Summarize research papers with Gemma

See Google's Gemma 3 in action as it summarizes and analyzes popular research papers. This video highlights Gemma's ability to handle long contexts, extract key information, and provide translations, streamlining the research process and making scientific insights more accessible.

Speaker: Tatiana Matejovicova
Products Mentioned: Gemma

[Mede_N]

What is the rationale behind setting the temperature to 0 here?

EDIT: OK, I did some digging. The most likely reason why they set the temperature to "0" is simply for reproducibility for that video (this means they can test it once in advance and only have to record once). While 0-temperature makes the model deterministic and decreases hallucinations, it also usually decreases its performance. The recommendations for low hallucinations and good performance are temperature values around 0.1 to 0.5. (Cao et al., 2023; Zhang et al., 2024). But according to them, the connections between temperature, performance, and hallucinations are not only model-dependent but also task-dependent. Interestingly, for some models, a very low temperature (0 or 0.1) even increased the hallucination rate compared to moderately low values (such as 0.3 or 0.5).

[BibhatsuKuiri]

can it handle math equations and images properly ? One important thing that is lacking in this PDF analyzer are- it misses important inferences from images and equations

[stanislav]

I wish there was more examples with the 1B model, especially locally using an MCP or something

[kimguerrette]

Which app are you using in this demo? Is Research assistant something you made in Vertex AI or AI Studio?

[lgmuk]

I was working on something very similar! Very com demo!

[Hukkinen]

So, I assume the summaries were correct? But how the user can verify the information quickly?

[Ali-kdx]

why would I use gemma instead of notebooklm?

[maraisdekker]

What is the difference between Gemma and Gemini? That’s unclear

[TorstenWerner]

Interesting. But I'd use notebooklm for such a task.

[alexisdamnit]

That time for inference though. I wonder how this compares to other open-sourced models. For anyone wondering -- this is still faster than OpenAI's fastest model gpt-4o-mini. OpenAI's inference time is awful

[coding-master-shayan]

Please decrease the rate limits for gemma 3 on aistudio, specifically requests per day (RPD).

[philips]

while on AI studio what's stopping us from using large models ... gemma's use case tutorials better be focused on running them on PCs with smaller computes and GPUs ...

[diptipriya]

I just remembered my English classes precis writing.
I could finally compare with mine 😂😂😂

[bpavuk]

Fix long chats performance in AI Studio first!