According to Andrew Ng (@AndrewYNg), improving large language models (LLMs) currently relies on a piecemeal, data-centric process rather than sweeping breakthroughs. Ng highlights that while LLMs demonstrate greater generality than prior AI systems, their ability to learn and adapt is still limited compared to humans. For specific applications—such as programming in niche languages or delivering accurate healthcare and finance insights—AI teams must curate, clean, and generate specialized data sets, often involving labor-intensive tasks like deduplication and paraphrasing (source: deeplearning.ai/the-batch/issue-332). Additionally, enabling LLMs to handle complex tasks such as web browsing requires building simulated environments for reinforcement learning. This data-centric approach shapes current AI development and signals significant business opportunities for firms specializing in domain-specific data engineering, annotation, and AI infrastructure. Ng forecasts that ongoing incremental improvements, rather than rapid leaps toward AGI, will continue to drive practical AI adoption and market growth in the coming years.