Anthropic researcher, context and memory systems
Karina Nguyen
Profile
Karina Nguyen is a researcher at OpenAI, where she leads post-training work in Model Behavior and has become one of the most visible voices on what it actually means to ship useful AI products. Before OpenAI, she spent about two years at Anthropic, joining as roughly employee 60 — hired as the first designer and front-end engineer — and leaving as a researcher when the company had grown past 500. Her career path is unusual: she came up through product and design stints at Notion, Square, Dropbox, Primer, and the New York Times, and did visual-forensics work at UC Berkeley investigating war crimes with Bellingcat and Amnesty International before ever touching a language model.
At Anthropic she worked across the stack on Claude 1, 2, and 3 — post-training, evaluations, Constitutional AI, cost-efficient models, and the 100K-context document uploads feature that shipped claude.ai’s early moat. At OpenAI she has been behind several of the products developers now take for granted: ChatGPT Canvas (the first fully synthetic-posttrained feature), Tasks, SimpleQA, and streaming chain-of-thought for the o1 reasoning models. Her current research bets are on RL environments, synthetic data, and the “behavioral engineering” of agents that co-create with people rather than just answer prompts.
What makes her worth following is the angle she writes from. She treats AI as an interface problem as much as a modeling problem — how a model remembers, what it refuses, how it shows its work, how it hands off to a human — and she’s one of the few practitioners who has been in the room for both of the leading labs’ post-training stacks. Her Semaphore essays (“Things I Learned at Anthropic,” “Things I Learned at OpenAI,” the reasoning-cost piece) are a rare inside view on how frontier teams actually make decisions.
For developers building with AI today, she matters because she’s working on the layer where capability turns into usability. The models keep getting smarter; whether they’re useful is a post-training and interaction question, and that’s her turf.
Key Articles & Papers
Things I Learned at Anthropic Things I Learned at OpenAI The Cost of AI Reasoning Is Going to Drastically Decrease Is Reducing AI Hallucinations a Capability or a Safety Feature? SimpleQA Discovering Language Model Behaviors with Model-Written Evaluations Question Decomposition Improves the Faithfulness of Model-Generated Reasoning Towards Measuring the Representation of Subjective Global Opinions in Language Models Cultures of WritingVideos
Spotify Podcasts
