KHALED AL KHAWALDEH (ABU DHABI)
DeepSeek was a hot topic among global AI experts and leaders who convened in Abu Dhabi for the first day of the AI Everything summit Tuesday, with many seeing the open-source Chinese model as a possible game changer.
The new software, which was reportedly produced with only $6 million for testing and uses a fraction of the GPU capacity and power needed by other comparable models like OpenAI's ChatGPT, sent markets into a tailspin last week, as investors reacted to its upending of expectations of the capital and infrastructure requirements needed to produce competitive language models.
Speaking to the crowd on Tuesday, HE Omar Sultan Al Olama, UAE Minister of State for Artificial Intelligence, sought to downplay some of the hype, while lauding the credentials of the Chinese model.
"There's the claim around the parameters that is being promoted and we're trying to constantly understand what the reality here is, there are so many questions, but the reality of the matter is, there are certain things that were extremely useful," Al Olama said regarding the validity of the claims being made about the DeepSeek's production cost and its GPU usage.
"What I think is, this is a key changer for one reason, because it's open source."
Al Olama warned that the push to make the software open-source could be a ploy to capture market share and said ultimately whether the model was a game changer would depend on whether future models are also be released as open-source, hence setting a new standard for global players.
"If it's just them trying to plant the flag and say we are doing things differently and we first order that we release is open source, then I think this is going to be a short-lived celebration," he added.
Al Olama said the model's usage of "distilling" a practice that involves transferring knowledge from a large model to a smaller one was notable alongside the use of innovative algorithmic models.
Other experts at the summit shared his views. Anna Gressel a partner at Paul, Weiss, Rifkind, Wharton & Garrison LLP, and popular US AI analyst, said the release of DeepSeek was a "moment" for open-source technologies, which could potentially benefit the UAE, which has been a leader in this domain.
"DeepSeek made certain changes to how they were doing the training… they used a mixture of expert's model which is essentially a bunch of little experts that specialise in answering questions, and it can make the model more efficient," she said.
"But they were not the first to do that. ATRC, which is based in Abu Dhabi, had a major release of Falcon that was based on mixture of expert models - that was at least six months ago."
"The UAE is leading on open source, and I think this is a testament to how fast companies can move, and countries can move when you have open source in the mix," she added.
Similarly, investor Ashish Puri, a partner at UK-based Lightrock, which manages over $5.5 billion in assets, many of which are in tech, said that DeepSeek had challenged the premise that AI needed obscene amounts of GPU capacity and power to be effective.
He believed that this would have a democratising effect on AI."What DeepSeek did was essentially take 60-plus billion parameter model and reduce the number of parameters you actually need at any point in time to be able to find the answers," he told Aletihad.
"If you think about it, if this is all true, then the number of chips and the amount of power you need is reduced and AI will be a lot more efficient. Or more importantly, I think it will democratise AI."
"What DeepSeek did is what Toyota did to the automotive industry, it made it more accessible," he added.
However, there was cause for concern, particularly around the model's data governance, with many European and Canadian AI leaders saying they would not allow the models use in their country for sensitive functions.
Australia on Tuesday issued an official order banning DeepSeek from all government devices as it seeks to block "an unacceptable level of security risk" presented by China's AI programme.
"After considering threat and risk analysis, I have determined that the use of DeepSeek products, applications and web services poses an unacceptable level of security risk to the Australian Government," Department of Home Affairs Secretary Stephanie Foster said in the directive.
And Ott Velsberg, Estonian Government Chief Data Officer and Ministry of Justice and IT, said: "On the question of DeepSeek, simply no, Italy yesterday banned its use. So, if a country cannot show adequacy level to that of EU, then you should not use it."
"There are serious concerns related to kind of personal data processing, also copyrights. So, I think that's an issue, it's not an open-source model in a complete sense."