AI Industry Shifts Focus to Efficiency and Adaptability Over Model Size
The artificial intelligence (AI) industry is increasingly prioritizing efficiency and adaptability in AI models, moving beyond the years-long pursuit of ever-larger systems. This shift is driven by the rising costs associated with deploying and scaling AI, particularly for inference. Experts argue that current AI models are often 'monolithic' and lack the ability to evolve, leading to significant inefficiencies and increased operational expenses for businesses.
After years of focusing on building larger AI models, the industry is now concentrating on making these systems affordable and practical for large-scale deployment. This changing priority was a key discussion point at Fortune Brainstorm Tech, where industry leaders highlighted the challenges of current AI infrastructure.
Sara Hooker, cofounder and CEO of AI lab Adaption, described most contemporary AI as 'monolithic.' She explained that once a model is trained, its knowledge and capabilities become fixed, failing to incorporate new information or user-learned insights automatically. Hooker emphasized the necessity for models that can evolve, warning that a lack of adaptability leads to considerable inefficiencies and contributes to escalating API costs for companies deploying AI agents.
Rodrigo Liang, CEO of AI chip company SambaNova, acknowledged that while scale remains important and large models are not disappearing, there is significant room for more efficient models. He noted that customers currently face difficulties with the high cost of scaling models, energy-intensive infrastructure, and securing sufficient AI expertise.
Hooker believes the industry is at an 'inflection point' requiring urgent changes to the model size curve. She suggested that applying the same massive model to all problems is inefficient, particularly for straightforward bulk processing tasks. Future AI systems, she argued, must continuously adapt to new data and modify their behavior without repeatedly calling a static model, thereby reducing compute and infrastructure costs.
Liang stated that a primary challenge for the industry is to operate today's massive models efficiently enough for real-world economic viability. He pointed out that trillion-parameter models are currently too expensive and power-hungry. SambaNova's strategy focuses on developing hardware specifically designed for large-model workloads, aiming to deliver faster inference with lower power consumption to help reduce overall costs.
According to Fortune, these insights were shared during the 25th annual Fortune Brainstorm Tech conference.
Advertisement
AdSense slot • inline
