Human Knowledge, Machine reasoning
The Blau Monuments are a pair of inscribed stone document from ancient Mesopotamia. Based on the proto-cuneiform script, these documents are dated by some authors as early as 3100 BC.
“Information: the negative reciprocal value of probability.”
In a quest to grasp the governing laws of the universe, each generation writes a chapter in the book of human knowledge. If you look at long gone civilizations like the Sumerians, they made inroads in their understanding of the universe and tried to preserve and pass on that knowledge. The natives of ancient Mesopotamia must have struggled with this fundamental question: is there an abstract “thing” that we can send forth that describes the achievements of our civilization. Symbolic language has been used as a vehicle to encode this knowledge. It's fascinating how much we know about ancient ancestors from what they've left behind that describe the features of their lives. Yet it's really hard when looking at archeological abstracts to decode what they meant. The notion of meaning changes from yesterday to today and there's nothing invariant about meaning.
Language has emerged as the mechanism through which we exchange knowledge among our human counterparts and across generations. I have a thought in my brain, I’m going to turn it into language or symbolic representation which will subsequently catalyze a thought in someone else’s brain. Yet, information loss during (d)encoding is an idiosyncratic trait of human languages. Can you recreate a thought so it's consistent with the original? Contrast that with the deterministic way in which machines make knowledge.
Knowledge is a human enterprise grown through language. It’s our attempt to capture a definition of the universe. We encode the things we care about, while leaving out all the configurations of the elements in the universe that don’t seem relevant. Knowledge does not characterize an arbitrary configuration of the universe, it's a well curated collection of elements our civilization has found most relevant. Through language, we transform the complexity of the universe into a symbolic representation of what is going on. That's a very human way of making knowledge. With the advent of AI, we have come to realize just how unskilled we are at communicating knowledge. Our animate companions can testify to this. Our inanimate companions, the computers, are even more baffled by our language as they retort back with a passive aggressive “Sorry, I didn't quite get that.”
With AI, we are trying to bridge what humans understand, what computers can parse, and what the universe looks like. But such bridges require lots of ingenuity when built against a complex landscape like life sciences.
The drug development process is a long-winded affair. One starts with an assumption about what might lead to a promising cure, and the more accurate the assumption the higher the success rate. Going from the initial assumption to a commercial drug takes years or even decades of work and piles of research reports compiled for posterity: clinical trial reports, drug development reports, FDA fillings etc. These reports are written to capture successes or failures, and pass down critical knowledge to the next Columbus who might undertake the journey through the land of molecular biology.
Just like Columbus lacked a map of the world he was venturing into, researchers lack a map of the molecular world they are looking to tame. A partial map can be assembled from all the research reports and articles published around the world. These reports are written by humans for humans and can only be decoded by humans at human speed. In theory we could leverage everyone under the sun to sift through these documents in real time. In practice one needs specialized domain knowledge to grasp the meaning of a research report.
In a time of pandemic, we come to understand just how critical AI is for language understanding. Imagine being able to instantly discover relevant facts from the world’s repository of pharma research (in English, Mandarin, Spanish, you name it) and get recommendations for the next research step. One can imagine this technology would enable a robust evolution of the collective knowledge and catalyze a faster solution to any new pandemic or rare disease.
Past pandemics left a dark page in the world’s history. But they forced the adoption of innovation as humans struggled to adjust to a new reality and prior techniques became redundant. This time around might be the right opportunity to unleash the power of AI and automate knowledge discovery and representation. What will AI automate? AI can automate many things but AI will not automate the goal setting. You can automate the doing but the assignment of goals is not an abstract notion. What is the goal of the universe? We don't have an abstract representation of that.
