AI Compute Overkill? Smaller Models Threaten Big Tech

Artificial intelligence companies might be constructing significantly more computing power than required, according to Tyler Saltsman, CEO of EdgeRunner AI. He suggests that current compute build-out is excessive and that the escalating cost of AI could lead to a decrease in data center demand.

Saltsman envisions AI evolving into a network of specialized models working collaboratively, rather than relying on single, colossal systems. He specifically stated that the world does not need Artificial General Intelligence (AGI), but rather domain-specific models on par with GPT-4. His company, EdgeRunner AI, develops specialized on-device AI models, such as a military assistant that operates on consumer-grade hardware with limited video memory.

By training and then compressing their models, EdgeRunner enables them to run directly on devices, eliminating the need for continuous cloud requests. This method could significantly impact the profitability of large AI firms by removing per-token charges. Saltsman also believes that the rapid progress of open-source models is a primary reason behind statements from leaders of Anthropic and OpenAI advocating for a slowdown in frontier AI development.

EdgeRunner focuses on creating smaller, military-specific AI models designed for deployment on soldiers' devices, functioning even without an internet connection. Saltsman contends that tailoring these models with military doctrine enhances their reliability for battlefield tasks compared to general-purpose large language models. He cited an instance where EdgeRunner's approach reduced error rates by 37% in a training exercise utilizing data from the Army's Artificial Intelligence Integration Center.