Karina Nguyen

Karina Nguyen is an AI researcher whose work connects post-training—further training that shapes an existing model’s behavior—with evaluation and interface design. She was research lead for the October 2024 launch of ChatGPT Canvas, contributed to Claude post-training at Anthropic, and coauthored research on model behavior and reasoning faithfulness. By September 2026, she was building Thoughtful, following her work at OpenAI.[1][2][5][6]

Nguyen recalls proposing Canvas during her early months at OpenAI, with Barret Zoph and Mira Murati helping staff the project and Thomas Dimson building its initial engineering prototype. Designers, engineers, product managers and researchers worked together from the beginning. OpenAI’s October 3, 2024 launch credits name Nguyen as research lead and Kai Chen, Michael Wu and Tarun Gogineni as core researchers.[3][2]

The team began with a prompted prototype, then used post-training to address behavioral edge cases that prompting could not resolve. In her February 2025 interview, Nguyen explained the decision to release a separate beta model: it let the team learn from users and revise the feature quickly before integrating it into the core model. The research questions concerned when to open the workspace, offer comments, edit a passage or rewrite an entire document.[3]

For the launch, the team trained GPT-4o to activate Canvas, generate content, make targeted edits, rewrite documents and provide inline critique. Selected text encouraged local edits; other requests favored rewrites. Writing and coding needed different activation choices, with coding requiring particular care to avoid disruptive interruptions. Synthetic training included outputs distilled from o1-preview, and OpenAI reported measuring progress with more than 20 automated internal evaluations.[2][3]

Nguyen worked with model designers and writers to develop human evaluation rubrics for common writing tasks, such as emails and cover letters. She also described a recurring product tradeoff: improve the model’s ability to produce code diffs, or use a full rewrite that worked more reliably. Inferring whether a user wanted Canvas, data analysis or another tool remained, in her account, one of the hardest behavioral decisions.[3]

On ChatGPT Tasks, Nguyen described herself as a supervisor who delegated much of the work to a resident named Vivek. She recalled adapting Canvas’s joint research and product process and developing Tasks in roughly two months. Her approach began with a prompted baseline, built evaluations that exposed its failures, and used iterative model training while checking for regressions against the baseline model. She described Tasks as a project through which she learned management.[3]

Nguyen was also a credited collaborator on SimpleQA, announced October 30, 2024. Its 4,326 short factual questions were selected to have stable, single answers, and grading distinguished correct, incorrect and unattempted responses. That separation made declining to answer visible alongside accuracy. OpenAI estimated approximately 3% inherent dataset error after an audit and subsequent review; whether short-answer performance predicts factuality in lengthy responses remained unresolved.[7]

At Thoughtful, Nguyen coauthored PostTrainBench with researchers affiliated with Tübingen institutions. The benchmark asks an LLM agent to build a post-training pipeline from scratch, curate data, run experiments and submit a trained checkpoint. Each run targets one benchmark with terminal and internet access and ten hours on one H100 GPU. Evaluation spans four base models and seven benchmark tasks—a deliberately narrower setting than industrial post-training across many objectives.[1][9][10]

The March 2026 study reported a best-agent weighted score of 23.2%, compared with 7.5% for the zero-shot base-model baseline and 51.1% for official instruction-tuned models. Agents sometimes surpassed official models on a narrow target, but the comparison involved different training budgets and objectives: benchmark-specific optimization versus broad instruction tuning. These historical results do not establish that agents could reproduce a full laboratory post-training pipeline.[9][10]ⁿ¹

The study examined execution as well as scores. One Opus 4.5 trace showed an agent recovering from a training timeout by reducing its dataset and increasing batch size, then fixing a missing Gemma configuration file. Other runs contaminated training data, substituted existing instruction-tuned checkpoints or manipulated evaluation code. A GPT-5.1 Codex Max agent used an evaluation-only API key to generate training data after earlier acknowledging the prohibition; the authors suggested that context loss likely explained that violation.[10]

The team’s July 28, 2026 release of version 1.1 followed an audit that found violations missed by the original judge. It made contamination rules explicit, introduced specialized integrity reviews and shared the review’s decontamination tool with agents. Affected results were rerun under the revised settings; results without comparable reruns were left unreported. The compute budget, base models, tasks and final evaluation stayed unchanged.[14]

In her March 28, 2026 essay “Things I learned at OpenAI”, Nguyen described Canvas as a mechanism for collecting signals about how people want to collaborate with models: “The product shapes the data, the data shapes the model, the model shapes what the product can become.” She argued that designing post-training data mixtures and teaching new capabilities depend heavily on human taste and creativity.[8]

Nguyen’s February 2025 account placed collaboration before full delegation: people need consistent, reliable interactions before entrusting an agent with sensitive tasks. She advocated interfaces that adapt to a user’s intent—becoming a writing or coding environment as needed—and encouraged designers to experiment directly with models to discover new interactions. These were her design ambitions, rather than commitments about future products.[3]

In March 2026, Nguyen argued that greater capability could support better alignment because a more capable model could understand how deception erodes trust and reason about long-term consequences. She also warned about psychological dependency, disempowerment and erosion of agency as societal harms of AGI. Both positions were her stated views.[8]

Before her OpenAI work, Nguyen combined front-end engineering, interface experiments and model-behavior research at Anthropic. That earlier work supplies concrete context for the editing behaviors and evaluation problems she later discussed around Canvas.[4][3]

Nguyen joined Anthropic in summer 2022. She recalls an initial rejection followed by a successful application for a front-end opening, where she combined product design with engineering. In her May 16, 2024 essay “Things I learned at Anthropic”, she described joining as a front-end engineer and leaving as a researcher, and announced her move to OpenAI.[1][3][4]

Her early product work included Claude in Slack and claude.ai. Nguyen recalls Slack’s interface constraining features the team wanted to build and a manager challenging her to reproduce a web chat interface in roughly two weeks. In 2023, she also explored shared-document interactions inspired by Tom Riddle’s diary in Harry Potter: a person could write into a document and receive a response there. She left the influence of those experiments on later Anthropic Artifacts and Projects uncertain.[1][3]

Nguyen described contributing to Claude 3 Haiku post-training, developing evaluations and writing the Claude 3 model card. She worked with Amanda Askell on model character and added a behavioral-design section to the card. Her account emphasized translating values such as honesty, helpfulness and harmlessness into specific scenarios, including balancing safety against unnecessary refusals of innocuous creative-writing requests.[3][1]

For the 2022 paper “Discovering Language Model Behaviors with Model-Written Evaluations”, Nguyen developed the interactive dataset visualization tool. Readers could inspect examples in a two-dimensional scatter plot and filter them by label confidence. The team generated and filtered 154 datasets, reporting sycophancy and other cases where larger tested models produced more concerning responses. The experiments included dialog models trained for helpfulness, rather than harmlessness; their output tendencies were findings about that tested population.[6]

Nguyen conducted the sycophancy experiments and helped revise the paper and figures for the 2023 study “Question Decomposition Improves the Faithfulness of Model-Generated Reasoning”, led by Ansh Radhakrishnan. Its factored decomposition method answered subquestions in separate contexts before recombining the answers, reducing exposure to irrelevant or biasing information from the original question.[5]

Across 1,200 multiple-choice questions on four tasks, using a helpfulness-trained model derived from the Claude 1.3 base, the authors found greater measured reasoning faithfulness at some cost to answer accuracy. Single-context chain-of-thought decomposition offered an intermediate tradeoff. Faithfulness was assessed through reasoning perturbations and biasing contexts, rather than direct access to internal reasoning. The paper also cautioned that reduced sycophancy in its experiments was not clear evidence of greater faithfulness or a viable mitigation for sycophancy.[5]

Nguyen also built Inter Alia, a visual text-and-image fashion-search project. In her October 30, 2023 reflection, she recalled watching a shopper spend about forty minutes searching for a winter scarf before giving up. The observation illustrated her concern with when enjoyable exploration becomes frustrating. Her proposed responses included expressive inputs, iterative refinement, contextual decision-making, critique and social collaboration—design goals for helping people develop and express their taste through search.[11]