AI has been in development for over sixty years. During this time, it has gone through several periods of excitement. In the past, there was considerable focus on looking at how human experts behaved and then codifying those practices as rules into a computer. In contrast, progress today is being driven by a subset of machine learning called deep learning. Deep learning involves feeding data to neural networks crudely modelled on how we think the human brain might work and using algorithms to have the computer learn from that data. Rather than trying to examine samples of information with preset rules, computers are now powerful enough to examine complete data sets and observe whatever patterns might exist within them. With the advent of unsupervised learning, this data does not even need to be labelled to allow the computer to make sense of it. Deep learning is enabling remarkable improvements in machine translation, natural language recognition and computer vision.
Moore’s law suggests that processing power doubles every two years, but other improvements in computing have actually enabled far greater progress. Cloud computing lowered the cost, while a shift from using computer processing units (CPUs) to graphics processing units (GPUs) has sped up deep learning by 10-20x. GPUs are well suited to deep learning because they specialise in doing a large number of simple calculations in parallel. Overall, computational performance in deep learning has improved a staggering fiftyfold in just three years.
The proliferation of the internet, particularly on mobile devices, has led to an explosion in the data necessary to feed deep learning algorithms. In 2016, it was observed that more data was created in the previous two years than in the whole of human history. This matters because today’s artificial mind is still rather dull. A human brain needs only a dozen images of dogs and cats to begin to distinguish between them. An artificial mind requires thousands of examples. If AlphaGo had only been able to practice as much as Lee Sedol, Sedol would have annihilated his artificial opponent. More data is creating smarter machines.
The most relevant near-term application of deep learning is natural language recognition. This has the potential to change how we interact with the internet and therefore how every internet company interacts with its users.
Li believes this represents the third era of the internet. The first was the desktop era, with keyboards as the input device. The second was the mobile internet with touch screen input. Li believes the mobile internet era is now coming to an end, replaced by the AI era with natural language recognition as the input device.
The application of deep learning to natural language recognition has fuelled significant progress. Error rates have fallen from over 15% in early 2015 to below 4% today. Moreover, a Stanford study using Baidu’s Deep Speech 2 recently demonstrated that speech recognition was 3x faster in English and 2.8x faster in Mandarin than using a touch screen. Furthermore, it found speech recognition actually reduced the error rate in English by 20% and in Mandarin by an incredible 63%.
It is because of these improvements that 10% of Baidu’s search queries are now made using speech recognition. Moreover, Ng predicts that by 2019 half of all Baidu search queries will come through speech and image search. Voice queries are more popular in China, in part because the language is notoriously hard to type and some users do not know pinyin (the system of writing Mandarin that uses the Latin alphabet). This could give Baidu a head start in voice recognition.
It is very early, but the implications of this shift strike me as hugely significant. In the desktop era, search was a dominant gateway to the internet. In the mobile era, it remained important, but less dominant. Users can now go directly to vertical search apps and app stores. For example, I can happily search for restaurants using Trip Advisor’s app, read news on the BBC News app and buy a new pair of shoes on Zalando’s app without ever going near Google. In contrast, with voice recognition, the search engine or voice recognition provider becomes the access point to all information and all functions, becoming both a search engine and a virtual assistant. Whoever fulfils this role effectively becomes the operating system (OS). This is why Microsoft and Apple are developing their voice recognition assistants, Cortana and Siri.
We already benefit from AI in our everyday lives. When Facebook auto-tags your friends in photos, it does so through deep learning; when you use Baidu’s or Google’s voice search, the natural language recognition is done by deep learning; and when you put your apartment on Airbnb, the recommended rate is generated by deep learning. Zalando has even used deep learning to improve warehouse efficiency and to power fashion design with Project Muze.
The applications may be felt first in the consumer internet space but they will extend well beyond. There are four big areas worth mentioning now.
This is an enabling technology. The potential impacts are therefore wide and of a similar breadth and scale to electricity, computing and the internet.
It is no coincidence that two of the leaders in the field of AI are search engines. Learning from data has been Baidu and Google’s core business for well over a decade. They should have a head start. Li highlights how Baidu have been investing in this area already for many years. Likewise at Google, when challenged “web search, for free? Where does that get you?” way back at their IPO party in 2002, Larry Page responded with incredible vision and foresight “Oh, we’re really making an AI”.
It is not a new insight that data can be a competitive advantage. But deep learning enhances that advantage because it allows you to extract even greater value from large data sets. This means data-rich and data-centric companies should see their competitive advantage further enhanced versus old world companies who have less data and whose strategies are not focused on data collection and usage.
Ng has spoken about how it is difficult for companies to derive real edge from algorithms. He notes that in the global top tier of AI companies no-one is more than one to two years ahead or behind in terms of algorithms. He therefore claims that “data is the defensible barrier … unique data assets are very difficult for competitors to copy or for us to get a competitor’s data assets”. This should mean that even as we transition towards an AI-first world with an AI-powered OS there is good reason to believe that existing large internet companies will be the ones that dominate this new era.
New AI companies will probably still emerge. Indeed, a new start up called Viv, which includes some of the team behind Siri, are currently working on their own assistant. Yet I can’t help but think that it is more likely such companies will have to sell themselves to data rich companies, or at best strike partnerships to gain data access, than succeed independently as search or OS providers.
Baidu and Google do lack various data sets. One long-term solution may be to commoditise deep learning and offer it as a platform. Google has TensorFlow, Amazon DSSTNE, Microsoft CNTK and Facebook has Torch for machine-vision technology. Baidu are opening up their own machine learning platform, PaddlePaddle, for free.
It is too early to fully understand what artificial intelligence could mean. This note is a very early attempt to grapple with the implications. I find it fascinating that data is becoming yet more valuable, to the benefit of data-rich and data-centric companies, thereby potentially enhancing the competitive advantage of online over offline commerce. It is also surprising that the most likely winners of this new era will be the same players that benefitted most from the earlier desktop and mobile eras.
What I particularly like about this opportunity is that the market seems to struggle with it. Does any broker assign any value to the AI opportunity? Concrete AI applications are three to five years away and the costs will of course be upfront. But the magnitude of this opportunity should not be underestimated. Li spectacularly mused to us: “In five years I’m not even sure search will still be our main revenue source”. Meanwhile, Ng remarked: “the advent of all-pervasive AI will be the single most important development for the global internet sector” and “whoever wins AI will win the internet”.
The views expressed in this article are those of Lawrence Burns and should not be considered as advice or a recommendation to buy, sell or hold a particular investment. They reflect personal opinion and should not be taken as statements of fact nor should any reliance be placed on them when making investment decisions.
This communication was produced and approved on the stated date and has not been updated subsequently. It represents views held at the time of writing and may not reflect current thinking.
Baillie Gifford & Co and Baillie Gifford & Co Limited are authorised and regulated by the Financial Conduct Authority (FCA). Baillie Gifford Life Limited is authorised by the Prudential Regulation Authority (PRA) and regulated by the FCA and the PRA. Baillie Gifford & Co Limited is a unit trust management company and the OEICs’ Authorised Corporate Director.
Baillie Gifford Overseas Limited provides investment management and advisory services to non-UK Professional/Institutional clients only. Baillie Gifford Overseas Limited is wholly owned by Baillie Gifford & Co. Baillie Gifford & Co and Baillie Gifford Overseas Limited are authorised and regulated by the FCA in the UK.
Persons resident or domiciled outwith the UK should consult with their professional advisers as to whether they require any governmental or other consents in order to enable them to invest, and with their tax advisers for advice relevant to their own particular circumstances.
Baillie Gifford Overseas Limited is licensed with the Financial Services Commission in South Korea as a cross border Discretionary Investment Manager and Non-discretionary Investment Adviser.
Baillie Gifford Asia (Hong Kong) Limited 百利亞洲(香港)有限公司 is wholly owned by Baillie Gifford Overseas Limited; and holds a Type 1 licence from the Securities & Futures Commission of Hong Kong to market and distribute Baillie Gifford’s range of UCITS funds to professional investors in Hong Kong. Baillie Gifford Asia (Hong Kong) Limited 百利亞洲(香港)有限公司 can be contacted at 30/F, One International Finance Centre, 1 Harbour View Street, Central, Hong Kong. Telephone +852 3756 5700.
Mitsubishi UFJ Baillie Gifford Asset Management Limited (‘MUBGAM’) is a joint venture company between Mitsubishi UFJ Trust & Banking Corporation and Baillie Gifford Overseas Limited. MUBGAM is authorised and regulated by the Financial Conduct Authority.
This material is provided on the basis that you are a wholesale client as defined within s761G of the Corporations Act 2001 (Cth). Baillie Gifford Overseas Limited (ARBN 118 567 178) is registered as a foreign company under the Corporations Act 2001 (Cth). It is exempt from the requirement to hold an Australian Financial Services License under the Corporations Act 2001 (Cth) in respect of these financial services provided to Australian wholesale clients. Baillie Gifford Overseas Limited is authorised and regulated by the Financial Conduct Authority under UK laws which differ from those applicable in Australia.
Baillie Gifford Overseas Limited is registered as a Foreign Financial Services Provider with the Financial Sector Conduct Authority in South Africa.
Baillie Gifford International LLC is wholly owned by Baillie Gifford Overseas Limited; it was formed in Delaware in 2005. It is the legal entity through which Baillie Gifford Overseas Limited provides client service and marketing functions in America as well as some marketing functions in Canada. Baillie Gifford Overseas Limited is registered as an Investment Adviser with the Securities & Exchange Commission in the United States of America.
All investment strategies have the potential for profit and loss, your or your clients’ capital may be at risk. Past performance is not a guide to future returns.
Any stock examples and images used in this article are not intended to represent recommendations to buy or sell, neither is it implied that they will prove profitable in the future. It is not known whether they will feature in any future portfolio produced by us. Any individual examples will represent only a small part of the overall portfolio and are inserted purely to help illustrate our investment style.
This article contains information on investments which does not constitute independent research. Accordingly, it is not subject to the protections afforded to independent research and Baillie Gifford and its staff may have dealt in the investments concerned.
All information is sourced from Baillie Gifford & Co and is current unless otherwise stated.
The images used in this article are for illustrative purposes only.
Ref: 31919 ALL WE 0078