by Tom Hayden
Ray Kurzweil has a fun chart in his book How to Create a Mind; It’s a typical exponential-growth looking chart with various labels on it for types of “minds” – insect, mouse, human and then eventually, all humans. A colleague of mine and I routinely poke fun at this when deploying a machine learning, asking ourselves, “Have we reached insect status yet?”, which reminds me of one of my favorite quotes:
A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects. -Robert A. Heinlein
Artificial Intelligence is slowly iterating its way up Kurzweil’s exponential curve but given all the tech hype around it, you would think we’re only a few years away from reaching the pinnacle: full human cognition. There have been tremendous improvements in some AI algorithms, particularly reinforcement learning and expert systems. We’re barely to specialized insect level and perhaps many decades or longer away from full human, but from reading the tech press, you wouldn’t think so.
Take this claim from a recent TechCrunch article:
It now appears that we will be able to achieve Artificial General Intelligence (AGI) sometime around 2025. Technology is clearly expanding at a faster and faster pace, and, by many accounts, most of us will be caught off guard.
They also cite that AI’s have been beating humans for some time—remember Deep Blue? Remember Watson? Or, take this recent Wall Street Journal op-ed on Universal Basic Income:
…it [the list of ______] also includes millions of white-collar jobs formerly thought to be safe. For decades, progress in artificial intelligence lagged behind the hype. In the past few years, AI has come of age.
The journalist cites Google’s Go-playing AI as example that we’ve “come of age.” but the technical leap to go from a Go-playing AI to say, an AI that could construct a building or even solve simple math equations is tremendous. These are specialized AI algorithms designed to solve a specific problem and are not yet easily extensible to other problems.
So if we can’t take Watson or the Go-playing AI and extend it to solving human cognition-like problems, what kinds of problems are we solving? One way to answer this is to look at the tools various big tech companies are using and in some cases, open sourcing.
- Alphabet – Tensorflow – A dynamic platform for building and training all sorts of models, particularly optimized for GPU intensive tasks such as neural networks. They’ve also built some customized hardware.
- Facebook – FBLearner Flow – An internal tool for engineers to plug in datasets, train models and deploy them into production for specific workflows. Additionally, their AI Research team has been open-sourcing packages related to speech processing and image recognition.
- Microsoft – Cosmos – Almost identical to FBLearner Flow, though Microsoft created Cosmos before FBLearner Flow. Engineers can plug in datasets and pick models to build and deploy AI.
- Amazon – Machine Learning – Essentially fblearner flow/cosmos for the general public. Amazon is working on integrating it with iot, and I expect more developments on this product in the near future.
- Tesla – Highly optimized AI models — Part of the “Master Plan” to create a self-driving car 10x safer than manual via “fleet learning.” In other words, Tesla is collecting massive data and using it to build highly optimized AI models.
What do these all have in common? They’re all tools for humans by humans to automate domain-specific workflows (driving a car, identifying a fraudulent transaction, converting speech to text, etc.) We need, at least for the foreseeable future, engineers and domain experts to make this all work.
In a space where AI is domain specific and tailored to simple repetitive workflows, who has the advantage?
- Big Tech companies that already have the machine learning workflow technology and can extend to more general cases. See above list.
- Cloud Tech companies who can build and support the layers required for this to work. I wouldn’t be surprised to see AWS drop some big products in this space before November.
- Consultants who automate workers by implementing some form of Machine Learning. This isn’t all that different from what has already been going on for the last 40 years—consulting firms have long sought to automate expensive workflows for big clients (remember the movie Office Space?) This is just another tool in their arsenal, which includes things like offshoring, robotic automation, etc.
Unless there is a serious breakthrough in the algorithms for generalized artificial intelligence or we find a way to extend the tools out of domain specific fields, I don’t yet see a world where AI can surpass the insect.
Tom Hayden was an engineer on the fraud team at Facebook, and built out the data infrastructure at GrubHub. Tom holds a BA in Telecommunications, Information Studies and Media from Michigan State University, and a Master’s of Science in Information, Incentive Centered Design from the University of Michigan. He was also an NU graduate student in theoretical computer science. Tom currently serves at an EIR at The Garage at Northwestern.