Is Bigger Always Better for Large Language Models (LLMs)?
Yasmeen Ahmad, managing director of strategy and outbound product management for data, analytics, and AI at Google Cloud, recently shared insights on this topic at VB Transform. Her straightforward answer? “Yes and no.”
While larger LLMs tend to perform better, they don't do so indefinitely. Ahmad emphasized that smaller models trained on domain-specific data can outperform their larger counterparts. “Data is at the cornerstone,” she noted, highlighting how industry-focused information empowers models.
This approach fosters creativity, efficiency, and inclusivity within enterprises. By leveraging previously inaccessible data, organizations can engage employees in innovative ways.
“Generative AI is expanding the horizons of what machines can create,” Ahmad said. “It blurs the line between technology and magic, potentially redefining what we consider magic.”
The Need for a New AI Foundation
Training LLMs effectively within a specific enterprise context involves two key techniques: fine-tuning and retrieval-augmented generation (RAG). Fine-tuning helps LLMs grasp “the language of your business,” while RAG connects the model to real-time data from various sources, such as documents and databases.
“This enables accurate answers crucial for financial analytics, risk analysis, and more,” Ahmad explained.
LLMs' true strength lies in their multimodal capabilities, allowing them to handle diverse data types like video, text, and images. Since 80 to 90% of enterprise data is multimodal, Ahmad stressed the importance of utilizing LLMs to harness this information effectively.
A Google study revealed a 20 to 30% improvement in customer experience when multimodal data was employed, enhancing companies' abilities to gauge customer sentiment and analyze product performance against market trends.
“It’s not just about pattern recognition anymore,” Ahmad stated. “LLMs can comprehend the complexities of our organizations by accessing all available data.”
Traditional companies face challenges with outdated data infrastructures that struggle to process multimodal information. The future of AI requires a new foundation built for this complexity.
Conversational AI: The Role of Interaction
Ahmad also highlighted the significance of question-answer interactions in successful LLMs. While chatting with business data seems appealing, it poses challenges.
For instance, if you ask a colleague for next quarter’s sales forecast without providing context, their answers may be vague. The same applies to LLMs, which require semantic context and metadata to deliver accurate responses.
Human analysis often involves back-and-forth dialogue to refine questions and gain clarity. Similarly, LLMs must foster coherent conversations, evolving from isolated, one-off interactions to "the next generation of conversational AI.”
“Think of it as a personal data sidekick,” she suggested. This tireless worker can engage in thoughtful exchanges, enabling transparency in queries so users can trust the results.
Ahmad referred to the emergence of “agentic AI”—systems capable of making decisions and pursuing goals. These models emulate human thought processes by breaking tasks into subtasks and developing strategic thinking skills.
With advancements in real-time capabilities, these developments are occurring at an unprecedented pace. “The future is here, and it’s giving rise to new breeds of business,” Ahmad concluded. “We are just at the beginning of what this technology can facilitate.”
Ranking
