Following up on the OpenAI Dev Day, here a few practical notes:
- Some terms used in the keynote are different from what developers see: “GPTs” vs. “Assistant” (or are there differences?), “GPT-4 Turbo” vs. “gpt-4-1106-preview”.
- Playground remains buggy, it feels like a rushed released. I continue to be unable to create an Assistant that would use that extra knowledge based on CSV or JSON.
- Low rate limits: 40 request per minute, 200 requests per day for GPT-4. For GPT-4V, it’s even just half of that.
-
Their RAG implementation is reportedly rather basic (top-k, no re-ranking). Their only finesse is adaptive chunk lengths. Pricing seems rather steep: $0.20 / GB / assistant / day.
- free until 11/17/2023
- source: https://openai.com/pricing#language-models
-
GPT-4 Turbo is reportedly more verbose than GPT-4. Implications:
- the quoted cheaper price is not the full truth
- evals need to be redone, and will likely lead to different results (depending on what is measured how)
- GPT-4 Proper may also be “stronger” than Turbo. (But remember the “ChatGPT is getting dumber” affair…)