3 ms·

What is Amazon doing in this space? I feel like they have the infra and skills to train their own model and serve it but I don’t see them creating anything.

mattnewton 3y ago
They’re selling a lot of steel and wooden handles to the shovel makers right now. No doubt if they see an open design like this one they like they’ll iterate on it and offer it too, that seems to be the “bedrock” branded stable diffusion and similar. Maybe they think, why train the foundational model from scratch when you can get some company to rent your sever to do it, and then license it to you for free anyways?
- ttul 3y ago
  Partly they do this to keep their ML researchers entertained after the glitz of their $1M compensation wears off. And partly they do this to develop products for their customers. Although they partnered with Anthropic, whose arguably SOTA models are vastly superior, AWS has an extremely diverse set of customers with simpler requirements that may be quite thoroughly satisfied with a 7B parameter model.
  - brucethemoose2 3y ago
    A 7B model must be cheaper and faster than Anthropic's huge model. To me, it feels good enough for retrieval/summarization tasks Amazon could sell, especially with grammar based sampling.
brucethemoose2 3y ago
You can say this about everyone. Stable Diffusion made a fuss a year ago. Everyone is hoarding A100s and H100s like they are the last GPUs ever made, and we have... Maybe 5 good downloadable base model LLMs, and ~4 good diffusion models? And only one from a really large company. shrug
albert_e 3y ago
Amazon is a bit late to the party but they are investing heavily to stay on the podium. They announced BedRock and made it GA directly after a gated preview of just a couple of months. Titan is their trained model. November is their annual gala conference time. Expect a lot of GenAI announcements resounding through late November.

mattnewton 3y ago

They’re selling a lot of steel and wooden handles to the shovel makers right now. No doubt if they see an open design like this one they like they’ll iterate on it and offer it too, that seems to be the “bedrock” branded stable diffusion and similar. Maybe they think, why train the foundational model from scratch when you can get some company to rent your sever to do it, and then license it to you for free anyways?

ttul 3y ago
Partly they do this to keep their ML researchers entertained after the glitz of their $1M compensation wears off. And partly they do this to develop products for their customers. Although they partnered with Anthropic, whose arguably SOTA models are vastly superior, AWS has an extremely diverse set of customers with simpler requirements that may be quite thoroughly satisfied with a 7B parameter model.
- brucethemoose2 3y ago
  A 7B model must be cheaper and faster than Anthropic's huge model. To me, it feels good enough for retrieval/summarization tasks Amazon could sell, especially with grammar based sampling.

brucethemoose2 3y ago

You can say this about everyone. Stable Diffusion made a fuss a year ago. Everyone is hoarding A100s and H100s like they are the last GPUs ever made, and we have... Maybe 5 good downloadable base model LLMs, and ~4 good diffusion models? And only one from a really large company. shrug

albert_e 3y ago

Amazon is a bit late to the party but they are investing heavily to stay on the podium. They announced BedRock and made it GA directly after a gated preview of just a couple of months. Titan is their trained model. November is their annual gala conference time. Expect a lot of GenAI announcements resounding through late November.