“everything is compression” is wrong, as it ought to be “everything is compression and execution”

asteriskarianism

georgism will fix everything people — conditional prediction markets fix everything people — daygame fixes everything people

hi ate us

The problem with the Repugnant Conclusion in UDASSA is that you run out of low-complexity creatures to occupy the short programs with

s h o r t t o r s o

Gwern must feel so lonely up there

“making infinite use of finite means”

distem, tetratem, extem, anatem, katatem, metatem, mesatem

transtem

And last but not least: Contests such as the underhanded C contest show that it is quite easy in sufficiently expressive formal systems) to create malicious outputs with high plausible deniability. Since training ML systems is even more expressive than that, the existence of such successful contests seems worrying

output format of AI systems doing alignment research (formal proof, set of heuristics, set of algorithms…)? AI systems will align their successor systems, repeatedly. Unless this process has 100% fidelity, errors in the alignment process will compound over time, similar to numerical instability. Similar to numerical analysis, could we make a useful statement about the “condition number” of this repeated alignment process?

lol @ the nauty source code: some functions are >1k lines long

are there any other industries in which we apply a similar safety standard as training of large neural networks? the closest that comes to my mind is animal husbandry, but I think we understand the genetics of that better than we understand neural networks, and don’t apply ever-more increasing optimization power to it. (although, to be fair, when humans began husbandry they understood it far less)

instead of e.g. new linkers. why is there only one ImageMagick, or one ffmpeg, but ~6k programming languages? status.

people creating new programming languages is a status thing

if you do something risky and potentially dumb please at least report on it so that we can learn from your mistakes

lean into reality ⋄ lean into irreality ⋄ lean into dissociation

trying to fit in with Gödelian set theorists by saying “V” when someone does something outrageously stupid

dril wrote for HiveSwap. It’s all connected

I am concerned about my outgoing debts, but not about small-ish incoming ones

Erz Herz zog

I think I’m getting more curious with age

Cormac McCarthys writing has strong Harvard sentence energy.

I’d like to garbage-collect some symbols in the math I’m currently writing, help

“You can’t take that! That’s my emotional-support Pareto improvement”

my experience reading Worm so far: 0%-20%: Taylor no! 20%-50%: lol. lmao 50%: gniihihihihi descriptionfucking

do not mix up creatine and ketamine. worst border control of my life.

draining the baby out with the bathwater

Shapley values are NP-hard to compute, right?

“nubile” should mean cloud-like

productivity under anxiety is a repulsive fixed point (you might even feel worse the closer you get to what you should be doing)

from a computational complexity perspective, completeness is the “worst” axiom to violate: only O(n²) to find a violation

no Eliezer you’re wrong Runaround is about tiling agents,

“…still working on the frugal priors though”

von Neumann: I came up with this new system that generalizes probability theory to consider convex sets instead of point estimates. I think that I could use this to prove regret bounds…

WAIT WHAT DID ZIZ GET HER NAME FROM ONE OF THE ENDBRINGERS FROM WORM‽

snitches get stitches, bitches get skritches

steelmanning benefits the steelmanner

roman buddhists

midjourney prompts: the buddha as a roman senator

gimli voice: they call it a site

e/acc: epistemic accelerationism

Unicode doesn’t distinguish between a dollar sign with one and a dollar sign with two strokes, which makes me sad.

performing a million moleFLOP on the common crawl to cultivate a new gradientspawn

faux-translating “estrogeno” from esperanto, it means master-hormone or master-chemical

How long is your fundamental OODA loop?

I don’t remember where I first read this, but: A government’s OODA loop is ~6 months long.

chugging the melatonin pills to keep my energy from making me explode

you might be thinking: “aha! so I should vote in elections, since even though under do()-calculus, the decision has a miniscule impact, there are many agents that are logically correlated with me, which means my influence is much higher!” A tiny problem is that the number of agents that are logically correlated because they base their decisions on logical correlation is, ah, not that big…