Karthik Dinakar SM ’12, PhD ’17 and Birago Jones SM ’12, former Media Lab students, developed a tool in 2010 to aid content moderation teams in companies like Twitter (now X) and YouTube. While getting ready to present this project at a cyberbullying summit at the White House, the pair discovered an issue with their model. It correctly flagged relevant posts, but failed to recognize teenage slang used within these posts. Dinakar nearly gave up, but Jones discovered the issue was not the model itself, but a disconnect with the language used by the posters they were trying to monitor.
This issue led them to develop point-and-click tools enabling nonexperts to construct machine-learning models, the foundation for their current venture, Pienso. Pienso grants people the ability to design comprehensive language models for identifying human trafficking, misinformation, weapons sales, and more, without any programming experience. They tested the first Pienso prototypes by collaborating with local students to train the models, discovering that these student-trained models outperformed anything they could create.
Dinakar and Jones continued improving Pienso while completing their Master’s degrees at the MIT Media Lab. Despite solid support from the White House, Pienso was a part-time project until 2016 when Dinakar completed his doctoral degree and machine learning’s popularity surged. Their partnership with SkyUK resulted in models that now process half a million customer calls a day and have purportedly saved the company over £7 million by reducing call times.
They were approached by government officials in 2020 to aid in understanding the emergence of Covid-19. Using Pienso, experts in infectious diseases and virology generated machine-learning models, sifting through thousands of research papers concerning coronaviruses. The researchers’ work bolstered critical supply chain routes for important drugs, including the antiviral Remdesivir.
Pienso works with both internal servers and cloud infrastructure. Dinakar describes it as an “Adobe Photoshop for large language models,” allowing users to import and refine data without coding, creating a refined model within 25 minutes. A new partnership with GraphCore, a faster, more efficient platform for machine learning, will decrease latency, making AI more accessible. Dinakar and Jones express Pienso’s goal as enabling experts in various fields to create specific AI models for their unique challenges, believing that understanding the data is key to successful AI models.
