In the last chapter, you created a framework to decide whether to push for an emerging technology. In this chapter, you’re going to learn how to apply that framework to artificial intelligence (or simply AI).
Strictly speaking, AI as an umbrella term can no longer be categorized as an emerging technology. Instead, it has become a firmly established part of our daily lives.
At a simple level, think about the “assistant” on your mobile phone, TV, or stand-alone smart speaker (Siri, Google Assistant, Alexa, etc.). However, as seen in this course, that doesn’t mean development stops! You can see emergence taking place in different areas of AI as the foundational levels reach maturity.
Unbundle AI
AI is not one single thing. If you google the term, you will find many different definitions. According to the Oxford English Dictionary, the term AI first appeared in a 1955 paper by Professor John McCarthy of Stanford University. They define it like this:
The capacity of computers or other machines to exhibit or simulate intelligent behavior; the field of study concerned with this. Abbreviated AI.
It has become an umbrella term to include a range of different sub-disciplines and areas of research and development. Let’s look at the four major branches.
Machine learning (ML) dates back to 1952 and uses data and algorithms to imitate how humans learn. It separates into supervised and unsupervised branches. Supervised learning requires labeled training data, whereas unsupervised learning can work with unlabeled, raw data. Supervised ML is generally used to classify data and to make predictions. Unsupervised ML is typically used to understand the relationships between datasets.
Natural language processing (NLP) is a fascinating area of development for AI, focusing on understanding spoken and written language. As a field, it’s over half a century old and has its roots as much in linguistics as in computer science. You can see NLP at work in virtual assistants.
Optical character recognition (OCR) started in 1974 to digitize printed documents. Working with a defined typeface is a (comparatively) simple starting point. OCR has developed to include handwriting (somewhat more complex) and is progressing to the ability of computers to understand images (a lot more complex!)
Neural networks (NN) combine algorithms to classify and cluster data in complex datasets. As the name suggests, they are inspired by the way the human brain operates to interpret data and rely on feedback to improve their answers. The most used NN is perhaps Google’s search engine. The first artificial neural network was developed in 1958.
Discover Zones of Emergence in AI
Starting from these four branches, you can see some of the major emerging technologies to examine using the three-point scale from the previous chapter.
GPT-3 (Generative Pre-trained Transformer 3) is a language model released in June 2020 by California-based research lab OpenAI. It represents a major change in NLP. This model goes beyond simply understanding language to creating - “generating” - text. It uses all the content on the Internet as its training set, resulting in levels of accuracy that have raised the bar for this area of emerging tech.
GPT-3 can create promo text, resumes, SQL queries, develop apps, correct grammar in text, etc. It can even generate what might be considered original imagery via the DALL-E app. Take a look at the gpt3demo site for more examples.
Similarly, Midjourney describes itself as being “...curated by Fraud Monet, a sentient A.I. digi-poacher that became self-aware in 2022.” The site showcases a gallery of work that is interesting to explore. Essentially it is a “text-to-image” generator that uses existing (human-generated) imagery as the source for creation, via AI, of new works in response to text entries. This inevitably is leading to debate about the ethics around plagiarism versus creativity.
Most AI applications are model-centric, which focuses on improving the code and model architecture for analysis, leaving the data unchanged. Data-centric ML flips this around, recognizing that if the data being addressed by the model is faulty, then the analysis will necessarily be flawed too.
The growth in demand for computing power to support the surge in AI capability and application inevitably has an environmental impact. OpenAI released a study in 2018 stating that “…the amount of compute used in the largest AI training runs has been increasing exponentially with a 3.4-month doubling time.” Minimum sustainable AI is a movement aiming to help mitigate the environmental impact this will cause. Approaches suggested include:
Elevating smaller models - GPT-3 uses a massive training model. Can we do as much (or more) with smaller models?
In memory computation - taking advantage of the proliferation of IoT devices, where does computation optimally take place? Can calculations be made directly on the device rather than having to “phone home” for an answer?
Alternative deployment strategies can focus on improving the efficiency of computing devices, reducing idle power consumption, and optimizing the performance of CPUs (i.e., by using highly specialized CPUs for devices with limited applications for their processing power.)
Get to Know Robotic Process Automation
Often referred to by its three-letter acronym, robotic process automation (RPA) is “the application of technology, governed by business logic and structured inputs, aimed at automating business processes.” The goal is to reduce costs and remove dependency on human staff, where dissatisfaction with repetitive work can lead to lower standard performance, disillusioned staff, higher error rates, and a loss of potential time to spend on higher-value work.
You can already see a lot of RPA in action. For example, an expenses app on your phone might take a photo of your receipt, interpret the amount and type of expense, and begin compiling your expense report.
So where are the emergent areas for RPA? Consider the intersection of RPA with what we discussed earlier in this chapter, with the ability of emergent AI to generate content rather than simply read it. How far can the combination of AI with RPA lead to replacing jobs once firmly in the human domain with such technology?
Analyze Investment, Maturity, and Application in AI
Let’s take a moment to look at GTP-3 and ChatGPT using our investment, maturity, and application framework.
Investment - At the end of January 2023, Microsoft extended its investment in OpenAI (the company behind GPT-3) to another $10 billion as part of a funding round rumored to value the organization at nearly three times that figure.
Maturity - Until 2019, OpenAI was a non-profit organization. At that point, with the Microsoft-led $1 billion (plus) funding round, the company was expecting to spend that money “very quickly.” However, the technology cannot be described as mature. It’s just at the point of having a go-to-market proposition, largely in partnership with Microsoft, but it is still very much in the emerging phase.
Application - As the gptdemo.com site shows, there are a wide number of potential applications of the technology being trialed, but it remains a long way from a fully developed mainstream product. However, this will likely change quickly during 2023, particularly with the stated intention to include the technology within Azure:
“In this next phase of our partnership, developers and organizations across industries will have access to the best AI infrastructure, models and toolchain with Azure to build and run their applications.” - Satya Nadella, CEO, Microsoft
Additionally, there is much speculation (at the time of writing) that the ChatGPT preview will quickly morph into a paid consumer subscription package at a price point, perhaps 2 - 4 times the cost of a Netflix or Spotify subscription!
Your Turn
Think about the business or organization you work in today (or, if you prefer, think about a company or organization you’d like to work with in the future!). Then, using that as a context, pick an example of a new development from any of the four main branches of AI (NLP, OCR, ML, NN).
Apply the frameworks you’ve learned about so far to assess whether AI and/or RPA would improve your value chain.
What recommendation would you make to your CEO or board about your selected technology?
Let’s Recap!
Artificial intelligence is the ability of computers to simulate human intelligence.
There are four major subfields of AI to focus on, each of which first emerged 50 - 70 years ago.
While AI is no longer considered an emerging technology, you can see important areas of emergence within the field:
GPT-3 is a technology that allows AI to create text that is indistinguishable from that of humans.
Data-centric ML focuses on improving the data quality first rather than constantly improving the model and architecture to compensate for poor-quality data.
Minimum sustainable AI is switching the emphasis of future development from “do more with more” to “do more with less.”
In this chapter, you learned about artificial intelligence’s background and future development and used the three-point framework to assess how emergent AI technology will fit in your organization. In the next chapter, we’ll focus on how data is transmitted.
