Real Estate and Third Wave AI: Hype or Hope?
In the PropTech start-up ecosystem, two different styles can be broadly identified. First are those which are led by the product, or more specifically, directly aiming to solve a challenge within the real estate sector. Second are those which are led by technology. In these cases, a technical founder or co-founder has developed a technological innovation, and has sought to apply it to a real estate challenge. Over recent months, I have noticed a subtle shift in the quantity of technology-led start-ups. One such technology is Third Wave Artificial Intelligence (AI). In this week’s article, we’ll unpack this concept, as well as its application to real estate.
Artificial intelligence is not a new phenomenon. Alan Turing asked whether machines could think in his 1950 paper titled Computing Machinery and Intelligence. Prior to that, notions of machines with humanlike cognition were common in popular culture (Maria from the 1927 film Metropolis is often cited in this respect). By the mid-1950s, the first AI programs were being developed, which was followed by a surge of university research funding on the subject.ᶦ Fast forward to today and the gradual development of artificial intelligence is being segmented into particular waves. Generally speaking, the first wave of AI involves programming logic. In other words, it is a fully coded technology which cannot learn. Examples include early computer chess, disease diagnosis and route optimisation. In logistics real estate, for example, route optimisation can be used to enable warehouse automation. The second generation of AI differs from the first insofar as it is able to learn. Some have therefore loosely classified second generation AI as ‘statistical learning’.
Over the last 10 years, tremendous progress has been made in second wave AI. An example is ‘natural language processing’, the ability of computers to process what we say or write (such as Alexa or Siri on your smartphone). Behind the scenes of Alexa and Siri, a new class of technologies allowed them to happen. These technologies are known as ‘transformer language models’, which are combined with the ever-increasing scale of cloud computing platforms and trained on huge volumes of data. These models fit into the category of ‘deep learning’, where models are built as ‘neural networks’ inspired by the structure of the human brain. The term ‘deep’ here refers to the practice of deploying multiple layers of ‘neurons’, which allows for more complex behaviour. The power of these models, roughly measured by the number of ‘parameters’ they include, is growing exponentially (see below graph). Similar progress has been made in computer vision, game playing, biological research, and so on.
In real estate, many of the applications of these models are in monitoring, for example analysing camera feeds from construction sites, directing robots, or monitoring building usage. There are also applications in analytics, such as advanced analysis of real estate markets.
Second wave AI technology suffers from some known limitations. In some circumstances it can be brittle, producing responses which are immediately recognised as not credible by a human. This has been identified as a common issue for ‘Automated Valuation Models’ — where identical homes next door to each other return substantially different values.ᶦᵛ Second wave AI also requires very large amounts of training data, which is not how humans learn. Second wave models are typically designed for a specific task, and cannot easily be deployed elsewhere. For these reasons, second wave systems are usually part of a broader system. A second wave computer vision system may be able to detect a construction worker and a hard hat, but the rule saying that the former should be wearing the latter will be coded by hand using a traditional programming language.
Third wave AI is often described as ‘explainable’, but can also be thought of in terms of its ability to develop internal abstract representations of the world. If a construction robot can represent the abstract concept of a room, and know that it has walls, a floor and a ceiling, it may well be able to learn how to paint the walls with less training data than if it thought in terms of undifferentiated generic surfaces. It might also be less likely to decide that a nearby tree is a wall to be painted if it’s accidentally switched on outside, and it may be easier to repurpose it for plastering or cleaning. Third wave AI will also enable new application areas. For example, much value in real estate finance seems unlikely to be unlocked without models which understand abstract economic concepts. This is on the radar of some real estate firms, and has sparked targeted acquisition activity recently.
The technology for third wave AI is still being developed. Some third wave characteristics are starting to emerge as second wave technologies scale. For example, a limited form of ‘transfer learning’ (the ability to use a model which was developed for one application and successfully use it in another) is now routinely deployed in natural language applications. An example of this is using a model designed to assess occupant sentiment and repurposing it for document summarisation. In addition, ‘zero shot’ techniques allow models to be adapted with very little training data. Novel technologies are also being developed, such as ‘graph neural networks’, which combine aspects of deep learning with first wave technologies such as knowledge graphs.
Technology in this space is moving very quickly, and many of the cutting-edge systems being deployed today didn’t even exist as prototypes or research projects five years ago. As things continue to develop, the Pi Labs team is cautiously optimistic of its continued impact on real estate. On one hand, AI advancements create both opportunities to improve existing solutions, as well as entirely new ways of solving challenges. On the other hand, some impressive technologies offer real estate practitioners few tangible benefits. For that reason, technical founders should be mindful of the distinction between an impressive technology and a viable product backed by impressive technology. At Pi Labs we depend on the deep expertise of our venture partners to understand the real commercial value of technology, and to provide support to our portfolio companies to help them stay focused on developing it into a commercially deployable product.
If you’d like to learn a little more about the three waves of AI, I’ve found this video from DARPA to be among the best resources on the topic:
Footnotes
i) If you’d like to read more about the origins of AI, Rockwell Anyoha’s Harvard blog informed much of this paragraph: https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/.
ii) Sanh, Debut, Chaumond and Wolf. (2020). DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. Available at: https://arxiv.org/pdf/1910.01108.pdf.
iii) Singer. (2021). The Rise of Cognitive AI. Available at: https://towardsdatascience.com/the-rise-of-cognitive-ai-a29d2b724ccc.
iv) You can read about the strengths and limitations of Automated Valuation Models (AVMs) in an upcoming white paper by the University of Oxford’s Future of Real Estate Initiative.
For more updates from the team and for a look into life at Pi Labs, follow us on Instagram and subscribe to our monthly newsletter.
