Karina Nguyen

Researcher — OpenAI Researcher (post-training and evaluation) — Anthropic Engineer — New York Times Designer — Dropbox Designer — Square

Listen — profile

0:00 / 2:58

Profile

Karina Nguyen is one of the more unusual careers on the AI frontier: a designer-turned-researcher who has shipped at the center of both leading labs, and who thinks about models less as artifacts to benchmark than as behaviors to be engineered. She joined Anthropic in 2022 as its first designer/front-end engineer and left it a post-training researcher — a trajectory that tells you most of what matters about how she works. She crosses the usual boundary between “the people who make the model good” and “the people who make the product usable,” and treats that boundary as the actual problem.

At Anthropic she worked on post-training and evaluation for the Claude 1–3 models, on model behavior questions like honesty, harmlessness, and hallucination, and on early products including claude.ai and Claude in Slack. She is closely associated with the 100K-token context window feature — the document-upload capability that, for a stretch in 2023, made Claude the obvious choice for anyone who needed to reason over long files. She also co-authored a run of influential behavior-evaluation papers (“Discovering Language Model Behaviors with Model-Written Evaluations,” the chain-of-thought faithfulness work) that helped make sycophancy and unfaithful reasoning legible as measurable phenomena rather than vibes.

In May 2024 she moved to OpenAI — a decision she framed publicly as hard but timely — and landed in the middle of the reasoning-model era. She contributed across post-training, reinforcement learning, and product on the o-series reasoning models (o1/o3), GPT-4o, and two of the interfaces that reframed what “using a model” looks like: Canvas, the side-by-side writing/coding surface, and Tasks, the scheduled-agent feature. She also worked on SimpleQA, OpenAI’s factuality-and-calibration benchmark. Her throughline is what she calls behavioral engineering: designing the model’s disposition and the interface around it as a single co-designed system, so that reasoning and agentic capability actually reach a person’s hands.

As of 2026 she has left OpenAI to build her own venture, Thoughtful (thoughtfullab.com), and her recent research — including PostTrainBench, on whether frontier models can automate their own post-training — points at where she’s headed: making the machinery that produces good models itself more autonomous. For developers, Nguyen is worth following precisely because she refuses the usual split. She’s a credible source on RL and synthetic data and on why the interface is where capability becomes useful — and she teaches both, from a UC Berkeley course on post-training to guest lectures at Stanford.