What if computers could understand what they read?
I don’t mean “understand” in the shallow sense of recent machine learning wizardry, where systems extract statistical patterns from millions of documents and learn to parrot back similar-sounding phrases. I mean a much deeper kind of understanding: the ability to read about geopolitical events, company policies, or personal symptom histories; to discuss with you what happened and why; and to help you explore your hypotheses and evaluate your options.
The ability to read, dialogue, and reason remains far beyond any existing technology. And yet it represents a long-standing hope for artificial intelligence dating back decades. Countless science fiction accounts envision a future where machines provide expert support for humans by engaging in fluent dialogue as hyper-informed, highly analytical thought partners.
This vision for AI is what drove me to start Elemental Cognition, a company dedicated to creating a new kind of language technology. We are building a system that learns about, understands, and explains the world in terms humans can recognize and engage with.
Through reading and dialogue, our technology continuously learns how the world works. Then it uses that knowledge as a foundation to understand new things it reads, reason about that information, and discuss it with humans. The resulting system is explicable and correctable: it can explain what it knows and why it thinks its understanding is correct, and it can take instruction from humans when it makes a mistake.
We believe that such a system holds enormous potential for humanity:
It will democratize expertise.In time, interacting with the system will become like talking to a human expert — except this expert will have the time to read any written body of knowledge, the patience to discuss what it knows for as long as you’d like, and the objectivity to check and confirm its own understanding.
It will turn computers into powerful thought partners. Instead of making inscrutable recommendations, machines will reason through our decisions with us, ultimately accelerating our understanding and creativity.
Our team has deep experience in AI, including natural language processing, formal knowledge representation and reasoning, linguistics, cognitive science, and machine learning. Most notably, many of us were key members of the team that developed Watson, IBM’s question-answering machine that achieved fame as a Jeopardy! champion.
In my experience building and leading the Watson team, I saw firsthand the power of modern statistical approaches to AI. These techniques catapulted Watson to its landmark success. But at the same time, I knew how little Watson truly understood. It was never designed to produce a deep understanding of the text in which it found and scored answers, and it certainly didn’t achieve the ability to read and understand accidentally.
Even as I moved to Bridgewater to apply explicable AI to managing investments and organizations, the vision of a system that really could understand language continued to dog me. After several years, and with Bridgewater’s support, I launched Elemental Cognition to pursue that vision.
Our name, “Elemental Cognition,” reflects the unique way that our technology works: it assembles its understanding from basic building blocks of human language, logic, and reasoning. The resulting intelligent system is compatible with human cognition — that is, with how we acquire and communicate our understanding.
We have a long road ahead of us as we continue to develop our approach to language understanding. In the meantime, I’ll be sharing some ideas here (and on Twitter and LinkedIn) about why popular approaches to AI are not enough and where the field needs to go to achieve the grand vision of AI as human thought partners.