Sasha Sheng is a cofounder and COO of TypeSafe AI. Previously a research engineer at Meta and FAIR, she coauthored benchmarks that test how AI systems connect images, language and sound.[1][3][4]
Start with Human-Adversarial Visual Question Answering →
X Background
Sheng completed a mechanical engineering degree at the University of Michigan. In a 2023 freeCodeCamp interview, she described a software career at Yahoo and Facebook spanning mobile applications and AI systems. After a layoff, she learned the emerging AI tools through hackathons. Chat Out Loud, one project linked by the interview, turns a conversation with ChatGPT into a spoken exchange.[2][5]
Career
At Meta and FAIR, Sheng worked on News Feed, AI experiences and research. TypeSafe lists her among its founders as COO. Her background combines work on deployed consumer software with research into the weaknesses of multimodal AI.[1]
Contributions
Questions that expose visual mistakes
Sheng is a coauthor of Human-Adversarial Visual Question Answering. People interacted with a visual question-answering model and tried to write image questions it would answer incorrectly. The resulting AdVQA benchmark exposed failures that ordinary test sets could hide: strong results on a familiar dataset did not translate into similarly strong performance on these deliberately difficult questions.[3]
A controlled world for multimodal research
MUGEN uses an adapted CoinRun platform game to create linked video, audio and text. Agents navigate the game and interact with objects, while the engine supplies additional annotations such as semantic maps. The team paired short clips with human descriptions and benchmarked retrieval and generation. A controlled game world makes it possible to vary interactions while retaining information about what actually happened in each scene.[4]
Tools, writing & talks
Writing, research & talks
Start here
A benchmark built from image questions people designed to make a model fail.
Start with the abstract and examples to see why a model can score well on a familiar image benchmark yet fail on questions a person deliberately makes difficult. Some familiarity with model evaluation helps; the motivating problem is accessible without implementing the benchmark.
Coauthor[3]
Tools
- Chat Out Loud GPT
github.comAn open-source voice interface for conversation with ChatGPT.