Hacker News Reader

Comment by temikus | Hacker News Reader

It’s not about ChatGPT replacing Google IMO but the fact that decades of very expensive AI research at Google produced relatively little apart from impressive looking papers. If I was a major shareholder I would be asking some uncomfortable questions to Sundar right about now.

EDIT: Yes, there is TensorFlow but that is not really a product of the core AI teams, rather than a need for better tooling from them.

View on HN

...Google Translate? Photos search? Gmail autocomplete? YouTube recs? Heck, Waymo?

I can't think of any other company that has produced more ML that I actually use daily. (Well, I don't use Waymo... yet.)

So I really don't know what you're talking about. And presumably they're not jumping into public versions of chat and image synthesis like ChatGPT and Stable Diffusion because they're busy figuring out how to apply them in ways that will actually make boatloads of money.

Oh, and pretty sure Search uses tons of ML as well... they just don't go around talking about it because they don't really need to.

I think we're probably really stretching the definition of ML with things like YouTube recommendations. I don't know about you but I find them to be transparent in how they work and actively useless (either irrelevant to my interests or the video I just watched, or a swathe of other stuff I've already seen) most of the time.

Algorithmic and ML are two different things.

Google Translate / Waymo is fair, I totally forgot about those ones.

Gmail Autocomplete, Photos Classification, YouTube recommendations are (relatively) simple models and don’t differentiate from the market much.

The thing that bothers me is that Google has been promising market-beating ML-as-a-service for ages which never materialised despite billions being poured into it.

Google Prediction API has been a great start in the area. Then Google stopped improving it and killed it with very little notice and without providing an alternative.

What do you mean? They also have sth like ChatGPT, called Lambda, and that was quite some time before ChatGPT was released. They basically have all the fancy large language model things as well. They are always at the state of the art and usually way beyond internally.

The bigger problem is, how to make this into a product. A product where you can show ads. A product which will not create all kinds of hate.

Throwaway for obvious reasons, but can confirm. I don't even work on the AI / search team and I remember trying Meena (-> Lamda) at least 2 years ago and being completely blown away by its ability to hold a conversation. There were even jokes on Memegen that we could let Meena write our perf packets and it'd be indistinguishable from real ones. I guess that's the Google-internal Turing test.

I believe you are giving too little credit for the state of the art to Google. Transformer is a Google research invention. The main criticism could be not turning these amazing tech into products.

But then nobody really did it either. ChatGPT, Dall-E are still solutions in search of problems.

They definitely have incorporated their impressive papers into their search engine. They've integrated BERT, a state of the art transformer model, into search engine, and it is used for almost every query now [1]. Google also has best-in-class computational photography (for their phones) and the best voice assistant, which is all ML.

1. https://blog.google/products/search/search-language-understa...

I've been wondering about the voice assistant, which I find absolutely appalling. It's so bad (and it's gotten worse) that my children noticed it and make fun of it! I don't understand why anyone would want to use it. In contrast siri works quite well, so I assume it's not because of the way I speak.

They have a bigger problem. If I find answers, I see less ads.

As the web has disappeared, Google’s focus changed from information to pure paid placement. You sorta get the information that you need, but not really.

ChatGPT will turn into an even bigger shitshow, as at its heart it’s a bullshit generator. But it will kill Google.

To me, this is the key thing. Brain has been getting one-upped by OpenAI despite being an absolutely enormous org including some of the most credentialed ML researchers in the world. It is sort of embarrassing that tools like Stable Diffusion and ChatGPT are coming from other places.

It's totally predictable that big behemoths like Google can't innovate with the same velocity and focus as more nimble orgs like OpenAI. Even in its current panic mode, Google won't be able to compete against OpenAI and other competitors.

Deep Mind? Protein folding?

I completely agree. A simple question, where are the models? Where is their superiority? After some point, put up or shut up

Reply on news.ycombinator.com