A CS Professor on Why Slow Learning Wins in the AI Era | CU Boulder, Tom Yeh

Main ideas, concepts, and lessons

• Learning should be about ownership and internal understanding, not just getting answers

  • Having an answer quickly (e.g., from AI) doesn’t mean you know.
  • Degrees/certificates can be purchased, but true learning is about what you internalize and own.
  • The value of knowledge is proportional to the time and effort spent acquiring it.

• “Slow learning” and learning by hand build real comprehension

  • Tom emphasizes that the brain often “gets it” only when ideas are mapped manually (drawing/writing out math and algorithms).
  • Teaching/learning should match human time constraints: you can’t write or process faster than you physically can.

• AI is a tool; humans must build stable foundations that survive tool changes

  • Certain computational/CS foundations are evergreen (the “core” stays useful even as tools/models change).
  • Even if specific models/technologies trend up or down (e.g., older deep learning hype), foundational concepts remain relevant.
  • “A foundation on solid rock” can be reused to build with new AI tools.

• Skill-building over time shapes identity

  • The way someone grew up acquiring skills (piano, soccer, etc.) reflects a learning process that doesn’t change.
  • That skill/identity transfers across future tools: you can apply your ability to learn difficult things to whichever AI comes next.

• Willingness to understand the black box matters more than memorizing formulas

  • What differentiates learners is not immediate recall of transformer math/attention mechanisms.
  • It’s the willingness to open the black box, persist through challenges, and invest effort.
  • Cheating and short-term success won’t build durable capability; people need incentives that encourage genuine learning.

• Cheating is a symptom of incentive problems, not just the presence of tools

  • When one cheating avenue disappears (e.g., Chegg-style solutions), people find another.
  • Even “AI cheating” is a symptom; the deeper problem is that systems often don’t encourage real learning and effort.
  • Hiring should prioritize traits like work ethic, problem solving, and teamwork, because those correlate with genuine learning behavior (including adopting AI appropriately).

• AI cannot change people

  • Hiring “the right person” (problem solver / team player) leads them to adopt AI naturally.
  • AI will not automatically make someone ethical, respectful, or a team player—people must change themselves.

---

Methodology / instructional approach (detailed bullet points)

Learning method: “AI by Hand” / slow, manual comprehension

• Write out the math and algorithms by hand

  • Don’t rely on instant output; manually draw and map the model.
  • Example framing: break down transformer processing at the token level (tokens as inputs represented by multiple numeric values).

• Use drawing as a comprehension tool

  • When concepts don’t “click,” write/draw the structure until understanding emerges.
  • Treat it like a “struggle-to-understand” workflow, not a fast consumption workflow.

• Match learning pace to human limits

  • Learn at the speed your writing/attention can realistically support.
  • Avoid teaching/learning that rushes beyond what students can follow.

• Maintain focus during learning

  • Move away from distractions (e.g., copying by hand reduces time spent switching to unrelated phone/computer activities).
  • Encourage active note-taking as part of the learning process.

Teaching method used in programming courses (blackboard approach)

• Teach an entire semester of C++ by writing on a blackboard rather than live coding.
• Rely on the benefits of:
  • Teacher pace constraint: you can’t go faster than you can write.
  • Student pace constraint: students can only keep up with what’s written.
  • Focus constraint: students using their hands to copy notes are less likely to be distracted (e.g., Instagram).

Career/skills strategy: build on evergreen foundations

• Identify a core foundation that remains useful across technological cycles.
• Treat new tools as replaceable surfaces:
  • If tools change, you keep rebuilding using the same foundational capability.
• Use your personal “learning process skill” (how you learn difficult things) as your durable advantage.

Hiring/education incentive lens (to reduce cheating)

• Focus hiring/assessment on:
  • work ethics
  • problem-solving ability
  • teamwork and communication
• Assume AI adoption will follow from the right competencies:
  • problem solvers will learn AI because it helps them solve problems
  • team players will learn AI because it supports collaboration
• Recognize that cheating persists unless incentives and evaluation methods promote genuine effort and learning.

---

Speakers / sources featured

• Tom Yeh (also referred to as “Tom Diet” in the subtitles; likely the same person in the recording):

  • Computer science professor at the University of Colorado Boulder
  • Founder of AI by Hand

• No other named speakers or external sources are clearly identified beyond general references (e.g., “Chegg,” “Khan too” as mentioned in subtitles, and examples like transformers/matrix multiplication/case references such as CGI/Jurassic Park and a historical Korean palace).