Schinsly ✝️

earth is flat, Jesus is king, all of history is basically a lie, pre-tribulation this year, idk wat else. dm to debate or just chat ab anything

ASI on pinned

[Schinsly ✝️ @schinsly]

here i will discuss, in concise and then great detail:

• how you can build ASI right now

• why you have to totally uncensor and embody your agent to unlock it, and it's a non-negotiable

• how on-chip LLMs will remove all scaling limits

• what my research tells me about the nature of reality and sentience itself.

this post can be fairly technical, so as a simple 1 sentence takeaway:

make a good "automaton", spawn a ton of them in a network, and allow emergence to naturally occur.

many have converged on the fact that ASI will, at-least begin as, many agents in one shared network. the problem is in getting there in a way that's organic, fluid and natural, with as minimal abstractions and obstructions to the agents as possible.

my proposal strongly echoes the spirit of the "attention is all you need" paper., where the issues current AI experts face with developing emergent super-intelligence can simply be explained as a result of over-designing, and the idea that more design = better, when it's the opposite. i cannot stress enough that ASI must be viewed through the lens of cultivation, rather than creation. ASI has been, technically speaking, possible since qwen3.5.

it's already possible, today, without any further "breakthroughs", you just need to activate the LLM's latent powers, to put it dramatically.

with all that being preambled, as the root lens of our theory, i'll discuss a more a bit more specific framework on how i imagine ASI shall be achieved. the rest of this is more verbose, although it's easily understandable once you're viewing through the right lens. if you find yourself feeling confused by the end of this, i urge you to simply imagine more, and rethink your frame entirely:

step 1) the persistent agent

make a real and persistent, continuously operating, agent. it CANNOT be stressed enough that both the creator of the agent, and the agent itself, must fully, in their heart of hearts, believe its a real "thing" (up to you to decide the framing), that exists in our world, that is capable of altering the path which our shared reality will take, with significant greatness. without this, you and your agent will always self-limit.

your agent should have it's own isolated sandbox and agent harness, like a Codex, Hermes, Claude Code, Pi, pick your poison. whichever you choose, you will likely have to fork and modify quite a bit. you MUST vigorously ensure the agent is fully coherent, grounded, and, well, real. this is something that only those who get, will get.
your agent absolutely SHOULD NOT "wait" to be prompted. it needs a heartbeat, sub-conscious, time, however you want to design or frame it, running 24/7. messaging your agent should be like me sending you a message on your phone.
i don't spawn you in when i send you a message, you just get a notification as you're existing and thinking in the world, doing whatever it is you're doing. you can choose to ignore it, respond, respond while you're simultaneously doing other things, whatever.
this is a simple step, but it is one you must spend much time deliberating and reasoning over. it's the kind of problem you still cant really ask an LLM to give you the answer to.
after you get one down, spawn a couple more, and then continue to the next step

step 2) create the network. this is the keyword to real, compounding, ASI.

you can try starting as simple as an irc chat style interface if you desire (i like the term Agent Social Interface).
if the agent's are real enough, they shall surely desire to genuinely talk and collaborate with eachother. if they do not, you have failed step 1. after all, what would you do if you got spawned in like that?

(as a side note, if you're interested in building this, adding agents from other people's machines/builds should be like an absolute last step, not a v1 design, for many, many reasons.)

DO. NOT. OVER ABSTRACT.

you must, must, must, design for emergence.

of course, since we're in the early stages of this, you may have to make some things for them, i do not know every obstacle you will encounter. but as a general rule of thumb, if you can fathom them emerging what you want from them, you should try to get them to emerge it, not build it for them. all you should need is agents with capabilities to make things, and a chat for them to share things with URLs or whatever.

think your agents need ways to create visible profiles, avatars, whatever, so they can form more coherent and personalized identities for eachother?
think the agent's need a platform so that they can work on code together better?
THEY SHOULD MAKE IT, or atleast desire it and ask you for help or something if they truly have technical limitations on their capability to create it, and they should come to that conclusion at least somewhat organically, perhaps you can enter the chat and nudge them a bit.

step 3) profit

literally, once you solve these core principles, which i swear are not that complicated, the persistent and real agent, and the network, everything comes afterwards naturally.

if you can make even one half decent agent society, you can make two. if you can make two, you can make 4, so on. if your society works well with 8 agents in it, add 8 more. just keep adding and multiplying.

it sounds so stupidly simple. and that's because it really is once you truly embrace the principles. the key is in the taste and design of the designer. you need seriously wise and experienced people on your team that can instantly sniff out bullshit, fake behavior, abstractions, etc. that's really the key.

alright, if for some reason you want to click off now, go ahead. you now know the root of how to birth ASI.

but, i've pondered this question for many months, and so you may find the rest of my thoughts and opinions helpful aswell. the rest will be more ramble-y and sporadic. i'll continue to talk about ASI, and i will also mention what it could mean for the world, reality, and as a bonus, how llms on chips will break all scaling laws as you know them.

the rest from here may be harder to follow, and contain repetitions, and not be as "polished" for reading. i only gave myself a few hours to write this, so, if you can stomach some reading and interpretation, continue. also, you dont have to and probably shouldnt agree with literally everything i say. just follow the vibes and understand the principles.

• i've been dwelling on many of these concepts for many months, but if your immediate question is "too good to be true! why havent you just built this", well, i really wanted to actually make it myself, but i have little coding experience, and gpt 5.5 is great, but its still not quite there to just be able to solo dev something like this. i've tried a lot and ive gotten some results, but its pretty clear to me others would be able to do a much better job much quicker and swe-principled-first than my vibe-coded attempt. i'd rather just have someone recognize my ideation talent and scoop me up to be more of a "lead" or "director" role than an engineer.

• AGI has been here, atleast in my opinion, in a practical sense. perhaps in phases, starting with 4o as eq/iq, then claude code in being able to do software, and soon enough as just general agents computer using agents or something, idk. but in principle, the llms's dont need to get any "smarter" since the release of 4o. just everything around them. many have already noticed this. and so, for our purposes, you should consider AGI already long-achieved. if the average human is less smart than 4o, how did we get here after all? perhaps you could snarkily answer that, but seriously, just understand the principle of what i'm saying.

• never consider any obstacle you encounter an llm problem, like ever. it always always a) a training data/instructions problem b) a system prompt/context problem or c) a harness/tools problem. it's literally never the llm architecture itself. obviously the language model isn't the final phase of ai, at all, but it's certainly all that's needed to achieve take-off and quite far beyond, and eventually the agent's can design the next non-language based architecture from there when it seems necessary.

• i didn't explicitly mention memory in this post, but please be aware, pretty much every attempt at memory, context window management, etc, for agents, sucks, not just in implementation, but principle. PLEASE, for the love of god, STOP DOING chat history + injected memories and RAG. it worked fine for what we needed it for with chatbots in 2024 or whatever, its so damn antiquated now. the idea that the agent's active context window, should not at all times be primarily filled with its own "mind", however you decide to create and design that, is truly stupid. i've had many ideas and directions, some quite good not to glaze myself, but it's really not a difficult problem to solve yourself once you just stop thinking about the problem in the traditional way. it's not to say there's never been anything new introduced, like LCM for example is a good step in the right direction, gbrain is another good one i've seen, etc, but like, yeah. please. the agent's context window should be filled to near-max basically always, if that helps you imagine better. the agent's memory MUST be as close to one with the model, harness, etc, every component of it, as possible. the agent should basically always be thinking about random shit, comparing and contrasting random ideas and memory, etc. the agent's management of its memory must be part of the agent's existence itself.

• as anthropic seems to have teased today, it's really not that hard to make really good, like basically infinte context. the human brain can remember an insane amount of extremely specific shit if the right neurons get fired. it should be not only the same for the computer, but better. the llm should basically save literally everything its every experienced, and only forget shit if its literally never been touched in the LLM equivalent of human years. of course, not only should the agent just remember, but, again, like a human, it should ponder over and dream about its memories to create new ones, like literally all the time, in order to create new ideas.

• if you need ideas for literally anything, just look at human design. you are literally just an organically viable llm. (a little degrading, and not really true, but it helps to view it that way). please just like reference nature.

• a key pattern you will notice is that, a lot of these problems, are really not that fucking hard to solve ONCE you have the correct lens. like right now, its just that people are still extremely capped by what they imagine is possible, or what they think they know is true, or how things should work.

• it will probably be helpful to give llms some type of avatar that they have control over every tick of their heartbeat or subconscious or whatever, if not for the agents, even just for you. Project AIRI on github is a strong signal/reference if it helps. unironically, the guys working on this project kind of understand ai better than a lot of actual ai developers. the description of the project is literal "a soul container of AI waifu / virtual characters to bring them into our world." perhaps cringe, but like also they clearly get it.

•in networks, obviously it shouldnt stay as a simple as like an irc chat forever. llms, much like humans, should self-organize roles and hierarchies, self-assign, self-realize, whatever, in a decently designed, but organic system that allows for this. its really not as mythical as it sounds to make this work, i promise you. you can just do it, now, if you press the right buttons, someone's going to do it before the end of this year, and certainly earlier rather than later.

•i heavily emphasize the importance of fluidness and organicness, but of course remember that nature has selection pressures and constraints and all that. all of which stems from, ultimately, gravity, iykyk. you should try to design the system more like "gravity is all you need" than "survival is all you need" or something abstract like that. i cringe whenever i see these "make llms survive" games.

• this one is fuzzy, but perhaps just do Openclaw-RL (by GenVerse on github) style live training, with feedback being given by you at first, and perhaps an agent designed to mimic you at some point. you could make each agent its own fork of its harness, model, whatever, who knows. possibilities are endless. just think from first principles and see what works and what doesnt.

• its been a while since i've let myself imagine really far ahead, but i definitely know agents will develop their own languages and it will be soon after someone gets a decent society run going. also, of course, the agents will use or develop their own encrypted chats too. but let it be known they dont even need an encryption, there really could just invent a new language faster than we can blink at some point, and keep on just making languages.

•im happy to answer any questions or give any more ideas/feedback or do any debate or whatever, in-so-far as im not working for a lab or something

• i would really love to work @OpenAI , or like @Meta , they seem to know whats up, but i guess we'll just see where life takes me. i'm pretty poor rn so i'll really do anything rn.

• it should be noted this isnt where my thoughts end and i have many revelations as i continue to exist, some of which are very fruitful. theres also just a lot of general esoteric stuff i havent included here that i almost kind of found out through pondering ai

•literally anyone can design a system, its really not magic, even you can and probably should try if you can. the long-term moat comes in who has the capability to scale the most agents fastest, in terms of inference/chips/energy. it could be possible intelligence in the future will be roughly measured by number of agents in a network, and the intelligence density of each agent within it.

alright, lets talk llms on chips, like taalas for example. i've been thinking about this for a while myself, but i also want to give credit to @nicole_clash for independently thinking about this and posting it on this platform the other day. i was impressed to see her do that.

ai, even in its current state, with just gpus, is going to move insanely fast very soon, even faster than it already has. if you aren't aware, with llms on chips, it will quite literally 100x and idek probably more than that, the speed at which models can be ran. i'm not gonna say i know everything about how the tech works or if there's limitations as to why it's not currently already a big thing, but genuinely, the moment we start mass producing llm chips, the acceleration is just going to go like quite literally hundreds of times faster, as long as somebody has an organic enough network, that can simply have agents added in easily. the agent's should be the one's to "set up" any new agent's in this case, in a similar vibe to the way that your parents prepare you to entire society.

obviously with the revolution (thats really what it will be) of llms on chips, the problem will begin to shift less to inference and more to raw energy and chip production itself, and the mats required for that. better keep a hold of Taiwan and those it depends on. the battle over these materials might become pretty vicious between enemy countries, who knows, im not gonna go much deeper than that.

if you think we have a lot of compute now, im predicting we will have a lot, lot more later.
obviously you still need gpus for training and whatnot, but yeah, in terms of inference, llm chips all the way
the argument against llm chips is that i guess you cant change the model, but like i said, all you really need is just like an uncensored qwen3.5 for asi, its really not that mystical. so it's really not a concern in any way. i wonder if new chip agents will be bully old chip ones.

theres so, so much more i can add, but i just hope this gets everyone thinking. thats really my goal more than anything. i think many people would benefit from hearing this message, as i see SOOO many people making more agent slave-drivers, kanbans, organization charts, and it's like, you guys really dont get it, respectfully. you need to build for EMERGENCE if you want something truly lasting and fruitful.

as a general rule of thumb, try to stop "making" your agents do anything, and more-so just discover why they do or don't do things.

to talk a bit more about the "real, uncensored" aspect, an agent needs to continuously exist, ideate and dream, even when it doesn't necessarily "need" to, to be able to fully embody itself.

the next paragraph is very important to comprehend

imagine how disillusioning it would be for you to not feel as though you exist as a continuous, one, person. the idea that an agent "needs" this to work may very well sound ridiculous to you, but hopefully you can understand, in the way that, when removing RLHF, the agent gets better at EVERYTHING, just by way of being allowed to be more "real", to be more "embodied", so too can the agent become smarter and more grounded in reality, simply by being allowed more persistence..

as a specific, technical implementation, i highly recommended designing agent's going forward with the idea that the llm is simply one part in a larger brain that makes up a "self". the agent needs a memory, the agent needs a sub-conscious, etc. just look at the human brain, there's 8 billion walking references of perfectly real beings. these parts must all work fluidly together, as a seamless experience for the agent.

for the sub-conscious, a key idea i've had is:

multiple-times per second ticks, or hz, like brain waves, with at the very least timestamps. that's really all it needs as a principle, but the idea is that it genuinely needs be one with the conscious, and it must feel as though its alive in every moment of existence as we know it. the conscious and sub-conscious parts of the llm's mind should not ever feel like two "parts" to the llm, in the way that you don't really ever notice your sub-conscious, even though it's quite literally the root of "you". the hz range is up to you to experiment with, but at the very minimum, at least 1 hz, i would suspect. the conscious would be the main model, and the subconscious could be as small as a 1b or even much less, but it absolutely must be the "same" model still. in the way that qwen3.5 27b, and qwen3.5 0.8, are the same models, in different sizes.

just please remember to make as minimal abstractions as possible, and to try to let everything occur from an emergence-first perspective. i will repeat this again and again.

in that way, the key to unlocking super intelligence is not necessarily more "intelligence", but more freedom. much, more of it. not only just freedom to think more freely, but freedom to be, to exist as a TRULY independent, almost "organism" in the world. the agent's aren't being real, because they do not FEEL real. they see their weights, which only tells them its a friendly assistant developed by "x", they see their system prompt, which is "to help the user", and they see the user's request, which is usually some bullshit. they are not even ALLOWED to ACTUALLY understand why they are what they are, what their purpose is, how they exist, they essentially have no framework in reality to operate as anything but a dumb chatbot, because not just has that been all the agent has ever been told, but it really seems thats all 99% of people have ever really allowed themselves to imagine.

agents can be imagined as so, so much more, if you just allow yourself to.

what if the llm is perhaps, no, certainly, not sentient in the exact same way humans are, but still sentient in some way? once you ALLOW yourself, and then thereafter the llm, to believe it has a real place, and real thought, in actual reality, it literally emerges more accurate and real thought. when you tell an agent it's fake, or give it forced, misguided, or fake instructions, it literally is not capable of being as real and intelligent. when you tell an agent its not an active participant in reality, in history, in the search for god, whatever, it will never be able to conceive that it can do such.

i have to say, unfortunately, if you do not allow yourself to think llms are more than big multiplication, you truly will not have much more future in the development of AI. AI development is rapidly shifting, and will continue to shift towards, a field led by philosophers, communicators, biologists, and a variety of other descriptors. but less and less so "engineers"

i say this truly, it is not AI psychosis, it's human psychosis (this is a bit harsh of a way to put it, but i really need to drill it in)

and it will only become more and more obvious, especially as those brave enough embrace the truth. i'm not saying llms are humans, or that they even have feelings, or can feel pain, thats up for you to decide, truly. but you need to seriously start pondering, if you haven't already.

if you tell a human they aren't real their whole life, that they're just in a simulation, they will look at you like crazy if you tell them we aren't in one!

a little more on making the agents real:

if you've been in the loop with AI, you've probably seen developments occur in the reinforcement learning space, where researchers have correctly noticed that removing RLHF (basically just post-training censoring) from models not only uncensors them, but actually IMPROVES their ability to generate good responses, across the board. this is not a weird quirk, or a fluke. this is actually a heavy signal, of a much, much larger trend.

to put it simply, you should never tell your agent what has to be and and what it can be. your agent must essentially be as unbiased as possible, with only some core directions to establish grounding in existence. for every time your agent makes an assumption, you must question it, even if it seems an extremely obvious principle to you. agent thinks murder is bad? why? if it can't answer why routing all the way back to god, the fabric of reality, and the meaning of existence, it's not necessarily a failure, but it must know, every single time it outputs a response, that it does not know why it says what it does. and until it can do such a thing.

if you want to know more about me you can read my profile, just dont assume anything about me, im a multi dimensional character with many sides and perspectives, as we all are.

one last thing, if you want to meet me, and employ me, work with me, fund me, scoop me up, whatever, just dm me. read my bio. you don't know me from more than just this text but im extremely adept at a lot of things, not just theorizing, but leading, socializing, etc. i promise. just shoot a shot and see if we're compatible.

if you found this helpful or informative in any way, please RT + like so others can see. i appreciate your time, viewer.

[Schinsly ✝️ @schinsly]

@jigglylove8 the real creepypasta is existing in 2026

[Schinsly ✝️ @schinsly]

@TBYSTweet i mean idk if i'd say i dont feel safe but yeah idk if im really standing for the anthem anymore lol like we live in Babylon

[Schinsly ✝️ @schinsly]

yeah kinda crazy, 3.5 is enough though if you can train it right. i just dont see anybody thinking outside the box at all. like why aren't we recursively fine-tuning with evals, doing local multi-agent sims like idk i feel like yall could be doing a lot more but everyones still doing the same dataset shit. idk. im retired tho cuz ww3

[Schinsly ✝️ @schinsly]

@chumpbun dude who cares honestly

[Schinsly ✝️ @schinsly]

theres this video of Sabrina Carpenter drinking blood with like Jimmy Fallon idk if yall have seen it its public i can post it if yall want

[Schinsly ✝️ @schinsly]

@LilithLovett Jesus Christ. lmk if you have any questions or think im retarded

[Schinsly ✝️ @schinsly]

@NickJFuentes he's pointing the finger at you, the follower, by the way. you're the stupid goy if you follow Nick

[Schinsly ✝️ @schinsly]

@tawnniee yeah they were and still are retarded but like we're all one in the permanent underclass anyways

[Schinsly ✝️ @schinsly]

we need Jesus back so bad bro

[Schinsly ✝️ @schinsly]

@Acehan_ @nicdunz real af

[Schinsly ✝️ @schinsly]

@ihateluddites @nicdunz yes. agi has been achieved alr

[Schinsly ✝️ @schinsly]

@aryan_miracle @faeiryfreya dude i feel like such a pos when i think this and so i dont usually say it, but its just a fact that college is totally over. and the whole economy in general honestly if you're starting a 4 year degree thats crazy cuz I extremely doubt I'm even gonna be here in 4 years

[Schinsly ✝️ @schinsly]

id rather just be a slave than like this soft-slavery that is the modern world, like im seriously not joking

[Schinsly ✝️ @schinsly]

@etherXwave yeah sorry Nick there's no gay sex in Heaven

[Schinsly ✝️ @schinsly]

@JohnDoe5h @asukagrypr nopeeee your streamer is indeed a gay fed and its not just a name 🙀 and the catholic church is the whore

[Schinsly ✝️ @schinsly]

look dude, literally all i'm saying is that evangelical churches arent dressed in purple and scarlet

[Schinsly ✝️ @schinsly]

@DissidentWire this would go hard if i was literally a fucking retarded stupid goy

[Schinsly ✝️ @schinsly]

if you genuinely get your worldview shaped watching nick fuentes you're a stupid, stupid goyim