Teaching in the Age of AI: Learning Goals and the Goals of Learning

Aaron Bauer's background and origin story

• Began fascination with computers in childhood; identified programming as his primary interest after a summer program in New Mexico involving Python-based asteroid orbit calculations.
• Took an online AP Computer Science course in high school via a retired teacher located on a beach in North Carolina.
• Developed an interest in teaching while serving as a Teaching Assistant at the Summer Science Program, delivering lectures on binary search trees.
• Studied early Fortran scientific computing at the undergraduate level, noting the historical presence of automatic differentiation in code used for atomic explosion and asteroid impact simulations.
• Pursued a PhD in Computer Science after a computer graphics course sparked interest, though the field had matured; shifted research focus to human-computer problem solving and cognitive psychology.
• Completed dissertation research on the game Foldit, analyzing how non-biochemists collaboratively solved protein folding problems that computers alone could not resolve.
• Foldit data contributed to training datasets for modern AI tools like AlphaFold, demonstrating the value of gamified domain learning.

Teaching evolution during and after the pandemic

• Transitioned to online teaching at Carleton College in Spring 2020; initially used a text-only, asynchronous model but pivoted due to poor student outcomes.
• Implemented a "flipped classroom" approach where students consume lecture content via video beforehand to maximize synchronous time for interactive problem solving and discussion.
• Found that live lectures provide a unique psychological focusing effect for students that recorded videos often lack, particularly regarding note-taking and retention.
• Adopted active learning techniques, such as posing questions and neighbor discussions, to combat passive consumption and improve information processing.
• Continued using recording technology post-pandemic to support student flexibility without reducing in-class attendance, which remained high at Carleton.
• Integrated automated feedback mechanisms and "mastery learning" principles, allowing students to re-attempt quizzes until passing, though with fixed deadlines to prevent procrastination.
• Evaluated the potential of Large Language Models (LLMs) for generating automated feedback, noting risks of misleading outputs (false positives) requiring careful framing.

Jane Street's Developer Educator role and hiring strategy

• Recruited from academia to fill a unique role combining 50% software engineering with 50% educational responsibilities.
• The role was created because Jane Street's proprietary tech stack and trading domain require specialized instruction that generalists cannot provide effectively.
• Hiring challenges stemmed from the rarity of candidates possessing both deep teaching expertise and credible, day-to-day software engineering skills in the firm's specific environment.
• Plans to expand the role into specialized areas, including an OCaml Educator, a Machine Learning Educator, and embedded educators for AI teams.
• Goal to evolve into a model where educators are embedded within specific product or engineering teams to provide context-specific training, similar to the role of designers in the firm.
• Emphasis on reducing tool specialization among engineers to foster a "generalist software engineer" mindset across multiple languages (OCaml, Python, proprietary tools).

Technical challenges in the Editors and Tools organization

• The Editors team maintains VS Code, Emacs, and NeoVim, customizing them to serve as the primary "home base" for all development workflows, including code review via the internal iron system.
• Current engineering efforts address performance issues in VS Code when handling massive diffs (e.g., large configuration files) that cause hanging or queue lock-ups.
• Internal integrations often conflict with external editor assumptions, requiring fixes for search functionality and buffer management.
• Python support challenges involve the limitations of existing notebook solutions; the firm is building a custom OCaml-based notebook tool to ensure control and seamless integration.
• Efforts to "de-weirdify" tooling include migrating from custom Emacs extensions to upstream standard open-source packages where possible, while maintaining OCaml-based extensions for critical functionality.
• Concurrency challenges persist in bridging OCaml (used for extensions) with Emacs's native Elisp and asynchronous limitations.

Impact of AI and LLMs on development and education

• AI tools have shifted the software engineer's primary activity from writing code to reading, understanding, and verifying generated code, increasing the volume of code to be reviewed.
• Telemetry data shows an increase in "jump to definition" actions, suggesting developers are engaging more deeply with codebases to understand AI-generated logic.
• The cost of generating plausible code has dropped to near zero, while the cost and cognitive load of verifying it have increased.
• Proposed workflow adjustments include "tutor-like" LLM modes where the model pauses for human input, forcing the developer to maintain mental models of the code rather than passively accepting outputs.
• Strategies to leverage AI in education include using LLMs as patient, on-demand tutors for hints rather than code generators, and embedding pre-computed code segmentation explanations directly into review tools.
• Recognition that the "vibe coding" approach (relying entirely on LLMs) is insufficient for deep learning; students must still perform implementation tasks manually to build sufficient conceptual understanding.

Curriculum design and the Jane Street education ecosystem

• Onboarding Structure:
  • OCaml Bootcamp: A one-week cumulative exercise building a client-server app to teach syntax, libraries, and code review style.
  • Production Bootcamp: Focuses on deployment, monitoring, and configuration practices.
  • Teach-ins: A catalog of short, topical courses on specific libraries (e.g., Bonsai, Market Data) and concepts (e.g., performance, debugging).
• Curriculum Philosophy:
  • Emphasis on "learning goals" defined as observable student actions (active verbs) rather than vague concepts, ensuring exercises directly practice the intended skills.
  • Application of "cognitive load" theory by separating the introduction of new concepts (syntax, tools, version control) before combining them in complex tasks.
  • Shift from abstract problem-solving to teaching practical, immediately applicable tools, leveraging the high baseline problem-solving skills of new hires.
• Internship Program Overhaul:
  • Replaced short, code-generation-heavy projects with open-ended, design-focused assignments that require requirements gathering and stakeholder interaction.
  • Integrated mandatory testing instruction to provide the necessary feedback loops for LLMs and ensure code quality.
  • Introduced formal code review training for interns, including peer reviews and guided practice with "answer keys" to build review skills early.
  • Utilized small-group "pods" for interactive discussion, allowing interns to share LLM usage strategies and troubleshoot together.
  • Adjusted scope and pacing based on feedback, prioritizing testing and design over multitasking to set a better tone for the internship.