Diseconomies of Network Scale And What They Might Mean For AI
Do networks suffer from diseconomies of scale?
Happy Sunday and welcome to Technically Sentient. I’m Rob May, a partner at PJC, where I invest in seed stage AI and Robotics deals, and I write this newsletter to highlight things I’m reading and thinking about in the area of Intelligent Systems.
— Interesting Links —
NXP launches an AI ethics initiative for edge AI. EETimes.
Reinforcement learning simulates a fairer tax policy. MIT Tech Review.
An evaluation of named entity recognition models. Anno AI Blog.
Machine learning with tiny data. MIT Tech Review.
How Google uses AI to improve search. Venturebeat.
The era of analog AI computing has arrived. Mythic. (video)
Choosing the right AI business model. Medium.
— Research Papers —
Analogical and Relational Reasoning With Spiking Neural Networks. Link.
Do’s and Dont’s For Human and Digital Worker Integration. Link.
Self Imitation Learning In Sparse Reward Settings. Link.
— Commentary —
Albert Wegner wrote an awesome post recently on how Innovation Upends Extrapolation. The gist of the post is that it is dangerous to extrapolate a trend into the future when you are dealing with complex systems. One of the reasons for this that isn’t directly mentioned in the post is the idea of “diseconomies of scale.” Most students don’t learn this in business school - they just learn the idea of economies of scale. But diseconomies of scale are real. They arise because, at some level of scale, a system gets so large that it becomes tough to manage, and the costs of management make additional growth less economically attractive, or more difficult.
Diseconomies of scales are one reason that, despite all the concern in the mid 2000s that Walmart was going to own the world, they didn’t. Similarly, the concern in the 1980s that Sears would own the world, on in the 1940s that A&P would - were all wrong. Why? Diseconomies of scale eventually catch up to most physical systems.
Facebook, Google, and Amazon though, have gotten so big because diseconomies of scale are much rarer in digital systems with lower marginal unit costs and network effects. I don’t see the natural drag of diseconomies of scale slowing these businesses down. But Wegner’s post made me wonder if we will see a related concept for network effect businesses, something like “diseconomies of network density.” As networks scale, are there things that operate as natural attenuators on their growth and power?
Wegner’s post looked at the urbanization trend, but denser populations in certain areas can make the world ripe for a pandemic like covid. There are downsides to network density. Is the same true for digital networks?
Well, we see in all the talk about the election and social media that disinformation could be one attenuator on scaling network effects for certain types of networks. When barriers to publishing information were higher, and distribution was more local and contained, the value of pushing disinformation was low. Networks like Facebook and Twitter changed that incentive and value structure. And now the results are starting to affect the usage and growth of those tools.
This ties back to where things are going with intelligent systems because, on the one hand, intelligent systems might fix some of these problems for digital networks by working as auto-arbiters that make decisions about these things as good or better than humans. But as intelligent systems learn more, they could have their own new knowledge based networks, which could have their own problems too. And I wonder if there will be attenuation forces that will slow down their power, or if they will grow until they too are manipulated. Will there someday be “diseconomies of knowledge scale”?
These are issues we should think about as we build intelligent systems.
Thanks for reading.
@robmay