In 2010, Karthik Dinakar and Birago Jones, while working on a class project at Media Lab, developed a tool targeting content moderation for social media. The project, aimed at identifying harmful posts on platforms such as Twitter and YouTube, landed them a presentation at a White House cyberbullying summit. They realized before the event, however, that their tool was not identifying certain posts because the developers did not understand slang or indirect language. This led to the idea that those who best understand the data should be the ones creating these machine-learning models rather than simply machine-learning engineers.
This insight resulted in the development of Pienso, an AI point-and-click tool which allows non-experts to create machine-learning models for various applications, including detecting misinformation, human trafficking, and weapons sales. Today, Pienso is being used to create large language models without the need of code writing.
Alongside Dinakar and Jones, local Cambridge students also helped in the initial training of the models, which led to a major breakthrough. The realisation that empowering domain experts was crucial to the process gave rise to Pienso.
Dinakar and Jones met at MIT’s Media Lab and started the work on what is now Pienso during their Natural Language Processing course. The project gained momentum again in 2016 when Dinakar concluded his PhD at MIT and deep learning gained popularity. They credit MIT’s Industrial Liaison Program (ILP) and Startup Accelerator (STEX) for early-phase partnerships.
An early partner, SkyUK, utilized Pienso to devise models to understand customers’ problems. Pienso’s models now process half a million customer calls a day, reportedly saving the company over £7 million by reducing call times.
During the outbreak of Covid-19 in the U.S. in 2020, government officials used Pienso to comprehend the novel virus better. Models were developed from thousands of research articles about coronaviruses, reportedly helping the government identify and bolster critical supply chains for drugs.
Furthermore, Pienso can run on internal servers and cloud infrastructure, providing an alternative for businesses that might otherwise surrender their data to other AI companies. The interface of Pienso, described as an Adobe Photoshop for large language models, streamlines the import, refinement, preparation, analysis, and structuring of data necessary for deep learning, enabling a fine-tuned large language model in around 25 minutes.
Pienso also announced a partnership with GraphCore, providing a quicker, more efficient platform for machine learning. The partnership is expected to reduce latency dramatically, making AI more accessible. The founders believe Pienso enables effective AI models developed for specific needs by the people who understand the related challenges best. They emphasize that no single model can address all needs, making the orchestration of a mix of models crucial, with those who grasp the data best leading the process.
