Live data from Hacker News

Gemini API File Search is now multimodal

blog.google

21–30 of 58 posts

Re: Gemini API File Search is now multimodal

#23
post #16
post #8

Earlier quoted context omitted.

It's incredible how far behind Gemini has gotten, both the product and the model. Even the ChatGPT plugin for Google Sheets blows away the native Gemini integration. Everyone thought Google was pulling ahead with Gemini 3. For a minute there they had the best language model, image model, AND video model in the world. But it's like they decided to pull over for a nap while OpenAI and Anthropic flew by.

3.1-pro is still very capable, and API is at competitive price vs e.g. Anthropic, they just can't seem to figure out RLHF and harness. It needs a lot of guiding, it tends to be lazy and poorly sticking to instructions by default. It just feels like many google products really, they are capable of really amazing things, it's just that nobody there seem to care. I would guess they are likely optimizing more for interna…

They optimize for making their SRE's lives easier, over quantizing models regardless of how negative an effect that has on the user.

Re: Gemini API File Search is now multimodal

#25

It’s a striking irony that the world's leader in search is receiving so much heat for poor search functionality and UX within its own flagship AI products

One of Googles core problems is internal silos of talent. The search team has likely never interacted with the Gemini app team or perhaps even the Gemini app.

For all intents and purposes Google Gemini is a totally separate company from Google search.

Re: Gemini API File Search is now multimodal

#27
post #8

This might be great and all but I am still miffed at how simple search on AI Studio is. You can only search the titles of your conversations and nothing inside them. On top of that they messed with the scrolling so Ctrl+F doesn't work reliably.

It's incredible how far behind Gemini has gotten, both the product and the model. Even the ChatGPT plugin for Google Sheets blows away the native Gemini integration. Everyone thought Google was pulling ahead with Gemini 3. For a minute there they had the best language model, image model, AND video model in the world. But it's like they decided to pull over for a nap while OpenAI and Anthropic flew by.

It’s still the best option for uptime, document analysis (on a cost basis), and Google is less likely to experience a significant cybersecurity breach than a less established company. They’ll be fine as long as they stay in the game even if they never have a Ferrari again plenty of people buy Toyota.

Re: Gemini API File Search is now multimodal

#28

It’s a striking irony that the world's leader in search is receiving so much heat for poor search functionality and UX within its own flagship AI products

One of Googles core problems is internal silos of talent. The search team has likely never interacted with the Gemini app team or perhaps even the Gemini app. For all intents and purposes Google Gemini is a totally separate company from Google search.

Google has ~190k employees. You can't have total collaboration across that many people.

Teams will cross collaborate, but they have to be for specific projects with specific people.

Re: Gemini API File Search is now multimodal

#29
post #8

Earlier quoted context omitted.

It's incredible how far behind Gemini has gotten, both the product and the model. Even the ChatGPT plugin for Google Sheets blows away the native Gemini integration. Everyone thought Google was pulling ahead with Gemini 3. For a minute there they had the best language model, image model, AND video model in the world. But it's like they decided to pull over for a nap while OpenAI and Anthropic flew by.

I have the opposite experience where Gemini (even the flash models) has the only useful model for my reverse engineering related use case. My hunch is Google utilizes its free access to entire Google search indices to train itself from niche non-English speaking community websites, much frequently and in a "relevant" manner, which in the end gives these models the most up to date info for this particular kind of work…

> for my reverse engineering related use case [...] Every other model is just either 10 years outdated with their answers

I've mostly been doing reverse engineering with Codex, mostly related to games, but not once has the "training data cut-off date" been in the way, the most useful part comes from handing it a binary/directory and letting it prod it until it finds the answer you're looking for, I don't even have web search enabled and sometimes it might take 30-40 minutes for it to find the answer, but I never saw it be unable to find the answer because it's training data was a couple of years old.

Post reply on HN