AI for Humans · Thursday, September 24, 2026
A common observation in the AI space is that models can degrade in performance after their initial release. This phenomenon is attributed to factors like scaling up compute resources and managing costs, leading to 'dialing in' the model based on global usage.
“They will release a model, week one, it is as impressive as it gets and slowly but surely as people flock, you know, to to switch to models or whatever, people come in and then they degrade the model, they slow it down, they cut back on it.”
“So, behind the scenes, there is an Oz at Anthropic and Open AI and even at Grok that is pulling the levers and saying, based off of worldwide usage, based off of how much RAM we have available, how much compute we have, let's just go ahead and dial this in.”
“Some of it is because they just have too many users and not enough compute. Some of it is because they want it to look as great as it can week one. And then week two and three they're like, hey, we got to make some margins. We do have an IPO in the future. So, there's a myriad of reasons why, but the reality is that like week one, these models are mind-blowing. I want to see other perform week three and four.”