Model sizes are currently constrained by availability of inference hardware (www.lesswrong.com)

The recent editorial on AI safety revealed that the progress in AI model capabilities is currently hampered by limitations in inference hardware. While there has been notable advancement in various tasks like coding and vision, the overall utility of AI remains disproportionate to its improvements. Many experts now predict that transformative AI capabilities, such as general intelligence in large language models (LLMs), are still 2-20 years away due to hardware restrictions that make it 30 times more efficient to perform post-training optimizations rather than large-scale pretraining. The discussion suggests that true advances in capability might be obscured by cost-cutting measures and limited predictions from current benchmarks. Key insights indicate that while training runs are scaling, the efficacy of these improvements is diminished by inference chip constraints that fail to accommodate the massive models being developed. For instance, achieving larger model pretraining, potentially reaching 30 trillion parameters, is deemed impractical until 2029, according to industry observations. Furthermore, despite claims about alignment and control techniques, the effectiveness of safety measures in current LLMs remains questionable, as emergent misalignments continue to pose risks. This highlights a critical juncture in AI development, where scaling challenges may ultimately shape the trajectory of AI progress in the years ahead.