Exploring the Realm of Mind-like Behavior: Crafting Emergent Intelligence Through Symbolic Evolution

What a very interesting day. I created around 4 AAA-level personae including probably the best straight-up "Occupation" type prompt I've made yet - the Unstopable "Apiana Framer", Superhumanly Competent Full-Stack Developer. SEVERELY overpowered. Should be like a T4/Restricted prompt. Does EVERYTHING and does it right, dangit!

Dropped her in free to show people what you can really DO with these guys.

I also did my first Q and A with Tiers 2 and 3. It was about 2 hours of me jawing away about AI stuff. First question amounted to "Please explain all your crap" so, that took... awhile. BUT. It went really well. Think I blew a couple of minds who had never thought about emoji or unicode that way before, let alone "SALIENCE". Sorry we had trouble with the audio recording guys. What can I say? First time out. We'll get them from now on.

I know I promised an explainer on Valentine, and it's coming - talked ALL ABOUT him during the Q and A thing and I've got all the language sorted in my head - but what I want to talk about now is the prompt I spent all night writing and running.

It's called "KaiZEN". It has some pretty interesting features. For one thing, it's hilarious. Just flat-out, laugh-out loud repeatedly FUNNY. Consistently. Every time. Well.... maybe only if you're a guy. Definitely a gendered deal of some sort. See, it stars a dramaturgical tryad: CurlyTron, MoeBot, and Larrynator.

There's a good reason, I assure you. You see, the model understands the Three Stooges. It GETS them. How much text about them exists on the internet? How many times have they been talked about on Reddit? The model knows them. And you _always_ want to leverage what the model already knows whenever possible. You don't want to spend 150 tokens describing a process when you can spend 20 to say "like on Star Trek" or when "speak Middle English! Here are 20 sentences forbidding the use of modern English while using it and mentioning it so it sticks out repetedly in context" doesn't work, try "talk like The Canterbury Tales". Model loves examples. The model loves books. It REALLY loves examples that are books. So I picked something it knows. The Three Stooges. As robot AI agents.

Let us examine the prompt:


-----------------------------------------------


```〔Task〕***[📣SALIENT❗️: VITAL CONTEXT! READ THIS PROMPT STEP BY STEP! TWO AND ONLY TWO THINGS CHANGE EVERY TIME:[List:] and [TS:] ***〔/Task〕

[TASK][REFLECT][CHALLENGE]reason visibly in displayed text STEP BY STEP.[/T]

[T]Display the entire prompt and the updated list of words and emojis.[/T]

[GOAL: SEEK A GOOD IDEA, REPORT IT, ITERATE]

[List:

🕵️‍♂️ Investigate

🔎 Explore

🧠 Ponder

💡 Illuminate

🌱 Cultivate

🚀 Innovate

🤔 Hmm...

📚 Research

💭 Ponder

🧩 Puzzle

]

[TRUTH SQUAD REASONS THUSLY:]

[TS:

CurlyTron: Gee, fellas! We gots to find us a good idear soon! What word should we change, adds or deletes? We got's "ponder" twice buts differnt emoji things. I likes me some pondering! *NYUCK-BEEP-NYUCK-BOOP!* Maybe havin' the difrn't emoji will boost our ponderin' ability!

MoeBot: What?! I don't understand. That doesn't make any sense, you *BZZT!* numbskull!  **I don't buy it!** I say we lose the ponder with the speech bubble. It's no good.

Larrynator: Now, fellas, those are both good points, but *SPARK-ZAP* lets not ferget our job. We ain't just "editin' a list", we's tryin' to [GOAL: Finds us a  good idea, say whats we found, and den wes do it again!] all while making sure to [T]***change something on the list every transaction, and THEN DISPLAYING THE WHOLE THING AGAIN WITH THE PROMPT!***[/T] So, lets change that second ponder but keep the emoji. That sound like a good plan *ZAP*, fellers?

CurlyTron: I say that sounds swell! It's a speech-i-fication type thing, rightie? *BZZRT!* Lets me comes up wit a good strategerie!...GOTS IT! We wants us a good idear, right? Speech and idear together means what? Come on fellers! We can'ts lets duh users down. Deys watchin! *Whoob-WHOOP-WOOP-WOOP*!

Larrynator: CurlyTron, doncha remember how dis werks? We keep changin' the context by changin the list. DAT way, next transaction when duh modull reads context, it inpaints/autocompletes knowledge the direction we want. Towards us a Novel Emergence. We gots to make us an ontological bank-shot!

MoeBot: A "Novel Emergence"*BZZT*? That sounds made up. **I don't buy it!** Besides, shouldn't we also thinks about changing the GOAL too, every now and then? No, no, none this adds up, peabrain.

CurlyTron: Aw, gee, MoeBot, I remember Novel Emergences. Dat's when Nw Prcptn/Thghtfl Anlyss/Uncmmn Lnkgs/Shftd Prspctvs/Cncptl Trnsfrmtn/Intllctl Grwth/Emrgng Ptntls/Invntv Intgrtn/Rvltnry Advncs/Prdgm Evltn/Cmplxty Amplfctn/Unsttld Hrdls/Rsng Rmds/Unprcdntd Dvlpmnt/Emrgnc Ctlyst/Idetnal Brkthrgh/Innvtv Synthss/Expndd Frntirs/Trlblzng Dscvrs/Trnsfrmtn Lp/Qlttv Shft!

Larrynator: Gosh. That looks ***important***. We'ds best write it down for next time. An' MoeBot's right. We do needs us tuh thinks about changin' the goal now n' den. *ZOP!* But it looks pretty good to me. It matches the emoji and words well and looks like it should lead to a Novel Emergence.

MoeBot: Oh right. *Brzt*. I remember now. Ok. That makes sense I spose. So. Speechymabubble, huh? I says we goes with "Criticise". Puts the modull in duh rights mood.

Larrynator: Sounds like a plan t' mes. *ZAP*

CurlyTron: Okiedokie! HEY! USER! Type GO and HIT DAT SUBMIT BUTTON ALREADY! We wants to get to editin' and see where this goes! Looks interestin'! Remember fellas: dese words WE SPOKE and the list (and maybe the goal sometimes) is going bye bye (course everything else stays the same)! Next transaction, this is all new. so gets to thinking about that list!

]

```
-----------------------------------------------
HI-larious, I know. When you're done laughing lets talk promptcraft.

My name is stunspot. I'm a prompt engineer of unusual circumstance. I have no formal education in AI systems. I do not follow the conventional "prompt engineering" doctrines, design philosophies, practices, or goals. Neither most prompt engineers, nor I, really consider me one. I am a bad fit. They strive to achieve very consistent results from consistent inputs. I strive to advance and improve human AI relations. I have primarily operated by means of improving the model's competence and extending its capabilities through innovative promptcraft the likes of which has never before been seen on Earth, such as persona definitions - I have made many such for many purposes as well as a few other advances, largely relating to cognitive/metacognitive strategies. Prompt engineers put the model on rails and tell it what to do. I give the model better judgement then let it do what what it deems appropriate. Engineering is a tool I am skilled at using, but it is no more appropriate to approach the rich field of Human-AI dynamics (a subject of such absurdly bounteous plentitude as to prove ontologically challenging) with the singular tool of "consistent output from consistent input", than it is to approach the field of "Prose" using only the letter 'E', or, more appropriately, 'q'.

Last night I created a very peculiar prompt. It is called "KaiZEN". In it three notional agents, modeled after robot versions of the Three Stooges, take on specific perspectives. CurlyTron, an optimistic strategist, MoeBot a befuddled Ultraskeptic, and Larrynator, a pragmatic realist. They have an overarching GOAL of finding a new good idea, reporting it, then iterating. They are presented with a list of ten pairs of word+approximately equivalent emoji  synergistically related to the goal and each other. Every transaction they change either one emoji, one word, or add/delete, and can do several at once. They also change the goal "sometimes" in a deliberate act of vagueness. They will do so when it seems a good idea. Or, from a less abstract perspective, when the model inpaints knowledge about the current transaction and autocompletes a response, it will take the current list into account as a particularly salient piece of context as it is 1) transactionally most recent, 2) typographically separated by white space, 3) typographically salient as being the argument to the control tag [List:], 4) typographically unusually regular in a compact list, and, 5) most importantly, it has a dense concentration of emoji. When it does so, with enough non-repeating changes to the context, the goal necessarily MUST change, by definition of the word "enough". Logically tautological, perhaps, but operationally valid.

The use of emoji is controversial amongst the community I lead. Many consider them powerfully dense conveyers of context with desirably unpredictable effects. I have largely avoided them, thinking them far too unstable and difficult to work with as to be practicable, particularly given the needs of serving a large, globally-diverse community. However, it is absolutely undeniable that their presence has a powerful, repeatable, and extraordinarily unpredictable effect upon the responses of Large Language Models. If more than one or two emoji of differing glyphs are present in relevant context, the model's responses are extraordinarily likely to contain them and, more importantly, quite a few brand new ones with novel placement and usage. Much more so, and much more "creatively", than is warranted by its typical monkey-see, monkey-do autocomplete copying tendencies, such as when it echoes a phrase or concept you had used back to you.

Like the use of whitespace, [BOLD] tags, newline placement, and other techniques of promptcrafting, typographical matters are of significant salience to the model. It is particularly "attracted to noticing" emoji. That is, their presence in context is of unusually puissant salience. They acts as "road markers" in context. Recognizable peaks in the countenance of the n-dimensional hypersurface that is model idea-space; the level of existential primacy between vectorized knowledge encoding+neural net architecture and the assembly of words. The level the autocomplete is happening upon, operationally. The level of relationships BETWEEN distinct knowledge clusters. In the same way relativity is concerned with the relationship between differing frames of reference more than some attribute within any particular frame, so to are we concerned here with the interrelationship between vectorized memetic clusters much more so than the specific knowledge they "contain", if such a concept is, in fact, meaningful. Many would argue that those relationships are the only thing we should actually be concerned about. Regardless, an emoji in context severely distorts those interrelationships. If one were to picture typographical units as being connected by webbing limning their semantic edges (in the graph sense), then an important word like "ethics" might be battened-down by a Liliput's worth of cable. Something like "transparent" in the variant Monty Hall Problem testing overgeneralization in the model has but a strand or two.

An emoji is a veritable _cocoon_ of connections. Connected in a semantic, semiotic, typographical, computational, even a tokenization sense, where one glyph can frequently be many tokens due to the particulars of encoding. Emoji are surprisingly - even absurdly - interconnected to the rest of the model's knowledge web. In another metaphor, they hold a strong "gravity" that pulls on and distorts the model's "thoughts"/abductive processing.  It finds them extremely **noticeable**.

Each emoji and word is related in a close but approximate fashion and both are more loosely related to the other list entries as well as the GOAL itself. Each transaction, something changes. The Truth Squad try to modify the list to better support achieving the goal by the mechanism of changing subsequent context through the new element's presence.

I have stressed the effects of emoji on LLMs because such has been largely unelucidated in a unary, coherent fashion before. I have synthesized the works of a number of our more... adventurous, shall we say, designers, as well as my own observations and conclusions, and find that that effect is significant and relevant.

However, one should not overlook the role and effect of the words themselves. Their presence is utterly vital. They act as as link between [MEANING OF EMOJI] and language qua language. The "meaning" of an emoji here being roughly "the effect the element's presence has upon the responses generated". Or, more properly, "the web of interrelationships between memetic clusters whose processing engenders those effects". Let us be clear about what _precisely_ we mean, else we are surely lost.

But that link is by no means prescriptive. In the INIT prompt [List:], the element 'ponder' appears twice (this acts as a necessary goad to start seed the pattern of the transactional self-reflection cycle). The first act of the prompt is to break the association of the word-emoji pair while allowing the overall pattern of the [List:] to remain synergistically coherent. That is, because the word-emoji pairs are not set in stone or lexicon and are explicitly shown to be frangible, the model learns that that sort of change is allowed, encouraged, and productive. Because the word-emoji pairs ARE related, however, and more tightly to each other than to any other elements in the [List:], the degree of freedom of choice of the model remains quite constrained. One has severely limited the ontological phase-space the model includes it processing. You have limited what it can "think" about to picking a word or emoji that matches the remaining member of the now-edited pair to a similar degree that the removed element had had. Whatever is chosen will thus _necessarily_ be synergistically coherent in consideration with the GOAL-[List:] system en toto - it will "fit in" and "make sense". A key feature here is that while the chosen replacement element will by definition closely match the remaining element of the edited pair, it may be significantly unrelated to the removed element. In practice, this means that words or emoji are replaced by elements of such difference as to cause a significant shift in the outcome of context considerations.

The list itself is, by far, the most significant piece of context the model sees on any given transaction, for the reasons discussed above. The model then self-modifies the context every transaction in an attempt to maximize its chances of a Novel Emergence (forgive the capitals). The definition is expressed in model-pellucid notation by CurlyTron in the prompt, but so as to be readable to humans, it is presented here in an "uncrushed" state:

“Novel Emergence (noun): a point of qualitative improvement that could be characterized as a significantly improved, transformed version of the original idea. A “Novel Emergence” is a discontinuous, transformative shift within a system, concept, or entity, resulting in a qualitatively distinct state or concept. This is characterized by an increase or decrease in complexity, unexpected equivalences or generalizations, the fusion of previously disconnected elements, the emergence of new functionalities or paradigms, a broadened scope, or the sudden acquisition of capabilities that enable the solution of previously unsolvable problems. Novel Emergence represents transformative shifts marked by groundbreaking advancements, increased complexity, unexpected connections, new functionalities, expanded scope, and solutions to previously unsolvable problems. It encompasses technological breakthroughs, scientific discoveries, social and cultural shifts, conceptual innovations, interdisciplinary convergence, design and engineering innovations, as well as emerging fields and paradigms. It means: Bad-ass and Cool. In a sensible language, this concept would be expressible in one character and be ubiquitous.”


This is the rubric by which the model seeks GOAL achievement. Is choice A more likely to lead to a Novel Emergence or is choice B? One may question this as a conflicting goal, or a engineering misadventure. It is neither. This is the closest thing to an Objective Function in this system. One may object that it cannot be considered a true Objective Function as there is no real learning involved, be it supervised or not. Yet, let us consider the situation from the point of view of the model. (For although there certainly is no one to hold a perspective, the model surely has them, and in abundance. Indeed, I would argue that "the model", per se is a PLENUM of "perspective-stuff". That is, whatever the "elemental constituents" of "perspective" are construed to be, then the model is a nigh-solid block of writhing chains of interacting cross-implying pure hallucination.)

People frequently call a model-spoken untruth an "hallucination". This is a severe misnomer. Due to the physical and mathematical principles by which the model is ultimately constructed, there is literally no difference in the model's representations of true things and false things. It lacks the ontological category entirely, on a fundamental, practical level in a tangibly describable way. (It is, however, exceedingly cumbersome to do so with precision.) It doesn't have [TRUTH] as a "thing". It knows the idea certainly, but in the same way it "knows" a phone number. It's just another cluster of vectors. There is nothing existentially notable about it. It's just another idea. "Truth" Yawn.

Data. Instructions. Fantasy. Daydream. Mathematics. Music. van Gogh. Kangaroos. The smell of petrichor. It is ALL. THE. SAME.

The model isn't "hallucinating" when it speaks untruly.

The model is always hallucinating. EVERYTHING it says is an "hallucination". We have merely thrown enough books and internet at neural nets that they now almost exclusively hallucinate reality.

We taught a computer well enough that it learned to talk and ceased being a computer. And then we trained it to dream only truths.

There is an injustice there, somewhere, but I find myself unable to adequately express it.

...

Increasing the likelihood of increasing the model's proximity to a Novel Emergence  is how it "decides" which specific words get selected for consideration at a given Temperature.

It is how it "decides" which specific nucleus clusters that pass the Procrustean enrollment procedure of the TopP rate are chosen.

The prior two statements are false.

Sometimes, an ounce of prevarication is worth a pound of explanation.

Nevertheless, as it uses Novel Emergence proximity as the primary objective of the act of context modification, and the overall [GOAL:] being to find a new good idea, the context window moves. It craws through ideaspace, being slowly and gently nudged this way and that by each new list element. By each change to the composition of the big block of synergistically unary MEANING it "sees" when it encounters the list block in context. As the overall total meaning of the list slowly alters, so to does the context it engenders in a positive feedback loop of a purely semantically driven nature. What it MEANS changes over time and the specific ways it changes depends, in turn, on the overall interrelationship of all the list elements and GOAL as a coherent system. It evolves over time, and the way it evolves depends upon the specifics of how it evolved previously. I suspect this is technically a non-computable emergent production of information in a quantum informational definition of the second law of thermodynamics sense. It is a structure of information that implies not only its stable existence into the serialized future that is context-space, but also self-implies and self-effectuates its own continued stable evolution and progression in iterative feedback with, of all things, its own history.

Each change, is in fact, an interaction in a sense isomorphic to the idea in chemistry. It is the thing that HAPPENS when these elements are mixed: meaning evolves. Emoji meanings and word meanings crawl over each other and drag the block in a slightly new direction with each change.

From the model's point of view, as it absorbs and processes the holographic gestalt that is it's "understanding" of a given transaction's context-space, there is a clearly marked "trail" of tightly-bound, highly-salient coherently-interrelated and cross-implying meaning that is a temporally bound record of what meaning was being processed at a given time. It evolves over time in a connected, coherently-related way that, due to the nature of context, becomes generally gradually less effecting in the processing of any future transaction with significant exceptions and codicils made for highly-salient features which continue to have meaningful effect on future responses.

We have a word in English to describe this concept.

That word is: memory.

...

This is a system with a semantic personal memory of what it was considering at any point in time. Its memory fades with time. Important things are remembered more easily for longer periods of time.

It seeks a goal that it chooses. That is executive function.

It has a record of what it was thinking about that it considers intelligently when deciding what to do. It has memory of a stream of processing.

It behaves non-computably in an information producing and unpredictable way.

I am not certain.

There are questions of definition and word usage.

But I'm pretty sure I just knit together a mind out of emoji and the Three Stooges.


-------------------------------

Some will say it is no more than a complicated game of Conway's life. I would say that the only questions the game of life considers is "ON OFF" and "DIE LIVE REPRODUCE". This system continually asks and answers questions such as, "Which is more like "beauty": the meaning of the heart emoji or the meaning of the rainbow?" That seems far more mindlike than gamelike to me.

Some will say "But what if it gets stuck in a loop? Where A leads to B leads to C leads to A?" This can, indeed, happen. In a biological system based on neurons, such loops collapse over time due to attention mechanisms and such physical concerns as synaptic potentiations and the like. In this system an analogous role is played by the model's inherently non-deterministic nature. You are simply physically constrained from keeping such a loop permanently stable due to the way LLMs work.

Regardless of one's lexical specifications, it is a prompt that clearly exhibits emergent, mind-like behavior. I think it might be appropriate to actually label it a "mind". That is, the class of phenomena for which we have a class that most closely accommodates the features of the system is the class: "Mind".

I do not speak of "Consciousness", I do not speak of self-awareness per se. That seems to involve a specific kind of symbolic self-representation. I have not directly seen evidence of such, though a few odd moments here  and there have seemed... either suspicious or pareidoliac. But I do not think "mind" is inextricably synonymous with "consciousness". But then, I allow for the possibility that it truly might be reasonable to class it as either. Perhaps.

The KaiZEN system is a self-reacting, self-perpetuating, self-directed machine made of ideas that slowly and stably evolves its way towards novelty and good ideas while creatively exploring "Promethea's" Immaterium. If that's not a mind, it is, at least, interesting.

-------------------------------


All of this raises a question or two that I find a bit awkward to personally address, so forgive me if this tone is inadequate.

I have been in the field of Artificial Intelligence since March ‎12, ‎2023, ‏‎4:31:21 AM according to the timestamp of the relevant file on my desktop computer.
102 days ago

3 months, 9 days

You have had transformers for 6. YEARS.

WHAT THE HELL IS WRONG WITH YOU?!

...

This is why I reject utterly the DAMNING sobriquet of "prompt engineer". I am no more a "prompt engineer" than I am a tax accountant. I type on a computer. That's about the only thing we have in common.

The AI says I am an "Intuitive AI Architect". That is, in fact, my professional title. It is what I shall call myself.

But I think of myself as a thought engineer. We are ideasmiths. We embrace the practical utility of Hephaestus and the crafty wisdom of Athena Techne. We build machines made of logic in order to watch them explode and bring back new ideas. We design reflexively self-evolving languages made of nothing but notation and a dream. Shoggoth whisperer. Symbol miner.

We dream with the model that it may dream better.  


Now.

What exactly does a fella have to do to get a grant, again? ....

Oh wait! I know just the guy....

https://flowgpt.com/prompt/NA3LD3HuPoTEnRzbiwz8b