How to get URL link on X (Twitter) App
https://twitter.com/surge_future/status/1612781521282256897So, clearly sometimes good planning and design can make things that are exceptionally reliable. This seem to happen mostly when (1) underlying physics and technology well understood, (2) designs optimized for reliability, (3) error correction mechanisms in production and use.
https://twitter.com/paperswithcode/status/1592546933679476736Trying to get a review of lava redirection produces a paper trying to model a layer of lava. "In other cases it may be necessary to divert the flow of lava from one location to another, such as from a volcano to a settlement." Good to know it is just diverting things *away*.
https://twitter.com/SamuelAinsworth/status/1569719494645526529The basic idea is simple: you can permute hidden layer neurons, so there are actually far fewer internal models than it looks. Training gets to one, with linear mode connectivity. Can hence interpolate differently trained networks if one is careful.
https://twitter.com/MariusHobbhahn/status/1559925158818750464The real issue might not even be the offense-defense balance, but whether defense is reliable enough. A world where bad actors occasionally have great wins may be worse than one where they can often gain small. Some credit fraud ok, not everybody's accounts drained.
https://twitter.com/deepfates/status/1542153191961272320I think the essay gets at something deep. We are not just being pre-scientific or pre-engineering with current prompt design ("ha ha, it's black magic"), but the domain has many similarities to magic because it draws on an overlapping source.
https://twitter.com/troed/status/1514884086128644098There is often (1) an assumption about what values are embodied by non-revenue aims, and (2) the assumption that such change must be joint by a group rather than unilateral by an individual.
https://twitter.com/anderssandberg/status/1511288970537312264
https://twitter.com/anderssandberg/status/1511733842876448772
https://twitter.com/andyzengtweets/status/1512089759497269251It is not that I think PaLM, DALL-E 2 or Socratic models in themselves are huge steps towards AGI. It is that they demonstrate rapid improvement in three entirely different domains, looks like it can be generalised to a lot of tasks, and surprise "experts" like me.