Skip to content
Dandelion seeds drifting over a meadow, linked by faint threads of light like a network

The science

How understanding actually forms.

Decades of research, in four charts.

What makes learning stick

Average effect size · 0.4 moves an average learner to about the 66th percentile

What makes learning stick: average effect size of each technique
TechniqueAverage effect size
Recalling, not re-reading0.50–0.61
Explaining in your own words0.55
Feedback (most when it explains why)0.48
Mixing problem types (less for math steps)0.42
Refuting misconceptions0.41
Working it out, not reading it0.40
Trying before being taught≈0.36–0.58
Dashed ends show a range. Lowest of all: rereading and highlighting. Rowland 2014 · Adesope 2017 · Bisra 2018 · Wisniewski 2020 · Brunmair & Richter 2019 · refutation-text meta-analyses · Bertsch 2007 · Sinha & Kapur 2021 · Dunlosky 2013

AI that answers vs. AI that asks.

Later exam score, vs. no AI

Later exam score compared with no AI, in standard deviations
GroupLater exam score, vs. no AI
AI that answers−0.19 SD
AI that guides−0.01 SD · about the same as no AI
High-school math, about 1,000 students. Bastani et al., PNAS 2025

Learning gains, and time it took

Learning gains and time taken, AI tutor compared with an active-learning class
GroupLearning gains · time
Excellent active-learning class1× · 75 min
AI tutor built never to give the answer2×+ · 49 min
Harvard intro physics, 194 students. Kestin et al., Scientific Reports 2025

Why we come back

Illustration, not data

How well you remember

No re-checks
1 day3 days7 days14 days

How well you remember

No re-checks
1 day3 days7 days14 days
Each re-check (dot) restores the memory, and it fades more slowly after. Spacing and sleep do the work. Cepeda et al. 2008 · Smolen et al. 2016

Inside your brain.

You have to produce itGenerated beats given.
New sticks to oldPrior knowledge speeds memory.
Spacing is biologySpread practice makes it last.
Sleep does the filingOvernight, learning settles in.
Curiosity opens the doorWanting to know helps it stick.
Surprise correctsConfident mistakes, once caught, fix best.

What we don't claim.

We don't model your neurons.

We don't sort you into “learning styles.”

We don't inflate results.

Sources

  • Adesope et al. (2017). Practice testing.
  • Bastani et al. (2025). PNAS.
  • Bertsch et al. (2007). The generation effect.
  • Bisra et al. (2018). Self-explanation.
  • Brunmair & Richter (2019). Interleaved learning.
  • Cepeda et al. (2008). Spacing effects.
  • Coe (2002). It's the effect size, stupid. Reading an effect size as a percentile.
  • Dunlosky et al. (2013). Effective learning techniques.
  • Gruber et al. (2014). Curiosity and memory. Neuron.
  • Kestin et al. (2025). Scientific Reports.
  • Metcalfe and colleagues. Hypercorrection of confident errors.
  • Pashler et al. (2008). Learning styles.
  • Refutation-text meta-analyses. Correcting misconceptions.
  • Rowland (2014). The testing effect.
  • Sinha & Kapur (2021). Productive failure.
  • Sleep and memory consolidation (Born lab and others). Nature Neuroscience.
  • Smolen et al. (2016). Nature Reviews Neuroscience.
  • Tse et al. (2007). Schemas and consolidation. Science.
  • Wisniewski et al. (2020). Feedback.
Next chapterThe manifesto →