For two years, telling people we ran production workflows on GPT-3 earned polite nods. Since November, everyone from my barber to Fortune 500 boards has an AI strategy. ChatGPT did in eight weeks what the technology itself had not done in three years: it made AI a mainstream assumption.
What actually changed
The models got genuinely better. GPT-4 class models are not an incremental step. Tasks we had written off in 2021, like reasoning over longer documents or holding structure across a complex output, now mostly work.
Expectations inverted. In 2021 the burden of proof was on us to show AI was useful. Now the burden of proof is on anyone claiming their process should stay manual. That flip changes budget conversations completely.
Talent got curious. The best people in our network are no longer asking whether to learn this, but how fast. We see it across the community: the AI sessions at this year's Women in Tech Global Conference were the most attended tracks we have ever run.
What did not change
A chatbot on a broken process is still a broken process. The failed AI projects I get shown all share one shape: nobody redesigned the workflow. They bolted a model onto the existing mess and got a faster, more confident mess.
Facts still have to come from you. The models are better writers and still unreliable witnesses. Retrieval helps, review remains mandatory.
The moat is still not the model. Everyone now rents the same intelligence. What differentiates is the workflow you redesign, the data only you have, and the distribution you already own. That was true with GPT-3 in 2020 and it is more true now that the models are commodities with a price war.
Three years in, my conviction has only moved in one direction. The next step for me is no longer using these models. It is building the systems around them. More on that soon.