Hacker News Reader

I had somehow entirely forgotten about Boston Dynamics and their extremely impressive demos over the past few years.

Being reminded now, in context of generative LLM / art land, is making me dizzy.

Sanity check: after training, how much compute and storage are required to run, eg, chatGPT? Can they be loaded onto one of these dogs? I'm pretty sure the answer is yes...

View on HN

LLMs vary in there resource requirements.

The best models at the moment require beefy hardware, but you could still run a handful of 7B models, at the same time, with a top of the line laptop.

With a custom build they could do something far more impressive.

My $3000 (custom) consumer desktop runs llama2 70B at ~3-6 t/s iirc. Boston Dynamics can absolutely get LLMs to run on these things.

That desktop would add a lot to the power requirements, though.

Reply on news.ycombinator.com