rime @rime@schelling.pt

@niplav OH MY GOD they added phoenix

Recipe: `🐦` + `Zero Width Joiner` + `🔥 ` = 🐦‍🔥 (Not supported on mastadon it seems, but works in gsheets!)

> "Approved in September 2023 as part of Emoji 15.1. Available via the latest Samsung devices, Google's Noto Emoji fonts, and iOS 17.4. Coming to more platforms throughout 2024."

I swear I've been looking for this when it didn't exist.

@niplav Btw, claude-3.5-sonnet seems surprisingly eager to quantify uncertainty. It frequently does so even when this is the phrasing with which I request it:

"""
And when no specific answer is available, I'd really appreciate if you quantify/bound your uncertainty. : )
"""

Whole system prompt below, in case usefwl to you. I'm embarrassed about not yet having invested the warranted amount of time&effort in optimizing&personalizing the upfront context I provide it, but… TODO.

@niplav Anyway, um… I have question for you. 👉👈

I want to ask an AI (via system/user message) to use subscript probabilities (or another non-clumsy in-line way to do it), but I'm not sure what the semiotically-optimal option is.

- Confidence interval? Idk how to write those, or how avoid subtle noob-mistakes.

- Or maybe point-estimates are fine? In percentages or odds? Log-odds?? Or maybe "share likelihood ratios, not posterior beliefs"? Hartleys then??

@niplav I've had this feeling at least two times, but not for an entire half-hour! (I can't recall credences, but the plausibility was just emotionally salient/urgent to me.)

@niplav Also, I feel like this thread started out with a tone that makes it appear like I was contradicting you. But I meant to confirm the thing you meant, while providing nuance re the sources of that non-locality (bon my current models).

@niplav ¹DCL ~is um.. smth like the expected amount by which two randomly chosen neurons have computationally-connected activity-levels at any given time, weighted by the distance between them.

…or smth. I'm trying to translate the concept from memory from where I saw² it defined for Ising models, and I think I failed.

²
https://youtu.be/vwLb3XlPCB4?si=MZzvj-L2vGQAzKQv&t=739

@niplav Globally-sensitive is another way to say critical brain hypothesis: the idea that the brain is constantly tuned for maximum Dynamic Correlation Length¹, which is achieved by maybe-something-like regionally renormalizing activity-levels so it borders "criticality" (ie, closeness-to-phase-shift) all the time.

Neuronal activity on several measures is lognormal, so a significant fraction of spikes have much larger effect on the rest of the network compared to others.

@niplav Or so says the module that is writing those words into a comment. By the time information from perceptual fields (by which I mean first-responders to external stimuli) becomes salient enough for you to *think about it*, it has already been synthesized with / filtered against everything else.

But yes, the field-as-perceived-by-us (where "us" refers to the obvious things I mean of course, whatever those are) is really globally-sensitive.

Could be interesting to combine it with a prompt like

> "Hello AI! Firstly, what does it look like I'm doing? Secondly, do you have any particular information you think I ought to be aware of wrt what it looks like I'm trying to achieve? An easier way I could go about it? Stuff that I don't know that I don't know, so can't even check for?"

Not sure what I wish to do with this information, but I do note that having an AI ~constantly scanning my monitor to try to infer what I'm up to is well within price-range.

Processing one full screenshot per 5m via claude-3.5-sonnet costs ~1.35 USD per day, excluding output-tokens. ~Same price for gpt4o & 〃-mini.

https://docs.anthropic.com/en/docs/build-with-claude/vision#evaluate-image-size
https://openai.com/api/pricing/

What a legend 🔥

"Hi everyone, I'm Tsvi-Benson Tilsen. Hope I'm saying that right."

the top shelf in any of my libraries (eg 'to-read', 'to-watch', 'to-listen') is always labelled "review". …I wish it were as shiny as the rest.

new things are ~always shiny in excess of their value. the Nth time we review a memory is worth expontentially more than the (N-1)th time, bc ➀ the Forgetting Curve, and ➁ total epistemic profit (the "rent paid" by the memory) depends on its longevity

we overestimate how much we remember, bc
➀ metamemories produce strong feelings of recognition (eg, I recall *that* I've read abt X ~indep of *what*), and
➁ survivorship bias wrt what we can even check (ie, we can't spot-check the uniform dist of what-we-used-to-remember)

@niplav TO BE CLEAR I'm liking this because it's a nice display of private-thoughts-in-public, not intending to communicate anything beyond that. I mean it!

@niplav It all adds up to normality. It may or may not add up to wayyy more than that too, but at least it can't add up to anything *less* than normality.

Just hope it doesn't mess with the ontology upon which my ethics depends. Reality done that too many times already. ❤️‍🩹

In a sense, it's my ethics which holds things together. I say "ouch", and I know all else has to adjust to accommodate the fact that I care about whatever-that-was.

So if we can compress the world as much as we already have, using tools (brains & GNW) for which there we have strong reasons to expect are very limited, that suggests to me that there's a wealth of untapped simplicity beyond the horizon. But it can only be accessed by top-down non-myopic deliberate practice.

One reason to expect brains to merely be scratching the surface is the deep inevitability of pleiotropy and build-up of technical debt (increased refactoring-cost) for all hill-climbing algorithms. "Consciousness" (i.e. global neuronal workspace) is the only adaptation on the planet that even marginally addresses it. But it too is powerless against the tide.