AI industry leaders' forecasts on model scaling have been questioned following reports of issues in developing next-gen models by OpenAI, Anthropic, and Google Gemini. The new focus is on 'inference scaling' which emphasizes using more computation during task performance. While the AI community is uncertain about the future of AI progress, inference scaling presents potential short-term gains though with unpredictable outcomes across different domains. The article advocates for reduced deference to insider predictions and emphasizes the need for product development to catch up with AI capability increases.
Table of contents
Is model scaling dead?Let’s stop deferring to insidersWill progress in capabilities continue through inference scaling?Is inference scaling the next frontier?Product development lags capability increaseConcluding thoughtsSort: