Google AI Studio – Frontier model with a huge context window
A few notes to start with. This is a “back-end” type interface. That means you need a little more knowledge to use it, but you also get a lot more control. Don’t let that scare you. Other than changing the model, everything else I leave the same. It CAN do a lot more than I will be talking about, and I don’t know anything else beyond the basic prompting it can do.
Why use Google AI Studio?
- Token Limit: 2 MILLION tokens. Input and output.
(To put it in perspective, ChatGPT has a 128k context length, but a tiny 4k output window. Claude 3, Opus has 200k token. You can see this is a much larger window.)
That token limit is very near the total Harry potter set of books being able to be processed at once. - Gemini (the LLM behind AI Studio) just reached the #1 spot on the Chatbot arena around 8/2/2024. (Click leaderboard at https://chat.lmsys.org/ for the latest). It’s never been better than ChatGPT or Claude to my knowledge.
- You can share it to others.
How do I use it?
I want to start with the prompt gallery link: https://ai.google.dev/gemini-api/prompts. It is on the left side in the AI Studio interface, but I wanted to bring attention to it. It will help you get started if you want some instructions.
I keep it simple, and I only change the Model (mostly)
- I just use “Create new prompt“.
- I use the latest Gemini pro model (as of this writing, Gemini 1.5 Pro Experimental – gemini-1.5-pro-exp-0801)
- Temperature 1 (default). This makes the model more deterministic (lower) or more creative (higher)
- Add Stop Sequence – blank
- Safety Settings – I generally leave this default, but I have moved all the sliders down before. Sometimes, the model guardrails are a little too strong for what I am looking for. They are still there, but lessened somewhat.
- Advanced Settings: I don’t open this. Though Top P, if the model allows it, shows just the most probable words to “come next”. Not all models allow its use.
- It saves your AI Query to Google AI Studios in Google Drive.
- You can move the AI Query from the default location to another folder easy enough.
- If you want to share the query with source material, I recommend sharing the drive folder, and move your query and reference materials into the same folder.
This portion of sharing is a bit of a pain. More difficult/involved than say a custom GPT. It may take a few tries.
What formats does Google AI Studio accept?
You can add a number of items to the model, but they pretty much all go to Google Drive.
While it accepted .xlsx, .docx, and .pptx, it did convert them to Sheets, docs, slides.
While they have a list of media they accept you need to experiment to check for other formats or confirm it does not accept that format.. It seems to accept all major versions of image, audio, video, and text.
I could not find a link to other documents it accepts, but I created a folder and tried a number of formats to see if it would accept them. It accepted these: I can get Google Docs, Sheets, and slides, as well as Microsoft Word Docs, Excel Spreadsheets, and PowerPoint Slides, plus PDF files to be accepted.
It DID reject the Microsoft XPS format and GIF formats.
One OTHER note: In my experience, Microsoft products will let you export to xps/pdf format. This is a bad idea. This is essentially an image format, and you lose the individual text into one image. Now, these models DO have OCR (Optical Character Recognition) and should be able to pull the text from the image, ok… but I have never had 100% success. You are better off exporting (or saving as) to a PDF format.
A quick/easy test: Open the document, and highlight the text. If the individual letters/words are highlighted, you are good. If the entire page highlights (like an image) try another way to export.