5 ms·

> I hope that this sharing will continue and increase It will continue until they pass us. They're coming from behind, and releasing models as open source esse

> I hope that this sharing will continue and increase It will continue until they pass us. They're coming from behind, and releasing models as open source essentially wipes the board of hundreds of American startups. Growth hackers and solopreneurs like @levelsio can be nipping at the heels of $300M funded companies who poured all of their points into training models, making expensive mistakes, and overpaying for compute. Companies like RunwayML are probably doomed at this point between Hunyuan, Wan, and the half dozen other open source foundation video models. They failed to raise their last round in 2024, and they definitely face a much bleaker world in 2025. To expound on that some more, while China might not be leading in LLMs just yet, they definitely are with media. Their video models are the best in class. This poses some significant questions. In some years, China might be able to unseat the US as the global cultural export powerhouse. As the price of content creation drops orders of magnitude, content creation will become hyper local. That'll be another ding to US soft power. China will want to commoditize this.

Zamiel_Snawley 2y ago
I think this is the best economic function of open source—it forces innovation by elimination of rent seeking. If you sit on your laurels, someone will make a solution thats at least 70% as good for free.
- wavefunction 2y ago
  or they will regulatory capture, that seems more likely
- Philpax 2y ago
  It's a nice sentiment and I agree, but I'm not sure we'll see this applied to AI going forward. These training runs cost a lot of money - at some point, every player in the game will realise that they need to charge something or they'll left without enough chips to play in the next round. DeepSeek's altruism has taken them far, but they have costs, too, and High Flyer / their personal warchest can only take them so far. And that's before any potential government intervention - it's very likely that this will become a natsec concern for all nations involved.
  - baby_souffle 2y ago
    I think there's always going to be somebody that doesn't want to bother with the complexities of setting up a model to run locally. Spending way too much time trying to track down a very particular version of a GPU driver or similar just isn't going to be worth it if you can make an API call to some remote endpoint that's already done the heavy lifting. Plenty of value in handling the hard part so your customer doesn't have to. I don't know how much of the current focus on local models comes from privacy concerns, but at least some does. Once there's something like the gdpr but for data provided for inference, I think even more people will put down the docker containers and pick up the rest endpoints.
  - tecleandor 2y ago
    Yeah well, but you should add OpenAI to that "altruism" first. They're the ones burning money like crazy that could find themselves in the end of their runway quite quickly. Isn't crazy to think DeepSeek could break even and make a good profit before OpenAI goes broke, if they don't find another investor with even bigger pockets.
  - ForTheKidz 2y ago
    There are models with open training as well as weights. It's not chatgpt, but that trained model can be fully audited and reused without further training. I think for most daily use I'd be fine using a greatly outdated model. Open source has always relied on contributors eating costs—even just the opportunity cost of contributing time. FLOSS software is slow, much slower than VC-funded explosive growth, but it's hard to compete with in the long term.
    - Philpax 2y ago
      It'll get increasingly difficult for open models to keep up with proprietary frontier models as the costs involved increase. Yes, there will always be older open models available, but at some point, they just won't be as competitively useful. Your code assistant that can output a file's worth of code will pale in comparison to systems that can create entire projects in seconds. If it costs billions to reproduce the latter, who's going to do it and give it away for free? There's a possible solution here in the form of distributed training, but that's still tentative and will always be at a lag to the centralised training the big players can do.
enugu 2y ago
Most startups are not building foundation models, but using them as an input. So commoditizing their complement can actually lead to more opportunities. Similarly, web hosting which required big investments previously is now commoditized after emergence of AWS and other cloud providers. Good point regarding movie localization. But, there is still a network effect in entertainment - people want to watch what is popular. So the top content made with lot of money can still have a big market.
csomar 2y ago
I find this a good thing as the end consumer. Not sure why I get a negative vibe from your comment but maybe I have misread?
seanmcdirmid 2y ago
Chinese companies have been diligently releasing a lot of models. I guess they could try shoving the genie back in the battle, but if it works out for them, I doubt they will just stop when/if they get ahead of everyone else.
therealpygon 2y ago
That, and they tend to not let things like copyrights and patents stand in the way of innovation. It’s hard to imagine how much better off consumers, and how many more advancements in technology there would be, if we had reasonable time limits on copyrights and patents. Instead, we have things like “I drew a mouse and now you can’t for the next 100 years” and “I was the first person to change this equation slightly, so you can’t compress video like this for the next 20 years.”

Zamiel_Snawley 2y ago

I think this is the best economic function of open source—it forces innovation by elimination of rent seeking. If you sit on your laurels, someone will make a solution thats at least 70% as good for free.

wavefunction 2y ago
or they will regulatory capture, that seems more likely
Philpax 2y ago
It's a nice sentiment and I agree, but I'm not sure we'll see this applied to AI going forward. These training runs cost a lot of money - at some point, every player in the game will realise that they need to charge something or they'll left without enough chips to play in the next round. DeepSeek's altruism has taken them far, but they have costs, too, and High Flyer / their personal warchest can only take them so far. And that's before any potential government intervention - it's very likely that this will become a natsec concern for all nations involved.
- baby_souffle 2y ago
  I think there's always going to be somebody that doesn't want to bother with the complexities of setting up a model to run locally. Spending way too much time trying to track down a very particular version of a GPU driver or similar just isn't going to be worth it if you can make an API call to some remote endpoint that's already done the heavy lifting. Plenty of value in handling the hard part so your customer doesn't have to. I don't know how much of the current focus on local models comes from privacy concerns, but at least some does. Once there's something like the gdpr but for data provided for inference, I think even more people will put down the docker containers and pick up the rest endpoints.
- tecleandor 2y ago
  Yeah well, but you should add OpenAI to that "altruism" first. They're the ones burning money like crazy that could find themselves in the end of their runway quite quickly. Isn't crazy to think DeepSeek could break even and make a good profit before OpenAI goes broke, if they don't find another investor with even bigger pockets.
- ForTheKidz 2y ago
  There are models with open training as well as weights. It's not chatgpt, but that trained model can be fully audited and reused without further training. I think for most daily use I'd be fine using a greatly outdated model. Open source has always relied on contributors eating costs—even just the opportunity cost of contributing time. FLOSS software is slow, much slower than VC-funded explosive growth, but it's hard to compete with in the long term.
  - Philpax 2y ago
    It'll get increasingly difficult for open models to keep up with proprietary frontier models as the costs involved increase. Yes, there will always be older open models available, but at some point, they just won't be as competitively useful. Your code assistant that can output a file's worth of code will pale in comparison to systems that can create entire projects in seconds. If it costs billions to reproduce the latter, who's going to do it and give it away for free? There's a possible solution here in the form of distributed training, but that's still tentative and will always be at a lag to the centralised training the big players can do.

enugu 2y ago

Most startups are not building foundation models, but using them as an input. So commoditizing their complement can actually lead to more opportunities. Similarly, web hosting which required big investments previously is now commoditized after emergence of AWS and other cloud providers. Good point regarding movie localization. But, there is still a network effect in entertainment - people want to watch what is popular. So the top content made with lot of money can still have a big market.

csomar 2y ago

I find this a good thing as the end consumer. Not sure why I get a negative vibe from your comment but maybe I have misread?

seanmcdirmid 2y ago

Chinese companies have been diligently releasing a lot of models. I guess they could try shoving the genie back in the battle, but if it works out for them, I doubt they will just stop when/if they get ahead of everyone else.

therealpygon 2y ago

That, and they tend to not let things like copyrights and patents stand in the way of innovation. It’s hard to imagine how much better off consumers, and how many more advancements in technology there would be, if we had reasonable time limits on copyrights and patents. Instead, we have things like “I drew a mouse and now you can’t for the next 100 years” and “I was the first person to change this equation slightly, so you can’t compress video like this for the next 20 years.”