Ed Chi, Vice President of Research at DeepMind, recently urged Taiwan to shift its focus from large language model (LLM) development to AI software innovation during a keynote at a DIGITIMES event. His remarks highlight a critical debate in global AI strategy: whether nations should invest in foundational models or leverage existing infrastructure to build specialized applications.

The Case for AI Software Over LLMs

Taiwan's tech industry, known for its semiconductor dominance, has recently ventured into AI research, including LLM development. However, Chi argues that the island nation would gain more by concentrating on AI software solutions that solve real-world problems. "Training LLMs requires massive computational resources and data," he noted. "Instead of competing with tech giants like OpenAI or Google, Taiwan can excel in creating niche applications for healthcare, manufacturing, and smart cities."

Key Advantages of Focusing on AI Software

  • Lower Barrier to Entry: Developing AI applications doesn’t require the same level of infrastructure as training LLMs from scratch.
  • Faster ROI: Software solutions can be deployed quicker, addressing immediate industry needs.
  • Specialization: Taiwan can leverage its strengths in hardware (like TSMC’s chips) to build optimized AI tools.

Taiwan’s Unique Position in the AI Race

Taiwan sits at a crossroads. While it lacks the data scale of China or the US, it possesses:

  1. World-class semiconductor expertise (crucial for AI hardware optimization)
  2. Strong engineering talent (ideal for software innovation)
  3. Strategic partnerships (with Western tech firms seeking alternatives to Chinese suppliers)

Chi emphasized that Taiwan could become a "global hub for AI-enabled industrial applications" by focusing on vertical integration—combining its hardware prowess with agile software development.

Risks of Overinvesting in LLMs

Training large language models is not just resource-intensive; it’s also highly competitive. The costs are staggering:

Expense Estimated Cost
Training GPT-4 ~$100 million
Annual Cloud Compute Tens of millions (for upkeep)
Talent Retention High (global AI talent war)

For a smaller economy like Taiwan, these investments might not yield proportional returns, especially when tech giants already dominate the LLM space.

Success Stories in AI Software

Chi pointed to examples where focused AI software has driven impact:

  • Israel’s Cybersecurity AI: Companies like Darktrace use AI for threat detection without building foundational models.
  • Germany’s Industrial AI: Siemens integrates AI into manufacturing workflows, enhancing efficiency.

Taiwan could replicate this in areas like:

  • Semiconductor yield optimization (using AI to reduce chip defects)
  • Telemedicine (AI diagnostics for aging populations)
  • Smart agriculture (precision farming tools)

Policy Implications

Chi’s advice aligns with broader trends in AI economics. Governments are increasingly recognizing that:

  • Not every nation needs its own LLM (see Singapore’s focus on AI governance instead).
  • Collaboration beats competition (Taiwan could partner with DeepMind or Microsoft for model access).

Taiwan’s Ministry of Science and Technology has already allocated $2 billion for AI development. Redirecting even a fraction toward software could accelerate practical deployments.

The Road Ahead

For Taiwan, the path forward involves:

  1. Investing in AI talent (specialized courses in applied AI)
  2. Encouraging startups (grants for software, not just research)
  3. Public-private partnerships (e.g., AI solutions for Taipei’s smart city initiatives)

As Chi concluded, "The future isn’t just about who builds the biggest model—it’s about who uses AI to solve the hardest problems." Taiwan, with its unique blend of hardware and software expertise, is poised to do exactly that—if it plays to its strengths.