Don’t Judge an Article by Its Title

I was heartbroken when Harvard Business Review editors decided to title my 2003 article “Don’t Trust Your Gut” (link), certainly a buzzier catchphrase than “How To Leverage Your Intuition With Analytics” or other possible variations on the ACTUAL topic of the article. That arrogant-sounding title tainted the piece, suggesting that I was supporting a thesis that was the almost exact opposite of what I wanted to convey. It also made it easy to dismiss: yet another zero-EQ engineer who wants people to become robots.

I did point out that intuition has severe limitations in complex, high-stake, non-stationary situations where the past is not necessarily a good predictor of the future. But it is also irreplaceable when experience, emotions or judgment are required. The article was an attempt at describing ways of correcting for the flaws and biases of human intuition and heuristics by carefully dividing labor between humans and machines.

Why am I so worked up about it NOW? Because the core tenet of that article, and of much of my work in the past 30 years, has been that there is a lot to gain from AI -Augmented Intelligence, if only we can design the right human-machine division of labor and interface. And I would argue that that is a very timely topic. See for example @ethan mollick’s brand new book “Co-Intelligence”.

What I outlined in the article was a simple (or simplistic) approach: first, in the context of a task, figure out what humans are good at and not so good at, what machines can do well and not so well; then figure out a way of outsourcing the human weak points to machines if machines can do it better, and keep the uniquely human value-add with humans. And I argued that, at a high level, humans are good at evaluating solutions but not so much at exploring alternatives, leading to an obvious synergy: use a machine to generate lots of (reasonable) solutions, have the human look at them and decide.

Remember, the year was 2003. Generative algorithms were very different then. But the idea remains: generative AI + human = superhuman, if done well. While I still believe that the exploration-evaluation framework has a lot of merits, exponential progress in #genAI in the last few years has the shifted the boundaries of that division of labor: the possible synergies between human and machine are not as clear cut as they were, which means that understanding their nature is even more crucial.

This is a recurring topic: remember the food replicator? The proposed machine was a generative algorithm connected to a 3d printer and the role of the human was to taste and evaluate printed items.

This young man is about to make a terrible decision.

The year is 1993, he is a 25-yr-old research engineer at what is now Orange Labs in Brittany, in the Neural Networks team headed by Daniel Collobert (Ronan Collobert’s dad), deeply hidden inside the Flat Screens Department. Because, you see, neural networks at that time sounded a bit like voodoo magic (nothing has changed except that voodoo magic is now a corporate priority) to be practiced under deep cover.

This young man joined the team in 1992 with a passion for neural networks and had the fortune of sharing an office with a wonderful, friendly neural networks postdoc from Québec, Samy Bengio. And the young research engineer was hearing again and again about how this Yoshua Bengio (brother of the aforementioned) was doing great work and that their friend Yann LeCun’s backpropagation thing was big. So one day, failing to see how anyone would ever muster enough compute power and enough training data to make it work, this young research engineer decided that other machine learning approaches were a safer bet.

He thought that perhaps evolutionary algorithms were more promising, or perhaps even self-organizing maps or Hopfield networks if one insisted on keeping with neural networks.The young research engineer shifted his attention to other biomimetic learning algorithms. It looked like a great decision for some time. For a little less than 20 years.

But boys were these guys persistent! For a little less than 20 years, algorithms improved and some important innovations emerged, but success by and large came from scale of data and compute. Once the barrier was broken, progress was exponentially fast. And enabled more structural and algorithmic innovation -attention, transformers, pretrained transformers, ... Today training data and compute remain major sources of competitive advantage.

The young man is a wee bit older. He made a 180 on his assessment around 2013. Perhaps there was enough training data and compute power to make it work after all.

Tell me I haven’t changed all that much.