AI development goes mainstream? The potential of OpenAI's ChatGPT.

Mainstream AI

At the beginning of December 2022, OpenAI launched ChatGPT, an AI model that ‘interacts in a conversational way’. ChatGPT responds to queries and requests with text content that takes a huge variety of forms – product reviews, book synopsises, news items and much more.

ChatGPT quickly went mainstream. I am used to reading articles about new AI through technology blogs and technology news sites, but ChatGPT quickly became used and reported on in other places. Along with all the usual places, I enjoyed reading a very considered piece about it by Professor Paul Taylor in the London Review of Books. And in my spare time I rock-climb and cycle. Both of my go to sites for news and stories in these areas - UKClimbing and cyclingnews - explored the scope for ChatGPT to write news and review articles, as well as for some of OpenAI’s other related AI services to generate images.

Perhaps this is reflective of the threat journalists perceive from ChatGPT (it’s important to note that this AI is not there as yet), but this is the first time I’ve experienced my wider interest sites reporting on AI in such a way. While ChatGPT could yet turn out to be more hype than impactful, there has certainly been a quick and wide response to it. By late December 2022, over 2 million users had signed up and tried it out.

Not a silver bullet!

In this blog I discuss some of the opportunities that AI services like ChatGPT open up. But they are not a silver bullets: these opportunities are not without their challenges. They include a lack of model transparency, ethical considerations and legal issues such as copyright. ChatGPT also cannot reason in the way humans can – a fundamental limitation which must be factored into the ways it is used. And some of the models and outputs it currently yields can also be described as bland – more on this later.

To understand and mitigate for these and other limitations is at least as important as seeking to benefit from the opportunities of ChatGPT and OpenAI’s other AI services. I am going to discuss this further in an ensuing blog. If all this sounds negative I don't mean it to - I continue to be very impressed by ChatGPT the more I use it and see the uses of it by others.

Accessible for experimentation and innovation

Probably the main reason ChatGPT and Open AI’s other AI services are becoming so prolific is how available they are. Register as a user and you can immediately start to explore.

Accessible interactively via a very intuitive front-end, ChatGPT will respond to a written request. ?For example, I asked ChatGPT to write a book synopsis about fluffy bunnies who don’t like carrots.

It’s also accessible programatically with minimal upfront development. I pulled the relevant source code from the github repo, followed the initial tutorial and had an app that synthesised pet names running in my browser in less than 30 minutes. This more than fired my imagination beyond this toy (and fun, my kids loved playing with it) example. Since then, I’ve been thinking and exploring some more, and have written apps that essentially build new AI services. These are currently pretty simple, but I’ll be trying lots more different things with it in the coming months.

This sense of innovation is of course pervasive across the tech start-up and SME economy. There are already AI content writing products on the market driven by OpenAI’s GPT-3 (the forerunner to ChatGPT). We can expect new products and services coming to the market driven by OpenAI’s updated services, and these are likely to explode in number given their wider potential and application base.

When I start to think of these novel products, coupled with the ways in which OpenAI’s products such as ChatGPT will evolve and improve as they are used to innovate, my head starts to spin. This really does feel like the beginnings of a step-change in the evolution of AI.

Utilisation by the Enterprise

What does this mean for larger companies? The big, long-established companies, who find it harder than the start-ups to harness data and AI for themselves than start-ups due to cultural and legacy challenges? This is on my mind as it’s Enterprises I’ve been working with for the past five years or so, and many of them face such challenges.

Such companies will be able to benefit from ChatGPT and similar AI in multiple ways including:

·?????Purchasing AI services developed by the disruptive start-ups (or perhaps buying the start-up).

·?????Increased automation of technology (such as coding, testing, DevOps and cloud optimisation) and data processes (including data management, governance, engineering and science) by ChatGPT and associated AI models and services.

·?????Internal data science & AI development and automation of standard business processes will all be accelerated by Chat GPT.

I think that, within a relatively short period of time, ChatGPT and similar AI have the potential to reduce the time it takes for a large Enterprise to fully leverage its data from 3-5 years to two years or less. That is a significant change in terms of the time taken to realise the benefits and of course the financial cost of achieving these benefits.

ChatGPT writes code

How will Chat GPT achieve this? Another thing ChatGPT does is write programming code. At the moment this can be a bit clunky but it’s being improved all the time. If we’re not there with some things already, it won’t be at all long until using AI generated code is a sound and viable option for the more routine tasks currently achieved within software development.

Synthesising AI within an ‘AI Ecosystem’

And the fact that ChatGPT can write code means it can, after instruction, build its own AI. Perhaps the biggest strides are being made here by integrating different ‘parts of AI’. By this I mean getting different AI services to interact directly with each other. Doing so using ChatGPT can yield results greater than the sum of their AI parts.

For example:

‘Hey Chat GPT, use Python to generate me a XGBoost model that will predict how many of my clients will use my new ‘build your own AI’ service. Tell me when they are likely to use it, how old they are and what their other interests are. Also generate me a report to show me that the model you build me is fair, unbiased and ethical.’

The above request can be broken down into a number of different kinds of AI

·?????Voice interpretation (natural language processing (NLP))

·?????Automatic coding of machine learning model(s) and diagnostics of these models

·?????Automatic insights generation and reporting

·?????AI Explainability

While initially at least there will need to be an ‘automating wrapper’ built around the above requirement to get them to interact (which itself is AI), in time all the above will happen in response to the voice request.

That is not to say the results will necessarily be any good! It may not be possible to build a fair and ethical model with sufficient predictive power with the data I have (in which case I would need the above process to tell me so). If it is possible to build such a model, to start with at least, the outputs of such automated processes will likely be vanilla: ‘ok’, maybe even ‘good’ but not ‘awesome’. There will need to be at least one creative human in the loop to achieve this.

And the above are ‘just’ data science examples. Improving the data science and AI capabilities of my Clients is my job at UST, so maybe I should not be surprised that this is my immediate thought when it comes to how I could utilise ChatGPT in my work.

However, the scope for doing similar interactions between different pieces of AI out there in the real world is significant. Interacting with apps on our phones or other machines using our hands will become less important. Instead we will just tell Siri, Alexa or other similar services what we need from the ecosystem of AI services they will interact with after interpreting our voice instructions. This is of course happening already. Chat GPT will accelerate it, and at a pace.