Hypothesis:
AI lacks an appreciation for simplicity that makes genius humans special.
Given only data up to 1600 our current AI would land on a geocentric model of the universe because it fit perfectly and would miss the heliocentric model because of small imperfections caused by noise/weirdness of space.
I see this in the code agents write, hillclimbing towards matching a spec rather than writing code that embodies the spec.
I wonder how this gets solved.
so we don't do just in case snapshots, but i don't think this is crazy. memory is fundamentally bound to the metal its running on at that instant, if power goes out (which it does even at AWS), RAM fails (which it does even at AWS), or the host OOMs/hits a kernel lock up (skill issue but it happens) its gone.
If you really really really care about the RAM, your options are either:
1) Network attached RAM via UFFD, slow and expensive af nobody does it
2) periodically back it up
We trust our host reliability and migrations enough that this strategy doesn't make sense for us, but i don't hate it as an approach.
400 virtual machines created in 2.8 seconds. San Francisco rebuilt in Minecraft. a simulated data-center outage.
we hosted four live demos at Vector Legal last night. thanks @freestyle_dev, @orca_build, and @rivet_dev for joining us!