The AI Crowd is Mad

The author argues that current discourse about large language models like ChatGPT is too optimistic and needs more nuance, while stating that they are not bearish on AI's potential as a technology. The post argues that bubbles are defined by investors' inflated expectations rather than by the technology itself, and it points to social media survivorship bias as a reason ChatGPT's performance looks better online than in practice. It also questions whether the costs of running transformer models can yield returns, citing a parameter-and-cost table of estimates, and it doubts that source referencing will be a simple "last-mile" fix.