OpenAI announced an update to ChatGPT intended to improve the system’s factual accuracy and mathematical capabilities. The announcement provided no additional details, but users could check the version shown below the prompt input box.
A new update with limited detail
The update was announced on January 30, 2023. OpenAI described its aims as better factuality and math skills, without explaining how the changes were made or offering information about their expected effects.
That leaves the announcement focused on the goals rather than evidence of results. The version label in ChatGPT offered a way for users to see which release they were using, but the article did not provide a benchmark or examples to measure the update’s impact.
Earlier changes focused on performance
A prior update, described in an article from December 16, 2022, was presented as a general performance improvement. One intended change was that ChatGPT would answer more questions and be less likely to refuse them.
The article said OpenAI was likely to have fine-tuned the model using additional human feedback from the beta test. It also noted that OpenAI co-founder Sam Altman had announced fine-tuning on Twitter in early December, anticipating a “significant improvement.”
These earlier changes point to two distinct concerns: how often the chatbot responds, and how dependable its answers are. The January announcement addressed factuality and math, while the December update aimed at general performance and answer availability.
Conversation history and user feedback
The December update also described plans for conversation histories. The feature would let users retrieve earlier chats and rename or delete saved conversations, with OpenAI rolling it out gradually.
OpenAI also aimed to collect more human feedback through a reward system. When users reached their request limit, they could provide feedback on ChatGPT responses to receive more capacity. This connected product access with a way to gather reactions to the model’s answers.
The article said users could identify the current version at the bottom of the screen, where it should read “ChatGPT Dec 15 Version.” The later January announcement pointed users to the area below the prompt input box to check their version.
A separate embedding model release
The December article also covered OpenAI’s embedding model, “text-embedding-ada-002.” It was intended to combine the five most performant models available at the time into one. Embeddings represent concepts as numbers, helping computers identify relationships among them.
That approach supports uses such as semantic search, content clustering and recommendation systems. According to OpenAI, ada-002 performed better than older embedding models on text search, code search and sentence similarity, and delivered comparable performance in text classification.
The model’s embeddings were one-eighth as large as those of “davinci-001,” and it could process four times the context, from 2048 to 8192. OpenAI said it achieved better or similar performance to davinci models at a 99.8 reduction in price.
Together, the announcements covered updates to ChatGPT’s responses and a separate model for turning text into numerical representations. The January release’s stated focus was factuality and math, though the announcement itself gave readers no further information about how those improvements would be assessed.