HN user

rrnechmech

27 karma
Posts10
Comments43
View on HN

To mitigate auto-regressive drift during inference, we corrupt context frames by adding Gaussian noise to encoded frames during training. This allows the network to correct information sampled in previous frames, and we found it to be critical for preserving visual stability over long time periods.

I get this (mostly). But would any kind soul care to elaborate on this? What is this "drift" they are trying to avoid and how does (AFAIU) adding noise help?

Fascinating to think about if and when FSD is perfected (or heaven forbid before) what a world with cars moving about with near zero reaction time would look like. That is to say, completely removing reaction time out of the equation at scale

retired senior aerospace engineer

Kudos, sir, and all due respect.

An admirable industry, and one that could nowadays use much of the values, approaches, and skills of your generation of engineers

I know this was just a crude crude hack but I once implemented something akin to what you describe using a bunch of domains on Namecheap and their "wildcard redirect" feature, where they would redirect allusernames@foo.bar (a catch all), forwarding all email, per domain, out of all places to a self-hosted Zulip instance, so each domain would forward to a Zulip channel (each Zulip channel gets an email address you can activate and forward messages to). Worked well. Pretty neat. Of course in this particular case only cared for the incoming mails, which may or may not be your case