carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

word report #3

terms i use, mostly pre-existing ones, whose meaning i want toclarify. see also word reports #1 and #2.

• "pretty much": i often need to say "either X, or almost X", and i've found "pretty much X" to be a nice way to express that by making more formal an existing expression, the same way i tend to use xkcd's definitions for "few", "handful", "several", and "couple". i just checked, and all uses of "pretty much" on my blog are meant to carry this definition.
• "universe": the set of things that have some amount of "regular" causal connection with us, our future lightcone, or our past lightcone. "regular" is meant to exclude weird things like aliens in parent universes suddenly interfering with our universe out of the blue.
• "cosmos": everything that exists. yes, this is meaningful; see 1, 2, 3.
• "demon": an agentic thing, typically unaligned from us. this can be an unaligned superintelligence, counterfactual unaligned agentic program in the solomonoff prior, aliens trying to acausally attack us, and arguably even malign agentic structures such as unaligned corporations. see also: are minimal circuits daemon-free (i don't make a distinction between "demon" and "daemon")
• "determining": i like to say "determining X" when i want to be ambiguous as to whether i mean "to make X" or "to find or figure out X" — typically because i don't know which i mean myself, or because i think the matter of which it is is poorly defined. though be aware that i haven't been super consistent with that use.
• "FAS": fully aligned singleton. see my outlook on AI risk mitigation.
• as i explain in what is value?, i use "core values", "axiomatic values", "terminal values", "intrinsic values", and "ultimate values" as synonyms; the reason i've been trying to favor "intrinsic values" is that it's the term wikipedia uses for that concept. in addition, when i say "values", i generally mean just intrinsic values, rather than both intrinsic and instrumental values.
• "RSI": as baffling as it is to me, many AI alignment researchers don't know that this stands for recursive self-improvement, the concept of an AI improving its own capabilities, including its own self-improving capabilities.
• terms i've started using quite a bit to characterize alignment schemes: wonky, formal, eventual & continuous.
• i've been failing to say "exfohazard" instead of "infohazard". i'll try to switch to exfohazard when i mean that, and perhaps "fohazard" to mean both.