3 ms·

We'll find out but certainly in the software industry many companies pay for those services I mentioned even when there are cheaper alternatives. It's one thing

by ericmay 5mo ago

We'll find out but certainly in the software industry many companies pay for those services I mentioned even when there are cheaper alternatives. It's one thing to download some "small" model and hack around, but that does not translate to efficiency at scale. The "small" models and such haven't proven to be profitable and so you can wind up with some, certainly powerful, small models but they'll be far enough behind that the bleeding edge models will continue to compound games. It's no different than any other positive feedback loop - you see it in chip design and software in general. If all China is doing is releasing models that are 6 months or more behind for free, forget it. Game over. That's 6 months of run time for a better model to exploit vulnerabilities or whatever. I think there's plenty of room for lots of approaches. But similar to how Open Office is nice and all, Microsoft Office makes however many hundreds of millions. I'm not really sure why folks believe that what we're seeing today will be any different. Bleeding edge compounds gains and advantage. American tech companies are run by some of the smartest people on the planet and from all over the planet. They're not spending hundreds of billions of dollars because they're wrong and because some 1-year behind model is a big threat.

verdverm 5mo ago
The big models have not shown to be profitable, likely the small models will first since it's on the order of 100x cheaper There is a likely world (imo) where Big AI produces mega models primarily to distill them down into economically viable sized models China also has mega models, US also has open models. Both groups are doing the same things. AllenAi is the most open but their models have not been competitive
- ericmay 5mo ago
  The pathway for the big models to be profitable is much clearer though. Some of the lack of profitability is simply distilled into CapEx. Generally I agree though - I think there will be lots of models, lots of business models, &c - I just think the arguing point about China is wrong which is what I'm interested in pushing back on.
  - verdverm 5mo ago
    I'm not sure the path to profitablity is so clear. They are just starting to come out of subsidized pricing. A lot of people are going to learn how expensive they are for the amount they are using them. I personally prefer the smaller models because they are much faster for almost the same quality. I prefer open weights even more for the freedom to use as much as my hardware will let me (~250M tok/day right now) and the expectations they will continue to improve over the coming years. We are going to be running experiments at work to determine what / how much it would take for self hosting a deep seek / kimi to have a nice ROI. A few of us now have the spark to get a taste and catch the bug

by ericmay 5mo ago

verdverm 5mo ago

The big models have not shown to be profitable, likely the small models will first since it's on the order of 100x cheaper There is a likely world (imo) where Big AI produces mega models primarily to distill them down into economically viable sized models China also has mega models, US also has open models. Both groups are doing the same things. AllenAi is the most open but their models have not been competitive

ericmay 5mo ago
The pathway for the big models to be profitable is much clearer though. Some of the lack of profitability is simply distilled into CapEx. Generally I agree though - I think there will be lots of models, lots of business models, &c - I just think the arguing point about China is wrong which is what I'm interested in pushing back on.
- verdverm 5mo ago
  I'm not sure the path to profitablity is so clear. They are just starting to come out of subsidized pricing. A lot of people are going to learn how expensive they are for the amount they are using them. I personally prefer the smaller models because they are much faster for almost the same quality. I prefer open weights even more for the freedom to use as much as my hardware will let me (~250M tok/day right now) and the expectations they will continue to improve over the coming years. We are going to be running experiments at work to determine what / how much it would take for self hosting a deep seek / kimi to have a nice ROI. A few of us now have the spark to get a taste and catch the bug