Lenny's Podcast: Product | Career | Growth

OpenAI researcher on why soft skills are the future of work | Karina Nguyen (Research at OpenAI, ex-Anthropic)

February 9, 2025

AI Summary

5 min read

Karina Nguyen, an AI researcher at OpenAI who previously led post-training at Anthropic, discusses building features like Canvas and tasks, model training nuances, and how AI shifts workplace skills. Drawing from her work on Claude 3 and OpenAI's o1, she explains practical mechanisms for iterating on models and why human creativity and management endure.

Model Training as Art and Debugging

Training large models blends data curation with software-like debugging, prioritizing quality over quantity to instill behaviors like helpfulness without harm. A key challenge is resolving contradictions, such as data implying physical actions (e.g., setting alarms) clashing with instructions that models lack bodies, leading to over-refusals or confusion. Post-training via reinforcement learning scales infinitely by teaching diverse tasks—web search, coding, writing—using synthetic data generated by prior models, avoiding a "data wall" since pre-training compresses world knowledge while post-training refines skills. Synthetic data excels for rapid product iteration but needs human or expert input for niche domains like biology.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

What you'll learn

1 **[04:45] Podcast Intro and Guest Background**
2 **[06:36] Model Training Misconceptions**
3 **[08:21] Data Walls, Synthetic Data, and Post-Training Scaling**
4 **[12:56] Building Canvas: Synthetic Training and Behaviors**
5 **[18:33] Day-to-Day: Evals, Prompting, and Product Work**
6 **[27:06] Developing Tasks and Operational Flows**
7 **[36:08] Future of Work: Skills and Trends**

+ Full timestamped outline available in the app

Show Notes

Karina Nguyen leads research at OpenAI, where she’s been pivotal in developing groundbreaking products like Canvas, Tasks, and the o1 language model. Before OpenAI, Karina was at Anthropic, where she led post-training and evaluation work for Claude 3 models, created a document upload feature with 100,000 context windows, and contributed to numerous other innovations. With experience as an engineer at the New York Times and as a designer at Dropbox and Square, Karina has a rare firsthand perspective on the cutting edge of AI and large language models. In our conversation, we discuss:

• How OpenAI builds product

• What people misunderstand about AI model training

• Differences between how OpenAI and Anthropic operate

• The role of synthetic data in model development

• How to build trust between users and AI models

• Why she moved from engineering to research

• Much more

—

Brought to you by:

• Enterpret—Transform customer feedback into product growth

• Vanta—Automate compliance. Simplify security

• Loom—The easiest screen recorder you’ll ever use