Google nears release of Gemini AI | Hacker News Reader

Google nears release of Gemini AI

reuters.com

20 pointsTempusFive20 comments

In the wake of the release of GPT4 there was so much pontificating about how Google was caught lacking and how its era of dominance was over. That definitely hasn't happened yet, and their broad and continuous release of AI-based products and features has been impressively nimble. Given their pioneering AI work, colossal user base, unfathomable data set, and world-scale computing resources, I would not bet against Google in anything related to AI.

I would bet against them. Google is too risk-adverse. They're likely going to put out an AI that is so neutered as to be useless. Even if it was good, it's not enough for a _Google_ product to be merely on-par or slightly better than the competition. They've killed too many services unceremoniously. Only a fool would trust Google to not pull the football once again. Their reputation is and has been in tatters for years.

Google has killed/retired/sunset numerous products for sure, but AI and supporting infrastructure is their core competency. I don't think they're going to cede any of that space without a fight.

Even if Google releases a decent AI, they won’t be able to monetize it as effectively as search ads. Further, training and serving an LLM is now much, much cheaper than running a real search engine. Billions of dollars worth of datacenters are not required. Hell, I can do it on my MacBook. They have no moat for a large percentage of search queries.

Bard is such garbage. Between Bard, and Google being Google, Gemini has an uphill battle that I just don't see as winnable.

Bard is not even comparable to GPT-3-turbo, let alone GPT-4.

It's so bad that I lose respect for people when they tell me they use it. In my personal testing of it, the contradictions between its multiple drafts are jarring. For example, in a comparison of three NPM modules, it claimed each of the packages was "fastest" in the three different "drafts" it produced.

And what does Google even plan to do with any sort of comparable model? Mine your chats for advertising data? Make chat responses gameable by "Chat Engine Optimization Experts"?

At least Meta makes their AI "open-source." But a closed source model by Google has exactly one possible outcome: killedbygoogle.com

Edit: removed unnecessary judgement about meta.

>For example, in a comparison of three NPM modules, it claimed each of the packages was "fastest" in the three different "drafts" it produced.

That's an odd benchmark, no? LLMs aren't omniscient; you shouldn't expect it to be an expert on something so specific, especially if there is not widely available material on the internet that clearly indicates that one choice is correct. They can only repeat what they've been trained on and reason about that information.

it's kinda laughable that bard sucks so bad when upstart anthropic/Claude actually does good it's own against gpt4, especially for creative writing, I haven't tried it much for code as I prefer using phind or code whisperer for code, but for creative ideas etc I go to claude 2.

This is Deepmind's first real attempt at a GPT-level LLM, isn't it? I'm very excited 1) to learn exactly how they've applied the MuZero framework to this problem and 2) to see how it performs. It could be underwhelming, or it could be a massive leap forward.

You see, this is how you make a good article title about Google product. It contains the word "release" which makes it immediately obvious it is a new product, not cancellation of existing product. Many times people post articles about Google products without single verb in title and everybody then has to read TFA to figure out if it was released or cancelled. As a rule, articles about Google products should always contain verb.

Google nears of release of end of Gemini AI

They ought to use it to make the search better. When the search is bad, the whole internet is bad. Even sites with useful information now have it buried amongst paragraphs of glorified keyword-stuffing... that they probably generated with the help of these LLMs google is busy releasing.

So Bard was just a placeholder until they could get Gemini released?

Bard is a product, not a model, the same way that ChatGPT is a product (ish) where GPT3.5/4 is the model.

Ah, ok, so Bard will still be around but with a different (hopefully much better) model underlying. So far Bard has been very lackluster.

Agreed, many issues with Bard, though I've had a better experience working with the underlying PaLM 2 directly.

Even given all of its faults, have you found a better LLM than Bard that can incorporate current events? Because for my purposes, Bing chat has been far more faulty than Bard.

I usually ask both Bard and Phind (which also does web searches to help generate answers, usually using GPT-4) the same questions, and I can't remember a single time Phind didn't blow Bard out of the water. Bard has done better when asking it creative stuff like to make up stories.

Correct. Bard is a public interface to their Pathways Language Model 2 (PaLM 2). It's anyone's guess exactly how they'll roll out Gemini. Sounds like might be initially through GCP through approved partners.

Has anyone tried the smaller version of this so far?

If it was that good they would just release it. openAI didnt make so much noise before GPT4.......

Reply on news.ycombinator.com