Excerpt from Irish Times article (2 April 2025) featuring our Founder, Maryrose Lyons.

The launch of the DeepSeek R1 reasoning model in January 2025, on the day of Donald Trump’s second inauguration as US president, rattled the tech industry and its investors, wiping billions off stocks.

High performing but cheaper to develop than blockbuster rivals such as OpenAI’s, DeepSeek’s development was a case study in the law of unintended consequences. After all, it was US moves to limit China’s access to high performing chips that forced the latter to innovate.

Developed at a fraction of the cost to build and train than US peers, and requiring significantly less power to operate, DeepSeek upended previous assumptions.

These were not just in relation to the cost of developing large language models but the expensive data centre infrastructure required to support them. DeepSeek reportedly uses just 10 per cent of the power required by its US rivals.

The fact that DeepSeek’s model is open source, providing developers around the world with access to the code, was only going to fuel further developments, changing the game everywhere.

“The major impact has been that the Chinese have developed a new way to create a frontier model, disproving the notion that you must invest a whole lot of money in computing power,” says Maryrose Lyons, Founder of Ireland’s AI Institute. “That they have created DeepSeek-R1-Zero and made it available open source has a significant impact because it reduces the gap in quality between open source and paid models. Having such a quality model available for all for free has massive potential for reducing inequality, globally.”

This can make open-source options more viable for enterprises, especially for running operations on confidential data.

“The fact that they were able to develop it in a different way is actually massively game-changing, because there are probably other research houses building by using the DeepSeek method now. They don’t have to have massive billions of dollars for computing power,” she says.

It’s also great to use and, she adds, can be used via Perplexity AI. “Perplexity has a version of the model running on an American server, so none of your data is going to China,” says Lyons, who believes the launch of DeepSeek has forced others, including ChatGPT, which recently brought out its new 4.5 model, to up their game.

Read the full article by Sandra O’Connell in The Irish Times, 2 April 2025.