Why the Environment Beats the Lesson Plan
In this lesson: Byron makes the case that a strong learning environment — doing, feedback, a tribe, and LLMs — drives roughly 90% of an engineer’s growth, while the formal curriculum contributes only 10%.
Top 3 takeaways
Environment beats the lesson plan
The 70-20-10 model shows formal training is just 10%; on-the-job doing and learning from others drive the other 90%.
Doing plus feedback compounds learning
Building beats consuming, and fast feedback cements it — like violinists’ 10,000 hours of focused, feedback-driven practice, not passive repetition.
Your tribe determines whether you finish
Cohort-based programs see 87–90% completion versus 3–15% for solo learning; shared struggle keeps people moving instead of quitting.
Byron Mackay
Director of Learning, Gauntlet AI
Byron Mackay is Director of Learning at Gauntlet AI, training hundreds of engineers to work AI-first. Previously a Principal AI Engineer, he led curriculum development at BloomTech.
Lesson notes
A written walkthrough of the lecture, covering the model, the practice, and the things that trip people up.
The 70-20-10 model
Byron opens with the Center for Creative Leadership's 70-20-10 model: roughly 70% of effective learning comes from on-the-job experience, 20% from interacting with others, and only 10% from formal training. His point as a professional educator: formal instruction is necessary but small, and the real acceleration comes from combining all three rather than living in a build-only silo.
Doing beats consuming
Reading books and watching videos produce ideas and direction but don't make you an expert. Doing does — and it applies to LLMs too. Gauntlet leans on coding agents but pushes engineers to do more with them (timed mini-hackathons, longer planning sessions) so they build new skills rather than just shipping for its own sake.
Doing + feedback
The strongest lever is doing plus feedback. Byron cites the violinist studies: the top performers weren't separated by raw hours but by 10,000 hours of focused, feedback-driven practice. This is why Gauntlet emphasizes "building in public," runs 3–5 feedback-generating deliverables per week, and prizes fast correction loops.
The tribe
Cohort-based programs complete at 87–90% versus 3–15% for self-directed learning (including solo Coursera-style courses). The people around you turn struggle into a shared problem rather than a reason to quit. Gauntlet's "tribe" includes cohort mates, instructors and office hours, engineering experts, and hiring partners — but tribes can also be built via Slack groups, Reddit, meetups, and networking.
Sharing ideas: architecture defense
A recent Gauntlet practice: five people get the same case study, each designs a solution within hours, then presents for five minutes while peers spend five minutes poking holes. It has produced a remarkable difference in the direction people take afterward.
Learning with LLMs
LLMs are a powerful but partial piece of the process. Byron references a fall-2023 Harvard physical-science study where an AI tutor (heavily grounded to avoid hallucination, with a control group that alternated weekly) produced higher engagement, better grades, and stronger explanations. Because LLMs converge on the mean, they get you to average fast — then you springboard. He notes scaffolded learning (quizzes, case studies), a system-design skill he built, and Gauntlet's internal scaffolded AI tutor, "Traverse." The TypeScript "discriminated unions" example illustrates why breadth of knowledge still matters: LLMs won't reach for lesser-known tools unless you know to ask.
The role of curriculum (the 10%)
Curriculum is the smallest slice but still essential — it provides focus, direction, and checkpoints so learners don't stall deciding what's next. Put a strong environment on top of a lesson plan and, in Byron's framing, there's no stopping you.
Q&A highlights
On agentic workflows: context is king. Rather than chasing frameworks, invest in well-placed CLAUDE.md (or AGENTS.md) context files across your codebase, build specialized agents per area (frontend, backend/database), then add an orchestrator once task-giving exceeds what you can manage. Recommended repo: awesome-skills; a favorite is the "grill me" skill for pressure-testing ideas and PRs. On logistics: applicants must be US citizens; the CCAT passing score is 40 (a speed test, ~16–18 seconds per question); graduates typically land AI-focused engineering roles via hiring partners after an in-person Austin interview.
FAQ
What is "Why the Environment Beats the Lesson Plan" about? +
What is the 70-20-10 model of learning? +
Why is doing better than just consuming or reading? +
How does feedback accelerate learning? +
Why do cohort-based programs have higher completion rates? +
How do you learn effectively with LLMs? +
How do you get an AI agent to perform well in your codebase? +
CLAUDE.md (or AGENTS.md) files across directories, build specialized agents for specific areas, and add an orchestrator on top once you have more task-giving than you can handle manually.What is an architecture defense at Gauntlet? +
What's next?
Keep building with the rest of Night School, or apply to Gauntlet — twelve weeks of technical intensity with the best AI engineers we can find.