News for hackers. Stuff that matters.

Ask HN: Why don't Google/Amazon/Meta/Apple even try to compete with OpenAI?

I believe in competition and right now OpenAI has some kind of a monopoly in commercial large language models. The longer GPT-3 and chatGPT remain the only players in town, the harder it'll be for others to compete. Why is it that no other big tech company has released a commercial model comparable to GPT-3 in the past 2-3 years? I know some of them have in-house models, but why not go public?
( 18 comments | 2 points | on HN )
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by ThrowawayR2 on Saturday February 04, 2023 @10:41AM (#34655257)
"The answer to any question starting, 'Why don't they—' is almost always, 'Money.'" --Robert A. Heinlein

1) How do those companies profit off of creating and commercializing such a model? It's not even clear that Microsoft's partnership with OpenAI will turn a profit for them.

2) What are the odds of success of them creating a commercializable model that beats the competition enough to turn a profit? Probably not that high for anyone other than Google since a large body of heavily curated training data is needed.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by smoldesu on Saturday February 04, 2023 @10:33AM (#34655185)
> right now OpenAI has some kind of a monopoly in commercial large language models

No? That's absolutely false. You can, using commodity hardware, host your own GPT-based text transformer. There is no "monopoly" just because OpenAI provides an API and hosts the stuff themselves. They have a "monopoly" on the research used to create the later versions of GPT, but that's no more of a "monopoly" than Apple's monopoly on the iOS source code.

Comparable products can be made and released right now. I presume Google and Microsoft are working on integrating AI into their software, but what's the rush? Nobody wants to integrate faulty AI in their software, especially if it's as consistently wrong as ChatGPT is. Tuning this stuff will take time, and making people trust AI will take something much more impressive than ChatGPT. And considering how expensive these larger AI models are to train, it's completely understandable why nobody wants to pour $12 million dollars into building an Infinite Liability Machine.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by behnamoh on Saturday February 04, 2023 @10:45AM (#34655285)
Training costs of large models are prohibitive. Only the big tech can afford to make insanely large models.
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by smoldesu on Saturday February 04, 2023 @10:48AM (#34655310)
You don't need to. As I said, there are freely-available, pre-trained GPT models you can run on your own computer: https://huggingface.co/EleutherAI/gpt-neo-2.7B

No one has a monopoly on LLMs. You just can't run OpenAI's latest-and-greatest, because... why should you?

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by behnamoh on Saturday February 04, 2023 @10:55AM (#34655353)
GPT-3's performance is much much better than any open-source alternative that I've tried (Bloom, GPT-J, etc.) which is mostly due to parameter size. Like I said, OpenAI has the monopoly in "large" language models, but of course you can have smaller models.
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by muzani on Saturday February 04, 2023 @11:38AM (#34655691)
GLaM, Gopher, LaMDA. Google alone has plenty of huge LLMs.

ChatGPT is a sample of where AI currently is at, and many people were just taken by surprise because they didn't expect AI to be at the current level. OpenAI had to raise funds and interest, and got a home run.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by seydor on Saturday February 04, 2023 @10:32AM (#34655176)
While ChatGPT is a nice toy, it has not solved any real world problems yet. I guess the most lucrative apps will be in legal / medical.
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by saluki on Saturday February 04, 2023 @11:09AM (#34655458)
People said the web and SaaS were toys.

It's going to be everywhere.

There are already apps that will write accurate PR summaries to save time for developers. It also recommends code refactoring.

I saw a demo where you can train a model on a single pdf and then allow users to ask questions about the content.

ChatGPT can already generate programming tutorials that are accurate and beneficial. Granted not 100% of the time but it's a great starting point for how to learn something new or troubleshoot an issue.

Instead of searching a website you're going to be able to just ask questions and get high quality answers and tutorials on the fly. This is going to be game changing.

This is the beginning of lots of $1M MRR SaaS apps built by solo developers leveraging ChatGPT.

Many $X00M companies built by small teams.

And several AI based unicorns.

I would recommend learning and leveraging AI.

It's time for another round.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by endisneigh on Saturday February 04, 2023 @05:02PM (#34658855)
> People said the web and SaaS were toys.

I'd love to see a source of a single notable person in technology who said this.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by ninethirty on Saturday February 04, 2023 @06:48PM (#34659770)
Radio DJs say something about the local sports team, the weather and the next song about to be played.

You could automate that whole job into a couple of API calls.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by RobotToaster on Saturday February 04, 2023 @10:29AM (#34655155)
Most likely they already do have comparable, or better models, we know for a fact google does. It's just not commercially advantageous to give anyone else access to it.
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by ninethirty on Saturday February 04, 2023 @10:44AM (#34655274)
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by behnamoh on Saturday February 04, 2023 @10:47AM (#34655305)
thanks for sharing.
Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by ChildOfChaos on Saturday February 04, 2023 @11:52AM (#34655851)
As another commentator mentioned, Google are announcing something next week.

Google have shown off a lot of stuff in papers and announcements without making it available to the public, I suspect they are sitting on a lot of stuff waiting to see how the market develops before they make there play.

Particularly when it comes to ChatGPT as this would dramatically alter there business if they were to introduce it into Google search, it's about to become interesting.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by kingkongjaffa on Saturday February 04, 2023 @12:32PM (#34656243)
Presumably the business case for doing anything more compute expensive, just isn’t there right now.

Alexa, cortana, Siri, Google assistant, are all pretty awful and fall back to simply doing a Google search for anything even mildly complicated and just do “here’s what we found on the web for x” without any synthesis or context.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by BlueTie on Saturday February 04, 2023 @12:47PM (#34656381)
It's cheaper to let OpenAI figure all the stuff out first - then pour your war chest of money to n+1 their model instead of doing all the work of experimenting/failing/correcting course yourself.

Give it a year and they'll all have competing and better models.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by lostdog on Saturday February 04, 2023 @04:10PM (#34658328)
Pretty much all of these companies have a similar LLM for internal use, but they've so far decided that releasing it externally doesn't make sense. OpenAI's main difference is making the model available for public demo, which they're willing to do because they're but willing to suffer reputational damage from the model saying something stupid, and they're more optimistic that public access will help them figure out how to make money off the model.

To understand the bigger companies, you have to ask "why go public?" Bigger companies just don't see a general purpose LLM as a product. You could argue that that's a big failure of imagination, but they're not lagging behind on the tech itself.

Re:Ask HN: Why don't Google/Amazon/Meta/Apple even try to co... by moomoo11 on Saturday February 04, 2023 @10:55PM (#34661672)
Thankfully AI is a lot more interesting and practical (than the last decade which was filled with crypto bullshit), so in time we will find novel uses for AI.

Maybe Google would turn GCP into the easiest to use solution utilizing prompts to deploy your infra unlike AWS where you have to break your sanity with the levels of access controls and AWS specific crap you have to learn. Also… No more twiddling with yaml files. Please.

Maybe meta can use all of its Facebook data (people talking to each other) to train models that help filmmakers and games developers make engaging content.

Let’s see where it goes!