Yes, AGI Can Happen – A Computational Perspective (danfu.org)

A new blog post by Dan Fu challenges the notion that advancements toward Artificial General Intelligence (AGI) are stagnating due to hardware limitations. While Tim Dettmers argues that the efficiency of current AI models is nearing a ceiling, Fu contends that today's AI systems remain significantly underutilized, with much room for improvement in both software and hardware efficiencies. He emphasizes that models like DeepSeek-V3 and Llama-4 achieve only around 20% of their potential FLOP utilization, despite newer hardware capabilities being available, suggesting that as hardware evolves, so too must the model architectures designed to leverage it. This discussion is critical for the AI/ML community as it highlights tangible paths forward in optimizing AI training processes and architectures. Fu identifies key areas for future research, including better co-design of efficient model architectures, advancing FP4 training techniques, and developing inference-efficient designs. These efforts could transform current computational capabilities, unlocking a wealth of potential for more effective AI systems that align closer to the principles of AGI. By focusing on improving efficiency rather than solely emphasizing hardware advancements, the community may accelerate the journey toward generally useful AI that meets varied definitions of AGI.