Ask HN: So Where's the Moat for AI? | Hacker News Reader

Ask HN: So Where's the Moat for AI?

10 pointsavbanks19 comments

Almost 2 years ago the infamous "moat memo" from Google leaked, it seemed like at the time this was heavily dismissed, however, as the months go on this seems correct. This current generation of GenAI seems like a race to the bottom. Apple seems a lot more justified in their reluctance to join this race. With the advancements that Deepseek has recently made is there any moat in AI?

https://news.ycombinator.com/item?id=35813322

View on HN

The “moats” that OpenAI may have are capital and regulatory capture. They can get laws passed to make it harder for other companies to ingest data for free to train their models and then pay for exclusive access.

The other moat is ChatGPT, the user facing product. It is the best end user product, they have brand recognition and a partnership with Apple. If they remain the default even when other LLMs are available for iOS, that still going to be a good customer acquisition funnel.

As far as the quality of the models, they all seem to be good enough for tte use cases I’ve come across from a development standpoint.

ChatGPT with how much debt though? It seems at best it will only become worse over time

What debt? As far as I am aware, all of their money comes from equity raises not debt. Compute will get cheaper over time and from what I read ChatGPT 4o is already much cheaper to run than 3.5 was when it was first introduced

Capital? With enough money you can buy a lot of NVidia GPUs and talents from top universities.

You’ve missed the news this last weekend. The Chinese made a model using non-NVIDIA GPUs and everyone is freaking out.

They used NVIDIA GPUs, just less of them

I didn't. I actually bought some nvda for 124, might not be a good call :/

I still think it's going to increase the sales of NVidia GPUs though, now smaller players can do it too.

It was done on NVIDIA chips.

And most likely they ran experiments and competing models on their other clusters. Using existing synthetic data.

Before running this one… and highlighting their ‘$5M’ cost.

Smoke and mirrors… all smoke and mirrors.

https://open.substack.com/pub/robotic/p/deepseek-v3-and-the-...

Could be marketplace/integrations. If someone comes out as a leader in price/product research and has exclusive access to APIs (think airlines, Amazon, movie theatres, etc) then the use case for that llm is unique and moat’d heavily.

IMO the current llms try to be everything, but that is an early internet (or non-American) idea. AOL, Yahoo, Elon Musk, etc all want to be the single place for everything. Yes it exists in India and China to some degree, but generally doesn’t seem to exist in the US. Why would llms be different? I could use the best ai agent for task A (coding at work) and the best for task B (content/social feed delivery at home). Is it likely openai can have Tik-tok level social algos and Claude level coding algos in the same product? Maybe, but more than likely not.

The Moat is the people.

Which people?

People with experience+talent in ML

That is only a moat if they manage to continuously improve the models. If improvements just slow down enough and the business gets disrupted by a mich cheaper good enough version it is game over for employing 1000s of 500k a year engineers

I believe a large number of the people working at the big AI companies are foreign nationals and Im going to assume at least some of them have mixed loyalties.

Tech businesses this decade are about building channels and then selling things down that channel. Moat is a bit of an odd analogy today; it's how defensible these channels are.

OpenAI's channel is ChatGPT. Google has well, Google. Microsoft has Windows and Bing. Apple has the iPhone, Mac, Siri.

Sometimes AI is the moat - if Bing has AI and Google does not, people might be using Bing more. Or that's the theory at least. What's scarier for Google is that the channel is Google itself and if people are searching via ChatGPT, then Google might go the way of Stack Overflow.

Smart people also factor in delays in metrics. People often say, "Look, everyone is still using Google! ChatGPT is hype!!!" But when did people stop using Yahoo? Many stopped using Stack Overflow 5 years ago shortly after they fired Monica, but the failure of Stack Overflow is due to the lack of quality answers since then, and yet has been blamed on ChatGPT.

People will stop using Google likely 5 years after the answer are lower quality than on ChatGPT. This happened years back. My default tech search is through Cursor today.

Once Google's moat dries up, everything else has trouble too - Gdrive, Gdocs, Gmail, Gmeet, Gcloud. If MS were to stop doing Windows, they wouldn't be as successful selling MS Office and MS Teams wouldn't have the dominance it has over Slack.

As for OpenAI, ChatGPT is the channel, and they're starting to sell things down that channel. Custom GPTs didn't work. Assistants and their built in RAG is actually pretty good. You can spend a month hiring someone skilled to make a fully custom chatbot. Or use theirs and have one up in a few days. It's more expensive but still less than $100, cheaper than an engineer or the CS people you'd use otherwise.

The magic isn't better AI products, it's products that can be used by a more mainstream market. Excel was one of the best consumer markets out there because it gave everyone low-code capabilities for work. Gmail made it unnecessary to know what POP3 and SMTP are.

tldr: Even if people can get higher quality for cheaper, most aren't going to install Deepseek and Deepseek doesn't have a channel to sell on.

OpenAi assistants are actually amazing, it would take a couple of hours to build a custom chatbot not a month. Just upload the docs and use one of the many services available to embed it to a website

> Sometimes AI is the moat - if Bing has AI and Google does not, people might be using Bing more

The issue is that LLMs are a commodity that anyone can integrate into their product.

To your broader point at least, you’re right. RIMs stock price peaked in 2010. Three years after the writing was on the wall with the iPhone and Android.

Smart people knew that Intel was in trouble over 10 years ago as mobile was taking off. They were propped up for awhile by server sales and less so by PC sales. Now they are both moving to chips manufactured by TSMC.

There is no moat.

https://semianalysis.com/2023/05/04/google-we-have-no-moat-a...

Reply on news.ycombinator.com