In Singapore, a government-funded artificial intelligence model
Converse in 11 languages spans from Indonesian to Lao. In Malaysia,
ilm chat
developed by a local construction conglomerate, claims it “knows which Georgetown you’re referring to.” Thus, it’s not a private university in the US, but the capital of Penang. Conversely, the Swiss Apertus
announced in September
that it can differentiate when to use “ss” in Swiss German instead of the “ß” used in standard German.
Globally, language models like these are integral to an AI arms race valued in the hundreds of billions of dollars.
dollars
Much of this is led by a few dominant companies in the US and China. As OpenAI, Meta,
Alibaba, and others invest billions in building more advanced models, middle powers and developing nations are closely monitoring the landscape and often making significant commitments of their own.
These initiatives are part of a movement loosely termed “sovereign AI,” where nations from the UK to India to Canada aim to create their own AI solutions and establish their positioning within this evolving ecosystem.
Yet, with hundreds of billions in play globally, can smaller investments yield substantial returns?
“U.S.-based firms, the U.S. government, and China can practically storm ahead in AI development, making it challenging for smaller nations,” noted Trisha Ray, a senior researcher at the Atlantic Council, a U.S.-based strategic think tank.
“Unless you’re a wealthy government or major corporation, creating a large language model from scratch is a considerable burden.”
Defense Concerns
Nonetheless, numerous countries are hesitant to depend on foreign AI for their requirements.
India, the second-largest market for OpenAI, has recorded over 100 million ChatGPT downloads in recent years. However, Abhishek Upperwal, founder of
Socket AI, highlights several instances where U.S.-made AI systems have fallen short. For example, a deployed AI agent intended to educate students in a remote Telangana village communicates in English but with a heavy, nearly incomprehensible American accent, while an Indian legal startup’s effort to adapt Meta’s LLaMa AI model encountered barriers, resulting in a mixed bag of U.S.-Indian legal advice, Upperwal explains.
There are also looming national security concerns. For India’s defense sector, any Chinese deep learning model is considered off-limits, according to Upperwal. “This could encompass untrustworthy training data claiming that Ladakh isn’t a part of India… Utilizing such a model in a defense context is absolutely unacceptable.”
“I’ve spoken with individuals involved in defense,” Upperwal stated. “They want to leverage AI, but they disregard DeepSeek and wish to avoid reliance on it altogether. Using U.S. systems like OpenAI is distinctly problematic since it risks data leaks from the country.”
Socket AI represents one of the few initiatives aimed at constructing a national LLM for India, supported by the IndiaAI Mission, a government-funded project that has invested roughly $1.25 billion in AI advancements. Upperwal envisions a model less resource-intensive than those produced by major American and Chinese tech firms, closely aligning with some from the
French AI company Mistral.
AI researchers have long contended that pushing the technology boundary to reach the often-elusive goal of artificial general intelligence (AGI) will necessitate considerable resources, including chips and computing capabilities. Upperwal emphasizes that India must compensate for its funding gaps with talent.
“In India, spending billions is not an option,” he asserts. “How can we compete against the $100 to $500 billion being invested by the United States? I believe leveraging core expertise and intellect is crucial.”
In Singapore, AI Singapore is a government initiative backing the SEA-LION project. SEA-LION is a suite of language models designed specifically for Southeast Asian languages that are typically underrepresented in U.S. and Chinese LLMs, such as Malay, Thai, Lao, Indonesian, and Khmer among others.
Leslie Teo, Senior Director at
AI Singapore, notes that these models aim to enhance rather than overshadow larger ones. Systems like ChatGPT and Gemini often falter with regional languages and cultural contexts, according to Teo. For instance, they may communicate in excessively formal Khmer or suggest pork-based recipes to users in Malaysia. Creating local language LLMs will empower local governments to code with cultural intricacies or at the very least become “smart consumers” of robust technologies developed abroad.
“I am very cautious with the term sovereignty. Essentially, we want better representation and a clearer understanding of how AI systems operate,” he states.
Multilateral Cooperation
For nations seeking to carve out a niche in an increasingly competitive global arena, collaboration is another option. Researchers tied to
Bennett School of Public Policy at the University of Cambridge have lately suggested forming a public AI enterprise distributed across a consortium of middle-income nations.
They refer to this initiative as
Airbus for AI, alluding to Europe’s successful efforts in establishing a competitor to Boeing in the 1960s. Their proposal envisages creating a public AI company that would unify the resources of AI initiatives from the UK, Spain, Canada, Germany, Japan, Singapore, South Korea, France, Switzerland, and Sweden, aiming to forge a formidable rival to the tech giants of the U.S. and China.
Joshua Tan, the lead author of a paper outlining the initiative, mentioned that the idea has garnered interest from AI ministers in at least three nations and several sovereign AI firms. While the emphasis is currently on “powerful middle powers,” developing nations like Mongolia and Rwanda are also reportedly expressing interest.
“There’s certainly less trust in the current U.S. administration’s commitments. Questions are arising about the reliability of this technology and what might occur if they withdraw support,” he remarks.
Tan’s proposal is optimistic about the potential for collaboration among nations. However, critics suggest that even a coordinated multi-country strategy could squander taxpayer resources on initiatives that may not yield fruitful results.
“I hope that those developing this [sovereign] AI model understand how far and how rapidly advancements are progressing,” comments Tzu Kit Chan, an AI strategist advising the Malaysian government.
“What’s the alternative? If governments pursue flawed strategies in crafting their own sovereign AI models, they risk wasting vast amounts of capital.”
According to Chan, a more prudent approach would be for governments like Malaysia’s to allocate these funds toward enhancing AI safety regulations, as opposed to competing with globally dominant products that have already captured the market.
“Walk down the streets of Malaysia, visit Kuala Lumpur, engage with your financial counterparts and inquire about the models they utilize,” he suggests.
“Out of 10, I doubt that more than 2 are employing a sovereign AI model. Most are using ChatGPT or Gemini.”
Source: www.theguardian.com
