{"id":20915,"date":"2026-07-26T08:38:39","date_gmt":"2026-07-26T02:53:39","guid":{"rendered":"https:\/\/eng.dragonmedia.com.np\/?p=20915"},"modified":"2026-07-26T08:40:48","modified_gmt":"2026-07-26T02:55:48","slug":"openness-the-real-alternative-to-americas-ai-dilemma","status":"publish","type":"post","link":"https:\/\/eng.dragonmedia.com.np\/?p=20915","title":{"rendered":"Openness: The Real Alternative to America\u2019s AI Dilemma<\/strong>"},"content":{"rendered":"

# Pu Yu Hai<\/strong><\/p>\n

\u201cArtificial intelligence will transform every industry, empower every company and be built by every country.\u201d<\/h4>\n

This sentence, posted by Nvidia Chief Executive Jensen Huang in his first message on X, is more than a prediction about the future. It raises a central question in the global technology competition: Will AI become a closed service controlled by a handful of large corporations, or a shared technological foundation that countries, institutions, universities and businesses can develop and use according to their own needs?<\/p>\n

This question has exposed a clear contradiction between the US AI industry and the American policymaking establishment. Leading technology companies warn that imposing sweeping restrictions on open-weight models too early could weaken competition, push innovation overseas and concentrate technological control in a small number of companies. Yet an influential section of the US political and security establishment increasingly views AI through the lenses of control, export restrictions and geopolitical rivalry.<\/p>\n

This is not simply a disagreement over regulation. It is a strategic dilemma over whether the United States should preserve its historic technological leadership through open competition or attempt to protect it through barriers and controls.<\/p>\n

Companies and organisations including Nvidia, Microsoft, Meta, Palantir, IBM and Hugging Face have urged US policymakers not to impose premature restrictions on open models through a joint letter titled \u201cOpen Weights and American AI Leadership.\u201d Their central argument is clear: American leadership in AI should not be measured by a single powerful frontier model, but by whether the country can build an open and competitive ecosystem capable of spreading the technology across industry, hospitals, schools, agriculture, small businesses and public institutions.<\/p>\n

This argument is consistent with America\u2019s own technological history. The internet, Linux, web standards and much of modern software development were built on open knowledge, shared technical standards and global developer communities. The open-source movement did not eliminate commercial companies. It created the foundations for new industries, services and markets. Even today\u2019s largest technology companies remain deeply dependent on open-source software.<\/p>\n

Restricting AI entirely within closed commercial structures is therefore not merely a security policy. It is also a question of political economy concerning who will control technological ownership, economic benefits, knowledge and social power.<\/p>\n

Open-weight and fully open-source models are not identical. In an open-weight model, the numerical weights of a trained system are made available, allowing developers and institutions to download, test, customise and operate it on their own infrastructure. However, a model may still be classified as open-weight even when its training data, source code and complete development process are not publicly disclosed.<\/p>\n

Even with these limitations, open-weight systems give users greater control than services built around closed application programming interfaces. Companies can run models on their own servers without sending sensitive data to an external provider. Universities can conduct independent research. Smaller countries can adapt models to local languages and cultural contexts. Startups can integrate models into their products instead of paying high fees for every use.<\/p>\n

In a closed-model system, by contrast, the provider controls access, pricing, service availability and terms of use. Institutional knowledge accumulated over many years may become tied to the technical infrastructure of a single company. This is known as \u201cvendor lock-in,\u201d or excessive dependence on a service provider. When a provider raises prices, changes policies, discontinues a service or restricts access in a particular country, users may have no immediate alternative.<\/p>\n

As AI becomes part of education, healthcare, communication, research and public administration, such dependence is no longer a routine commercial inconvenience. It becomes a matter of data sovereignty, institutional autonomy and national security.<\/p>\n

The rise of open models developed in China has intensified this debate. Chinese company Moonshot AI\u2019s Kimi K3 model, which is distinct from DeepSeek\u2019s models, is a recent example. According to the company, it has 2.8 trillion parameters, native visual understanding and a context capacity of up to one million tokens. Its significance lies not only in its scale, but in the strategy of making a powerful model for long-form coding, knowledge work and complex reasoning openly available.<\/p>\n

China\u2019s open-model strategy should not be understood solely as an expression of technological idealism. It also serves clear industrial and geopolitical objectives. When open models are widely available, developers around the world become connected to Chinese models, tools and technological ecosystems. This gives China an opportunity to expand its global influence through software, applications and technical standards.<\/p>\n

But the strategic interests of a competitor do not eliminate the value of openness. American closed models are not free from commercial and institutional interests either. The central question is not which side is selfless, but which structure gives users, researchers, smaller companies and developing countries more choice, control and capacity for adaptation.<\/p>\n

The rapid progress of Chinese models has also intensified debate in the United States over model distillation and intellectual property. Distillation uses outputs from a powerful model to train, improve or evaluate another model. It is a widely used method in AI research and development.<\/p>\n

Distillation is not inherently illegal. However, legal questions arise when it is conducted by violating terms of service, collecting data without authorisation or misusing protected intellectual property. That is a question about the legality of how the technique is used, not about the legitimacy of the technique itself.<\/p>\n

A clear distinction must therefore be maintained between proven violations and normal technological competition. Declaring a rival model to be the product of theft merely because it has become powerful is not an evidence-based assessment. Proven misuse can be addressed through legal and commercial mechanisms, but imposing sweeping restrictions on an entire technological method or a country\u2019s AI industry may be a disproportionate response.<\/p>\n

US controls on advanced chips, manufacturing equipment and computing capacity are intended to slow China\u2019s AI development. Such restrictions may create short-term difficulties, but they can also encourage competitors to build more efficient models with fewer resources, develop alternative architectures and strengthen technological self-reliance.<\/p>\n

This is the historic contradiction of technology controls. Restrictions may buy time, but they do not guarantee permanent leadership. Knowledge, research, talent and software are difficult to contain behind borders in the same way as physical goods.<\/p>\n

This does not mean open models are free of risk. Powerful systems may be misused for cybercrime, fraud, disinformation and other harmful activities. Once model weights are released, controlling and monitoring modified versions becomes more difficult. Establishing responsibility for misuse can also become complicated.<\/p>\n

Yet the claim that closed models are automatically secure is equally misleading. Closed systems can be hacked, secretly abused or produce harmful and inaccurate results without meaningful scrutiny from outside researchers. Concentrating AI power within a small number of companies does not eliminate risk. It may instead create centralised points of failure.<\/p>\n

The proper response is therefore not to reject either open or closed models entirely. Regulation should vary according to the level of risk, the capability of the model and its field of application. Highly capable models may require independent testing, safety assessments, cybersecurity evaluation and clear accountability for misuse. Imposing the same level of restriction on limited research, educational and language models, however, could obstruct innovation.<\/p>\n

For Nepal, this debate has direct significance. Nepal cannot immediately compete with the United States or China in developing frontier models, but it does not have to remain a passive consumer of foreign AI services. Open models can create opportunities to develop locally relevant systems for Nepali and other national languages, agriculture, education, tourism, disaster management, public health information, government services and journalism.<\/p>\n

Dependence on closed foreign services may cause Nepali data to be stored abroad, local languages and cultural contexts to be misunderstood, and national use to be affected by a provider\u2019s pricing and policies. If open models can be operated and adapted within Nepal, they could strengthen data sovereignty, language preservation and local innovation.<\/p>\n

Access to downloadable models alone, however, does not amount to self-reliance. Nepal also requires computing infrastructure, reliable Nepali-language data, researchers, software engineers, cybersecurity capacity, adequate electricity and sustained public investment. Even an open model offers limited autonomy when the infrastructure needed to operate it remains controlled by foreign cloud companies.<\/p>\n

When developing a national AI policy, Nepal should prioritise open standards, interoperability and access to multiple sources rather than aligning itself with the technology position of a single country or company. Public agencies procuring AI services should ensure control over their data, the ability to audit systems, the right to change models and the capacity to transfer services to another provider.<\/p>\n

Shared computing centres could be established for universities and research institutions. Legally usable datasets in Nepali and other national languages should be developed. Independent mechanisms are also needed to evaluate model accuracy, bias, security and linguistic capability. In journalism, open AI can support translation, archive searches, data analysis and local-language content production, but its use must remain tied to copyright protection, source transparency, fact-checking and human editorial responsibility.<\/p>\n

Neither the United States nor China will write the future of AI alone. It will be shaped by countries, companies, universities, communities and developers across the world. This does not mean every country must build a massive model from the ground up. It means every country should possess the capacity to test, adapt and use AI according to its own language, culture, economy and national priorities.<\/p>\n

To preserve its leadership, the United States should invest in research, talent, infrastructure, security and open technological ecosystems rather than building walls against competitors. The same conclusion is relevant to Nepal. Instead of attaching itself to one technological pole, Nepal should build its own capacity by drawing on knowledge, models and technologies available across the world. Openness must be balanced with security, rights and accountability.<\/p>\n

Openness is not a risk-free path, but building walls around knowledge is even more dangerous. The future cannot be stopped by barriers; it is welcomed through an open courtyard. Confidence in participation is stronger than the illusion of control.<\/p>\n

\n\t\t\t
\n\t\t\t <\/div><\/div>\n\t\t