In 2010, Karthik Dinakar and Birago Jones, students at MIT Media Lab, built a tool for their class project aimed at aiding content moderation teams in companies like Twitter and YouTube. The duo was invited to demonstrate their invention at a cyberbullying summit at the White House. The night before the event, Dinakar found that he couldn’t make the tool properly identify concerning posts on twitter because he was unable to spot the use of teenage slang and indirect language by the posters.
Jones and Dinakar then realized that such models should be designed by those who best understand the data involved. Thus, they went on to develop point-and-click tools enabling non-experts to create machine-learning models. This tool emerged as Pienso, a system used today to develop large language models that can detect misinformation, human trafficking, weaponry sales, etc.
The team got help from Cambridge high school students for data verification while demonstrating an early version of their system at the White House. Jones and Dinakar decided that the makers of these models should not be AI engineers, but those who comprehend their data best.
Pienso developed as a part-time venture until 2016 when Dinakar finished his Ph.D. at MIT. The two entrepreneurs credit MIT’s Industrial Liaison Program and Startup Accelerator for connecting them with early partnerships. One such partnership was with SkyUK, which used Pienso to build models to understand and manage their customer call center, saving them over £7 million to date.
During the Covid-19 outbreak in 2020, Pienso was used by government officials to glean information from coronavirus research articles. The system helped the government tighten critical supply chains for drugs, including remdesivir.
Pienso, able to run on both internal servers and cloud infrastructure, offers businesses an alternative to giving out their data in order to use AI services from other companies. The tool enables quick data refining and preparing for deep learning analysis, providing a fine-tuned, large language model within 25 minutes.
In partnership with GraphCore, Pienso aims to lower the entry barrier to AI use by cutting latency considerably. Jones and Dinakar, through Pienso, foresee a future where the development of AI models for specific applications is spearheaded by people who best understand the problems being addressed. They suggest that different problems require different models, and that the data best be understood by those creating the models.
