In 2010, Media Lab students Karthik Dinakar SM ’12, PhD ’17 and Birago Jones SM ’12 began a project to develop a tool to help content moderators at Twitter (now X) and YouTube spot troubling posts. While preparing for a cyberbullying summit at the White House, they realized that their model was missing nuances in teenage slang and indirect language. This led them to a key insight: machine-learning models should be built by those who understand the data the best, rather than just engineers.
The researchers then developed point-and-click tools that allowed non-experts to construct their own machine-learning models. These tools became the foundation for their company, Pienso, which today aids in the construction of large language models that detect misinformation, human trafficking, and weapons sales, without the need for coding.
The pair’s efforts began as part of their masters’ degrees at the MIT Media Lab and continued until 2012. They returned to the project in 2016 when deep learning began to gain popularity. The MIT campus and its Industrial Liaison Program (ILP) and Startup Accelerator (STEX) helped connect them to early partners and shape their philosophy.
The team’s collaboration with SkyUK, a leading British telecommunications company, showcased their approach with SkyUK’s customer service team using Pienso to build models deciphering the most frequent customer issues. This application aided in processing half a million customer calls daily, reportedly saving the company over £7 million by reducing call length.
The onset of the Covid-19 pandemic in 2020 marked another milestone for Pienso. The founders were contacted by government officials to use Pienso to analyse the emerging disease better. The tool helped experts in virology and infectious disease create machine-learning models to assess thousands of research articles on coronaviruses, aiding the government in identifying and bolstering critical supply chains for drugs.
Pienso’s web-based platform allows users to import and refine data without coding. It can structure unlabeled or unannotated data, which can then be used to fine-tune large language models in roughly 25 minutes. Compatible with internal servers and cloud infrastructure, Pienso serves as an alternative for companies that don’t want to share their data with other AI services.
In a recent development, Pienso has partnered with hardware manufacturing company GraphCore to improve the efficiency of their computing platform for machine learning, reducing latency.
Dinakar and Jones view their work as paving the way for a future where more effective AI models are created by those who are most acquainted with the problems they seek to solve. The ultimate goal is to assemble a collection of models, each designed for specific applications, brought together to collaborate and orchestrated by those who understand the data best.
