Andrew NG’s “The State of Artificial Intelligence” reviewed

David Pereira
Jan 9, 2018 · 7 min read

Andrew NG gave this speech about the state of Artificial Intelligence last November 7th, 2017. During his almost 30 minutes speech, the until recently chief scientist at Baidu, founding lead of the Google Brain team, and co-chairman and cofounder of Coursera, presented his vision of the current situation of AI and what to do to become an “AI-first company”.

While I do not pretend to be nothing close to Mr. NG’s vision on the subject matter(no doubt he is one of the world’s leading leaders in the field of artificial intelligence) I will try in this article to break down the main ideas of his speech and analyze more detailed implications. So, let’s start one idea at a time.

#1 — AI is the new electricity

This is probably one of the most well-known statements by Mr. NG. He uses this comparison to explain that AI is going to transform every major industry. He goes on acknowledging the hype around AI, and I absolutely agree as I stated in one of my previous articles. Afterwards, he describes how supervised learning has become the most relevant AI category in terms of real business outcomes. He ends stating that “anything a typical person can do within less that a second of thought we can probably now or soon automate” and explaining the potential impact in jobs.

After introducing the above concepts, Mr. NG explains why now AI and specially deep learning have become so popular in the last years, stating it in terms of performance vs. amount of data needed, later presenting other AI techniques (transfer learning, unsupervised learning, reinforcement learning) and its PR hype vs real outcome relationship. Another important remark during this section is what he calls the “hunger for data” of some of these techniques, which make them poorly valuable in real business contexts, putting into side all the marketing generated by some very specific achievements.

Although I agree that AI can be a “game-changing” technology and that it will transform every major industry, I find it a far more complex concept than electricity, for several reasons:

  • While electricity is a relatively complex physics problem, and can also be complex in terms of generation and distribution, AI sits on top of it and, in fact, on top of some other layers like hardware and software, with the later requiring the development of very complex mathematical models. All of these layers need to evolve together in order to produce significant breakthroughs (unless some black swan appears in one of those categories making it enough to produce a huge leap in scalability and performance).

Actually, AI has been around for so many years and it has not been until faster and scalable computing (hardware), new mathematical models (software) and of course, data, have been available at the same time that it has become disruptive for business. So AI is a very complex form of electricity, one that requires a combination of advanced technology, and at the same time very high skilled people to make it actionable through a lot of research.

This article exemplifies this vision through a concrete example: the improvements of the AI playing chess from 1980.

  • Although AI is being used for physical processes, one of its main promises is to transform the cognitive space. That is, transforming and automating cognitive tasks now executed by humans (as Mr. NG expects, every one of them executed with less than a second of thought). Here, the complexity stands on how big organizations evolve to be mature enough to embrace a transformation process like this, not only from a technology perspective but from a people and processes perspective. That is why I think Digital Transformation skills and capabilities are a prerequisite to build an AI practice.
  • As Mr. NG explains, the required amount of resources (data, but also electricity — one single AlphaGo game requires $3000 in electricity bill) to be able to effectively apply AI continues to be enormous in some cases.

Despite all of this complexity, one thing is clear. No matter the hype, real breakthroughs are taking place and AI is going to be one of the most competitive advantages in the market in the near future. It is hard to predict how fast its adoption is going to be, or which markets are going to be affected the most, but every organization should have a plan, not to apply AI, but to continuously adapt and think how AI can help in the journey.

#2 — There is an excess of PR about the importance of data

Mr. NGs next idea is the positive feedback loop that takes place whenever a product or service is defined from a sufficient data set, feedback is collected from users and is later used to improve available data.

While this makes perfect sense for new companies defining and refining its value proposition, it is far more complex for large organizations with long term stablished products and processes. Why? because most of the times, relevant data is difficult to collect, due to a huge variety of reasons: lack of governance, departmental boundaries, poor quality, amongst others. Even if large companies find suitable data to launch new products and services, the complex part tends to be how to scale the firsts prototypes. My advice here would be not underestimating complexity. Don’t put the cart before the horse. If you don’t have a strong and agile organization in terms of data governance, processes, mindset and talent, maybe you should take care of that before embracing AI. As stated before, companies should move towards being a digital company before being an AI company.

#3— A Technology company + Deep Learning does not make an AI company

Mr. NG starts by describing what a true internet company looks like, and later describes what elements make a company to truly be an AI company:

  • Strategic Data Adquisition. Is the company capable of generating or acquiring data to create a defendable business?
  • Unified Data Warehouses. Is data accesible across the organization or it is siloed?

This two first characteristics described by Mr NG. suit very well the problem of scale and maturity that I described previously. Most organizations are facing the risk of building AI solutions with feet of clay.

  • Pervasive automation. How good is the organization is applying automation to internal processes? Is it a cross-company goal?
  • New job descriptions. Have skills, mindset and roles evolve to support the creation of AI solutions?

This two characteristics of a true AI company as seen by Mr. NG are also a result of a high level of digitalization, in my humble opinion. Digital companies are good at applying technology to automate backoffice processes while transforming others by completely redesigning them. A good redesign might lead to the example pointed out by Mr. NG in his presentation. Maybe some customer care process is evaluated to be more effective while delivered via a chatbot. This leads to the necessity of new roles and even processes (e.g. how to stablish an effective methodology to deploying new chatbots — taking into account, amongst others, problem and language localization, exception handling, etc.).

So, I would summarize by stating:

A Digital Company + Deep Learning/ Machine Learning = AI Company

An AI Company acts as a Digital Company, looking how to benefit from AI to solve its transformational challenges

#4 — Build a centralized AI team

In his last point, Mr. NG gives advice on how to incorporate AI talent to the organization. Given that AI profiles are scarce, he recommends to create a unified AI unit across the organization, so different BUs do not fight for these scarce resources. While I agree on this approach, I think that the vast majority of organizations will still struggle to find enough profiles to tackle all of their challenges around AI. My suggestions, here, would be, in addition to Mr. NG’s:

  • Learn about Cloud AI offerings. Analyze how they fit in your strategy. Balance how core the problem you are trying to solve is versus things like differentiation and vendor lock-in.
  • Work on a DataOps strategy. Think about what DevOps can bring to your AI lifecycle. What tasks can be automated, how to generate reusable assets (e.g. algorithms), so your needs for specialized profiles do not scale proportionally to the number of AI projects underway.

David Pereira

Written by

Head of Data & Intelligence for Europe at everis, an NTT Data company. All opinions are my own.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade