AI In Action
200 ways AI is being used to solve real problems
So far we have been focusing on the theory of AI:
- In Module 1 we developed answers to a range of hard metaphysical questions which provide a philosophical context for artificial intelligence
- In Module 2 we explored the theoretical foundation upon which the whole field of AI sits
- In Module 3 we identified a set of 10 ideas which represent the technical building blocks used to create modern AI systems
We’re now going to park the theory and move into the real world by reviewing 200 examples that show how AI is being used to solve practical problems today.
The approach will be to use the accompanying eBook to clearly explain each of the 200 problems and describe how the solution works, drawing on the insights provided in Modules 1 to 3.
The coverage in this module will be non-technical in nature, which means that I will be explaining the application to you as I would if I was speaking to someone who had no knowledge of the technology, market segment or specific problem.
Given the constraints imposed by the podcasts, we will not be able to discuss all 200 applications in detail. Instead, we’ll group them together into about 20 classes and then discuss each problem class with practical examples on a podcast
If you want to learn about all 200 applications then you can do that by downloading the accompanying eBook (premium content).
The material in this module is based on:
- Discussions that I have had with the CEOs or senior technical staff of AI startups
- Media reports that have caught my eye and which I have then investigated in more detail
- Material published by AI solutions providers in the form of case studies or product applications or provide details on how client problems have been solved with AI
- Academic research papers that focus on solving industry problems, such as how to identify the players on a basketball field using their motion profile
- Conversations with friends and colleagues which have unearthed problems that are ideal candidates for an AI-based approach
In this Module will see that AI is already being used right across the economy in most, if not all vertical markets.
You can regard an AI as a narrow slice of human cognition that has been commoditised
A good way to view AI is to regard it as a commoditised form of human intelligence – meaning a domain-specific cognitive ability that can be accessed via the cloud at minimal cost (depending on the data volumes the word ‘minimal’ might have to be replaced with ‘high’…).
Technically, this is not really correct because as we saw in Modules 1 and 2, AI is not ‘real’ intelligence anyway. Nevertheless, the metaphor that AI represents a commoditised form of human intelligence is useful to appreciate what’s going on.
Clearly, in markets where domain-specific AIs exist – say for analyzing X-Rays or MRI scans for the presence of breast cancer – then we would expect adoption to be very strong because the business case, when measured in purely economic terms should be compelling when the cost of the technology comes down.
AI can also be viewed as a kind of high-level software application that is horizontal in nature and can ‘run on’ any existing software stack – which means that AI can make any existing form of complex software work better.
An example of this would be a service like Amazon Rekognition Video which can analyze video content to identify the presence of a cat, city, beach, wedding or even loosely-defined, temporal ideas like ‘blowing out a candle’.
This AI software like this can be used in any industry vertical it is horizontal in nature.
These two ways of viewing AI help see why AI will eventually become pervasive and deeply embedded in all computational systems.
But we should keep our feet on the ground
Although some AI systems are very powerful and are already performing at ‘above human’ level this is only the case in very specific situations – like identifying melanoma or playing the board game, Go.
The reality today is that outside of very tightly-defined tasks, AI systems are still very weakly capable when compared with the equivalent human cognitive capability.
The main reason for this is that the technical approach taken so far is to use a ‘brute force’ approach to create a type of performance that seems to indicate the presence of real intelligence, when the system concerned is not intelligent at all and utterly incapable of understanding anything.
And the reason for this is that so far nobody has found a way to codify an abstract idea like ‘understand’, at least not in a way that properly reflects how what the word ‘understand’ means for human cognition.
We have no way of imparting real understanding into an AI system, or equipping the system with the ability so that it can develop its own understanding.
This means that AI systems are highly tuned for specific problems and when they are moved too far off their intended operating point they fail spectacularly. And even within their intended operating domain they can be easily fooled.
In addition, even if we restrict ourselves to use cases that can be addressed using AI, there can be serious practical problems, say related to access to data.
For instance, you might imagine that AI would find a natural home in a situation where users of a travel website want immediate answers to questions; say whether 6-year old children are allowed in the fine dining restaurant in a particular hotel.
We could imagine that it is likely that the hotel concerned had been asked this question before. And let’s imagine that for whatever reason the hotel has chosen not to put that information on the website.
In such a situation we can imagine building an AI system that could intelligently analyse the content of email changes between hotel staff and customers with a view to providing those same answers to new customers.
We could further imagine that prior customers would have asked many other questions, such as – how far is it to the beach, whether you need shoes to walk to the beach, where is the nearest doctor, what is the typical pool temperate etc. – and then we would have a way using a cloud-based AI system costing $10 a month to answer previously-answered questions while the expensive human staff focused on new questions, or something else entirely.
But AI is still struggling to reliably execute such use cases, even in a situation like this where digital information needed exists.
The problem here is not just building and training an AI system that has the requisite performance in natural language processing and Q&A but also involves regulatory, competition and privacy considerations.
For instance in Germany, the website would need to be very careful about simply uploading the content of email exchanges with prior customers to a cloud-based AI system: did the customer consent to this? where will the data be hosted? what software is needed to delete the data if the customer withdraws consent?
It is tempting to dismiss such concerns as ‘execution details’ but the reality in a post GDPR world is that these details can make the difference between a use case being doable, or not.
If the travel website concerned was a marketplace where many different hotels advertised their services, then the problem becomes even more complex because you now have to persuade rival businesses to share their private email communications with the company operating the marketplace and also the AI service provider when all players might be operationally active in different jurisdictions.
Developing a value proposition that makes sense to all stakeholders in such as situation would be complex, and might prove unfeasible.
In such a situation what started out as a bounded AI problem quickly becomes a complex business problem.
But given that well-executed AI systems can improve performance (i.e. deliver cognitive performance at ‘above human’ levels) and reduce cost (i.e. reduce the number of human resources needed) then it seems clear that in spite of the problems with data privacy and security which are ultimately solvable, AI systems will be adopted more strongly and, eventually, pervasively.