My worst take of all time

Quick Overview

The speaker discusses Dave Shapiro's tweet about his "worst take" regarding OpenAI's role in popularizing Chain-of-Thought (CoT) reasoning, admitting he underestimated its importance, and then shifts to discussing the controversy surrounding OpenAI's "Pulse" feature pricing model, contrasting the enterprise focus with consumer needs and the high utilization of GPUs overnight.

Key Points: Dave Shapiro admitted his tweet claiming OpenAI's Strawberry model lacked Chain-of-Thought (CoT) reasoning was his 'worst take' one year later, conceding he underestimated the development's importance. The original CoT paper was published at NeurIPS in January 2022, predating OpenAI's widespread adoption of the technique. The speaker analyzes the internal controversy at OpenAI regarding the 'Pulse' feature, where infrastructure teams highlighted that expensive GPUs were idle overnight, prompting finance to question utilization. Infrastructure staff reportedly scoffed at the idea of running GPUs overnight because 'users are sleeping,' suggesting a disconnect between enterprise cost focus and consumer usage patterns. The speaker notes that despite OpenAI's dominance in the consumer market, their current focus seems heavily weighted towards the enterprise market, which has different needs (like guaranteed service levels). The speaker argues that high GPU utilization times, such as 11 AM to 5 PM Eastern Time, are when models should be running, rather than sitting idle overnight. The speaker concludes by stating that the core issue is that finance teams often do not understand what a GPU is, leading to poor cost optimization decisions.

Context: The video centers on a discussion analyzing two recent tweets by AI personality Dave Shapiro. The first tweet is Shapiro retracting a previous statement about OpenAI's model capabilities, while the second tweet details an internal conversation at OpenAI regarding GPU utilization and pricing for their 'Pulse' feature, highlighting a conflict between infrastructure cost-saving measures and actual user demand patterns.