Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Gemini AI

deepmind.google

Post reply on HN
Print view View original

41–50 of 1001 posts

Re: Gemini AI

#41

Post by kolinko »

obiefernandez wrote: post #7

> For Gemini Ultra, we’re currently completing extensive trust and safety checks, including red-teaming by trusted external parties, and further refining the model using fine-tuning and reinforcement learning from human feedback (RLHF) before making it broadly available. > As part of this process, we’ll make Gemini Ultra available to select customers, developers, partners and safety and responsibility experts for ear…

It won't be available to regular devs until Q2 next year probably (January for selected partners). So they are roughly a year behind OpenAI - and that is assuming their model is not overtrained to just pass the tests slightly better than GPT4

Re: Gemini AI

#42

Post by albertzeyer »

So, better than GPT4 according to the benchmarks? Looks very interesting.

Technical paper: https://goo.gle/GeminiPaper

Some details:

- 32k context length

- efficient attention mechanisms (for e.g. multi-query attention (Shazeer, 2019))

- audio input via Universal Speech Model (USM) (Zhang et al., 2023) features

- no audio output? (Figure 2)

- visual encoding of Gemini models is inspired by our own foundational work on Flamingo (Alayrac et al., 2022), CoCa (Yu et al., 2022a), and PaLI (Chen et al., 2022)

- output images using discrete image tokens (Ramesh et al., 2021; Yu et al., 2022b)

- supervised fine tuning (SFT) and reinforcement learning through human feedback (RLHF)

I think these are already more details than what we got from OpenAI about GPT4, but on the other side, still only very little details.

Re: Gemini AI

#43

Post by bigtuna711 »

philomath_mn wrote: post #31

This is very cool and I am excited to try it out! But, according to the metrics, it barely edges out GPT-4 -- this mostly makes me _more_ impressed with GPT-4 which: - came out 9 months ago AND - had no direct competition to beat (you know Google wasn't going to release Gemini until it beat GPT-4) Looking forward to trying this out and then seeing OpenAI's answer

Ya, I was expected a larger improvement in math related tasks with Gemini.

Re: Gemini AI

#44

Post by rolisz »

What is up with that eval @32? Am I reading it correctly that they are generating 32 responses and taking majority? Who will use the API like that? That feels like such a fake way to improve metrics

Re: Gemini AI

#45

Post by logicchains »

NewsaHackO wrote: post #26
Oras wrote: post #17

Feels more like an Apple post "the best fastest blabla-est". How about making it available to try without the fluff?

The articles seems to report some data points which at least make it seem comparable to GPT4. To me, I feel as though this makes it more objective vs fluff.

There are some 7B weight models that look competitive with GPT4 on benchmarks, because they were trained on the benchmark data. Presumably Google would know better than to train on the benchmark data, but you never know. The benchmarks also fail to capture things such as Bard refusing to tell you how to kill a process on Linux because it's unethical.

Re: Gemini AI

#48

Post by code51 »

wiz21c wrote: post #27

The improvement over ChatGPT are counted in (very) few percents. Does it mean they have entered a diminishing returns phase or is it that each percent is much harder to get compared to the previous ones ?

> We’re already starting to experiment with Gemini in Search, where it's making our Search Generative Experience (SGE) faster for users, with a 40% reduction in latency in English in the U.S., alongside improvements in quality.

This feels like Google achieved a more efficient inference. Probably a leaner model wrt GPT.

Post reply on HN