Reply to thread

Message

<blockquote data-quote="Jfdlsjfd" data-source="post: 9650897" data-attributes="member: 42856">TBH those are mostly (rather small-scale) modification to base models. The cost of training is however lowering for base models: that was the reason the Chinese text model Deepseek got a lot of attention. Not because it was better than the existing top players, but because the compute time cost was only 6 millions USD instead of the (presumably) billions of OpenAI's.A model image would need less training. AuraFlow is the leading model for prompt adherence and was developped by a little guy (student at the time) (although helped by a small compute grant).&nbsp; Pony v7 cost 300,000 USD of compute time and is a community-funded initiative paid by little guys. And it's an ambitious base models. The cost might also be declining in the long term because one doesn't need to start from scratch by using a decent base model, though. As you pointed out, a good base model might only need limited finetune to be improved, for which a gaming computer is enough.Also, a H200 right now starts at 100,000 USD. People routinely pay that price for cars. They are not &quot;little&quot; guys but the type of system needed went down from extremey costly to within the means of an enthusiast. And it won't become more expensive with time.</blockquote>

[QUOTE="Jfdlsjfd, post: 9650897, member: 42856"] TBH those are mostly (rather small-scale) modification to base models. The cost of training is however lowering for base models: that was the reason the Chinese text model Deepseek got a lot of attention. Not because it was better than the existing top players, but because the compute time cost was only 6 millions USD instead of the (presumably) billions of OpenAI's. A model image would need less training. AuraFlow is the leading model for prompt adherence and was developped by a little guy (student at the time) (although helped by a small compute grant). Pony v7 cost 300,000 USD of compute time and is a community-funded initiative paid by little guys. And it's an ambitious base models. The cost might also be declining in the long term because one doesn't need to start from scratch by using a decent base model, though. As you pointed out, a good base model might only need limited finetune to be improved, for which a gaming computer is enough. Also, a H200 right now starts at 100,000 USD. People routinely pay that price for cars. They are not "little" guys but the type of system needed went down from extremey costly to within the means of an enthusiast. And it won't become more expensive with time. [/QUOTE]

Verification