More Hanselminutes with Scott Hanselman episodes

AI Preceptorships with CMU's Dr Michael Hilton and Julia Gao thumbnail

AI Preceptorships with CMU's Dr Michael Hilton and Julia Gao

Published 5 Nov 2026

Duration: 00:35:15

Microsoft's PRAISE program uses preceptorship to counter AI's impact on junior hiring, shifting focus from code writing to architectural judgment and critical review.

Episode Description

What does it mean to learn software engineering when AI can already write much of the code? And how do we make sure the next generation of developers...

Overview

The discussion centers on Microsoft's PRAISE program (Preceptorship for AI Augmented Software Engineering), which pairs experienced preceptors with new engineers in an apprenticeship model. The program responds to the concern that AI will collapse the early-career pipeline, as organizations may hire fewer junior engineers if they believe AI can handle the work. Unlike a standard internship or mentorship, the preceptor's explicit job is to develop senior engineers rather than simply ship code.

A key theme is the shifting role of the software engineer. As AI tools handle more of the actual coding, engineers increasingly become reviewers and architects rather than typists. The conversation explores how this changes what deep understanding means: context, judgment, and the ability to critique code become more valuable than raw coding speed. The bottleneck shifts from writing code to reading it, validating it, and ensuring it fits within a broader system.

The discussion also addresses how AI is reshaping computer science education. Instructors report that students use AI for assignments they don't care about while avoiding it for work they do. The conversation proposes that reading, critiquing, and understanding code should be taught as explicit skills, analogous to how literature courses teach reading and critique without requiring students to write the original text.

What If

  • What if you restructured your daily workflow around code review and validation rather than code generation?

    • Move: Propose dedicating your first work block each day to reading and critiquing AI-generated code before writing anything new, treating review as the primary activity.
    • Why Now?: The source identifies that "me reading the code and me waiting for the build are the two things" that now define the bottleneck, not the coding itself.
    • Expected Upside: You could develop the judgment and "code smell" sense the source identifies as the differentiating human skill, positioning you as a reviewer rather than a typist.
  • What if you built a structured preceptor relationship with a senior engineer in your field?

    • Move: Propose a weekly one-on-one where you bring specific design questions and implementation decisions rather than asking for code help, mirroring the PRAISE model.
    • Why Now?: The source describes how the preceptorship let Julia "pick the brains of my preceptor" and learn "how he would think about it and his view," gaining context no AI provides.
    • Expected Upside: You could gain the broader system context that "no infinite context window can possibly understand," accelerating your engineering judgment beyond what prompting alone delivers.
  • What if you treated reading and critiquing code as a standalone practice, separate from writing?

    • Move: Propose setting aside time to read and annotate codebases you didn't write, writing short critiques of design decisions, similar to how literature students read and critique the Odyssey without writing it.
    • Why Now?: The source argues that "reviewing code and critiquing code and understanding code is itself a separate skill from writing code" and that we "need to think about teaching, reading, and critiquing as a separate skill that we directly teach."
    • Expected Upside: You could build the explicit reading and critique muscle the source says is currently assumed to be picked up incidentally, making you more effective as AI handles more of the writing.

Takeaway

  • Make code review your primary daily activity: read and validate AI-generated code before writing new code, since "me reading the code and me waiting for the build are the two things" that now define the bottleneck.
  • Practice "code smell" and judgment explicitly by critiquing code you didn't write, since "a million monkeys and a million typewriters can eventually write software, but it's up to us to decide what's good and what's bad."
  • Map the broader system context before delegating to AI, because AI is "very good at solving that one problem and one repo, but I don't think it is very well currently at situating itself across the broader context."
  • Treat reading and critiquing code as a separate skill to practice deliberately, not something you pick up incidentally while writing, since "we need to think about teaching, reading, and critiquing as a separate skill that we directly teach."
  • Own your PR regardless of who generated the code: take full responsibility for shipping, since "your PR is in your name" and accountability for production code remains with you.

Recent Episodes of Hanselminutes with Scott Hanselman

30 Jul 2026 AI and the Software Supply Chain with Endor Lab's Varun Badhwar

"Software development faces growing security risks from AI-generated code, open-source dependencies, and citizen developers, demanding zero-trust environments, early security integration, and hybrid AI-driven defenses to combat rapid exploitation and supply chain threats."

18 Jun 2026 The space between the Commits with Zed and DeltaDB's Nathan Sobo

The text critiques the dominance of monocultures in tech tools like Chromium and Electron, highlights the limitations of web-based IDEs, introduces ZED as a native alternative, and presents Delta DB, a version control system that tracks granular development processes for real-time collaboration and privacy-focused code evolution.

11 Jun 2026 Braille Is Freedom with Bristol Braille's Ed Rogers

Braille's declining literacy, tactile advantages over screen readers for precision and numeracy, and modern open-source innovations like the Canute Console face ongoing challenges of cost, complex data access, and integration with digital tools, emphasizing the need for community-driven solutions to sustain tactile accessibility.

More Hanselminutes with Scott Hanselman episodes