Best of: The future of computer-aided education

Episode Summary

Executive Summary: Chris Peach argues that AI should complement, not replace, teachers by improving motivation, feedback, and assessment in education. Through Stanford’s Code in Place and related research, he shows that near-peer human teachers outperform GPT-4 for course completion, while AI can still help with grading, interactive feedback, and even medical testing. The future he envisions is a hybrid system where humans provide relationship, judgment, and empathy, and computers scale feedback and evaluation.

Main Topics: Human teachers as the core of learning (Priority: 5/5): Peach emphasizes that the strongest force in education is the relationship between learner and teacher. AI is valuable, but it should augment the human connection that drives motivation and joy in learning. Near-peer teaching and large-scale classroom design (Priority: 5/5): He describes a model where recently trained teachers, only a few weeks ahead of students, lead small groups effectively. This structure leverages empathy, humility, and shared struggle to improve learning. AI and generative tools in programming education (Priority: 4/5): The conversation covers how generative AI like ChatGPT can write code, changing how coding is learned and practiced. Peach sees these tools as making programming more enjoyable and expanding what learners can create. Assessment, grading, and fairness (Priority: 5/5): Peach explains that computational assessment can help evaluate student work more accurately and consistently, especially for coding. He stresses the importance of fairness, bias awareness, and using grading to understand students rather than just assign numbers. DreamGrader and open-ended feedback (Priority: 4/5): He introduces DreamGrader, a system that interacts with student creations like games or apps to generate feedback on open-ended work. This is presented as a step toward grading creative, interactive projects rather than only multiple-choice answers. Generative grading as a new model (Priority: 5/5): Peach defines generative grading as reasoning from student misconceptions to possible work products, which is easier for humans and potentially useful for AI. Teachers currently outperform neural networks at this task, suggesting a future of human-AI collaboration. Extending educational methods to medical testing (Priority: 4/5): The same feedback and assessment ideas are applied to medical diagnostics, such as eye tests, where better questioning can yield more accurate and faster measurements. This shows the broader relevance of learning science beyond the classroom.

Key Arguments: Motivation and joy are central challenges in education, not just content delivery or tooling. The most powerful learning mechanism remains the teacher-student relationship; AI should preserve and amplify it. Recently trained near-peer teachers can be exceptionally effective if trained and supported well. Large-scale experimentation showed that 15 minutes with a near-peer teacher improved completion by 10 percentage points. GPT access in that same class made students 4 percentage points less likely to finish, even though the conversations were healthy. AI is already transforming programming by making coding faster, more creative, and more enjoyable for professionals and learners. Assessment should move beyond multiple choice toward open-ended, motivating work with rich feedback. Fairness varies by task: coding can be relatively universal, while essays and resumes may embed more bias. DreamGrader shows that interactive student work can be evaluated by having an algorithm 'play' the work and generate concrete feedback. Generative grading works by imagining student misconceptions and predicting resulting work, a task humans currently do better than AI. The likely future is hybrid: AI scales knowledge and feedback, while teachers supply context, judgment, and understanding of the student.

Data Points: Students in experiment: 10,000 - Stanford/Code in Place class used for teaching and grading experiments Teachers in experiment: 1,000 - Near-peer teachers leading small groups in the large class Group size per teacher: 10 students - Teachers were assigned to lead small groups Teacher proximity to student learning: about 6 weeks ahead - Teachers were recently ahead of students in the material Teacher proximity described by host: about 1 hour ahead - Host’s analogy for effective mentorship Course completion improvement from near-peer teaching: 10 percentage points - 15 minutes with a near-peer teacher increased chance of completing the class material Effect of GPT access on completion: 4 percentage points less likely to finish - Early access to GPT-4 reduced completion compared with no access in the experiment Teachers studied: about 4,000 - Total teacher experience in Code in Place used to learn what makes a good teacher Countries represented: 150 - Coding class participants came from many countries, used to argue coding is relatively universal Three-year-old example: 1 book - Peach used LLMs to co-create a printed book with his child Research target in assessment: Open-ended work - DreamGrader and related work aim to grade tasks more complex than multiple choice

Pivotal Quotes: "the single thing that I've seen have the biggest impact on learners is that relationship building with a teacher" — Chris Peach: Explaining why human connection matters most in education "we are really underestimating how great amateurs could be" — Chris Peach: Discussing how recent learners can become effective near-peer teachers "the future is probably going to look like a hybrid" — Chris Peach: Concluding that AI and teachers will work together rather than AI replacing humans

Implications: Education is likely to become more personalized, interactive, and feedback-rich, but not less human. Schools and edtech tools may increasingly combine AI scaling with teacher judgment, while grading and assessment expand to open-ended work and even medical diagnostics.