An AI model deprecation dilemma
OpenAI said with the general availability of GPT-4 it would be deprecating several old models. It won't be the first time this happens.
As OpenAI continues to upgrade to newer language models (which is a good thing, to be clear) it increasingly faces the thorny proposition of how to manage moving its customers and developers onto those new models.
OpenAI said this week that it was making GPT-4 generally available, as well as several other models. In that same post, however, it announced it was deprecating a number of old models. It also said developers who have fine-tuned those models that will be deprecated will also have to do that again.
“As part of our increased investment in the Chat Completions API and our efforts to optimize our compute capacity, in 6 months we will be retiring some of our older models using the Completions API,” the company said in its blog post. “We plan for future model and product improvements to focus on the Chat Completions API, and do not have plans to publicly release new models using the Completions API.”
The developer migration problem is one that many companies with existing developer tools have faced before (like Google migrating to updated versions of TensorFlow). But the pace of development and sensitivity these models have makes it a new and difficult proposition.
And migrations to newer and updated products also expose another one of the big challenges foundation model developers offering their products through an API will have to deal with: their customers have to recognize don’t own their own models.
"At MosaicML, we believe that the ultimate control over models should be with users, not big companies,” MosaicML chief scientist Jonathan Frankle said.