The INSEAD Chaired Professor of Management Science has an unusual achievement: Scientific publications aside, he has published thousands of crossword puzzle books in about 20 languages. According to GROK, Philip M. Parker is a “pioneer in the field of natural language generation” and one of the first to bring generative AI to the fore.
For our first edition of INSEAD’s AI Leaders series, we sat down with INSEAD’s very own AI pioneer to get his views of the field.
Q: Let’s start with the beginning. When did INSEAD make its first foray into this generative AI world?
Well, let’s start by broadening the definition of generative AI. It encompasses all forms of human-machine interaction where the output is content – especially original content – based on some human input, or prompt. Pocket calculators of the 1970s fit the definition. What is the square root of 345? You prompt the machine, and it responds with “18.574". These rule-based generative methodologies have existed for decades and are everywhere around us.
For INSEAD, the most impactful application was probably Markstat, pioneered by our emeritus professors Jean-Claude Larreche and Hubert Gatignon. The multi-player simulation of market forces was used in the classroom, pitting one team against another as they fight for market share. Each team receives computer-generated research reports that respond dynamically to not one, but multiple player inputs. This was a massive breakthrough in the field of education, and was scaled to business schools around the world.
This period of history coincided with another INSEAD pioneer, emeritus professor Spyros Makridakis, who was recognised for creating the field of forecasting – the antecedent to supervised machine learning.
Q: What about your journey in this space?
I was heavily influenced by my grandfather and mother who were mathematicians. But mostly, I was inspired by my brother David B. Parker who formalised the backpropagation method for training artificial neural networks and introduced the word “learning”, as in “machine learning”, in his 1985 paper. From there, others like Yann LeCun in 1985 and Geoffrey Hinton in 1986 started the momentum in this area. Against this backdrop, I joined INSEAD in 1988, dipped my toe in this space and literally crashed the network.
At the time, only researchers at schools with supercomputers or large computer science departments were estimating neural networks. But distributed parallel processing was the rage, so I had a go by inserting some code into PCs across INSEAD’s Fontainebleau campus, running them overnight to minimise disruption. That failed; Everything crashed. I needed an alternative approach.
Around 1990, John D.C. Little from MIT visited INSEAD and presented his famous paper on CoverStory, detailing a method to identify facts and trends from databases (e.g. grocery store barcode scanner data). This concept of coupling databases with natural language is really the birth of natural language generation. But rather than using neural networks, I focused on methodologies that would work on a basic personal computer. By the late 1990s, I got it to work in a semi-automated way, and in 2000, I found a way to scale it.
Q: What were the earliest applications?
Our then-Dean of Executive Education, Arnoud DeMeyer, gave me his blessings to create customised teaching handouts, where each learner was given a tailored version, be it finance in the oil and gas industry or marketing in the fashion industry. The idea is that I could be teaching a core concept, like pricing, to everyone in the same way, but the course notes would vary from person to person. This worked extremely well, but scaling it in a fully automated fashion was the issue. If I could crack that, I could potentially narrow the “content divide”.
Q: Can you explain “content divide” and its implication?
In the 1980s while I was doing consulting, my projects took me to Africa, Asia and Latin America. I noticed then that foreign direct investment might be hindered by a lack of market data, especially for obscure products made in factories in, say, Thailand. Some products, like 6-inch copper nails, are too obscure for any market research to exist. And if there's no reliable estimate during the due diligence phase, there's no investment.
Automation could achieve this at scale, and at a level of quality that passes the Economic Turing Test. To my relief, the algorithm I developed passed the test, meaning it can perform economically valuable work that is indistinguishable from a human doing that work. “Economically valuable”, in this respect, means it could generate sales or replace a paid employee.
I distributed my first-generation research reports, covering topics with an obvious content divide, such as the market for wooden toilet seats in India. Requests to speak to the author followed – but of course, the author was a computer program! By the early 2000s I had generated tens of thousands of research reports this way, distributed through Wall Street intelligence firms. Eventually, books generated the same way – from bilingual crosswords to language learning books and healthcare guides – were made available on Amazon.com. About five years later, the patent was approved. By the time a New York Times article appeared on the subject in 2008, I had created over 200,000 titles. Today it stands at over 1.6 million titles.
Q: Why patent the methodology?
It’s a fun story. The patent application was submitted in 2000, back when patent filings remained hidden from public view. It took about six years, which was not unexpected, given what must seem, at the time, an outrageous claim: that a computer program could generate any form of human-produced content, regardless of its format (text, video, images, games, etc.) or content type (newspaper articles, research reports, books, movies, television, whatever).
I was invited to Alexandria, Virginia to demonstrate the technology to the patent examiners. So, with just a Toshiba laptop, and no internet connection, I Within a few strokes, reports were automatically generated, formatted and ready for distribution. The patent was approved shortly after.
Like my brother Dave who came up with the backpropagation algorithm, I thought patenting a method is rather goofy. I did it purely as a defensive nature, so that no one could come along later and make claims over my research. In fact, I never enforced the patent, and granted full licenses freely to anyone who asked.
Q: Your work has captured a lot of interest, having presented at the White House, the World Economic Forum, the G8 summit and the World Bank. What are some recent developments since?
Shortly after the patent was announced, I was contacted by various universities to work on projects with the goal of scaling knowledge to underserved communities, as I did with my language learning books for smaller language communities. With funding from various governments and foundations, my TotoGEO lab at INSEAD created scientific summaries in botany, reports for farmers, weather forecasts for remote villages who received news via radio and many others. My TedX talk opened more doors to collaborate across various domains.
Also, I’ve recently announced a few initiatives, one of which is Botipedia.org, a platform of encyclopaedic materials. Some of the 400 billion entries are generated on the fly, without using LLMs, and others are performing data reduction using natural language programming methods. Focusing on underserved subjects and languages, amounting to over 200 languages and all known topics across all encyclopaedic sources, Botipedia complements projects like Wikipedia and Grokipedia.
Currently, I am working with INSEAD alumni on solving the “news desert” problem with a “newsroom in a box” concept, and creating McKinsey in a box to provide small and medium enterprises with affordable access to high-end insights. I’m also scaling “big tech” to smaller countries, allowing them to achieve AI sovereignty (i.e. “Google in a box”).
