Jeremy Howard on AI: Augmentation vs. Atrophy

The Psychology of Human Flourishing

Jeremy Howard opens by framing the current AI discourse through the lens of Self-Determination Theory (SDT). He distinguishes between 'hedonia' (frictionless, easy pleasure) and 'eudaimonia' (the actualization of human potential). While modern AI tools often market themselves as productivity enhancers that remove friction, Howard argues that this convenience can lead to 'dark flow'—a state where users feel productive but are actually experiencing an illusion of control. This mirrors the design of gambling machines, where the dopamine hit of 'progress' is decoupled from actual skill acquisition or meaningful output.

The Trap of 'Vibe Coding'

Howard highlights a growing trend of 'vibe coding,' where developers use AI agents to generate massive amounts of code without deep understanding. He cites anecdotes from experienced engineers (including the creators of Flask and the comma.ai system) who found that while they could generate 95% of a project quickly, the final 5%—and the ability to debug or maintain the system—remained elusive. This creates a dangerous feedback loop: the more one relies on agents to bypass effortful practice, the more one loses the foundational mastery required to verify or improve the system's output. He warns that corporate incentives often push engineers toward this 'decay' model because management prioritizes short-term token-based output over long-term engineering quality.

Augmentation as a Historical Tradition

To counter the trend of replacement, Howard situates the future of AI within the lineage of human-computer augmentation, citing Ivan Sutherland’s Sketchpad (1963), Douglas Engelbart’s 'Mother of All Demos' (1968), Kenneth Iverson’s APL, and Brett Victor’s interactive explorations. These pioneers viewed the computer as a 'tool of thought' meant to amplify human intelligence through synergistic structuring. Howard argues that AI should be used to deepen our connection to our work, not to outsource it. He demonstrates this with his tool, 'Solvit,' which uses AI to help users actively engage with complex research papers, forcing them to pause, verify, and replicate concepts rather than simply summarizing them.

Reclaiming Agency in Engineering

Howard concludes that the choice is ours: we can use AI to atrophy our skills by letting it 'do the work for us,' or we can use it to support 'effortful craft.' He emphasizes that because most AI vendors and employers prioritize output metrics, engineers must take personal responsibility for their own autonomy and mastery. True progress in AI engineering, he suggests, is found in tools that act as partners in thinking, allowing the human to remain the primary agent in the creative process.