In a sense, I think everyone knew it was coming. Historically speaking, in US at least, there initially may be a notable number of upstarts, but those numbers quickly dwindle to form oligopolies. That is how we end up with 3 million cereal types, but 2-3 companies controlling them. The same repeats throughout most industries with an annoying consistency.

I actually use both gpt and local qwen and I personally think the question tends to be of what one needs ( speed? privacy? lack of corporate silliness? ).

Still, it is sad that we can predict the future so well based merely on what happened before.

Edit: I certainly hope I don't have to explain to you why the above is relevant. No market participant would dare to argue this if the competition was healthy.

i think Sam Altman is wasting oxygen

Got piles of dram and consumer GPUs racked locally generating 100% of my tokens with complete privacy and sovereignty on my own property.

Sam Altman can eat my entire ass.

The look on his face looks a lot like Elizabeth Holmes in You Can See Everything

Two statements can be true and he stated one of them. Sam is not wrong.

1. Local inference can cause a large under-utilization of resources from users that do not run local inference hardware at capacity over the hardwares lifespan. This is under-utilization and under-utilization is not optimal from an economics perspective.

2. The concentration of inference hardware to singular providers can aggregate power to single systems or organizations, and their incentives may be at odds with individual users.

The same power dynamic and economics struggle has existed nearly everywhere for all of history. For example, if you want to decide to own or rent a piece of heavy construction machinery for business, you will find a long list of do-not's on the rental contractual terms, but you are more than welcome to buy your own and operate your own machinery at several times the price for the hours used and use it however you want.

It would be interesting if there were some kind of holographic LLM. You could send part of the encoded prompt and receive generated tokens that only you can decode. Then we could use the distributed capacity of the world and everyone could be paid for marginal token generation capacity. Even open models are profitable to run on almost all electricity.

Altman: You are welcome to invest a little bit of your fortune in making major software developers write more efficient code that stops wasting our RAM. Start with internet browsers and the Windows 11 app that wastes one gigabyte showing the weather.