Post

I agree that the plot does not really convey why this is huge. But for robotics ... this is huge.

But what evidence is there that this is huge?

Well if it isn't entirely fake, this is the ultimate replacement for decision-trees that need to perform close to real-time. Do you have some experience in robotics?

it’s not about the performance it’s about the magnitudes of the cost & latency, and the shape of the output

The plot has two axes Accuracy Cost Accuracy as they explain is measured by tasks they chose; cost is a parameter chosen entirely by them.

yeah, i mean take it with a grain of salt. most people here just care that it’s “good enough”, because the cost is so low. Plus the latency enables all sorts of new things. the hard part is getting people to understand what it is and why it’s significant

The interesting to me is thing to me is that it's shaped very differently from an LLM, like it's just a structuring machine. But then yeah, hard to compare head-to-head with other things.

I mean, this is a model class that has existed before. But it’s pretty sweet markwting—look how edgy and cool it looks!

my conclusion after playing with it is that i agree with you Maxine. it's fun marketing, has some cool webpages. it's a nice little model which i suspect is a finetune-ish variant of deepseek or qwen or something. the alternate interface of choices+bools and stuff is very fun to play with.