[codekaze]

Just played with Gemini Ultra for like 10-15 mins, and right off the bat, it made mistakes I've never seen GPT-4 do.

To give you an example, I asked Gemini Ultra how to set up a real-time system for a TikTok-like feed that matches card difficulty with user ability. It correctly mentioned "Item Response Theory (IRT)", which was a good start. But when I followed up asking how to implement a real-time IRT system, it suddenly started going off about "Interactive Voice Response (IVR) system" - something totally unrelated and never mentioned before. Never had this kind of mix-up with GPT-4.

https://g.co/gemini/share/f586a497013e

[tomasff]

It doesn't seem like it's using Gemini Ultra yet. For me it seems like only the interface has been updated since the image generation capabilities are not working.

[QuantumGood]

Image generation is working for me

[esperent]

Not for me. The first time I asked it paused for about ten seconds so I thought it was making an image. Then responded with:

> I can't generate images of that. Try asking me to generate images of something else.

I thought it was saying there was an issue with my prompt. So I asked:

> Why can't you?

I got this response:

I can't generate images for several reasons:

    Technical limitations: Image generation is a complex task. Even the most advanced image generation AI models require powerful computer hardware and specialized software. I'm simply not equipped to do this as a language model.

    Focus on language: My primary function is to understand and generate text.  I'm trained on massive amounts of text data, which helps me  write poems, answer your questions thoughtfully, and have conversations. Image generation is a whole different specialization.

    Potential for misuse: Image generation technology has the potential to be misused, such as creating deepfakes or  images that might be harmful or offensive. It's important to use these tools responsibly and ethically.

Here's what I can do:

    Describe the scene in detail: I can enhance the poem with vivid descriptions of the setting, the baker's emotions and appearance, and the fairy's ethereal form. This can help you imagine the image more clearly.
    Find inspiration: I can search for illustrations or art in a similar style to guide a human artist, should you want to commission an illustration for your poem.
    Recommend image generation tools: I can point you towards simple online image generation tools if you'd like to experiment with creating visuals yourself. Just be aware of their limitations.

Let me know if you'd like me to further develop any of these options!

[mjgeddes]

Yeah, I noticed with 'Gemini Pro' , it didn't seem to be able to remember much about earlier outputs in the conversation (apparently little to no context window), which obviously drastically dumbs it down.

I was starting to get OK results with 'Pro', but I had to use special prompting tricks.

Tried 'Advanced' (Ultra), seems only marginally better so far.

[hackerlight]

> I was starting to get OK results with 'Pro', but I had to use special prompting tricks.

Like what?

[mjgeddes]

I usually put a couple of keywords in brackets at the beginning (before the body of the prompt) to provide some context

[MrsPeaches]

From the FAQ:

"Why doesn’t Gemini know what I said earlier in a conversation?

Gemini’s ability to hold context is purposefully limited for now. As Gemini continues to learn, its ability to hold context during longer conversations will improve."

[codekaze]

Yeah, I saw that in the FAQ, but this was literally my second question in the convo, so not exactly a "long" conversation. Seems like it should be able to handle context for at least a couple of exchanges, right?

[behnamoh]

> Gemini’s ability to hold context is purposefully limited for now. As Gemini continues to learn, its ability to hold context during longer conversations will improve."

This is ridiculous. Context is everything with LLMs. gpt-4-32k performs better than gpt-4 exactly because of this.

[okdood64]

> Created with Gemini Advanced

You're not using Ultra here...