What Will Stop AI’s Momentum?
It’s not a question of if but when
If you like this article, check out another by Robbie:
Why Artificial Intelligence is Different from Previous Technology Waves
In an earlier article, I wrote about how artificial intelligence is different than previous technology waves regarding its potential for innovation. It centers around what I refer to as a fully distributed innovation model that is unique among major technology trends in recent memory.
In addition to having nearly unlimited potential, AI has had a tremendous amount of hype in the past few years. Some of the hype is warranted, but much of it has been overstated.
When you are in the middle or nearing the top of a technology’s hype peak, it can be difficult to imagine problems ahead. Given the prolonged hype cycle we’ve seen with artificial intelligence and related technologies, it feels like we should be nearing a peak, but I’ve felt that way for the past 18 months. Gartner’s 2016 Hype Cycle agrees (see Machine Learning):
It is possible AI, on the whole, has already started down the Trough of Disillusionment. Timing these things is not an exact science since it is largely a subjective exercise. Whether we have or not, it is all but inevitable that we will see some backlash or “correction” to the AI hype. It’s not unexpected, or even a bad thing especially since in a lot of cases the hype has far exceeded the current state of the art.
But I’m not worried about that. I’m more interested in the long-game. Can some of the impressive achievements of the last few years be sustained and continue to follow a fully distributed innovation pattern? Or will things go in the other direction and end up in another AI winter? It’s hard to predict, but there are a variety of factors that can impede the current rate of innovation, which I’ll review next.
Patents become preferred over Papers
One of the big advantages that AI has over previous technologies is its strong research background. In computer science, innovation is measured in terms of published papers and that has carried forward with AI. Google, Facebook, and the other large AI-focused firms hired away top AI talent from universities, and that talent has continued to publish papers.
While academic papers are often unnecessarily difficult to read and jargony, they do contain a blueprint for how to implement a new idea or algorithm. That means most of the innovation that’s occurred with AI has been in broad daylight for everyone to see. In some cases, within a few days of a new paper coming out, there are example implementations already on github.
Inside of large enterprises there is a constant tension on whether to publish patents for a particular idea. Given the pace at which new machine learning advances are coming out, often researchers want to get their papers submitted as quickly as possible to stake a claim on “owning” an idea. The Google’s of the world have not held ideas hostage so far despite patenting concepts like word2vec. Hopefully, that continues or it could severely hamper progress in the AI community.
A few big companies hire ALL the AI talent
Much has been said about the tech titans gobbling up AI talent from the major universities. Many of the high profile acquisitions in the past couple years have been trying to land AI-centric talent. While this is a good thing in that it drives up the value of AI skills, it is a bad thing in that the talent gets concentrated in a small number of companies.
Google, Facebook, Amazon, Baidu, Microsoft, Tesla, Apple, and IBM have all been very aggressive in establishing AI centers and hiring as many Ph.D. students as they can find. These companies have agendas, and while most have played nicely in the AI research community to maintain their prestige, clearly their freedom to operate is not what it would be if they were still in the university setting or starting their own companies exploring their own ideas.
If a dozen companies have the lion’s share of AI talent in the world, we won’t see as much diffusion of ideas or new solutions. Most big companies are not known for being the place to go to iterate quickly and build new things. We need a healthy startup ecosystem to inject new ideas.
An extended period without good results
When everyone hears about AI on Good Morning America, a mental stopwatch of sorts starts. The hype can go for only so long until there are results to back it up. Eventually, people will stop believing in the promise of something new and relegate AI to the dustbin of overhyped fads.
What that means is if we don’t hear about new solutions or new results (beyond beating Atari games) that are significant, people will turn their attention elsewhere. This has a tendency to happen anyway as the next new shiny object comes out. AI will lose its luster.
Several high profile failures
A likely unavoidable scenario that will slow down AI’s momentum is high profile failures. Whether it is a company in the name of AI not delivering on the value they promised or a new product launch that falls flat on its face, this kind of public failure plant seeds of doubt in the public. Much like there not being any press-worthy results to share, high profile failures like MD Anderson canceling their contract with IBM Watson hurt the perception of AI and turn a once overly eager public into skeptics.
Tensorflow becomes too complex and too dominant
It may seem counterintuitive, but not having a large dominant platform at the core of AI has been a good thing. While by no means easy, implementing new AI advances has been approachable for a small team or even individuals working alone. It’s not like working on a big monolithic piece of software like an operating system that requires significant coordination with lots of developers. As a result, when a new paper is published, competing implementations are often built in a matter of weeks if not days.
Tensorflow is the most popular machine learning framework, and it continues to get more powerful and complex. If it gets to the point where it becomes so complex that it is difficult for individuals to improve on, or if Google loses interest in updating it so quickly, it could restrict innovation. While in the near-term putting all our eggs in the Tensorflow basket can be beneficial, in the long-term it can be problematic.
We reach a local maximum with Deep Learning
Deep Learning has reached mythic proportions. It can now help software see better than doctors. It can cook, it can clean, it can do the dishes. Ok, maybe not those things (yet), but so much attention has been put on deep learning that a real concern becomes that attention won’t be spent in other areas that might lead to the next breakthrough. Deep learning has limitations, so we need to keep exploring new ideas and new concepts.
Societal and/or political pressures impede progress
My bet on the primary reason that AI loses momentum has nothing to do with the technology or the ability to create innovative solutions, but instead revolve around the public’s interest in seeing those solutions deployed.
It just takes an occasional high profile incident to undue years of trust being built by the public. There will be more incidents/accidents and even if those accidents happen much less than what would have happened with humans, AI will still get a black-eye.
Slowing down will be a good thing
I look forward to the days of AI being less mainstream again. While at a certain level all the attention it receives is a good thing because it makes individuals and companies interested in working on new advances, it also brings a lot of distractions in the form of hype and bandwagon followers that take and never give back to the community. Also, I worry that the greater the hype, the bigger the fall eventually.
If you like this post, give it a ❤️ below so others may see it. Thank you!