As someone who isn't a machine learning expert by any means, are these "System One" models effectively a "refreshed look" at prior art in the space, or is it truly novel new ways of using models? I'm deeply invested in trying to dissect more from this space, as I think inserting a small {classifier,decider,extractor,...} model at any point in the SDLC can drastically improve the state of affairs. More specifically, I'm thinking that since these models are so small, they're able to fit into Docker containers, which are effectively run ~everywhere for CI / CD, and cost basically whatever is the cost of storing a Docker image.
Biggest thing is that these are made from very capable LLMs trained on the whole internet. So they work out of the box on almost any classification task. Much, much different than struggling to train your own brittle classifier.
I’m surprised this hasn’t had more attention, it seems like it should be a really solid half way house between the Jev-alikes and full-fat LLM calls. Hopefully ‘soon’ means ‘days’ for the open weights release.
As someone who isn't a machine learning expert by any means, are these "System One" models effectively a "refreshed look" at prior art in the space, or is it truly novel new ways of using models? I'm deeply invested in trying to dissect more from this space, as I think inserting a small {classifier,decider,extractor,...} model at any point in the SDLC can drastically improve the state of affairs. More specifically, I'm thinking that since these models are so small, they're able to fit into Docker containers, which are effectively run ~everywhere for CI / CD, and cost basically whatever is the cost of storing a Docker image.
Biggest thing is that these are made from very capable LLMs trained on the whole internet. So they work out of the box on almost any classification task. Much, much different than struggling to train your own brittle classifier.
I’m surprised this hasn’t had more attention, it seems like it should be a really solid half way house between the Jev-alikes and full-fat LLM calls. Hopefully ‘soon’ means ‘days’ for the open weights release.
It's interesting that they haven't published any benchmarks yet.
Love this trend of specializing models for specific tasks and making them better at them
Looking forward to this being reinforced by smaller models as well
Exciting stuff happening
Seems like it might excel at NER and intent-slot tasks, too.
Alright, I definitely want the open weight version of this.
[flagged]