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

221–230 of 1001 posts

Re: Gemini AI

#221

Post by willsmith72 »

tikkun wrote: post #60

One observation: Sundar's comments in the main video seem like he's trying to communicate "we've been doing this ai stuff since you (other AI companies) were little babies" - to me this comes off kind of badly, like it's trying too hard to emphasize how long they've been doing AI (which is a weird look when the currently publicly available SOTA model is made by OpenAI, not Google). A better look would simply be to sh…

they have to try something, otherwise it looks like they've been completely destroyed by a company of 1000 people

Re: Gemini AI

#222

Post by renewiltord »

Interesting. The numbers are all on Ultra but the usable model is Pro. That explains why at one of their meetups they said it is between 3.5 and 4.

Re: Gemini AI

#224

Post by n2d4 »

kolinko wrote: post #41
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

> and that is assuming their model is not overtrained to just pass the tests slightly better than GPT4

You are assuming GPT4 didn't do the exact same!

Seriously, it's been like this for a while, with LLMs any benchmark other than human feedback is useless. I guess we'll see how Gemini performs when it's released next year and we get independent groups comparing them.

Re: Gemini AI

#225

Post by arnaudsm »

albertzeyer wrote: post #42

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 o…

I miss when ML scientific papers had actual science in them. Now they all feel like ads.

Re: Gemini AI

#226

Post by visarga »

SeanAnderson wrote: post #218

Not impressed with the Bard update so far. I just gave it a screenshot of yesterday's meals pulled from MyFitnessPal, told it to respond ONLY in JSON, and to calculate the macro nutrient profile of the screenshot. It flat out refused. It said, "I can't. I'm only an LLM" but the upload worked fine. I was expecting it to fail maybe on the JSON formatting, or maybe be slightly off on some of the macros, but outright ref…

That's what they taught it "You're only a LLM, you can't do cool stuff"

Re: Gemini AI

#227

Post by modeless »

onlyrealcuzzo wrote: post #187
modeless wrote: post #113

"We finally beat GPT-4! But you can't have it yet." OK, I'll keep using GPT-4 then. Now OpenAI has a target performance and timeframe to beat for GPT-5. It's a race!

Didn't OpenAI already say GPT-5 is unlikely to be a ton better in terms of quality? https://news.ycombinator.com/item?id=35570690

I don't recall them saying that, but, I mean, is Gemini Ultra a "ton" better than GPT-4? It seemingly doesn't represent a radical change. I don't see any claim that it's using revolutionary new methods.

At best Gemini seems to be a significant incremental improvement. Which is welcome, and I'm glad for the competition, but to significantly increase the applicability of of these models to real problems I expect that we'll need new breakthrough techniques that allow better control over behavior, practically eliminate hallucinations, enable both short-term and long-term memory separate from the context window, allow adaptive "thinking" time per output token for hard problems, etc.

Current methods like CoT based around manipulating prompts are cool but I don't think that the long term future of these models is to do all of their internal thinking, memory, etc in the form of text.

Re: Gemini AI

#228

Post by jasonjmcghee »

SeanAnderson wrote: post #218

Not impressed with the Bard update so far. I just gave it a screenshot of yesterday's meals pulled from MyFitnessPal, told it to respond ONLY in JSON, and to calculate the macro nutrient profile of the screenshot. It flat out refused. It said, "I can't. I'm only an LLM" but the upload worked fine. I was expecting it to fail maybe on the JSON formatting, or maybe be slightly off on some of the macros, but outright ref…

Sounded like the update is coming out next week- did you get early access?

Re: Gemini AI

#229

Post by sockaddr »

SeanAnderson wrote: post #218

Not impressed with the Bard update so far. I just gave it a screenshot of yesterday's meals pulled from MyFitnessPal, told it to respond ONLY in JSON, and to calculate the macro nutrient profile of the screenshot. It flat out refused. It said, "I can't. I'm only an LLM" but the upload worked fine. I was expecting it to fail maybe on the JSON formatting, or maybe be slightly off on some of the macros, but outright ref…

> I just gave it a screenshot of yesterday's meals pulled from MyFitnessPal, told it to respond ONLY in JSON, and to calculate the macro nutrient profile of the screenshot

> Not impressed

This made me chuckle

Just a bit ago this would have been science fiction

Post reply on HN