CHAPTERS
- 0:00 – 0:30
From “try everything” to proving ROI: the token-burning temptation
Matt Beane frames 2025 as a year of broad AI experimentation and 2026 as the year executives demand measurable returns. He warns that “burning tokens” (using lots of AI compute) can look like productivity while masking inefficiency and low-quality outcomes.
- •Shift from AI experimentation to ROI accountability from boards and executives
- •The “burn tokens” metric can incentivize waste rather than value
- •High AI usage doesn’t guarantee effective work—could be “running in circles”
- •Sets up the risk of confusing activity with progress
- 0:30 – 2:02
Don’t let AI set your quality bar: resisting the B+ default
He argues most AI output trends toward “B+”—good enough to ship, but not excellent—creating a subtle slide in standards. Leaders and experts add value by exercising restraint: choosing what not to do and refusing mediocre directions.
- •AI can flood teams with passable output that erodes taste for A+ work
- •Expertise often means saying “no” to the wrong work
- •Quality selection becomes harder when output is cheap and abundant
- •Organizations need discipline to pursue only the highest-value ideas
- 2:02 – 2:33
The B+ trap becomes deskilling: losing the ability to notice what’s wrong
Beane explains how relying on AI for thinking and writing can prevent learning and even degrade existing skill. When you haven’t internalized the expert process, you miss hidden quality problems and gradually stop developing good judgment.
- •AI-assisted output can conceal quality defects novices can’t detect
- •If you don’t practice the craft, you don’t build the ability to evaluate it
- •Skill loss happens subtly through reduced cognitive engagement
- •The danger is long-term erosion of expertise and judgment
- 2:33 – 3:33
A trillion-dollar deskilling risk: incentives create “fighters” and “coasters”
He scales the concern to an economy-wide issue: without active safeguards, AI adoption can deskill current workers and the next generation. Whether people protect their learning depends on personal disposition and, crucially, workplace incentives.
- •Deskilling is framed as a massive delayed economic cost
- •Individuals vary: some fight to learn; others coast with AI
- •Incentives and social structure strongly shape behavior
- •Output metrics (lines of code, tokens) can unintentionally reward mediocrity
- 3:33 – 4:04
Burning tokens vs. creating value: set the A-/A+ threshold
Using the Jensen Huang quote as an example, he challenges token spend as a proxy for productivity. The real first question should be whether the work is worth doing—and whether it clears a quality threshold.
- •Token usage is an unreliable productivity measure
- •High activity can still be low-value (“calories burned” analogy)
- •Teams must classify work as B+ vs. A-/A+ early
- •Quality gating and idea selection matter more than volume
- 4:04 – 4:34
Reward the B+ killers: leadership must celebrate saying ‘no’
Beane argues leaders should visibly reward people who stop B+ ideas from advancing, not just those who ship more. High standards plus patience enable real innovation and help teams relearn what “great” looks like.
- •Create recognition and promotion incentives for rejecting mediocre ideas
- •Reinforce a culture of restraint and excellence
- •Innovation requires patience, persistence, and repeated “not good enough” calls
- •One A+ win can reset norms and expand what teams believe is possible
- 4:34 – 6:05
Shadow learning: how people rebuild skill when tech disrupts apprenticeship
Drawing from robotic surgery research, he introduces “shadow learning”—rule-bending ways novices learn when new tech blocks normal participation. As experts can do more independently, novices lose access to hands-on practice and mentorship.
- •New tech disrupts not only work but how people learn at work
- •Novices typically learn via participation with experts; tech reduces that access
- •Most struggle within the old model; a few invent “shadow” paths to learn
- •The pattern appears across many occupations, not just surgery
- 6:05 – 7:07
What shadow learners do (and why it’s diagnostic, not a blueprint)
He shares examples: residents operating without a senior surgeon present and consuming far more surgical video than peers. These practices aren’t endorsements; they reveal what learners are desperately trying to preserve to keep developing skill.
- •Shadow learners use unconventional, sometimes inappropriate learning methods
- •Examples: unsupervised operating; extreme use of recorded video resources
- •These behaviors signal missing conditions for healthy skill growth
- •Shadow learning is a diagnostic window into what work environments lack
- 7:07 – 7:37
The Skill Code: the Three C’s that protect skill development
Beane distills the common ingredients behind shadow learning into Challenge, Complexity, and Connection. These are the conditions people try to recreate when formal systems no longer support learning on the job.
- •Three C’s: Challenge, Complexity, Connection
- •A framework derived from patterns across shadow learners
- •Focus is on preserving learning while maintaining productivity
- •Sets up practical implications for individuals and leaders
- 7:37 – 9:07
Challenge: learn at the edge—with expert support for frustration and failure
Skill grows when work is difficult enough to stretch you near the edge of capability. An expert’s presence matters because they help interpret inevitable failures and keep frustration from shutting learning down.
- •Learning requires tasks that are hard, focused, and slightly stressful
- •Expect small failures; they are necessary for growth
- •Mentors contextualize setbacks and track progress over time
- •Without support, challenge turns into discouragement rather than development
- 9:07 – 11:09
Complexity: see the whole system, not just the task
He argues experts-in-the-making engage with the broader system around the focal task—people, tools, constraints, and workflows—and protect time to reflect. Leaders can enable this through structural choices like job rotation, which also improves resiliency.
- •System understanding improves adaptability, quality detection, and idea discovery
- •Reflection time is essential but squeezed by performance pressure
- •Job rotation builds broader exposure and cross-functional competence
- •Leaders often adopt rotation for resiliency, with skill growth as a major benefit
- 11:09 – 13:11
Connection: trust and respect as a core learning infrastructure
Beane emphasizes that deep learning is often tied to a “who”—a mentor relationship where trust and respect flow both ways. This bond fuels motivation, feedback acceptance, and access to the next opportunity.
- •Trust/respect increases intrinsic motivation to improve
- •Mentorship creates opportunity pipelines (functional, not just relational)
- •Senior people also gain meaning from developing juniors
- •Connection is a learnable, protectable condition for skill growth
- 13:11 – 15:12
What leaders must do now: model AI use, learn in public, hire juniors
He calls for leaders to get direct experience using AI, including publicly sharing failures and waste to normalize learning. He also warns against cutting junior hiring and advocates “inverted apprenticeship,” where AI-native juniors teach seniors while seniors teach craft and judgment.
- •Leaders should gather firsthand data by building with AI themselves
- •Share wins and failures openly to reduce stigma and accelerate learning
- •Junior hiring slowdowns are shortsighted in a general-purpose tech shift
- •Inverted apprenticeship enables bidirectional learning and future readiness
- 15:12 – 16:38
If AI eventually beats humans at everything: act together to shape the transition
Beane argues it’s increasingly plausible AI will outperform humans across all tasks, even creative and interpersonal ones, within decades. Because institutions may adapt slowly, he urges immediate collective action to make AI broadly beneficial rather than painfully disruptive.
- •AI surpassing humans across domains is no longer a fringe possibility
- •Institutional adaptation may lag even with long timelines
- •Proactive coordination can make AI a solution rather than a disruption amplifier
- •Goal: a future where benefits are widespread, not concentrated
