Advances in deep learning have led Elon Musk and others to start preparing for the AI apocalypse. And indeed, by feeding terabytes into neural networks, computers are now able to understand voices, recognise faces and sift through data with unprecedented accuracy. And yet, advances in so-called unsupervised learning - which finds the structure or relationships in data inputs without training in the way that a child learns from experience - are almost non-existent.
In recent years, Yann LeCun of Facebook, Geoffrey Hinton of Google and Yoshua Bengio from the University of Montreal have made significant advances in machine learning through their use of deep neural networks and other learning techniques. For example, Yaniv Taigman, one of my co-founders at face.com (which was acquired by Facebook in June 2012), recently published that the company achieved a 97.25 per cent accuracy rate for face recognition, just 0.25 per cent below human perception. The Google car, sensor networks, robots and all the other things that will supposedly replace humans are all packed with machine vision and computer-based control systems.
State-of-the-art deep learning requires tens of thousands of examples to learn a pattern - it's as if a child needed to be electrocuted more than ten thousand times just to understand that he shouldn't push a screwdriver into a power socket. AI is hard. The human brain 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 explained by rules.
AI is also engineering that can't fail: it has to be perfect - and most software engineering projects are not perfect. Unlike a bridge that can't be 90 per cent done, software can be "good enough". Mark Zuckerberg is often quoted as saying that done is better than perfect - that works in software, not in bridges, and probably not in AI. The reason done is better than perfect in software is because the initial 80 per cent of the reward has 20 per cent cost. The remaining 20 per cent of the reward would take 200 per cent of the cost and time (and software projects are rarely on time).
In a world where entrepreneurs build products that provide delightful services, human labour is both a margin problem and an opportunity. The downside is that you have to pay humans hourly wages. On the upside, humans are light years ahead of computers in handling situations they are not used to.
There is an alternative to full AI, where humans can still do better and cheaper, through a new breed of companies one might call Shared Economy 2.0 - humans doing bite-size Amazon Mechanical Turk tasks. These companies will provide much of AI's promise, but will be faster to build and capture market share well before their AI counterparts reach operating conditions. The knowledge their human labourers generate will be accumulated and used in teaching for their AI engines. They can then innovate from a dominant market position.
These companies will take advantage of a global workforce and cost arbitrage. To put human labour cost into perspective, median annual income in India is $616 (£475). The median annual income for Holmes County, Mississippi, is $13,794, whereas in the Hamptons, Long Island, it's close to $100,000. With such a source of lower-cost labour, increasing high-speed connectivity and remote sensing, we can leverage people to do the work of the AI-endowed robots. And the humans will be significantly faster to market and better at it. People working remotely can drive the Uber car way before the autonomous code will be ready - estimates range from anywhere between two to 15 years' time.
For example, if you take Uber's latest San Francisco pricing, the UberX service costs $1.85 per km, which after 20 per cent commission means the driver nets $0.92. If the driver uses a Prius, at 25kpl and $1.06 per litre, the cost of fuel per kilometre is $0.017. This means the driver is netting (excluding depreciation) close to $0.37 per km. For a remote driver in India, this is more than half a day's worth of a salary. For each kilometre!
Needless to say, providing these kinds of services remotely presents significant challenges, such as reliability, latency of communications, the cost for deployment of remote equipment and a raft of regulatory issues. This is not unlike the current issues that Airbnb, Uber and other companies are working through. It might be that the unit economics for a service will initially be negative, but every service provided by a Shared Economy 2.0 company will undergo margin optimisation over time, as more components of the service get automated.
Facebook is already combining humans and AI to provide the service of M, its virtual assistant. Nexar is leveraging humans to tag traffic accident reports to make our streets a safer place. Other startups can already use humans as software building blocks through services like crunchable.io, whereby an application can send a request using the machine-to-machine JSON protocol. Humans can easily answer requests such as whether an image contains violent content, a video frame is of a child swimming or drowning in a pool, or understand hotel feedback like, "We stayed at that fancy schmancy resort - I will definitely recommend it to my mother-in-law."
Much like the needs of big data drove the development and adoption of the open-source software framework Hadoop, and turned the AWS cloud into a major money-spinner for Amazon, there is a basic human-resource platform that will need to be developed. This is a real-time, synchronous version of Amazon Mechanical Turk, combining humans and AI into a mesh that is able to provide services on a global basis. There is a new human services cloud, just waiting to be built.
Eden Shochat is a partner at Aleph VC, in Tel Aviv. He co-founded face.com in 2009
The WIRED World in 2017 is WIRED's fifth annual trends briefing, predicting what's coming next in the worlds of technology, science and design
This article was originally published by WIRED UK
