Professor Mark Parsons: Supercomputers, AI and the dangers of a digital 'dark age' 

In a nondescript grey building, nestled between the rolling hills and green trees of Penicuik on the outskirts of Edinburgh, sits one of the most important computing facilities in the UK. 

It’s an unassuming location for such an influential place, but the Advanced Computing Facility (ACF), operated by the EPCC (formerly the Edinburgh Parallel Computing Centre) at the University of Edinburgh, is the nerve centre of the UK’s past, present and future in supercomputing.

Housed at the site is ARCHER2, one of the UK’s most advanced supercomputers, capable of solving complex equations in a fraction of a second. Plans are in motion to build its replacement, a next-generation supercomputer that the UK Government is funding to the tune of £750m. 

Heading up this operation is Professor Mark Parsons who, as director of the EPCC, is responsible for implementing and running a project that could be crucial in ensuring the UK doesn’t fall behind in the computing arms race.  

“I'm really proud of the fact that EPCC has survived and flourished since we were set up in 1990,” says Parsons. “Getting that £750m of funding is really important because we were going to end up as a country without any national supercomputing resource and that would just be wrong. There'd be no European country like it.” 

Inside the walls of the ACF, rows upon rows of black cupboards house thousands of computer chips, all cooled by a complex system of pipes that circulate cold water into the computer and hot water back out. The pipes are used to keep ARCHER2 as cool and efficient as possible while it works away solving any number of complex problems that a standard system would flounder at the metaphorical sight of. 

The future system at the ACF, which is predicted to come online in 2027 after next-generation chips from Taiwan are purchased, is predicted to be 50 times faster than ARCHER2. This will open the door for a new level of computational processing that allows for even larger and more difficult problems to be solved.  

“I've got a big ten-year project with Rolls-Royce that we're halfway through, where we're starting to model whole gas-type turbine engines,” says Parsons. “We can't do that on ARCHER2. It doesn’t have the memory and it's not fast enough. But we'll be able to do that on the next system.” 

Taiwan dominates the production of semiconductors, the single most crucial component needed to make a supercomputer. These semiconductors are used to make the silicon chips that act as the “brains” for modern computers. Companies based in Taiwan produce at least 90 per cent of the world’s most advanced semiconductors and generate over $125bn in annual revenue from the sale of them to countries around the world.  

“I don't think we as a country have, at the moment, the infrastructure to produce silicon chips,” says Parsons. “But we will [in the future] have enough significant infrastructure that can deliver the sort of projects and new activities that our public sector, industry and academia need to do. So, our computer is just part of that path.” 

The new supercomputer at the ACF will be the centrepiece of a UK-wide scheme, announced by the UK Government and costing £1bn, that should increase the country’s computing capacity 20-fold in the next five years. In comparison, China has announced a plan to invest one trillion yuan (£105bn) in a government-backed fund focused on increasing the country's global competitiveness in technology such as quantum computing and the ever-evolving world of artificial intelligence (AI). 

The emergence of AI as a significant factor in our lives has been powered by the supercomputers that facilities like the ACF host. These systems have been used to train large language models (LLMs) like ChatGPT, a system used daily by over 800 million people to request answers, write poems and generate silly little images of cats in party hats. Despite this, Parsons thinks that many of the big investments into computing taking place around the world aren’t focused on the training of AI for future usage but more on the running of current AI chatbots to answer queries on known datasets. 

“It's fantastic we've got large language models, they're just fascinating things to interact with,” says Parsons. “But actually, the use of AI is much, much broader than that and many of the benefits of it will not be through a large language model.” 

The potential for AI to increase productivity, cut costs for businesses and create new opportunities for growth in the economy is balanced by real fears over its implementation at the cost of human jobs and society. In polling published by the UK Government, 33 per cent of respondents said that they expected the future impact of AI on society to be a negative one.  

“It's clear that when we've had moments in time where there have been big changes in technology, the Industrial Revolution is the biggest example of that, where people's jobs have changed,” says Parsons. “But if you look back through history, the number of jobs has not got less generally, people have just changed. So, I think for many people, AI is just going to help them with what they do currently.” 

This doesn’t mean the technology is without flaws. Studies show that AI systems like ChatGPT can “hallucinate” answers to questions that it does not know, providing users with incorrect information. AI systems can also be influenced by the data used to train them if that data is incorrect or lacking context and depth. 

“My father is in his late 80s now and was a famous geologist,” says Parsons. “I sat with him and we were playing with ChatGPT, asking it really complicated geology questions, which he was amazed that it was getting it right. But then we asked about one of his areas of research that he really focused on in the 1970s and it knew nothing about it at all. And it's because those papers have never been digitized, they're just available on paper.” 

Without the full context, Parsons worries that we could be entering a world of “bland internet culture” that lacks the local knowledge available in undigitised resources. A world of information where answers are given on a global scale, without taking into consideration the in-depth, region-specific context that is necessary to provide useful and accurate information to users. 

“There's a pre-internet dark age where none of that information is in these models,” says Parsons. “Some really old information is because there are electronic versions of the works of Shakespeare and Jane Austen. But there are big bits of our knowledge that have never been digitised and I think that's a danger with where some of the models are going.”