Iris Coleman
Dec 17, 2025 06:09
Dan Fu from collectively.ai argues that synthetic normal intelligence (AGI) is achievable by optimizing software-hardware co-design, enhancing present chip utilization, and overcoming perceived {hardware} limitations.
The talk surrounding the potential for reaching synthetic normal intelligence (AGI) is intensifying, with Dan Fu, Vice President of Kernels at collectively.ai, offering an optimistic outlook. In keeping with collectively.ai, Fu challenges the notion that developments in AI are being stymied by {hardware} limitations. As an alternative, he posits that present chips are considerably underutilized and {that a} strategic strategy to software-hardware co-design may unlock substantial efficiency enhancements.
Present Limitations and Future Potential
Because the AI panorama evolves, considerations about reaching the bounds of digital computation have gotten extra prevalent. Some consultants counsel that {hardware} constraints, significantly in GPUs, would possibly impede progress in the direction of growing typically helpful AI. In distinction, Fu presents a extra hopeful perspective in his publication, “Sure, AGI Can Occur – A Computational Perspective,” which argues that the ceiling has not but been reached for AI capabilities.
Underutilization of Current {Hardware}
Fu highlights that state-of-the-art AI coaching runs, akin to DeepSeek-V3 or Llama-4, usually obtain solely about 20% Imply FLOP Utilization (MFU), with inference utilization typically within the single digits. These figures counsel a big alternative to reinforce effectivity by higher integration of software program and {hardware}, in addition to improvements like FP4 coaching.
Developments in Computational Fashions
Present AI fashions are based mostly on older {hardware}, and the potential of newer computational sources has not been absolutely realized. Fu emphasizes that large clusters of the most recent technology GPUs, numbering over 100,000, have but to be absolutely built-in into AI improvement processes, indicating a promising horizon for future developments.
Current-Day Utility and Future Implications
Regardless of the perceived limitations, current AI fashions are already revolutionizing complicated workflows, akin to writing high-performance GPU kernels with human help. This transformation underscores the quick utility of AI applied sciences and hints on the huge potential for future functions.
For these within the intersection of programs engineering, {hardware} effectivity, and AI scaling, Fu’s evaluation supplies useful insights. The total evaluation may be accessed on the collectively.ai web site.
Picture supply: Shutterstock