OpenAI o1: Incremental Progress, Not Revolution
September 14, 2024 by Vincent Schmalbach
o1 shows some improvements over GPT-4o. It fixed a small bug in my code, returning a working solution that was on par with Claude 3.5 Sonnet and better than GPT-4o.
But o1 isn't the leap forward that OpenAI's marketing might suggest. At its core, it's using enhanced reasoning to call itself multiple times and produce a "best attempt" output. This approach works, but it's not revolutionary.
I estimate about 80% of o1's functionality could be replicated using the existing API. You'd need to implement multiple calls, instruct the model to reason step-by-step, and synthesize a final result. o1 essentially wraps this process into a single call.
This release raises questions about OpenAI's progress. They're promising GPT-5 and AGI, but delivering incremental improvements.
