Humans Back In Vogue
“Artificial Intelligence is not the science of building artificial people. It’s not the science of understanding human intelligence. It’s not even the science of trying to build artifacts that can imitate human behavior well enough to fool someone that the machine is human, as proposed in the famous Turing test…AI is the science of making machines do tasks that humans can do or try to do. “ — James F. Allen, Professor of Computer Science, University of Rochester, from “AI Growing Up,” AI Magazine, Winter 1998
At this year’s D conference, Uber CEO Travis Kalanick saidthat human taxi drivers will be replaced by autonomous cars. Basically, cars will soon be driven by artificial intelligence (AI). Hours later, after an uproar, he said it would take decades for it to happen. Seemingly unrelated, the Centers for Disease Control and Prevention (CDC) has reported there are about 10 drowning fatalities a day, many of which happen in private pools. And yet, there are no lifeguards in private pools. The only operating machine-vision based virtual lifeguard costs upwards of $100,000 and requires the deployment of multiple cameras in an electricity averse environment.
These two stories highlight two conditions the emergence of AI and replacement of humans. There is an alternative where humans can still do better and cheaper through a new breed of companies one might call “Shared Economy 2.0”. These companies will provide much of the promise of AI but will be significantly faster to build and capture market share well before their AI counterparts reach operating conditions. They can then innovate from a dominant position in the market.
Science is advancing AI. In recent years, Yann LeCun (Facebook), Geoffrey Hinton (Google) and Yoshua Bengio (University of Montreal) are leaders of significant advances in AI and machine learning due to their introduction of deep neural networks and other deep learning techniques. Yaniv Taigman, my genius partner at face.com, recently published that they achieved a 97.25% recall rate for face recognition, just 0.25% below the human perception. That is an incredible advance in machine learning. Obviously, the applications and implications of these significant advances are vast and the press is gaga about them. The Google car, sensor networks, robots — and other things that supposedly will replace humans in an Asimovian world — are all packed with machine vision and autonomous, computer-based control. It makes a good press story. Man vs. Machine. Human intelligence vs artificial intelligence. We’ve more recently been hearing more sensational predictions from Elon Musk and Stephen Hawking about the development of AI’s possible future.
I would suggest we all calm down and the press may need to take a cold shower. Slate recently suggested that “The autonomous Google car may never actually happen” and Michael I. Jordan one of the key machine learning professors saying “we are no further along with computer vision than we were with physics when Isaac Newton sat under his apple tree”. I agree. It depends a lot on the emergence of AI both absolutely and relatively, and studies have shown that BOTH experts and laymen are very bad at predicting when AI will be good enough.
AI is hard. The brain is the product of millions of years of evolution. It has developed intrinsic capabilities to handle what programmers call “corner cases” — situations not defined by any well understood rules. This is because some learnings can never be fully explained by rules. AI is also engineering that can’t fail. It must be perfect and most software engineering projects are not perfect. Unlike a bridge that can’t be 90% done, software can be “good enough.” Mark Zuckerberg is often-quoted as saying that done is better than perfect. That works in software but not in bridges, and probably not in AI. The reason done is better than perfect in software is because the initial 80% of the reward has 20% cost. The remaining 20% of the reward would take 200% of the cost and time (software projects are rarely on time).
This is where economics comes in. There is a large supply of drivers and lifeguards. However, they are not in the right place for their need. Remember our lifeguard example above, a Tel Aviv lifeguard may be able to earn more working for a furniture store in Tel Aviv so he is relatively underpaid. However, today we live in a world of a global workforce and global cost arbitrage. Median annual income in India is $616. Within the US, median income for Holmes County, Mississippi is $13,794 whereas on the Hamptons it is close to $100,000. The other trend sweeping the global economy is increased remote work-force availability. According to a study by the Global Workplace Analytics, work from home is expected to increase by 63% in the coming 5 years.
With this source of low cost labor, increasing high-speed connectivity and remote sensing, we can actually use people to do the work of the AI endowed robots. And the humans will be significantly faster-to-market and better at it. People can drive the cars — the same way drones are flown from a ground station, be virtual lifeguards or drive the robots to find perfectly ripe tomatoes via a remote video stream. Companies like Drivecam are already using remote analysts in south-east Asia for video analysis of driving behavior. We will sooner be able to do that in real-time than have perfect AI.
One aspect worth analyzing is how much of Uber’s cost structure accounts for the local driver of the car. If you take Uber’s latest San Francisco pricing, the UberX service costs $1.3 per mile, which after 20% commission means the driver nets $1.04. If the driver uses a Prius, at 60 mpg and $4 per gallon, the cost of fuel per mile is $0.07. This translates into the driver netting (excluding depreciation) close to $1 per each mile. For a remote driver in India, this is more than half a day worth of a salary. For each mile!
Every service provided by a “Shared Economy 2.0” company will undergo margin optimization over time, as more components of the service get automated. A virtual lifeguard may initially only watch a single stream of video in an initial, simple and fast go-to-market implementation. However, as time progresses, the lifeguard may graduate to monitoring dozens of customers once an AI feature analyzes incoming video feeds to person came close to the vicinity gets added to the system. This scenario could be even further optimized when an additional ability is implemented that recognizes a person has entered the pool.
Needless to say, providing these kind of services remotely presents significant challenges, like reliability and latency of communications, the cost for deployment of remote equipment and a myriad of regulatory issues like whether a remote service provider has the required state drivers license or lifeguard certification. This is not unlike the current issues that Airbnb, Uber and other companies are currently working through.
Much like the needs of big data drove the development and adoption of Hadoop, and turned the AWS cloud into a major money-maker for Amazon, there is a basic platform that will need to be developed. This is a real-time, synchronous version of “Mechanical Turk”, combining humans and AI into a mesh that’s able to provide services on a global basis. Travis is right. It will take decades. Humans will drive the cars because it will be cheaper for the humans to drive them and they intuitively will drive cars better than AI for a long while. It’s only a question of where will these humans be.
Thanks to Sundeep Peechu, David Rowan, Sam Lessin, Eran Shir, Michael Eisenberg for reviewing & commenting on the many (more verbose) drafts of this.
