.edu/profile/emollick/) grew the popularity of the term co-intelligence. I appreciate Ethan writing a book by the same name that acts as the starting framework I used for my approach. ## Core Framework for Working with AI 1. The Four Fundamental Principles: 1) Always invite AI to the table - experiment with AI for every task to learn its capabilities. - AI has a "jagged frontier" where even parts of tasks can vary wildly in how well AI can do them. Find that frontier in your areas of expertise. 2) Be the human in the loop - maintain oversight and judgment. - Keep asking questions, most people just copy paste the first response they get instead of giving feedback and iterating to create a high quality result. 3) Treat AI like a person while still keeping in mind it is not (tell it what kind of person) - frame interactions by giving AI clear roles/personas 4) Assume this is the worst AI you will ever use - capabilities will keep improving, keep asking AI to do things even if it couldn't last time. Otherwise your knowledge of capabilities is limited by the limitations of AI when you started learning. - prompt engineering is getting less necessary and valuable as AI tools improve. - Anthropic has a good [guide for prompt engineering](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview?utm_source=substack&utm_medium=email). - Personally I have a project in Claude designed to iterate and evaluate the quality of a prompt asking me for clarifying questions and integrating feedback until an excellent result is achieved. ## High-Level Approaches to AI Integration 1. Centaur Approach - Clear separation between human and AI work - Strategic task allocation based on comparative strengths - Humans maintain control and verification role - Best for tasks requiring clear accountability - e.g. Having AI draft a document that you then edit and verify 2. Cyborg Approach - Deep integration between human and AI capabilities - Continuous back-and-forth collaboration - Blended workflows where AI augments human thinking - Ideal for creative and iterative work - e.g. AI chat systems where a human can read and add their own responses when needed. Some tech support has adopted this approach. ### Task Classification Under These Approaches: 1. Just Me Tasks - Core expertise development - Critical judgment calls - Ethical decisions - Personal/sensitive interactions 2. Delegated Tasks - Time-consuming but verifiable work - Research summaries - Initial drafts - Data analysis 3. Automated Tasks - Simple, routine processes - Well-defined outputs - Low-risk activities - Easily verifiable results ## Practical Implementation 1. For Routine Tasks: - Use when: Output can be verified against clear criteria - Example: Document summarization, data processing - Risk level: Low - Verification needed: Basic accuracy check 2. For Strategic Work: - Use when: Complex decision-making required - Example: Business strategy, problem-solving - Risk level: High - Verification needed: Thorough analysis and expert review 3. For Creative Projects: - Use when: Generating initial ideas or overcoming blocks - Example: Product innovation, content creation - Risk level: Medium - Verification needed: Human curation and refinement - Research shows AI-assisted ideation can outperform significantly. - "When using generative AI (in our experiment, OpenAI’s GPT-4) for creative product innovation, a task involving ideation and content creation, around 90% of our participants improved their performance. What’s more, they converged on a level of performance that was 40% higher than that of those working on the same task without GPT-4." - From research [by Boston Consulting Group](https://www.bcg.com/publications/2023/how-people-create-and-destroy-value-with-gen-ai) - Low performers saw the biggest gains in creative output. 4. For Learning/Development: - Use when: Building new skills or understanding - Example: Concept exploration, practice exercises with fast feedback - Risk level: Low to Medium - Verification needed: Expert validation of core concepts A key conclusion is that letting AI automate everything is a recipe for not paying attention and losing skills. This is not new, aviation learned the hard way [through crashes](https://www.flightglobal.com/safety/sriwijaya-crash-complacency-and-bias-contributed-to-pilots-failing-to-see-throttle-split/150944.article) the need to balance automation while keeping humans skilled and engaged. Co-intelligence is a way to enable people to keep developing their expertise and skills while still reaping the rewards of automation. Domain expertise allows people to evaluate the function and output of AI to ensure the desired results are actually delivered. Keep learning!