This article argues that the current approach to building AI tools is fundamentally flawed, deskilling humans and hindering collaborative learning by attempting to automate the very cognitive processes humans excel at. Drawing on theories of human learning—specifically retrieval practice, process-based learning, and cumulative iteration—the author proposes a "human-first" paradigm for AI development. This paradigm reframes AI as an "absent-minded instructor" rather than an "auto-magic" button, advocating for tools that explain, demonstrate, guide, and enhance human actions (the EDGE method). Using examples from incident management, code generation, and cross-functional collaboration, the article illustrates how AI should augment and reinforce human reasoning loops, leading to a positive feedback cycle where both humans and AI continuously improve. The ultimate call to action is to prioritize building AI for collaborative learning and human excellence, ensuring humans remain the central "loop" in problem-solving and innovation.
Summarizing the Transcript: Stop Building AI Tools Backwards
Critique of Current AI Tools: The author contends that the tech industry is building AI tools "backwards," leading to poor design, ethical concerns (e.g., energy waste, unethical LLM development), and ultimately, deskilling human critical reasoning rather than enhancing it.
How Humans Learn and Innovate:
Retrieval Practice: Humans learn effectively by expending effort to recall information, not just by passively receiving it.
Process-Based Learning: The most effective "thing" humans learn is process, akin to learning to bake a cake by doing, not just memorizing ingredients.
Cumulative Iteration: Sustained solo innovation is rare and less significant than collective, iterative improvement. Humans are "turbo optimized for communities," learning via copying, mimicry, and building upon prior work (standing on the shoulders of giants).
Innovation = Problem Solving: Effective problem-solving, propagating learning, and integrating collective knowledge are key to true innovation.
The Problem with "Click AI button -> ✨ magic ✨": Current AI tooling often bypasses crucial human retrieval, task initiation, process reinforcement, and collective knowledge transfer, automating the very actions humans are best at and AI is poor at. This creates a negative feedback loop where humans deskill faster than AI improves, leading to ineffective systems.
Building Better AI Tools: The "Absent-Minded Instructor" Analogy:
AI should be seen as an instructor that guides and helps humans learn how to learn, rather than an intern or co-worker. It's a "very overconfident rubber duck that exclusively uses the Socratic method."
The goal is to build tools that help humans think, not think for them; AI should amplify human effectiveness, not obfuscate it.
The EDGE Method for AI-Augmented Learning (Explain, Demonstrate, Guide, Enhance):
Explain: AI suggests missing steps or pulls up incident process guides (e.g., "have you tried turning it off and on?"). Bad interactions remove human retrieval (e.g., "click this button to perform an action," "explain this error" tooltips).
Demonstrate: AI translates human queries into system syntax (e.g., "top 10 slowest endpoints" into a specific query) or provides short, dynamic demos of task execution. Bad interactions include "auto do the thing" buttons that deskill.
Guide: AI asks clarifying questions, suggests next steps based on human plans (e.g., "You seem stuck on X. Do you want to try investigating Y?"), or prompts for documentation. Bad interactions provide unasked-for information or correct humans authoritatively.
Enhance: AI suggests incremental improvements during or after actions (e.g., filtering by time range -> suggest "five-minutes-before-the-alert-fired"), reveals UI shortcuts, or proposes process/instrumentation refinements. All enhancements should add more recall prompts, embedding micro-learning.
Application to Code Generation: Instead of AI generating code first, it should help with "backwards" processes: generating rough documentation, architecture diagrams, testing plans, and tests before stubbing and generating code. This ensures human validation is based on a structured approach rather than a blank slate.
Untapped Potential: Cross-Functional Collaboration: AI can bridge communication gaps, such as providing immediate rough drafts to customer support during an incident that developers can then review and refine. This allows technical teams to respond in jargon-heavy ways, which AI can then translate and refine for different audiences.
Conclusion: Humans Are the Loop: The fundamental issue is that AI is replacing, rather than augmenting, human strengths like cumulative learning and collaboration. To build effective, valuable tools, the emphasis must be on humanity, reinforcing positive feedback loops where both humans and AI continuously improve. The author stresses that "Don’t just keep humans in the loop, remember that humans are the loop."