Those models are still outrageously expensive to train. They may be free for you to use, but they aren’t free to make and once the existing hype train crashes no one with the resources to create this stuff is gonna wanna spend money training the next Mistral or Llama or Deepseek.
The existing models will continue to exist, sure, but they’ll be frozen in their current state and updated versions with newer training data won’t be released. The ecosystem will go stale — at least until someone figures out how to make training cheap which probably isn’t even technically possible.
When community models were first getting formed (going back to the pygmalion days), there were distributed systems that allowed you to “contribute” gpu power from your personal computer to help update the model or provide compute for others - I would imagine we would return to something similar to those roots.
This would not be deepseek - this would be the niche models that were made by hobbyists for hobbyists.
Those models are still outrageously expensive to train. They may be free for you to use, but they aren’t free to make and once the existing hype train crashes no one with the resources to create this stuff is gonna wanna spend money training the next Mistral or Llama or Deepseek.
The existing models will continue to exist, sure, but they’ll be frozen in their current state and updated versions with newer training data won’t be released. The ecosystem will go stale — at least until someone figures out how to make training cheap which probably isn’t even technically possible.
When community models were first getting formed (going back to the pygmalion days), there were distributed systems that allowed you to “contribute” gpu power from your personal computer to help update the model or provide compute for others - I would imagine we would return to something similar to those roots.
This would not be deepseek - this would be the niche models that were made by hobbyists for hobbyists.