In 2010, Karthik Dinakar and Birago Jones, two students from the Media Lab at the Massachusetts Institute of Technology, collaborated on a project aimed at supporting content moderation teams for social media companies such as Twitter and YouTube. Their ambition was to develop a tool capable of identifying concerning posts online. Despite some initial struggles, the pair eventually succeeded.
Their breakthrough occurred when they discovered that while the tool they had designed was proficient at flagging potentially problematic posts, it was less capable of correctly interpreting posts featuring teenage slang or indirect language. This realization led the pair to refocus their efforts on developing a solution that could better understand its target data.
They produced point-and-click tools designed to aid non-experts in machine-learning model construction. These tools form the foundation of Pienso, a system today that aids users in building large language models for various purposes, such as detecting misinformation or identifying instances of human trafficking and the sale of weapons. Importantly, Pienso allows users to achieve this without writing any code.
The founders of Pienso attribute the project’s success to the benefit of involving those best versed in understanding specific subject areas in AI development and implementation. They found that models trained by these so-called domain experts, often with more specific knowledge in certain areas than AI engineers, performed better. The pair see this as an important step in diffusing AI capabilities into a wider range of fields.
One early partner of Pienso was SkyUK, a British telecommunications company. They used Pienso to construct models to understand their customers’ most frequent issues. These models now help process hundreds of thousands of customer calls each day.
In 2020, as the Covid-19 pandemic took hold, government officials reached out to Pienso to help understand the virus and its effects. Pienso was used to set up models that mined thousands of research papers for information on coronaviruses. The work helped the government identify critical supply chains for vital medication.
Pienso’s software operates on internal servers and cloud infrastructure, providing data protection for business users wary of donating their data for use by other AI companies. Earlier this year, Pienso announced a partnership with GraphCore, a provider of innovative computing platforms for machine learning, with the aim of lowering barriers to using AI by greatly reducing latency.
Dinakar and Jones envision a future where more effective AI models are developed for specific uses by those with the most experience and understanding of the problem. They argue that every application, need and dataset differs. Therefore, the best approach is to use a range of tailored models, and those who understand the data best should be involved in their construction.