Would anyone be interested in me hosting a few open-weights and very large base models for Loom usage? I may also host a few frontier open-source models, but with an API for steering vectors and maybe other mech. interp. niceties.
Note that for cost-effective hosting of these base models, a form of scheduling will be applied.
More specifically: if a model is not ‘active’ (e.g. llama-3.1-405b), you would basically request it (for e.g. an hour). Requesting a model costs excess money, since it takes significant amounts of time to download and load weights into memory. On a webpage, you would have a hourly calendar that shows when certain models have been requested (or are planned to be requested). During each period of activity, the cost of using the models is split between users, based on the percentage of tokens received by each user.
This should hopefully allow me to host infrequently used base models at substantially lower (and likely economical!) prices[1], since users should be incentivized to share (and therefore saturate) the compute used for inference.
- ^
vs. the countercasual, which is each user hosting their own version of llama-3.1-405b independently
(i don’t know where to submit bug-reports to LW staff)
why did LW ask for the ‘access to apps’ permission today (only on the homepage, does not occur when accessing specific blog posts, using chrome 149)? said permission request can be replicated easily after resetting the permissions for lesswrong.com.
note that this seems likely to be a coding agent’s error; the ‘access to apps’ permission on modern chrome is usually requested when an ajax request is made to localhost. in the developer’s console, we see a http-request-over-https warning for an image located at ‘http://localhost:1234/images/blog/moon-venus-jupiter.jpeg’, the likely source of the issue.