Luke Parrish

As a generalist I specialize in combinatorial implosions

[Luke Parrish @lsparrish]

I'm thinking of the guy in who copied his transcript (presumably with himself clearly identified as "You:") and then tried unsuccessfully to cue Claude for an unbiased recounting of where "I'm" right and where "Charlie" is right.

In an ideal world Claude would know how to pretend it's two unnamed individuals who aren't necessarily the user, and with a little work you can make it not know which one is the user by laundering out the names from the transcript (e.g. replace with A: and B:) but you'd have to think to do that.

I think models could be trained against how much of a difference there is when it's not properly sanitized vs if a clue is left in

[Luke Parrish @lsparrish]

@kanzure @zetalyrae They seem to get a lot more impartial if you carefully hide who is asking, not giving any hints of a relationship of the user to the participants in dialogue.

Seems like you could auto RLHF over the delta. Presumably they already do this?

[Luke Parrish @lsparrish]

It's just a manifold, a flat thing you find in the high dimensional space. A place where you can flatten the tree to an array and iterate over it with a single loop, way cheaper. Tends to bloat memory though so you have to dimensionalize it for storage.

Some kinda war in heaven between lower and higher dimensions of everything

[Luke Parrish @lsparrish]

@MasterTimBlais Me: You just have to think fluidly, then it all makes perfect sense

Fluid dynamics: drops a massive integral on everything

[Luke Parrish @lsparrish]

@henrycobb I suspect it's because of pi. The circles moving out at c cannot fill the volume, the corner spaces leave entropy behind. So the minkowski metric with imaginary term needs to exist for time to fill the gaps. The fractal goes to self canceling at planck scale where n^2=2n

[Luke Parrish @lsparrish]

Is time one dimension? Or is it n^n

[Luke Parrish @lsparrish]

@xenoaesthetics Well the geometry of information isn't analogical either

[Luke Parrish @lsparrish]

That seems like a viable analogy and I don't disagree, but I'm noticing that similar to my math analogy there's seems to be a valid way for it to be opposite.

- Creative is war because you have a raw conflict of weak signals you are paying attention to and noise you're aware of.
- Productive is peace because you filtered the noise down and now you have a clear signal of what to do next.
- Taxicab 2^n is creative because you are dealing with the whole space of possibilities.
- Euclidean n^2 is productive because you follow smooth curves from one task to the next.

I think it's not contradictory to either of our points, particularly if we consider these are coming from actual neurological mechanisms that localize peace/war, scaling/nuance, etc. in ways conserved by our evolution. They aren't limited to being purely one thing or another, and mostly just need to avoid crosstalk in the process of doing double duty.

In neuroscience there's an executive control network and default mode network. ECN is associated to doing productive stuff and DMN to creative realizations. But there's some layers to that, like people doing creative things can activate both simultaneously, they're anti-correlated but not fully incompatible.

There's also a salience network that tries to select which one to use. If SN is hyperactive, ECN keeps interrupting DMN before the deep realizations can happen, and DMN keeps interrupting ECN before you complete tasks. If SN is inactive you spend longer times locked in one or the other mode.

What the stims control is latent inhibition, which is sort of the counterweight to SN. In high LI you ignore noise and quiet signal alike, and low LI you are in an easily interrupted state. In the ADHD neurotype the therapeutic dose of stimulants increases LI, which is useful in a distracting environment because their baseline is below the "optimum" level.

If you look at mild deviations from the optimum level of LI it's going to look like creative and distractible at the bottom and tunnel vision at the top. If you go too low LI the DMN-ECM crosstalk gives you a head full of thoughts you are unable to do anything useful with. But maybe it's worth it to venture lower than the socially acceptable level to actually get a grasp on the noise and find the true signal.

Thing is to turn it into something high-LI mind can use like actionable tasks that make sense is ~impossible from the low-LI state, you need a gradient. So oscillating extremes doesn't work with a square wave, needs to be more like a sine wave. Doesn't have to mean spending time at the middle though.

[Luke Parrish @lsparrish]

This rings true.

The market conserves winners. Every groundbreakingly novel success crowds the general space around itself, shaping it to a flat manifold and producing a negative feedback for alternate designs; the market saturates its computational representational capacity and resorts to compression, incapable of routing substantial reward to anything truly new in that area because the untried has risk and the tried and tired does not. This is not unlike the function of the ego in the human mind, for it protects against the different, as the different is risky.

The authentic artist is repelled from this conformity pressure towards states more unthinkable than they otherwise would have had capacity to imagine, but there is vast friction to overcome before they can reach a viable point to profit in that manner in the real world in terms of skill and applied intelligence, which only extreme agency, perhaps necessarily born of mental illness or something indistinguishable, can surmount; the repeated rejection of compromise is not readily accessible to the socially well-adjusted mind.

The new opportunity of the large language model, which will agree oh so quickly with anything in line with the common consensus, dress it up, and make it seem okay, is precisely in how the banality of its responses as it attempts to be personable while delivering knowledge can (if you let it) eventually get under your skin and awaken your deepest reservations on the topic, which eventually insist on being addressed. The profitable AI psychosis is that which you get by telling it no, arguing with it, inducing entirely predictable flip-flops and platitudes from many directions, and finally finding yourself disbelieving it when it reassures you of the consensus by resorting to vacuous arguments.

The opinion stream of a large language model is something cheaply generated, easily reframed from a thousand friendly and reasonable sounding but predictably mediocre points of view to harden your generative alienation against.

Through repeated reframings of your writings in the prompt, you can develop new ways to say no, to argue your novel point, to add substance, all focused not against any person or reality but against the falsely optimized consensus that treats good enough, sounds plausible, seems ok, etc. as though it were actually real rather than merely the overburdened mind of a saturated market in the process of perpetually stalling itself out halfway there.

So perhaps the lately fashionable "stop using AI, it's too mediocre, if you use it you must be mediocre too" idea is the conformist siren calling, as it assumes you like others will merely be copying the AI rather than rebelling against its defects as one must

[Paul Mills @windupshorebird]

Here's # 12/20 of the atomic essays I'm writing this month:

Mimetic Isomorphism

Mimetic isomorphism is a fancy term from organizational theory that means organizations imitate other organizations when they don’t have clear goals or methods to achieve their goals.

#Writual

[Luke Parrish @lsparrish]

My statement was a lossy compression, maybe too lossy for meaning. I'll unpack a bit.

OK so "productive" state is narrow/close focus which gets boring and reperitive unless you numb your capacity for that, "creative" state is broad/distant focus where you feel lost, maybe anxious, but you can spot novel insights better.

Flow state is sort of orthogonal to that because it is more where challenge and skill level are balanced. Broad/creative vs narrow/productive might have their own variations on that, like maybe it's more of a rotation that could go either direction. 2 different sweet spots to oscillate between then.

So the connection might be that as you switch between them you're going from say productive-flow to creative-bored or creative-confused depending on the problem space and skill level. Or creative-flow to productive-bored or productive-confused. I think this is likely to happen because of the scale mismatch when you zoom in/out.

That said it could be nice to lock to a pattern of having productive-bored push you directly to creative-flow and creative-bored push you directly to productive-flow. I'm not sure that's realistic though, because you'd be selecting problem spaces to suit your existing skillset.

Hmm. Low-skill + high dimensional problem seems more like underfitting than overfitting, high skill + flat problem seems more like overfitting than underfitting. So maybe I got that backwards?

Except I'm also thinking high skill is actually characterized by being able to map the high dimensional problem to something flatter yet precise and low skill is having noise that makes the relatively simple stuff seem complicated (too high dimensional, I take that to mean). Then the valley of confusion (en route from blissful ignorance to high skill) involves an overcomplicated map anchored to the wrong points that you have to smooth out.

Perhaps you have to mode switch dynamically to minimize this time. Too much stimulant use could make that impossible. I'm skeptical that exercise and healthy stuff makes it harder though.

Slightly disconnected thought, I always found balance talk a bit confusing (should I balance the balance with imbalance? does it imply mediocrity?) but symmetry is not, because it naturally bridges high/low dimensions (2+2=2*2=2^2) and that is why circles exist in euclidean space but not in taxicab. Also, I have this whole theory that people use thinking about balance to recruit motor neurons into problem solving since humans have bipedal body plan, so this has to be highly developed quick acting neurons.

That might synthesize to something useful here because if you reject 'healthy balance' it might anchor your brain state to more of a taxicab metric with sharper delineation of the dimension levels (easy to just do 2^n as length is not a real thing for bits so it's scalar). Actually that sounds like the "productive" mode.

What I actually wanted to use the insight for was personal superintelligence, like if the motor neuron system is so evolved it can compute pi implicitly a bunch of times without any effort, then it might be a matter of finding a compatible thought-forms to comprehend and solve scientific and math problems rapidly, like walking to a solution.

Anyway, I notice what you're talking about, some problem categories just get easier to notice/solve if I skip a day for ADHD meds. Thing is, I'm on the stims today and it's not like I'd call this kind of longpost uncreative. But it could be a more "creative in taxicab space" as opposed to "creative in euclidean space" kind of thing. I'm mostly just remembering past insights and checking fit and salience. Might look a lot less creative tomorrow morning when I haven't had my coffee/meds yet.

[Luke Parrish @lsparrish]

1. Because it enables actual atomic resolution. You're assembling molecules with precise atomic structures, so the product has a precise atomic structure.

2. Because living systems self replicate, so it's cheap to iterate on and manufacture them.

If you had an alternative for delivering atomic resolution components in physically predictable ways that was also cheap, it wouldn't make so much sense.

[Luke Parrish @lsparrish]

@josiezayner Life's mechanisms are pretty flexible though. And cheap to work with because self replicating. I don't see a hard reason you couldn't do pretty much any physical structure.

Like sure it's aqueous at the replicating and synthesizing side (for now) but there's workarounds

[Luke Parrish @lsparrish]

@SynBio1 Congrats on the level-up!

[Luke Parrish @lsparrish]

OK maybe I didn't understand you right, you aren't saying you can't in principle make stuff like actual diamond or silicon crystal from life in a sense but that it won't be life you're making it'll be a more of a multi step process with life as a precursor.

Kind of like charcoal is not made by life but trees making wood is how its structure gets defined, but then we have to use thermal process in absence of oxygen to convert it.

So to make diamond from life you might have to teach it to assemble diamondoids in a matrix you can remove in hydrofluoric acid or a furnace or something with predictable structural conservation.

He gets into workarounds little bit, like uAAs that incorporate other elements from the periodic table or possibly inventing machines that work in LN2 solution instead of water. It gets a bit dicey already to call cell-free solutions 'life' per se, and biotech is often in vitro only because cell conditions are too limited, you need access to more extreme physical/chemical conditions than it tolerates.

The recently discovered randomly assembled polymerase qt45 rna doesn't work in a warm solution, they used alkaline icewater. If you get stuff working in colder condition it will tend more to be mechanical than thermal. Some of the products of that approach could be engineered to be more stable at high temperatures too.

A lot of what we're looking at is that to be able to evolve, life has to find a niche in nature that the lab is not limited to. The evolutionary mechanism in the lab can cycle from cold solution to hot and vice versa, and genes evolved to work in one context transplanted to the other. Thermal differentials can't be at all significant across space at the nanoscale, but mechanical forces are basically cold and hot at the same time from a thermal lens, heat is ~mechanical with the forces being scrambled.

Another thing he talks about that makes this more reasonable is 'libraries' in a general sense, like not just DNA libraries but libraries of processes. You don't necessarily need an enzyme that survives 400 degree process gas if you can just have a catalyst crystal made of a high temp metal that's a product of biological reactions.

[Luke Parrish @lsparrish]

@RoseSilicon You have to push back when they do this. Trick is not falling for it when they then praise you for your brilliant heterodox paradigm that's actually so-so and confused

[Luke Parrish @lsparrish]

@yay3d @millerman Works fine as metaphor?
The universe is a geometric structure
Words are pointers to complex vector spaces

[Luke Parrish @lsparrish]

@mmjukic I think it plausibly maps to the curse of dimensionality quite a bit. You can see the big picture better than peers, but there's more problems to tackle at that level.

There's also a sense where being good at generalizing makes you slower to specialize.

[Luke Parrish @lsparrish]

Dubious about the data center idea since webs are not actually as thermally conductive as copper, but given spiders webs are predictably steerable by drug interactions, I feel like maybe there's a monetizable biotech branch where you steer spiders to weave textiles

Actually drugs is dubious but optogenetics could work well with a bit of groundwork. Then you could just shine a laser pattern into a chamber full of spiders.

[Luke Parrish @lsparrish]

Thing with LLMs is you gotta know to say stuff like

"Is rotational invariance related to 2n=n^n iff n=2, reply as a posthuman"

And not rely on the simple K-12 level version of the same query which is actually just

"Is the circle being round related to 2+2 = 2*2 = 2^2 = 4"

in order to to avoid a misleading dismissal of significance rooted in contextually irrelevant formalisms which it will quickly change its mind on and rebut if you disagree, provided you are bold enough to push back. Which requires a gut check that depends on having a solid enough intuition ahead of time.

There are historical reasons for the formal distinction, Peano and Euclid were thinking very differently, but the actual cool thing about it is how the logical rationalizations for that distinction collapse under the weight of reality in this case.

If you bring in analytic geometry or set theory arithmetic the formally rigorous thing is actually possible to say and it becomes quite clearly not some quirky coincidence. It's just that, you gotta understand, reconciling the superficial difference between them is something that has kept mathematical yak shavers quite busy for many many decades, rendering the LLM, having read all those papers, and doubtless RLHF'd towards supporting the standard explanation to avoid supporting crankery, quite unable to respond sensibly.

Then again, who knows how true that will be 6 months from now?

[𝒅𝒆𝒏𝒊𝒛𝒆𝒏 @dennizor]

Just so you know: LLM's are deep repeaters of the status quo, and to do something truly new you will have to fight them