4 ms·

Can anyone explain what "extra knowledge" means specifically? Is that like fine tuning? How much data can I give it to learn? How much can it retain? Can it be

Can anyone explain what "extra knowledge" means specifically? Is that like fine tuning? How much data can I give it to learn? How much can it retain? Can it be updated over time?

minimaxir 3y ago
The keynote used a .txt file of a lecture that the user uploaded as a data source the model can select from. From a technical perspective, it's anoher data source for retrieval-augmented generation (RAG) doing a vector search in the background: https://platform.openai.com/docs/assistants/tools/knowledge-retrieval https://platform.openai.com/docs/assistants/tools/knowledge-...
- tracerbulletx 3y ago
  Ah ok, thank you.
- singularity2001 3y ago
  "optimizes for quality by adding all relevant content to the context of model calls." So for their own profits they maximize recall and let GPT handle precision.
Racing0461 3y ago
I doublt it's fine tuning (actually changing the model weights). It's more like "im going to paste a text blob then in the following chats i will ask questions about it) type inner prompt.
- singularity2001 3y ago
  For longer documents it uses vector embeddings
  - Racing0461 3y ago
    How's that different from pasting the text in the first chat and running the vector embedding step on the text on the server (maybe at least bypassing the chat text limit)? Does this fix the amnesia issue where the info from chats longer than the context length is forgotton because the document isn't baked directly into the weights like fine tuning?
ankit219 3y ago
They showed a demo where you could upload a file while creating an agent. As others have answered. I think it's about configuring an agent in a way that you give it material on some specific topic (one file, multiple files) and it uses Retrieval to augment the answer based on the source material. Google launched Notebook LM[1] a while back which does a similar thing conceptually. It allows you to import a Google drive folder with docs of the stuff you would want to understand and then just chat with it. It's a good product but restrictive in the sense that it only allowed Google docs. [1] https://blog.google/technology/ai/notebooklm-google-ai/ https://blog.google/technology/ai/notebooklm-google-ai/
vunderba 3y ago
The demo just showed a file dialog box where they could upload a set of static files. What I'd really like to see is the ability to sync a source integration (for example into a GitHub repo or a notion account), and it would always pull relevant information using a RAG architecture.

Can anyone explain what "extra knowledge" means specifically? Is that like fine tuning? How much data can I give it to learn? How much can it retain? Can it be updated over time?

minimaxir 3y ago

The keynote used a .txt file of a lecture that the user uploaded as a data source the model can select from. From a technical perspective, it's anoher data source for retrieval-augmented generation (RAG) doing a vector search in the background: https://platform.openai.com/docs/assistants/tools/knowledge-retrieval https://platform.openai.com/docs/assistants/tools/knowledge-...

tracerbulletx 3y ago
Ah ok, thank you.
singularity2001 3y ago
"optimizes for quality by adding all relevant content to the context of model calls." So for their own profits they maximize recall and let GPT handle precision.

Racing0461 3y ago

I doublt it's fine tuning (actually changing the model weights). It's more like "im going to paste a text blob then in the following chats i will ask questions about it) type inner prompt.

singularity2001 3y ago
For longer documents it uses vector embeddings
- Racing0461 3y ago
  How's that different from pasting the text in the first chat and running the vector embedding step on the text on the server (maybe at least bypassing the chat text limit)? Does this fix the amnesia issue where the info from chats longer than the context length is forgotton because the document isn't baked directly into the weights like fine tuning?

ankit219 3y ago

They showed a demo where you could upload a file while creating an agent. As others have answered. I think it's about configuring an agent in a way that you give it material on some specific topic (one file, multiple files) and it uses Retrieval to augment the answer based on the source material. Google launched Notebook LM[1] a while back which does a similar thing conceptually. It allows you to import a Google drive folder with docs of the stuff you would want to understand and then just chat with it. It's a good product but restrictive in the sense that it only allowed Google docs. [1] https://blog.google/technology/ai/notebooklm-google-ai/ https://blog.google/technology/ai/notebooklm-google-ai/

vunderba 3y ago

The demo just showed a file dialog box where they could upload a set of static files. What I'd really like to see is the ability to sync a source integration (for example into a GitHub repo or a notion account), and it would always pull relevant information using a RAG architecture.