Karina Nguyen

Karina Nguyen is an AI researcher who has worked at both Anthropic (post-training and evals for Claude 3; 100K context windows; file uploads; early Claude-in-Slack prototypes) and OpenAI (Canvas; Tasks; Frontier Product Research). One of a very small number of people to have worked at both frontier AI labs at a senior research level. Originally a front-end engineer and designer (New York Times, Dropbox, Square); switched to research after realising Claude would surpass her coding skills.

Key ideas

• Model training is more art than science. Data quality, self-knowledge contradictions, and debugging model behaviour require craft as much as method.
• Synthetic data for product behaviours. Canvas was built by identifying three core behaviours and training them synthetically using o1 — cheap, scalable, and generalises from defined examples.
• Form follows function. Product form factors (file upload, notification, calendar) make model capabilities accessible; capability alone is not enough.
• Soft skills outlast hard skills. Aesthetic judgment, creative reasoning, management prioritisation, and emotional intelligence are where models remain weak.
• Personal model trajectory. The evolution from chatbot → collaborative agent → asynchronous agent → personal model that learns and predicts user preferences.

Appearances

See also

• Reinforcement Learning from Human Feedback
• Evals
• Context Window
• Agentic Engineering
• LLM OS