It is probably cheaper to simply integrate with OpenAI or Anthropic or whoever might unseat them in the future, than spend $50B on training a model. Not only is it cheaper, but it also gives them the flexibility to ride the wave of popularity, without ceding hardware or software sales.
The way I see it, models were always predicated on openness and data-sharing. That, too, will be the competitive downfall of those who poured billions into creating said models.
They won't stay caged up forever. Ultimately the only thing OpenAI has between itself and it's competitors is some really strong computers.
Well... anybody can buy strong computers. This is, of course, assuming you don't believe the promise of ever-increasing cognition and eventual AGI, which I don't. The people going fast aren't going to be the winners. The second movers and later on will be. They get all the model, with 1/100th the cost.
Ultimately models, right now, for most consumers, are nothing more than novelties. Just look at Google Pixel - I have one, by the way. I can generate a custom image? Neat... I guess. It's a cool novelty. I can edit people out of my pictures? Well... actually Apple had that a couple years ago... and it's not as useful as you would think.
It's hard to see it because we're programmers, but right now, the AI products are really lacking for consumers. They're not good for music, or TV, or Movies, or short-form entertainment. ChatGPT is neat for specific usecases like cheating on an essay or vibe coding, but how many people are doing that? Very, very few.
Let me put it this way. Do I think Claude Code is good? Yes. Do I think Claude Code is the next Madonna? No.
There's a lot of difference between OpenAI and, let's say, Facebook (Llama). The difference between them is not only strong computers. There's architectural differences to the models.
It will be a sign of a maturing market when we see vendors actually say "bad for X" and pulling away from general-purpose messaging.
You see it a little bit with the angling for code-specific products, but I think we're nowhere near a differentiated market.
This is the real death knell people should focus on. Apple buried their AI R&D to rush complete flops like Vision Pro out the door. Now that the dust has settled, the opportunity cost of these hardware ventures was clearly a mistake. Apple had more than a decade to sharpen their knives and prepare for war with Nvidia, and now they're missing out on Nvidia's share of the datacenter market. Adding insult to injury, they're probably also ~10 years behind SOTA in the industry unless they hire well-paid veterans at great expense.
Apple's chronic disdain for unprofitable products, combined with boneheaded ambition, will be the death of them. They cannot obviate real innovation and competition while dropping nothingburger software and hardware products clearly intended to bilk an unconscious userbase.
Think of how important it is for any AI model company to be the go-to model on the iPhone. Google pays Apple billions to be the default search engine on the iPhone.
Am I supposed to keep waiting until that changes one day?
Unlike OpenAI, Apple doesn't need to charge a subscription to get AI-based revenue. It just needs to properly integrate it into its products that billions of people are already using, to make those products more useful so people continue buying them. At that point most users don't care what model is powering it - could be GPT, Claude, Mistral etc.
And also to hop off without any penalty if/when the wave collapses.
Edit: Yes it exists, seems to be built off qwen2.5 coder. Not sure it proves the point I thought it was, but diffusion LLMs still seem neat
Why? Because everyone else is doing it (and not making a profit btw)?
Something about incentive for people to buy a phone that looks and acts identical to a 5 year old phone otherwise
Source?