Update January 30, 2023:
OpenAI has released another update for ChatGPT, which should provide improved factuality and mathematical capabilities. The update has just been announced, no further information is available yet. You can see your ChatGPT version below the prompt input box.
Article from December 16, 2022:
OpenAI releases ChatGPT update and new embedding model
OpenAI announces several updates to ChatGPT and a new embedding model.
ChatGPT gets better: OpenAI introduces a “general performance” improvement that should, among other things, make ChatGPT give more answers to questions, meaning that it is “less likely” to refuse to answer questions.
For this, OpenAI is likely to have fine-tuned the AI model with further human feedback from the beta test so far. OpenAI co-founder Sam Altman announced this fine-tuning on Twitter back in early December, anticipating a “significant improvement” at the time.
ChatGPT gets conversation history
Conversation histories will also soon be available, allowing users to retrieve past conversations and rename or delete saved conversations. OpenAI will roll out this feature gradually.
In addition, OpenAI aims to collect more human feedback for ChatGPT with a reward system: To this end, OpenAI limits your requests. If you have exhausted your requests but want to make more, you must first provide feedback on ChatGPT responses. As a reward, you will then receive more capacity.
You can see if you are using the current ChatGPT version at the bottom of the screen: It should say “ChatGPT Dec 15 Version”.
OpenAI introduces a more performant embedding model for less money
Also new is OpenAI’s embedding model “text-embedding-ada-002”. This model is supposed to merge the five most performant models so far into one.
Embeddings are the translation of concepts into numbers so that computers can more easily establish relationships between different concepts. They are used for semantic search, content clustering, recommendation systems, and many other applications.
Ada-002 outperforms older embedding models in text search, code search, and sentence similarity tasks, and achieves comparable performance in text classification, according to OpenAI. However, the embeddings are only one-eighth as large as those of “davinci-001” and the new model can process four times the context (2048 to 8192).
According to OpenAI, the new model achieves better or similar performance compared to davinci models at a 99.8 reduction in price. More information about ada-002 is available at OpenAI.