One of the influential people in AI who I follow is Andrew Ng. In the past, he has headed AI functions at both Google and Baidu and co-founded Coursera. Last December he was back on the stage of MIT Technology Review conference EmTech discussing the present state of AI. I found his presentation very inspiring and picked the following insights for those who didn’t have the half an hour to listen to him.

What is AI now good at?

Andrew has for some time defined the capacity of current AI as follows:

Anything that a typical person can do in less than one second AI can learn. This is an imperfect rule, but holds pretty well.

Jobs and manual procedures which can be decomposed into these simple, constituent jobs can probably be automated in the near future.

Nowadays there are good examples of using AI to do market automation, loan decisions, speech recognition, and even to steer an autonomous vehicle. The technology behind the majority of these opportunities is “standard” AI, otherwise known as supervised learning.

99% of value created by AI comes out of supervised learning – mapping from A to B [identification, categorization].

The deep learning is the fancy new variant of AI repeatedly discussed in the media. Deep learning is finally improving and it provides superior performance in comparison to “old” AI technologies (SVM etc.) when the number of available data increases. Old learning solutions could not benefit from larger data sets, whereas neural networks can benefit from increasing datasets. The bigger the network, the more data can be poured in with a performance increment.

Andrew lists different techniques based on their current business impact

  1. Supervised learning
  2. Transfer learning
  3. Unsupervised learning
  4. Reinforcement learning


“Reinforcement learning PR excitement is largely disproportionate with its impact”

The most valuable thing for AI-based businesses is an exclusive data asset

Leading AI company don’t only have great data scientists, but unique data assets. Andrew says that data assets make AI-based businesses defendable in a competitive landscape. Although he has worked with leading search engines and knows intimately how they work, he would be unable to create a competitive product without similar sets of user data. To build a defensible business, a company must build a positive feedback loop that allows to accumulate more data from users.

Data assets allow leading web search companies to provide more relevant results.

What is an internet company and what is an AI company?

Andrew introduces the notion of an AI company, a digital business set apart by their unique power derived from utilisation of AI. But what defines this type of a company? Let us compare it to the picture of an internet company.

An Internet company is not just about selling stuff over the internet. Based on Andrew, the advantage of internet companies is to have distributed decision making which can’t depend upon centralized decision making (or the Hippo, cf. Lean). They do testing (AB) and have short cycle times and are able to ship product improvements frequently.

In comparison, an AI company is not just a company which uses neural networks on top of traditional technology products. AI companies do strategic data acquisition, which allows them to build defensible data-based business. They have unified data warehouses which allow fluid flow of data from application to application, across any superficial silos. They are good at spotting pervasive automation opportunities, including those under the one-second threshold.

New requirements for product management

To run an AI company or manage an AI-heavy product, visual representation of the new product is not enough. To deal with AI capabilities, product managers must meet AI developers in their terms, for instance, present annotated datasets which describe how the product should behave, in terms of matching A’s to B’s. Traditional specifications such as wireframes do not suffice when trying to crack this equation.

How to incorporate AI into a corporate structure

The final theme Andrew touches is upon is the integration of AI know-how in large organisations. First, he recognises that AI is not a mature capability. As such AI capabilities are currently best integrated as centralised AI teams which help the whole organisation to integrate AI functions (in a matrix fashion). Later on, when the practices and methods of AI work mature, individual business units may hire their own talent as has happened with UX and mobile developers, for instance.


“Common teams, common standard, company-wide platforms of AI”

Find out about our data-driven solutions delivering business intelligence here.