Sarah Pan, teenage AI wizard

2025-03-18 · Source: Jeremy Howard · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

Sarah Pan, a high school student, achieved a NeurIPS publication by developing an RLHF pipeline for multi-step reasoning, building on OpenAI's reward model verifiers. Her work, initiated in early 2023, improved mathematical and logical coherence in large language models by grading individual steps in a reasoning process. This approach predated similar advancements seen in OpenAI's 01 models. Pan's journey into AI began at 13-14 with fast.ai courses, fostering a non-linear learning path. She later joined the MIT Primes program, mentored by Vlad Leen, and became an Answer AI fellow, contributing to WebGPU puzzles with Austin Huang. Her experience highlights the value of self-directed learning and early, practical engagement in AI.

Key takeaway

For AI students or aspiring researchers considering traditional academic paths, your demonstrated portfolio and self-directed learning are paramount. Focus on building practical skills and contributing to real-world projects, as this can open doors to top-tier research opportunities and fellowships, even without advanced degrees. Embrace non-linear learning to accelerate your expertise and impact in the field.

Key insights

Early, self-directed, top-down learning in AI can lead to significant research contributions and career opportunities.

Principles

• Top-down learning accelerates complex skill acquisition.
• Process-oriented feedback improves AI reasoning.
• Online communities foster self-driven learning.

Method

Sarah's research extended OpenAI's reward models by integrating process-based verifiers into an RLHF pipeline to update a completion model for improved multi-step reasoning.

In practice

• Adopt top-down learning for new technical domains.
• Seek mentorship for early-stage research projects.
• Utilize online forums for community support.

Topics

• Multi-step Reasoning
• Reinforcement Learning from Human Feedback
• Large Language Models
• AI Education
• WebGPU
• Answer AI
• NeurIPS

Best for: Research Scientist, AI Student, AI Scientist, Director of AI/ML

Related on AIssential

• If I Were a Student Today, I Would Learn These 5 AI Skills — Focus on practical, outcome-driven AI skills to future-proof careers, rather than broad, unfocused learning.
• It’s the Lessons We Learned Along the Way. Or, Is It? — AI agents can efficiently complete complex data tasks, but may diminish human understanding and critical thinking development.
• Guide me to become an AI Engineer — AI tools can personalize learning roadmaps and provide self-assessment for career transitions.
• Almost 80% of Australian uni students now use AI. This is creating an ‘illusion of competence’ — AI use in education can improve short-term performance but undermine deep, long-term learning through cognitive offloading.
• The Most Expensive Mistake AI Beginners Make — Action and rapid iteration in AI development are more valuable than extensive preliminary research.
• Harvard Professor: CS50, What Matters More Than CS, Lecturing Well | David J Malan — Foundational CS education, emphasizing problem-solving and engaging pedagogy, remains vital despite AI's rise.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Jeremy Howard.