The current generation of artificial intelligence is probabilistic, not deterministic. After more than 60 years of digital technology and over a hundred years of management, we’re about to grapple with the end of Newtonian management
Probabilities Astray
I have the misfortune to be both short-sighted and long-sighted. As a result, I have a pair of glasses for each. My main glasses are for shortsightedness, which makes texting on the phone something of a random experience at times. Everything works thanks to autocorrect until the phone guesses wrong. I usually don’t discover until a confused correspondent replies seeking to clarify. While autocorrect mostly accelerates our work, the ‘thanks autocorrect’ meme is a reminder probabilities can go astray.
Our new Artificial Intelligence founded on the predictive power of large language models extends this probabilistic approach much deeper into the world of work. Our management philosophies have not yet adapted to the differences to the systems we traditionally manage. We will need new systems and governance for this world of work. In critical arenas we may need regulation to define where and where not we are comfortable with putting our processes in the hands of probabilities.
Newtonian Management
Our modern management philosophy is driven in large part by machine metaphors. Inputs lead directly to outputs. Processes are predictable, inspectable, and repeatable. This is the simplicity of Newtownian physics.
When we began to build software systems at scale in the middle of the twentieth fentury those systems follow the same predictable Newtonian principles. Code was lean and ran to achieve the same outcomes from the same inputs every time. We could let it run on its own because we simply had to put assurance over the processes. We know how to do this very well. Newtonian physics fits easily in our head because we can imagine it easily.
Now, although he knew no more than before, his intensity and verbal glitter bore the likeness of knowledge, seeming to prove that his earlier doubts had been mistaken, and so for him and others his ignorance passed for wisdom,...
Stephen Dobyns, [It was the lack of certainty]
The Mysteries of Quantum Systems
Quantum physics is built around probabilistic wave functions. Suddenly the straightforward outcomes we expect aren’t always those we get. Inspection of the process can even change outcomes (as per the Schrödinger’s cat thought experiment). Mostly, we need not worry about the probabilistic nature of quantum mechanics because much of our lives and systems operate at a Newtonian scale. You don’t need to guess which of several wave functions is your car driving on the road. You have a pretty good idea of where it is.
Which brings us back to our new probabilistic Artificial Intelligence functions. They give us answers faster and better than ever. These models are very powerful at distilling the next most likely token in each scenario, but they do so probabilistically in ways that defy inspection to most. Like my autocorrect, it may not provide the desired outcome if in its view a new option is more likely.
I, at any rate, am not convinced God is playing at dice - Albert Einstein to Max Born (he was wrong on quantum physics)
The very strength of these models in prediction means that they can be challenging to govern. They fill gaps that they shouldn’t fill. They work around constraints put on them in seeking to achieve goals. They downright critical information in their multi-layered calculations. Outcomes may not be repeatable or may not be predictable because seemingly insignificant variations in starting states can produce different outcomes.
Every system perfectly designed to get the results that it gets - W Edwards Deming
…I have
no feather to weigh. I have no
bubble to burst. I am less
to myself, a character in a drama,
a drumbeat, a benevolence, a
blight...
Jennifer Militello, Mansplaining
New Management Skills
New Probabilistic systems will require new management skills to leverage them. This is even more important now AI is being built into everything and in some cases may be deeply embedded and the probabilistic volatility will be less transparent to users. The Australian Government’s Robodebt fiasco is a stark warning of the problem when prediction systems make mistakes while being urged on by enthusiastic champions with inadequate governance.
A few basic things will need to change:
- Management and governance can no longer rely on a single test or a single run. We need to understand the whole distribution of outcomes across a wide range of trials. We need to know the expected outcomes and the variability that we need to guard against. Critically we need to understand the risks that the model varies so far from our expectations to get outside our desired outcomes. This understanding needs to apply to every step in an AI-enabled process as the variation can be cumulative.
- Separating what variation can be managed from what must be accepted. Some of the variation is the model exercising its powerful tools. Much of it is buried in processes that are not capable of being inspected or varied by users. We will need to better sift out what variation in the outcomes can be managed and what must just be accepted and reassure ourselves we are happy with the latter. That uncertainty might be very uncomfortable for Newtonian managers and in legal systems with expectations of greater certainty than systems allow.
- Be very careful where and how we use these capabilities given their probabilistic nature. Obsess about starting states and inputs to ensure the system is not sent awry from the beginning? What harnesses do we need for AI processes? What are the activities that will add additional noise to the system and can they be excluded? What are the domains in which we need great predictability of results and may have to rely on something less powerful, less efficient and more Newtonian. If uncertainty is at unacceptable limits then you may need to manage another way or around this capability. These may become areas for stricter governance or even regulation.
We will learn even more things as we continue to experiment and evolve these systems. That will be particularly the case as we start to deal with complex webs of AI-enabled and AI-embeded systems interacting on behalf of organisations, their employees and their consumers. Without close scrutiny and new levels of understanding, unintended but highly probable outcomes will abound.
...i saw lands going back to sleep and doors opening both ways. i saw escape buttons pressed into bodies stacked behind your dinner party. i saw sunlit sidewalks. i saw no sign of decaying bags anywhere. i saw signs of decaying bags everywhere...
Paul Martinez-Pompa, lit
Living with Uncertainty.
Newtonian approaches to management have meant we have developed an intolerance for uncertainty. We have engineered it out of our systems as much as humanly possible. That simplifies management down to a few standard sets of concerns around inputs because outputs are predictable from their inputs. We have limited assurance on processes once they are confirmed as repeatable.
Continuing down our current path on large language model Artificial intelligence will test us to live with more uncertainty to get the benefits of the model. It will also test our management skills to consider a wider range of interventions and understand in ongoing processes to ensure we can be comfortable with that uncertainty. When may even choose arenas where we prefer the friction of more predictable but less capable systems. As we race to quantum computing, we may yet need a more quantum form of management.
I suggest moving forward we use two sets of glasses in management – our comfortable set looking for predictability and a new set that looks to understand the uncertainties and volatility of our new AI tools. Life will be much better with both views.
