Karina Nguyen

Timeline

1. By 2022

   Data science at Berkeley

   Originally from Ukraine, she studies data science, with a concentration in geospatial systems, and public policy at UC Berkeley. Working with Berkeley's Human Rights Center, journalists and lawyers, she applies open-source intelligence, data visualizations and machine learning to war crimes investigations. She also interns in software engineering and product design at companies including Dropbox, Square and The New York Times before joining Anthropic in 2022.

2. 2022

   Front-end engineer at Anthropic

   She joins Anthropic in summer 2022 as a front-end engineer and product designer. Her early work covers Claude products such as claude.ai and Claude in Slack, before she moves into research.

3. December 2022–October 2023

   Model-written evaluations and faithful reasoning

   In December 2022, she coauthors Discovering Language Model Behaviors with Model-Written Evaluations, which generates 154 datasets and finds new cases of inverse scaling where language models get worse with size. Her July 2023 paper with Ansh Radhakrishnan and colleagues shows that decomposing questions into subquestions improves the faithfulness of model-generated reasoning. In October 2023, she helps Thomas Liao train the models for Anthropic's Collective Constitutional AI study, which aligns a language model with a constitution drafted by about 1,000 Americans.

4. By 2024

   Post-training Claude 1 to 3

   By 2024, she has spent two years at Anthropic on post-training for Claude 1 through 3 and on model behavior research into honesty, harmlessness and reducing hallucinations. She leads post-training and evaluation work for the Claude 3 models and creates a document upload feature with 100,000 context windows.

5. May 2024

   Moving to OpenAI

   In May 2024, she leaves Anthropic for OpenAI, joining as a Member of Technical Staff to continue working on AGI. At Anthropic, she grows from a front-end engineer at a company of about 50 people to a researcher at one of more than 700.

6. 2024–2025

   Research lead for ChatGPT Canvas

   At OpenAI, she leads research for ChatGPT Canvas and coauthors SimpleQA, a factuality benchmark that measures whether language models answer short, fact-seeking questions. By February 2025, she is a research manager and post-training lead in Model Behavior, working on Canvas, Tasks and streaming chain-of-thought for o1 models.

7. 2026

   PostTrainBench and a best paper award

   In March 2026, she introduces PostTrainBench, a Thoughtful benchmark built with the ELLIS Institute Tübingen, the University of Tübingen and the Max Planck Institute for Intelligent Systems that tests whether AI agents can post-train base language models within 10 hours on one H100 GPU. The paper wins a Best Paper Award at the ICLR 2026 Workshop on AI with Recursive Self-Improvement. Her March 2026 essay “Things I learned at OpenAI” argues that post-training is where the next frontier of AI progress happens.

8. As of October 2026

   Founder of Thoughtful

   She is the founder of Thoughtful, an AI product-research company that develops collaborative AI systems, and an angel investor and advisor to startups and neolabs. In July 2026, she and colleagues at Thoughtful Lab and Paper Instruments release DiligenceBench, a set of 150 equity-research tasks for evaluating AI agents.