An Interview with Senior Technologist Adam Leon Smith
Adam Leon Smith CITP FBCS is a senior technologist and an advisor to multiple technology startups. He is an SME in international AI standardisation and provides expert advice to governments, regulators and the private sector on the topic. He is also Chair of the Fellows Technical Advisory Group BCS, The Chartered Institute for IT.
Let’s start at the beginning Adam(pun intended). Is there a recognised starting point for what we now refer to as “AI ” in computing? And have there been recognisable eras of AI technologies?
You could argue there have been 3 eras to this point. There was the period before we invented neural networks. mimicking the way our brains work when they learn from facts. The first artificial neural network was invented in 1958 by psychologist Frank Rosenblatt. Called Perceptron, it was intended to model how the human brain processes visual data and learned to recognize objects. Then, there was the period before the amount of data available in machine-readable format was large enough to enable machine learning to evolve, and then there was the time before all our devices were constantly connected through the internet. These are probably the main milestones that have enabled AI development. Arguably, the success of LLMs in 2023 may turn out to be another key milestone, it certainly is from the public’s perspective.
…and from a technical perspective, are there commonly held definitions and categories of AI? Or is it quite a nebulous term, even in engineering circles?
Artificial intelligence is difficult to define for several reasons. Firstly, we don’t really understand human intelligence well. So attempting to define an artificial reproduction is doubly difficult. Second, when we perfect a technique we no longer regard it as AI (computer playing chess, for instance). Finally, we tend to anthropomorphise because of our exposure to science fiction and its portrayal of futuristic AI. Where AI will make the biggest impact in the next few years will be things like a computer program classifying medical images. It won’t have hands, feet, a voice or even a chat interface.
The best definitions, in my opinion, are ones rooted in the assertion that the output of the system is a prediction, recommendation or decision. This is a more technical definition which is less dependent upon our definition of intelligence, the evolution of technology and our personal opinions about its uses.
What are the core programming languages and technical disciplines that make up AI?
Python and statistics, but increasingly there are more people working in the space focused primarily on the data, moving it around to suit the AI’s needs (data engineers). There are risks around AI that are different to conventional technologies due to the way they learn information. Therefore, there is also a growing need for those with expertise in managing AI risk, from a technical and legal perspective.
For someone who’s been in the AI space for many years, 2023 must have been quite odd, with AI thrust into the public consciousness in a way that hasn’t happened previously, chatGPT will have come somewhat out of the blue to most people outside of the AI community.
Yes somewhat, however the transformer models that ChatGPT (for example) are based on are from 2017 and have evolved more gradually from a technical perspective than is appreciated. These are models that predict the next occurrence in a sequence, whether it is words, images or video. They have been getting gradually better over the last few years and now have become really effective at catching people’s attention.
Could you talk a little bit about how generative AI and Large language models came about? Are they a new technology?
Generative AI includes transformer models and also generative adversarial networks which have been rising in popularity for a few years. Generative AI is a category of AI technology rather than a technical approach per se. Defined by its ability to generate things based on other things. For example, one type of generative AI is LLMs (Language Models) that predict the next most likely word based on their training. The big technology firms are using these technical core approaches combined in combination with supplementary techniques focused primarily on correcting errors and implementing safeguards.
What have you made of Open AI and Chat GPT? Is the technology revolutionary or simply first to market? Have you been surprised by the reaction/adoption? Do you expect bigger players like Google to catch up and overtake them?
It has been fantastic to watch the public see the technology in a new way. I use ChatGPT for many things, and I do find it better than using search engines. However I don’t rely on it for things like code or building a factual basis for a concept, it just isn’t good enough.
There are lots of applications though, where it is good enough – for example, I think businesses will start to implement LLMs to manage internal institutional knowledge.
How will generative AI/ LLMs evolve in the near future?
I doubt they will very much, in terms of core technology. However, lots of amazing ways to use them, or ways they can be trained on different, potentially private data sources, will create ongoing business opportunities.
Over time I think there will be a small number of LLMS, and I think they will be open-source. I suspect most of the intellectual property relating to systems like ChatGPT is in the error correction processes. I also think the sheer cost to run these systems will mean they are soon going to be more expensive to use than they are now, I’m not sure ChatGPT can afford a free version for much longer.
What are the most impressive use cases of LLMs you are aware of currently?
I think they have a lot of promise doing things like summarizing meetings and notes. This use case can be applied a million ways. I’ve seen AI systems that have been trained on the output of LLMs, which is a very interesting development to watch and poses new risks. For example, if you are training an AI system to prioritize written communication for a human to read, you can train it using lots of examples of communication generated by an LLM. Instead of text originally written by humans. This might be because you can’t get the consent you need to use human text.
What’s going to be the next big moment/moments in the evolution of AI?
Quantum computing.
Quantum means rather than computers thinking in 0 or 1, they think in an infinite amount of states. We will reach “quantum advantage” in the optimization of machine learning models in the coming years. By that I mean that quantum computers will be more efficient at optimizing AI systems than classical computers. Thus, increasing our ability to train better AI systems.
What are the biggest risks and ethical questions these new technologies pose? And how do individuals, organizations and even governments navigate this?
The risks are manifold to firms and depend on the context. Some significant and general ones are accidental discrimination or accuracy failures leading to customer and regulatory ill-will. Then there is a lack of explainability leading to an inability to gain confidence in an AI system. Or to investigate its behavior.
European and the UK have already taken a different path to most of the US on data privacy for example, and data privacy law is actually frontrunning on tackling AI. Early bans of ChatGPT came in European countries and led to rapid change.
We are already seeing in industries like acting and publishing, a pushback against the use of AI, and fierce debates around access to and use of data for training models. What is your sense on where regulation will land?
As I say, data privacy is providing some protection for consumers and citizens. That’s not necessarily the case with an employee though, especially someone whose ‘image’ or ‘voice’ is very obviously being used for business purposes. I’m not sure how rules will develop in that space, but I imagine they will mostly be the result of employment law and contract negotiations rather than regulatory action or legislation.
Nvidia, the chip designer, became only the 5th company ever with a trillion dollar valuation, (the others being Googe, Microsoft, Amazon and Sadi Aramco) Is this hype or are they going to be central to the next phase of AI? and why?
Nvidia and a number of other companies like ASML are key to GPU chips, which we currently need for efficient AI. The geopolitics of this space are complex though, encompassing everything from the Inflation Reduction Act in the US to sanctions / export controls on Chinese companies and risk of war in Taiwan. The chip manufacturer that will be most successful will be the one who significantly reduces resource consumption efficiency.
What should companies in industries like marketing be doing now to prepare for the opportunities that AI will bring?
The answer to this question is the same for almost all industries. Adopt new tools that harness the power of AI to solve current challenges and then move further up the value chain. In marketing, for example the production of low quality SEO copy might now be redundant, but how about generating copy that chatbots give higher “ranking” to? How do you make your marketing services standout in a post search engine world?