Source item

The new OpenAi o3 reasoning model is mind blowing. DeepSeek could sort of compete with o1-mini but not necessarily o1-pro but the new o3 is well ahead - in terms of analytical capabilities and deep hard science problem solving - than anything DeepSeek can currently do. And inference rather than training will be where most of the compute time takes (30 minutes to write a complex economic model paper). So only hyperscalers - rather than the DeepSeeks pretending to require little compute time - will be able to manage the onslaught of data (mostly proprietary and synthetic) and the inference compute demands that are needed in this brave new world. It re-enforces my point that DeepSeek is an evolutionary change not a revolution in AI. It helps gradually over time to increase global TFP growth but most of the benefits still go to US producers of AI - more US than Chinese hyperscalers - and eventually to US and global consumers of AI . Over time a. Net positive for US and global stock markets