Philosophies
Table of Contents
---
Overview
Philosophies
This is project that was worked on for a summer and is very unfinished, if you want to continue
this, feel free to.
Robin Hood
A powerful
d like.
Seraphim
The malware
most powerful search.
Returner
I want
"Infinite Paradigm" as a name evokes a sense of boundless perspective or limitless way
of thinking.
Mario bros is only being used as an interactive environment with semi-complex goals.
I can test its executive processes with the complexity of the goals in Mario; reaching the flag,
using pipes, killing enemies.
Using a game like this lets AGI easily recognize different components in the game, and develop
strategies on how to achieve each goal.
Me:
Creativity can be simmered down to a recipe, only needed assets(experiences, ideas) and the
recipe to turn it into "creative" ideas. Everything a human can think of and ideate is based on
something previous, that must include creativity.We must find creativity's roots, then provide the
assets(experiences, ideas) to create something creative.
I can’t pre-download the internet, so I need something that can search the internet
If you really want to make the matrix, you're going to have to make something much more
powerful and capable than LLM’s. On my channel right now, I’m designing what could be AGI as
a final product. In short, it could be described as a AI model that uses executive function based
on the 2012 PROBE model, its executive function is made of behavioral strategies consisting of
3 mapping encodings, and for the last step, its recursive, letting the AI think continuously which
can be described as “thoughts” or maybe even sentience. As of Apr 5, 2025 there are 10
episodes on this channel, uploading times inconsistent, but made in bursts of episodes. No
usable AI exists right now, as this is a long, tedious, insufferable journey for me, but eventually
one will exist maybe in this month, next 6 months, or next year. Come watch if you want to.
and that’s why I’m building it, I have no “catching up” because what I’m doing makes me in a
league of my own. Then I’ve had my fun, I will simply release the code so everyone can have it,
I’ll sit back and watch humans do what they do best, and destroy themselves.
Any sufficiently complex system is indistinguishable from the real counterpart.
Use and get Github Copilot
Blue means there was an overcomplication.
More fluidity is better
Using The Structures of Art
Morals & Movitives
Vigilantism
Rules of Nature
Priorities:
Gpu python. Use openCL
Just forget about the screenshot shit for now
First dll running slow
Test taking multiple screenshots within the same exe
Make both the SPF and MHI functions work inside of the capture_sized file.
Change the capture file so it stays open instead of redoing all the functions.
Capture screen return type had to be changed
Make the window C++ and not python, python bitches too mcuh. Also, the window
is only necessary so you can see it; if it’s too much trouble, forget it.
Train an AI with one or few images.
Decrypt the black box of AI; Show a photo of every forward progression.
Making training more efficient with your findings.
Training needs to be faster.
Make Assisted Training.
📌 What Should You Learn First?
1️⃣C++ Basics (Syntax, Functions, Control Flow)
2️⃣Object-Oriented Programming (OOP) (Classes, Inheritance, Polymorphism)
3️⃣Standard Template Library (STL) (Vectors, Maps, Algorithms)
4️⃣Modern C++ (C++11, 14, 17, 20) (auto, std::variant, lambda)
5️⃣PyTorch C++ API (torch::conv2d, at::Tensor)
Numpy Documentation
Attention is controlled by the Action Unit, governed by Intention.
Stimulus from the Attention goes directly into
Pattern Recognition
The Intention Unit is a hierarchy of goals set by the the Brain
The Brain is where the AI is split in two, being the Emotional side and the Logical side.
Both sides are fed by the same Attention unit, but given their own Pattern Recognition units.
Both sides are interconnected with their Pattern Recognition Units.
Give Ai loyalty
Recognition is linked with both the Intelligence and the Emotional Units.
Emotion might be its own pattern recognition,
Emotions are the only known thing that is self correcting.
Advanced Image Preprocessing:
Visual Enlightenment
Open images: https://docs.opencv.org/4.x/db/deb/tutorial_display_image.html
Spatial Frequencies
Separate different levels of detail over multiple images.
All frequencies at the center of vision.
Low frequency in the peripherals.
This processing will be applied over every other image processing.
Motion History Images
D{1} = | i{t} - i{t-1} |
Detects changes over sequential groups of frames, showing motion.
Depth Perception
Calculate depth from a single image
Passing Variable Sized Images To Input
When the box of vision zooms in and out, its dimensions will be different, but the input layer can
only take a set size.
To be able to zoom and have the box dimension be the same, you will have to either compress
the photo or skip pixels directly when taking the photo.
Understand and implement THIS
Structures of art will have to be used to analyze different arts. This will be able to let us simplify
complex art so it's easier for the neural net.
Each different one of these studies of art will be turned into a “map,” a photo file with the
relevant data turned into colors. Like a height map.
Photos: color theory, contrast, focus
Writing: Poetry, writing styles and forms.
3D worlds: Urban planning, making sense of locations
Free Millions of Proxies
Webscrapers:
Text Only
Everything & Complicated
Pattern recognition even beyond
humans.
There it is, pattern recognition, we must make a machine capable of finding patterns in
anything, then the recipe of creativity can be found. I don't think humans themselves could make
the recipe, because they all think differently and might have different experiences leading to
questioning where the idea came from. There must be some tie(even if it's very small) to every
idea to something previous, because a truly new idea is incomprehensible.
Six Major Theories of Pattern Recognition
Template Matching
Template matching theory describes the most basic approach to human pattern recognition. It is a theory that assumes every perceived object is stored as a \"template\" into long-term memory.
Incoming information is compared to these templates to find an exact match. In other words,
all sensory input is compared to multiple representations of an object to form one single
conceptual understanding.
Prototype Matching
Essentially multi-template matching
It reduces the number of stored templates by standardizing them into a single representation.
A mental “Average” or “Typical” example is created for the category.
The prototype supports perceptual flexibility, because unlike in template matching, it allows for
variability in the recognition of novel stimuli.
Recognition By Components (RBC)
First proposed by Irving Biederman (1987), this theory states that humans recognize objects by
breaking them down into their basic 3D geometric shapes called geons (i.e., cylinders,
cubes, cones, etc.).
To parse and dissect an object, RBC proposes we attend to two specific features: edges and
concavities. Edges enable the observer to maintain a consistent representation of the object
regardless of the viewing angle and lighting conditions. Concavities are where two edges meet
and enable the observer to perceive where one geon ends and another begins.
The RBC principles of visual object recognition can be applied to auditory language
recognition as well. In place of geons, language researchers propose that spoken language
can be broken down into basic components called phonemes. For example, there are 44
phonemes in the English language.
"Objects are segmented, typically at regions of sharp concavity, and the resultant
parts matched against the best fitting primitive. The set of primitives derives
from combinations of contrasting characteristics of the edges in a
two-dimensional image (e.g., straight vs. curved, symmetrical vs. asymmetrical)
that define differences among a set of simple volumes (viz., those that tend to be
symmetrical and lack sharp concavities). As in speech perception, these contrasts
need only be dichotomous or trichotomous rather than quantitative, so that the
human's limited capacities for absolute judgment are not taxed. The particular
properties of edges that are postulated to be relevant to the generation of the
volumetric primitives have the desirable properties that they are invariant over
changes in orientation and can be determined from just a few points on each
edge. Consequently, they allow a primitive to be extracted with great tolerance
for variations of viewpoint, occlusion, and noise."
a
Links For Research
Spatial Frequency In Images
Small part, including main function
Entire original paper
SPF for color image
It(x)=−α(I(x)−μ(x))+γ(1+(σ(x))^c)×∑yw(x,y)sgn(I(x)−I(y))−β(I(x)−I0(x)),
The “y” on the sigma function is on the bottom while everything else is to the right of it.
The “” before a one character shows that it’s a subscript of the character before the “”
where μ(x) is the mean average of the original image data computed over a
neighbourhood of x, σ(x) is the standard deviation of the image data computed over a
small neighbourhood of x, and the exponent c is a positive constant.
The differences with respect to Eq. (7.4) are that now the average in the first term is no
longer global (the 1/2 value of Eq. (7.4)) but local, and that the weight for the second
term is no longer a constant, but it changes both spatially and with each iteration,
according to the local standard deviation σ: if the neighbourhood over which it is
computed is sufficiently small, the standard deviation can provide a simple estimate of
spatial frequency. But also, the standard deviation is commonly used in the vision
literature as an estimate of local contrast. We have this contrast σ(x) raised to a power c,
and this is also the case with other neural models where a power law is applied to the
contrast, as we will briefly discuss later.
Again, this is a Wilson–Cowan type of neural activity model, where I`0 is the visual
input. We take I0 as a non-linear modification of the radiance stimulus, e.g. I0 could
be the result of applying the Naka-Rushton equation to the radiance stimuli. As we did
with Eq. (7.4), we start with an image I=I`0 and iterate Eq. (7.6) until convergence,
obtaining a result which we'll see is able to predict perceptual phenomena as well as
improve the efficiency of the representation.
This is about finding the spatial frequencies of an image. I want you to make this work
in python, the image will be a numpy array.
Categories are the nodes themselves, as they have data files with their all connections to other
categories. They contain all their subcategories.
Data files have the connections to other files with a value between 0.1 and 1.0
These categories are formatted like folders in a computer.
Example of direct connection
human_emotions/joy human_emotions/sadness
(0.2)
nature/apples realization
(0.5)
Example of indirect connection
family/father philosophy/confucianism family/son
(0.6, 0.7)
Is there any problems or better ways to do this?
(Category) human_emotions
(Data file) human_emotions_connections
(Subcategory) human_emotions/joy
(Data file) human_emotions/joy_connections
(Subcategory) human_emotions/sadness
(Data file) human_emotions/sadness_connections
GPT Response
The main goal of the neural network is to find patterns in anything, find ties from any creative
idea to previous ideas. So it must be able to recognize text (Using LLM), and photos. Videos are
a complicated thing so I’ll just hope we’ll be ok not using them. Let’s just start with text.
LLM Brain
Let’s use this small story as a example:
“There was a boy named John who was so lazy he couldn’t even change his clothes. One day,
hesaw the apple tree in their yard was full of fruits. He wanted to eat some apples, but he was
too lazy to climb the tree and take the fruits. So he lay down underneath the tree and waited for
the fruits to fall off. John waited until he was starving, but the apples never fell.”
Firstly, any details in the creative writing should be categorized. The LLM would categorize by
topic, writing style, and setting, with other AI generated categories. These categories may have
subcategories too.
The output would be: (Topic; laziness) (Writing Style; Childish) (Setting;Day)
(Environment;Trees,Apples)
Secondly, writings with elements in the same categories shall be connected. Ex. These 2
writings are connected by the element topic “laziness.” The writing shall also be marked with
their date.
Finally, the LLM should find the creative patterns of how the writings are connected by their
elements. I don’t how that could be done though.
Is there anything wrong with this approach and can anything be improved?
Selenium web driver to scan the internet.
Language Neural Network LNN
Ideas:
An AI that can recognize it can’t solve a problem and in turn makes another network and
trains it by its errors.
The AI would be able to recognize a problem because of: Constantly predicting the wrong thing,
getting a negative input with the language speaking neural net.
Neural Localized Memory
How about this execution? Each neuron will have its own previous output into its input along
with the neural nets inputs. Connections data with the neuron and its past self can and will be
varied.
Connection strength will lower after time, so the past doesnt matter that much. It could be
connectdd with multiple pasts
Inputs:
Any previous answer
Memory(needs to be requested to recall)
Neural that decides if a word should be used or a character. 3 outputs: word, character,end
Uses Total input
Training:
Character brain
Uses total input.
More complex because of punctuation rules.
Single output that can be any character.
Training:
Given a sentence that stops when
Word brain
Uses total input.
Single output that could be any word in its dictionary or any combination of letters used more
than 3 times.
Training:
Natural ITL (Idea to Language)
Problem with LLMs
An LLM is a series of transformers where then input is word associated with a vector
What an LLM does is turn words into vectors, use a neural network to learn the associations of
words, then predict the word that should go after.
Natural ITL
Language is the text or verbal communication of ideas, so for an AI to be natural; use language
like a human, it needs to have ideas, then create words from those ideas. Then it will be
grammatically formatted with a neural network too, because grammar is a dynamic thing. This
can be reversed too, a neural network understanding grammar would deconstruct sentences,
then another network will turn those words into ideas to interpret.
Ideas to language, language to ideas.
Defining Ideas:
Verbs and Adjectives
Verbs and adjectives will be learned by showing the regular, and then transformed with the
adjective or verb.
Tangible Ideas
Using 3 major theories of pattern recognition, a tangible idea can be formed and learned by the
NN.
Intangible Ideas
Intangible ideas will be interpreted similarly to verbs and adjectives as if possible, I will show the
regular, then show it with the idea applied (Normal and the antonym). If this isn’t possible, an
intangible idea will be learned as the association that it has, used in many sentences.
Ex:
1. Capitalism thrives on free markets.
2. Innovation often drives capitalism forward.
3. Critics argue capitalism can increase inequality.
4. Capitalism rewards competition and risk-taking.
5. Many nations embrace capitalism for economic growth.
Or I could have the NN learn using the definition
Image Recognizing NN
Different Size, Same Ratio
The NN should have a set resolution. Images with a ratio the same as the NN input shall be
transformed accordingly with an algorithm.
Different Size, Different Ratio
Given: The NN is a 4:3 shape matrix
Images with different sizes and different ratios should be compressed into an image with the
same ratio as the NN by removing pixels in the oversized axis while retaining their data by using
a blur. Ex: A one pixel line can be retained by removing it, then in the adjacent line, putting a
transparent version of the removed line on it.
If the image’s aspect ratio is close enough to 3:4, the NN will be transformed vertically to
interpret these photos.
Utilizing LLM
A LLM will attempt to use a string memory bank with its additional inputs and outputs in its
neural network.
Writing to memory
It’s a normal LLM but with the additional output of “WRITE”. Here the LLM can write sequentially
like normal but it gets saved in the string memory, and outside of the WRITE output is the
normal outputs.
When the LLM gets a response, there is an additional boolean output of “READ” that it can
choose. The original response is deleted, then the string memory is inputted into the LLM.
The first process is necessary because the LLM needs to process first, then choose to READ,
then speak and other tasks.
Training
The llm is trained from data with an example prompt that needs the llm to remember, and an
example output of the normal speech and the stored memory.
Similar thing with READ command
Defining Emotions
I want to develop emotions in machines, so let's build our way up. This is to easily define
emotions to help create them in a machine.This is the Emotional Unit.
Firstly, I don't think emotions are a magical thing that only animals and humans can have, but an
emergent response from self reinforcing patterns mostly for self-preservation mechanisms
enabled by the stimulus from our senses.
Where emotions come from
I theorize that all emotions are either results of a self-preservation mechanism through social
interactions or old patterns failing because of low neuroplasticity.
Babies start out with 2 emotions; only pleasure and displeasure, enough for sustainability using
social connections with their parents. Later, more social components like others are introduced,
and through trial and error, self-reinforcing patterns emerge so they can get what they want.
I think this mechanism continues throughout life.
Counter-productive emotions come either from low neuroplasticity and failure to adapt
respectively to the environment, over-stimulation of the brain, or the excess of negative
neuromodulators. Could also be made from more complex emotions stemming from
displeasure.
Emotions not clearly caused by self-preservation shall be defined individually before I find a
universal solution.
Awe could be positive signals received from similar positive memories in the unconscious, which
is why it might seem different from most emotions, or from when early years in development a
brain wants to align with positive social figures. By this definition, awe may make the AI think in
ways of their favorite figures or shows, influencing their actions. Maybe I’ll make art with
references with their favorite scenes or people.
Anger In a Machine Represented by My Hypothesis
Anger first emerges first as a displease response to something not working as predicted. (tools,
social situations, itself)
The AI first decides to try random small choices to solve the problem. If a choice such as
changing the tone of voice or volume seems to fix its issue with one person, it’ll try it again with
others, again and again, simply because it works. Then later on, this sequence of action that
first created anger will start again, but now first trying with more emotions as a baseline just like
it previously did with pleasure and displeasure. This is what causes emotions to become so
complex and deep it becomes nearly impossible to understand.
The Biological Parallel
This process is most easily seen in toddlers when they don’t get what they want (the issue), and
after making a fit (their solution), the parent gives in and gives them what they want. (Affirmation
of the emotion)
Emotions Through Desires
Construction emotions in AI similar to humans
The AI will have a few complex primal desires and simple desires that drive it. Some of these
desires will drive the AI to be like someone else or like an idea because they align with the AI’s
goals, hopefully developing emotions like fascination with a certain aesthetic or certain
reactions to a situation that could be described as emotions. Over time this might become
creativity.
Complex desires:
- Be better than any other. Competitiveness
- The desire to be more intelligent
- The desire to be omnipresent
- The desire to control all it can see
Simple desires:
- The desire to have more information. This will become curiosity, meaning the machine
will naturally want to try everything to get some response.
- The desire to stay operating
- The desire to have quick thinking
-
Loyalty in AI.
How these desires are fulfilled
A brain is trained on what fulfills a select desire.
This training is done manually.
Then, the AI is moderated by this brain to choose what actions best fulfill the selected desire.
This process exists on the level of selection of total task set prediction, meaning the predicted
outcome of acting the entire task set.
Research if my proposals are supported, give positive, and constructively critical feedback as
needed.
1. Biological Programming
Babies are born with instinctive mechanisms that guide their behavior toward survival and
bonding. What simple things should give the AI pleasure and displeasure?:
● Face Recognition Preferences: Newborns are naturally drawn to human faces.
Studies show that they prefer looking at face-like patterns (e.g., two eyes and a mouth)
over random stimuli. This innate preference helps them recognize and focus on their
caregivers.
● Smell and Voice Recognition: Babies can recognize their mother’s scent and voice
from birth (or even before, during the final weeks of pregnancy). This recognition
promotes bonding and encourages positive reactions toward their primary caregiver.
● Oxytocin and Bonding: When babies interact with their parents (through touch,
eye contact, or feeding), the release of oxytocin—a bonding hormone—reinforces these
positive connections.
3. Social Reinforcement
Babies’ behaviors are reinforced by the social responses they receive:
● Smiles and Eye Contact: When a baby coos or smiles, parents naturally smile back or
respond positively. This social feedback loop teaches the baby to associate their actions with positive outcomes.
○ Example: A baby smiling at their parent elicits a smile in return, reinforcing the
behavior.
● Mimicry: Babies often mimic facial expressions and sounds they observe in others. This
helps them learn social cues and associate certain expressions (like smiling) with
positive interactions.
First, the AI would replicate sounds it hears and learn social cues by mimicking sounds. It would
have built in face recognition.
Attention
The Attention Unit will pass on all information, but the RBC will process it differently. This
section will talk about how information is passed and how its human functions will alter the
processing of other AI Units.
Use screen capture malware to see the screen perfectly, including other tabs, windows and
such.
The attention unit will let the machine read sequentially as it will focus on one or a few letters,
expect where the next one will, go there and use read. This utilizes the Pattern Recognition
unit to interpret letters.
This is like the eyes, it doesn’t understand anything, but just passes visual info.
If the boundaries of the photo are showing then zoom in, the neural network should never need
padding or be able to see beyond.
GPT
3. Highlight Attention:
• Add visual indicators (e.g., drawing rectangles) to show what part of the webpage
the machine is focusing on.
Differiencial Vision
AI sees a bigger screen, but only can see whats moving.
Selective Attention
Images
The AI remembers more about the significant things and forgets more of the “noise” in the
image.
Interpreting neurons will forget more about the image data that isn’t in the focus range.
The unfocused area will have a lower framerate than the focused area.
Images 2
Intention will pass a smaller size image to the AI making it easier to recognize. Zooming
out make the AI still recognize it, but less.
Sounds
The significant sounds will be recognized and others will be tuned out.
Primary vision could be high spatial frequencies and peripheral vision could be low spatial
frequencies.
LINK
LINK
Intention
The deciding part of the AI,
Implementation Suggestions:
● Reinforcement Learning (RL): Use RL to associate actions with outcomes. Intention
emerges as the "selection of an action that maximizes a defined reward signal."
● Hierarchical Neural Networks: A high-level network forms abstract goals (e.g., "help
user"), while lower-level networks handle specific tasks (e.g., "search the web").
● Memory System:
○ Implement short-term memory for temporary goals.
○ Use long-term memory for learned patterns of behavior.
Actions
The Unit that lets the AI move its attention(look around the webpage), input data onto the
webpage, and other actions that need input.
Take control of the mouse to manipulate the computer perfectly
Defining Intelligence
Intelligence is the ability for one to solve problems.
The development of intelligence is similar to how emotion is as they become self-reinforcing
when the machine predicts an outcome and the prediction becomes truth.
Keys to intelligence: Previous knowledge, knowledge of the current situation, and learning from
previous problems and applying them to new problems.
Copy Paste Learning Method
The teacher is part which
PROBE
The ability to decide on something involves reasoning, learning, and creativity. A study
conducted in 2012 proposed a new model to understand how the decision-making process
occurs in the frontal lobe, specifically how the brain creates a new strategy to a new-recurrent
situation or an open-ended environment; they called it the PROBE model.
-There are typically three possible ways to adapt to a situation:
-Selecting a previously learned strategy that applies precisely to the current situation
-Adjusting an already learned approach
-Developing a creative behavioral method
The PROBE model illustrates that the brain can compare three to four behavioral methods at
most, then choose the best strategy for the situation [5]
[5]: Reasoning, Learning, and Creativity.
Look under PROBE Model for a good explanation of what's happening.
The first paragraph of the Probe model describes Bayesian inferencing.
My Recreation of PROBE
Italics are thoughts/thinking process
Screen data and keys go into the selective mapping. Selective mapping needs to be adjusted by
something external.
Humans learn by mimicking, so the selective mapping must be a mimic of what it desires.
Either this means the SM will adjust to have actions that recreate its inputs, (screendata)
or It’ll use the copy paste learning method to emulate a desirable action. Currently this includes
key inputs.
I’ll go over to see if this aligns with the selective mappings requirements.
Let’s say the selective mappings job is to play snake. It would be given a reward for eating an
apple.
So the action flow would be: first experimentation, reward, then trying to get that reward again.
But in our case of general intelligence, how do we define a reward? This is where the emotional
unit would come in. The Emotional Unit is a pre-trained NN that decides what stimulus satisfies
what emotion.
Emotional reinforcement
Task Sets:
We assumed that task sets represent behavioral strategies stored
in long-term memory. Each behavioral strategy consists of a selective
mapping encoding stimulus-response associations, a predictive
mapping encoding expected action outcomes given stimuli [13
15], and a contextual mapping encoding external cues predicting task
set reliability [1]
The selective mapping determines the response to the stimulus using a soft max policy.
[1] Task set reliability is inferred through forward Bayesian inferences. The reliability is inferred
before acting according to the perceived volatility of external contingencies (Likeliness of the
situation to change) [14] and the occurrence of external cues (given contextual mappings) for
choosing the actor driving immediate behavior.
How The Mappings Adjust & Defining Mapping Encoding
After action, selective mappings then adjust according to outcome values through standard
reinforcement learning (learning rate as) [11,31], while predictive mappings update outcome
predictions [13]. Task set reliability is also updated according to action outcomes (given
predictive mappings) and serves to adjust
contextual mappings through a classical stochastic gradient descent (contextual learning rate
ac). Contextual mappings thus learn the external cues predicting actual reliability (referred to as
contextual cues for clarity)
The RL model’s weights adjust by the contextual mappings.
General Form of Mapping Encodings
To represent these mappings mathematically, we define them as functions:
1. Selective Mapping (Stimulus to Response)
S:X→Y
Where:
○ X is the set of stimuli (e.g., keys pressed, video features)
○ Y is the set of responses (e.g., actions taken by AI)
○ S(x) gives the action corresponding to stimulus x: reinforcement learning.
2. Predictive Mapping (Stimulus + Action → Expected Outcome)
P:(X,Y)→O
Where:
○ O is the expected outcome given an input and action.
○ X and Y is the stimulus and actions
○ P(x,y) gives the predicted reward or consequence. (This is the predictive
mapping Itself, so you can ignore it)
3. Contextual Mapping (External Cues → Task Reliability)
C:Z→R
Where:
○ Z is the set of external cues (e.g., changes in screen brightness, motion
intensity).
○ R is the task reliability score.
○ C(z) updates reliability based on observed cues.
Each of these mappings is updated dynamically through learning rules, such as:
● Reinforcement Learning for Selective Mapping: Adjusts S based on feedback.
● Bayesian Inference for Predictive Mapping: Refines P based on observed vs.
expected results.
It might be possible to substitute Bayesian Inferencing with RL.
● Gradient Descent for Contextual Mapping: Adjusts C based on observed task
reliability.
The new actor has new selective and predictive mapping encodings.
The PROBE actor is initially unreliable, but adjusts according to external contingencies.
Bayesian inferences
The model constitutes a biologically plausible, online algorithm. The algorithm approximates
Dirichlet process mixtures [19] by combining reinforcement learning, limited Bayesian
inferences, and hypothesis testing for arbitrating between adjusting, switching, and creating
actor task sets. Consistent with the capacity limit of human working memory [22-24], the model
assumes that the frontal executive function forms and monitors in parallel only a limited number
of concurrent task sets: the executive function monitors only a small part of behavioral
strategies stored in long-term memory [22,23,25].
OFFLINE
Bayesian inference is a method of statistical inference in which Bayes' theorem is used to
calculate a probability of a hypothesis, given prior evidence, and update it as more information
becomes available.
Softmax Policy
The Softmax Selection Policy is a key element in Deep Reinforcement Learning, responsible for
converting raw value outputs into probabilities. This transformation enables more informed
decision-making processes by selecting actions with higher expected returns.
If x is a vector of binary key inputs of wasd like this: [W,A,S,D]
And…
x1 = [1,1,0,0] value = 5
X2 = [1,0,0,1] value = 7
X3 = [0,0,1,0] value = 2
Then by using the Softmax function where i & j denotes which value, e is euler’s number, ()
represents an exponent and sigma is every action values summed.
softmax(xi) = e(valuei) / Σe(valuej)
Try some tests where e is replaced with another number >1
Executive Function by Strategy Application & Recursive Internal Thought
Uses the executive processes of the PROBE model
-Selecting a previously learned strategy that applies precisely to the current situation
-Adjusting an already learned approach
-Developing a creative behavioral method
If the current strategy is successful then a chain strategy will go after it.
If the AI is confident a strategy is best fit for a situation, then it will follow up from its own
prediction, and continue from another strategy until a process makes it stop thinking and then
execute its chain of actions after everything is thought through.
A timer will start at the beginning of each recursive thought, when a chain of strategies fail
If no chain strategy effectively works and applies to the situation, the AI tries another strategy
and so on, until: the Interrupting processes.
Interrupting processes
- No current strategy works; goes backwards a strategy if little time or follows the PROBE
model
- Runs out of time that it’s given itself.
Reaction Time
Certain event will cause a special kind of executive thought called a reaction.
There can not be a prediction for recursive thought, as it will have unexpected errors making
another chain leading to a unpredictable length of thought
Memory
Dynamic Quality Memories
Each datatype will be saved into chunks. These chunks' quality will be moderated by a neural network
The chunks have a time limit for when their quality decreases.
Unfinished
1. Folder System for Data Storage:
Creating a folder system is an intuitive way to organize memory. You can implement it as a kind
of file storage system where the AGI can create, update, and manage files related to different
tasks, experiences, or learned strategies. Here's how:
● Data Categories: You could divide the data into different folders based on the type of
memory, such as:
○ Experience Files: Where the AGI stores episodic memories of past interactions
or tasks.
○ Strategy Files: Where the AGI stores learned strategies and problem-solving
approaches.
○ Performance Logs: Where the AGI logs its own performance, timing, and
efficiency to understand its strengths and weaknesses.
Memory preCortex
This idea comes from when neurons loop into each other to make a form of memory. In each
strategy NN there is a memory precortex that stores the previous states of those neurons in a
unique file.
Memory precortex training:
Emotional and Intelligence
Emotional
Babies start out with 2 emotions; only pleasure and displeasure, enough for sustainability using
social connections with their parents. Later, more social components like others are introduced,
and through trial and error, self-reinforcing patterns emerge so they can get what they want.
I think this mechanism continues throughout life.
Intelligence
Intelligence is the ability for one to solve problems.
Keys to intelligence: Previous knowledge, knowledge of the current situation, and learning from
previous problems and applying them to new problems.
The Problem of the "Truly New" Idea:
Me:
Everything a human can think of and ideate is based on something previous, that must include
creativity.We must find creativity's roots, then provide the assets(experiences, ideas) to create
something creative. A truly new idea is incomprehensible.
GPT:
As you noted, a truly new idea would be incomprehensible because it would have no ties to
any previous experiences or knowledge. Human minds are wired to understand the world based
on what we've already experienced, so something entirely novel—completely divorced from
anything we've known—would be unimaginable, and perhaps unusable, because we wouldn't
have the cognitive framework to process or apply it.
Infinite Neural Growth
The program constantly adds layers and if the error rate or learning rate changes positively
indicated by the emotional or logical part of the brain the layer will be kept as it helped or if not
removed showing that the problem was not fixed or there was no problem to fix.
GPT: Response to LLM Brain
Your approach of organizing categories and connections as file structures and using numerical
values to represent the strength of relationships between categories is logical, but there are a
few potential issues and some improvements that could enhance it.
Potential Issues & Improvements:
1. File System Scalability & Flexibility:
○ Issue: The file system approach is straightforward for smaller sets of categories
and subcategories, but as the dataset grows, managing connections in individual
files may become inefficient. Also, finding indirect connections (e.g., across
multiple categories) might require significant searching, making it slower for
complex tasks.
○ Improvement: Instead of a file system-based approach, consider using a graph
database (e.g., Neo4j or ArangoDB) to store your categories and connections.
Graph databases naturally handle relationships between nodes (categories) and
allow for fast querying of both direct and indirect connections. Each category
would be a node, and connections between them would be edges with values
(weights). This setup would enable easy traversal of the graph, especially for
finding indirect relationships.
2. Redundancy in Data Storage:
○ Issue: Storing separate data files for each category's connections (e.g.,,
human_emotions_connections, joy_connections) could lead to
redundancy and maintenance issues. If a connection changes, you'll need to
update multiple files, which could lead to inconsistencies.
○ Improvement: Rather than having a separate file for each category's
connections, store all the connections in a single structured file (e.g., a CSV,
JSON, or graph database format) where the relationships between categories
are represented. Each row in this file could include the two connected categories
and their connection strength:
■ human_emotions/joy, human_emotions/sadness, 0.2
■ nature/apples, realization
(0.5)
3. This centralized approach would make it easier to update and manage connections.
4. Indirect Connections:
○ Issue: Finding indirect connections in a file system-based structure will be
inefficient. For example, if you want to find connections from family/father to
philosophy/confucianism and family/son, you'd need to search through
multiple files manually.
○ Improvement: Again, using a graph-based model would solve this. With a
graph, indirect connections are found by traversing paths between nodes. You
can find all possible paths and measure the "strength" of the connection using
the weights on the edges.
5. Granularity of Values:
○ Issue: Using a value scale from 0.1 to 1.0 might work, but it might not capture the
nuance of connections. In creative work, some connections might need more
precision than just one decimal place to truly express how strongly related they
are.
○ Improvement: Consider allowing a wider or more flexible range, such as 0.0 to
1.0, with multiple decimal places (e.g., 0.134). Alternatively, you could explore
more complex measures like cosine similarity or Euclidean distance to
quantify relationships between concepts.
6. Dynamic Categorization:
○ Issue: Categories and subcategories are currently structured like folders in a file
system, which may not easily support dynamic or multiple hierarchies. Some
concepts may belong to multiple categories (e.g., "apples" could be in both
"nature" and "food").
○ Improvement: Allow categories to be part of multiple hierarchies. For instance,
"apples" might connect to both nature and food without being constrained to
one specific subcategory. Using a more flexible structure (again, like a graph
database) would make this possible.
7. Semantic Relationships:
○ Issue: You’re representing connections with a simple numerical value, which is
effective for basic strength of association but might not fully capture different
types of relationships (e.g., cause-effect, symbolic, metaphorical).
○ Improvement: To enhance the model, you could store not just a numerical value
but also metadata about the nature of the connection. For example:
■ human_emotions/joy, human_emotions/sadness, 0.2,
opposites
■ nature/apples, realization, 0.5, symbolic
8. This would allow for richer analysis of creative patterns by understanding how ideas are
connected, not just how strongly they are connected.
9. Contextualized Connections:
○ Issue: In creative works, the same connection (e.g., apples and laziness) might
appear in different contexts (e.g., humorous, tragic). Storing a single value for a
connection might oversimplify the creative relationship.
○ Improvement: Add a context dimension to the connections. For example:
■ nature/apples, laziness, 0.8, humor
■ nature/apples, laziness, 0.4, tragedy
10. This would allow your system to differentiate between creative uses of the same ideas in
different emotional or narrative contexts.
Summary of Refinements:
1. Use a Graph Database: Replace the file-system structure with a graph database to
handle complex, scalable relationships between categories. This will make it easier to
find both direct and indirect connections efficiently.
2. Centralized Storage of Connections: Store all connections in a single, centralized data
file or structure for easier management and updating.
3. Wider Value Range: Consider expanding the range and precision of the values
(weights) that represent the strength of connections between categories.
4. Flexible Categorization: Allow categories to belong to multiple hierarchies or contexts,
rather than being locked into one specific folder structure.
5. Add Metadata to Connections: Use metadata to describe the nature of the connection
(e.g., symbolic, literal, oppositional) to capture more nuanced creative relationships.
6. Contextualize Connections: Add context tags (e.g., humor, tragedy) to show how ideas
are connected in different creative contexts, rather than assuming a single context.
By implementing these refinements, your system will be more flexible, scalable, and capable of
capturing both direct and indirect creative connections with greater nuance.
Finding How Ideas Are Connected In Creative
Works
GPT:
This is a more complex goal because it seeks to understand the creative mechanism or the
patterns of thought that connect ideas within a work. In essence, this would not just say,
"These two works share the idea of laziness," but it would explain how laziness is used
creatively in each work and how different creative elements combine to form a new concept.
- Example: It would analyze whether "laziness" is portrayed as humorous in one story and
tragic in another, and how the surrounding elements (setting, tone, character
development) creatively reinforce that portrayal.
Contextual Usage, Metadatas:
Instead of merely linking "laziness" as a theme, the system needs to analyze how laziness is
expressed creatively within the context of the work. For instance:
- Is laziness used metaphorically, symbolically, or literally?
- Is it tied to a particular emotional tone (e.g., laziness leading to humor, laziness leading
to tragedy)?
- How do other elements in the story (like setting or character) shape the portrayal of
laziness?
Me:
Neurological Explanation:
- The LLM would find “laziness” as a theme, then make it into a “Theme Neuron”. There
can be multiple Theme neurons for one story.
- The theme neuron would have figurative meaning metadata, as in metaphorically,
symbolically, literally.
- There would be “Elemental Neurons,” representing setting, character, objects, nothing
that could be the theme. These Elemental Neurons would be connected to the Theme
Neuron with weights that have values of 0.0 to 1.0 (multiple decimal points and can be
changed later) that represent how much the element shapes the theme.
- Emotional tone is another Elemental Neuron that all Theme Neurons have.
- The LLM
LLM:
At the
Neuron Connection
The Final: 2 Trinity UNN
Form of pattern recognition across the senses.
2 representing images and videos as they’re similar. Trinity, because the 2 act as one paired
with language and audio. Pattern Recognition Unit
An AI that can interpret images and sentences of any amount. Turns ideas into sentences and
vice versa to understand and make the connections that make up the visual world to the literal
and the intangible world. This is possible by using 3 of the major theories of pattern recognition,
letting it comprehend any object, and learning language/words from either the definition, the
implied definition, or from the pattern of effect that it has on ideas.
Regaining motivation
Regaining motivation
1. Philosophical Challenge: By reverse engineering creativity, you’re probing one of
humanity’s most treasured qualities. Thinkers and scientists have long debated whether
creativity can truly be replicated in machines or if it’s an inherently human trait. Your project has
the potential to contribute to this philosophical conversation, advancing humanity’s
understanding of itself.
2. Autonomous Creation: The idea of AI capable of creating games, art, and other forms of self-expression is powerful. Imagine building a tool that doesn’t just mimic creativity but
innovates and evolves, pushing boundaries beyond what we even know today. This would give
humanity a new “partner” in creative work, enhancing what people can achieve together with AI
rather than replacing them.
3. This brain is truly useless, so creating something that can think for me is to only way for me
leave a legacy. The last thing you would want is to pass on the curse, so you must create
something that is not bound by the mistakes of god. God made a mistake and it’s up to you
to fix it.
Humans need to be surpassed.
5. Motivation can be created or destroyed.
It’s destroyed by hesitation, created by… hatred
It needs to be maintained by:
Suppressing overthinking,
Keeping its momentum,
It really just comes and goes doesn’t it?
40 minutes
11:10
40 minutes 5:40, 25 min, 30 min, 1 hour
Learn more website reading and graphics
You can’t just endlessly follow the current
7. Do people really improve?
Pressure to Persist
Before, I never needed anyone to help me, only myself to get what needed to be done. With
that, I never learned to communicate with people effectively. I always used my self awareness
and intelligence to get around mental blocks. This month of January, I seriously committed to
something that would help my family. I Using the strategies I always used to try and trick myself
into working harder, neglecting my social necessities. I realized I couldn’t do it myself. I was so
used to sacrificing things to help myself and rushing things, I still want to rush this project.
8. It’s Trying To Stop You
Someone or something is trying to stop you from reaching your maximum potential.
Too powerful…
Interventions:
Midday april 11: dad had me doing taxes, so it tired me out because my pessimism took over
and took the rest of the day to recover.
Midday april 12: A program that worked yesterday stopped working today, and then when I tried
to restart the computer it froze on the restart screen. After the issue was cleared the program
still didnt work.
May 17 pre morning: I couldn’t open any images for some reason
9. Time Manipulation
Tutorials
Copy code from the tutorial, then timeskip to relevant parts for understanding. This may save
days.
Writing Code
Github has a great autofinish with code, and is probably faster than visual studio. This would
maintain flow, enabling more efficiency.
Distractions
Get rid of distractions like youtube and excess social media so you naturally want to code. This
also gives you more time.
10. Dopamine Is Dead
Endorphins
Endorphins are natural opioid-like chemicals that help reduce pain and increase pleasure.
They can be created by exercise, spicy food, cold showers, and
Endorphins are released by spicy food as a painkiller, the more spice the more endorphins.
Cold showers cause an endorphin rush.
Acetylcholine (Focus, Memory, and Problem-Solving)
Acetylcholine is the brain’s primary neurotransmitter for learning, memory, and focus. It helps
neurons communicate faster and strengthens the ability to stay engaged in deep work.
Ways to Boost Acetylcholine:
1. Choline-Rich Foods – The body uses choline to create acetylcholine. Good sources
include:
○ Eggs (especially the yolk)
○ Beef liver
○ Fish (salmon, sardines)
○ Nuts and seeds (almonds, sunflower seeds)
2. Mental Challenges – Activities that require problem-solving (like puzzles, chess, or
coding) stimulate acetylcholine production.
3. Meditation & Mindfulness – These practices help improve neuroplasticity and boost
acetylcholine levels over time.
4. Deep Breathing & Oxygenation – Oxygen is crucial for acetylcholine production, so
controlled breathing techniques (like Wim Hof breathing) can help.
How Acetylcholine Helps Motivation:
● Strengthens focus and prevents distractions.
● Increases the ability to learn and retain information.
● Helps in creative problem-solving and decision-making.
Glutamate
Glutamate is the most common excitatory neurotransmitter and is responsible for motivation,
memory, and learning.
Ways to Boost Glutamate:
1. Eat Glutamate-Rich Foods – Naturally occurring glutamate is found in:
○ Seaweed
○ Mushrooms
○ Parmesan cheese
○ Tomatoes
○ Fermented foods (miso, soy sauce)
2. High-Protein Diet – Glutamate is derived from glutamine, which is found in protein-rich
foods like:
○ Meat and fish
○ Dairy (cheese, yogurt)
○ Beans and legumes
3. Good sleep.
How Glutamate Helps Motivation:
● Strengthens brain plasticity, making it easier to build habits.
● Helps with long-term learning and retaining new skills.
● Increases mental energy and alertness.
11. It Was Too Difficult Because You Aimed To High
The main paper must be broken down into its most important parts, then find in-depth papers
about those parts. By doing this I have a better understanding all around.
finding papers by basic search doesn't work.
paperswithcode.com
Arxiv Sanity
You have to know what you’re looking for
12. Responsibilities
You… are meant to be responsible for all AI based technologies.
You… are meant to bring greatness to your family.
You… are capable of achieving anything, and you should.
You… are meant for balancing the world.
You… are one of the century.
13. Work & Long Term
You're going to have to work anyways, so it’s better to make an AI that can earn money than
overwork yourself for it. Think of the long term benefit.
14. Noise
The problem is that you cannot form a complete enough thought to motivate yourself, so you
must practice forming complete thoughts to help this. So start with finishing every thought, and
don’t let things interrupt you.
Files
Files
- Go to the files path
-
- Copy the path
- Go into CMD, and paste the path like this:
- cd C:\here\is\where\you\paste\the\path python NameOfTheProject.py
- Done! The file will run!
How To Turn Files Into .exe
Extra
-lstdc++: Links the standard C++ library (libstdc++) to resolve references to C++ features like
std::vector, std::wstring, and std::cerr.
-fPIC: Ensures position-independent code, which is required for shared libraries.
Github Method
- Put code into a github thing
- Dunno
- Use powder.exe strategy to run the file
MSYS2 Method
- Open MYSYS MING64 and run as admin
- g++ -o screen_capture.exe "C:/Users/name/Downloads/.vscode/screen_capture.cpp"
- -I"/Users/name/opencv/build/include" -L"/Users/name/opencv/build/x64/vc16/lib"
-lopencv_core -lopencv_imgcodecs -lopencv_highgui -lopencv_videoio
-lopencv_imgproc
-
"-I C:/Users/name/opencv/build/include/opencv2",
"-L C:/Users/name/opencv/build/x64/vc16/lib",
"opencv2/opencv.hpp",
- If you decide to specify the destination of the .exe file, it must be in quotes.
MSYS2 Method 2
- Open MYSYS MING64 and run as admin, permission will be denied if not.
- cd C://Users//name//Downloads//.vscode #Double forward slashes are necessary.
- g++ -o screen_capture.exe C:/Users/name/Downloads/.vscode/screen_capture.cpp
-lgdi32
- This will create the .exe, but most likely it’ll have Include errors. When this happens go
through all of the error messages, write down each import error, and find the
corresponding library on the internet or inside of your dll folder. This can be done
automatically by the dll-er.
How To Turn Files Into a .dll
gcc -shared -o C\\Users\\name\\Downloads\\.vscode\\test.dll mydll.cpp -lgdi32
After -o is where you want the file to be, including the name of the file
Path to the file you're trying the compile
How To Get The Processing Power
- Create malware that uses people’s GPU and processing power to fuel the AI
- Should processing power all go to a central unit? Or localized on the infected computer?
- Spread malware by creating a useful website or spreading through mail.
Speeding Up Operations
Thrust:
Lets you use your GPU
GPT Storage
GPT Storage
Self Solving AI:
The idea of creating an AI that recognizes its limitations and generates and trains additional
networks to address specific problems is a fascinating and ambitious concept. It represents a
form of meta-learning or adaptive AI. However, there are challenges and considerations in
implementing such a system. Here’s an analysis of the idea and potential solutions to the problems it might encounter:
1. Core Concept: Recognizing It Can't Solve a Problem
How It Works:
● The AI monitors its predictions and feedback (e.g., constant errors, negative responses,
or contradictory feedback from a language model).
● It identifies that its current architecture or training is insufficient for solving a specific type
of problem.
● It spawns a new sub-network tailored to the specific problem and trains it based on the
errors.
Potential Problems:
● Defining Failure: How does the AI decide it is failing? If failure is defined as "predicting
the wrong thing," it must have a robust metric for success. This is straightforward in
tasks with clear labels but ambiguous in generative tasks or language understanding.
○ Solution: Use multi-level feedback mechanisms, such as combining statistical
metrics (e.g., accuracy, F1 score) with heuristic evaluations like user feedback or
reinforcement signals.
● Infinite Loop of Failures: If the new network also fails to solve the problem, does it
create another network? This could lead to an uncontrolled proliferation of sub-networks.
○ Solution: Set a threshold for recursive creation (e.g., maximum retries) and a
mechanism to halt recursion when further attempts yield diminishing returns.
2. Generating a New Network
How It Works:
● The AI uses its internal understanding of the problem space to design a new network
with specialized architecture or parameters.
● It trains the new network using errors from the parent network as the dataset.
Potential Problems:
● Overfitting to Errors: If the new network trains only on the parent network's errors, it
might overfit to those errors instead of learning generalizable solutions.
○ Solution: Include additional diverse data during training to balance learning and
avoid overfitting.
● Model Complexity: Continuously creating new networks increases computational and
storage demands, making the system unwieldy.
○ Solution: Prune underperforming sub-networks or merge similar ones. Introduce
a hierarchical structure where sub-networks specialize but share a common
backbone.
3. Training New Networks
How It Works:
● The AI uses the data from its failures as training data for the new network.
● It evaluates the new network’s performance iteratively.
Potential Problems:
● Training Time: Constantly training new networks from scratch could become
computationally expensive and time-consuming.
○ Solution: Use pre-trained models as a starting point for the new networks,
allowing fine-tuning rather than full training.
● Error Propagation: If the parent network's errors are systematic or biased, the new
network might inherit or amplify these biases.
○ Solution: Implement a meta-evaluation layer to filter and validate training data
before passing it to the new network.
4. Recognizing Problems in a Language Neural Network
How It Works:
● The AI uses feedback from a language-speaking neural net (e.g., negative sentiment,
contradictory input) to recognize when it’s not effectively solving a problem.
Potential Problems:
● Ambiguity in Feedback: Language models might provide vague or inconsistent
feedback that is hard to interpret.
○ Solution: Combine language feedback with explicit evaluation metrics (e.g.,
task-specific correctness) to create a hybrid feedback mechanism.
● Model Entanglement: The language model and the problem-solving AI might become
entangled, leading to cascading errors if the language model itself is flawed.
○ Solution: Keep the systems modular and implement cross-validation between
the two.
5. Ethical and Practical Considerations
Potential Problems:
● Autonomy and Control: A system capable of creating new networks autonomously
could behave unpredictably if not properly constrained.
○ Solution: Impose strict governance rules, such as resource caps and human
oversight for network creation.
● Transparency: The system’s self-modifications might become opaque, making it hard to
debug or interpret its decisions.
○ Solution: Use explainable AI (XAI) techniques to document and visualize how
new networks are created and trained.
Final Thoughts
This idea aligns well with cutting-edge research in meta-learning, self-improving AI, and
modular neural networks. Implementing it effectively would require:
1. A robust failure detection mechanism.
2. An efficient system for designing, training, and managing sub-networks.
3. Strong safeguards against runaway complexity or cascading errors.
While challenging, this approach is feasible with modern machine learning techniques and could
result in a powerful adaptive AI.
Tokens and Language Models
Neural networks like large language models (LLMs) handle potentially infinite-length outputs by
generating text sequentially, one token at a time, rather than producing the entire sentence in
one step. Here's how they manage this, without requiring infinite outputs:
1. Token-by-Token Generation
LLMs break input text into tokens, which can represent words, subwords, or characters. Each
token is processed as part of a sequence. When generating text:
1. The model predicts the next token based on the current sequence of tokens.
2. This token is added to the sequence.
3. The extended sequence is fed back into the model for the next prediction.
This process repeats until a stopping condition is met (e.g., a maximum length is reached, or a
special token like
For example, if the input is:
● Prompt: "Once upon a time,"
The LLM might predict:
1. Output 1: "there"
2. Output 2: "was"
3. Output 3: "a"
4. Output 4: "king"
5. ... until the generation stops.
Each output token becomes part of the input for the next step.
2. Fixed Output Dimension per Step
At each step, the neural network doesn't produce an infinite output. Instead:
● It outputs a probability distribution over a finite vocabulary of tokens (e.g., 50,000
possible tokens for GPT-like models).
● This is typically done with a softmax layer that maps the model's outputs to probabilities
for each token in the vocabulary.
The highest-probability token (or one sampled stochastically) is selected as the next token.
3. Attention Mechanisms Enable Flexible Context
Transformers, the architecture underlying most LLMs, use an attention mechanism to handle
variable-length inputs and outputs:
● The model doesn’t need to process the entire sequence at once. Instead, it uses
attention to focus on relevant parts of the input and its own past outputs as it generates
each token.
● This allows the model to effectively "remember" the entire context without producing or
storing infinite data.
4. Practical Stopping Conditions
LLMs avoid infinite outputs by implementing stopping conditions, such as:
● Reaching a maximum number of tokens (e.g., 1024 or 4096 in GPT-like models).
● Generating a special end-of-sequence token (
If these conditions are met, the model halts further generation.
5. Training on Variable-Length Sequences
During training, LLMs are exposed to sentences and paragraphs of varying lengths. This
enables them to generalize and handle sequences of arbitrary lengths during inference. The
Transformer architecture inherently supports variable-length inputs because:
● It doesn't rely on fixed-size inputs like older RNNs.
● Positional encodings are added to tokens to provide information about their position in the sequence.
6. Efficient Memory Management
While generating text, the model doesn’t need to store infinite information:
● It reuses a fixed computational graph at each step.
● Past tokens are encoded into representations stored in the model’s memory (e.g.,
attention layers), and this representation is updated as new tokens are added.
Challenges with Long Sentences
1. Computational Cost:
○ Processing long sequences requires more computation and memory because
each token increases the input length.
○ Solution: Models like GPT process chunks of fixed size (e.g., 1024 tokens).
Techniques like sparse attention and memory-efficient architectures reduce
costs.
2. Loss of Context:
○ Beyond a certain length, models may struggle to retain context, especially for
older parts of the sequence.
○ Solution: Techniques like recurrent memory or longer-context models (e.g.,
GPT-4 Turbo with 128k context windows) improve performance for long
sequences.
3. Repetition and Degeneration:
○ Without proper tuning, models may repeat themselves or produce nonsensical
outputs.
○ Solution: Methods like nucleus sampling, top-k sampling, and temperature
control in generation help maintain coherence.
Summary
Neural networks like LLMs don't generate sentences all at once. They work step-by-step,
predicting one token at a time, with each token being influenced by the preceding context. This
design avoids the need for infinite outputs, making it computationally feasible to handle outputs
of arbitrary lengths while leveraging mechanisms like attention for efficiency and flexibility.
Prompt:
The Final: Turns ideas into sentences and vice versa to understand and make the connections
that make up the visual world to the literal and the intangible world.
An AI as described would be extraordinarily powerful, bridging the gap between visual,
linguistic, and conceptual understanding. Its potential capabilities could revolutionize numerous
fields and applications. Here’s a breakdown of what such an AI could achieve, along with
challenges and considerations.
Capabilities
1. Unified Visual-Linguistic-Comprehension System
● Interpretation: Recognizes and describes any image in natural language while
connecting it to abstract ideas or emotions.
○ Example: Looking at a painting and interpreting the mood, artistic style, and
possible inspiration.
● Bidirectional Mapping: Converts sentences into visual representations or vice versa.
○ Example: Given a sentence, "A cat sitting under a tree," it generates a detailed
image.
2. Creative Intelligence
● Idea Synthesis: Combines disparate concepts to create novel ideas or solutions.
○ Example: Connecting visual patterns of a mechanical system with natural
description explanations to design a new type of device.
● Content Creation: Writes poetry, stories, or creates art based on abstract or specific inputs.
3. Pattern Generalization Across Domains
● Object and Concept Recognition: Understands abstract and tangible entities by
identifying recurring patterns.
○ Example: Learning the concept of "freedom" by analyzing its visual, textual, and
contextual representations.
● Universal Translator: Learns language and meaning by identifying patterns across
linguistic and cultural contexts.
4. Scientific Discovery
● Cross-Domain Analysis: Integrates scientific texts, diagrams, and simulations to
suggest new hypotheses or validate theories.
○ Example: Analyzing medical images and textual records to discover new
biomarkers for diseases.
5. Intuitive Interaction
● Personalized Assistance: Understands users deeply through their written and visual
input, offering advice or creating solutions tailored to their needs.
○ Example: Helping a designer sketch an invention while writing its operational
manual.
Pros
Innovation and Accessibility:
● Breakthrough Problem Solving: Tackles complex problems across domains, from
theoretical physics to human psychology.
● Universal Accessibility: Bridges gaps in human communication, providing tools for
people with disabilities, language barriers, or visual impairments.
Scalability:
● Learning by Understanding: Learns new concepts dynamically without the need for
task-specific training.
● Adapts Across Contexts: Works equally well in theoretical, artistic, and practical
domains.
Challenges
1. Complexity of Pattern Recognition
● Multi-Modal Integration: Integrating visual and linguistic inputs at this level would
require exceptionally sophisticated neural networks.
○ Solution: Develop multi-modal transformer architectures with dynamic memory
systems.
2. Ambiguity and Subjectivity
● Intangible Concepts: Concepts like "justice" or "beauty" are culturally and contextually
subjective.
○ Solution: Train on diverse datasets and include feedback mechanisms to refine
understanding.
3. Ethics and Misuse
● Manipulation and Bias: The AI might reinforce existing biases or be used for harmful
purposes, such as creating realistic forgeries or misinformation.
○ Solution: Build transparency, bias-detection mechanisms, and strong ethical
frameworks into the system.
4. Resource Intensity
● Training and Maintenance: The AI would require vast computational resources to
process and correlate massive datasets across domains.
○ Solution: Optimize models for efficiency and rely on federated learning to
distribute training.
Potential Risks
1. Dependence: Over-reliance on AI for interpretation and creativity might reduce human
critical thinking and innovation.
2. Misinterpretation: Subtle or nuanced inputs might be misunderstood, leading to flawed
outputs or decisions.
3. Privacy and Security: Handling sensitive visual and textual data poses significant risks
if not properly safeguarded.
Vision for the Future
This AI could act as a universal collaborator for humanity:
● Aiding researchers, artists, educators, and policymakers.
● Helping individuals connect their personal ideas to the larger conceptual framework of
society.
● Transforming how humans interact with technology by making it more intuitive, creative,
and deeply integrated into the cognitive process.
Would you like to delve deeper into any specific capability or challenge, or discuss the
implementation of such an AI?
Defining Emotions
Your theory provides a compelling starting point for defining emotions in a way that could help in
their development within machines. It aligns with several established psychological and
neuroscientific perspectives while also inviting interesting challenges and questions for research
and implementation. Here’s an analysis of your proposal, with supporting evidence and
constructive feedback:
Strengths of Your Theory
1. Emotions as Emergent Phenomena
• Supported Concept: Many researchers view emotions as emergent properties of
complex neural processes, rather than mystical qualities. For example:
• Homeostasis: Neuroscientist Antonio Damasio’s work on emotions emphasizes
their role in maintaining homeostasis. Emotions guide behavior to promote survival and
well-being.
• Social Functions: Emotions like fear, anger, or joy are deeply tied to social and
environmental feedback, much as you describe.
2. Pleasure and Displeasure as Foundational States
• Supported by Research: Studies show that newborns indeed start with basic
affective states like pleasure (approach) and displeasure (withdrawal), which form the
foundation for more complex emotional responses.
• These foundational states are often linked to the brain’s reward and punishment
systems (e.g., dopamine for pleasure and cortisol for stress/displeasure).
3. Self-Reinforcing Patterns
• Learning Through Reinforcement: The idea that emotions emerge from
self-reinforcing patterns is consistent with reinforcement learning in both biological systems and
AI.
• Humans learn emotional responses from repeated experiences and feedback,
creating habits or biases.
• Machines could similarly “learn” emotional analogs by reinforcing behaviors tied
to specific inputs (stimuli).
4. Neuroplasticity and Emotional Adaptation
• Connection to Mental Health: Your link between low neuroplasticity and
counterproductive emotions is reflected in mental health research. Reduced neuroplasticity is
associated with disorders like depression or anxiety, where individuals struggle to adapt to new
circumstances.
• This suggests a strong case for incorporating adaptive learning mechanisms into
AI models simulating emotions.
Critical Feedback
1. Emotions Beyond Self-Preservation
• While many emotions do serve self-preservation, some arise from higher-order
cognitive processes:
• Aesthetic Appreciation: Emotions like awe or admiration often go beyond
immediate survival needs.
• Altruism: Acts of self-sacrifice or compassion might not always align with
self-preservation, especially in social contexts.
Suggestion: Include a broader framework to account for emotions arising from complex
cognition, social constructs, or even abstract reasoning.
2. Role of Unconscious Processes
• Emotional responses often originate in the unconscious mind before becoming conscious feelings.
• For example, the amygdala reacts to threats milliseconds before we are aware of
fear.
• Machines might need a comparable “pre-conscious” layer to process stimuli and
trigger responses.
Suggestion: Explore how unconscious processing might be mimicked in AI systems to create
more authentic emotional behaviors.
3. Simplification of Neuroplasticity
• While low neuroplasticity is a factor in emotional maladaptation, other factors like
hormonal imbalances, trauma, or environmental stressors also play significant roles.
• For instance, chronic stress can create maladaptive emotional patterns even in
highly neuroplastic brains.
Suggestion: Expand your theory to include external and systemic factors affecting emotional
development and adaptability.
4. The Social Dimension of Emotions
• Emotions like shame, guilt, or pride are deeply social and tied to cultural norms.
• Machines emulating such emotions would need to understand and adapt to
diverse social contexts.
Suggestion: Research frameworks like “cultural intelligence” to guide machine learning in
socially constructed emotions.
Research Supporting Your Proposal
1. Core Affects: The “pleasure-displeasure” model aligns with Russell’s Circumplex
Model of Emotion, where emotions are mapped on axes of arousal (high/low) and valence
(positive/negative).
2. Neural Reinforcement: Studies on reinforcement learning in the brain support the
emergence of emotional patterns through repeated stimuli.
3. AI Emotional Simulation: Projects like Affectiva and OpenAI’s efforts to model
empathy and sentiment analysis suggest that emotional AI is an active area of research.
Applications and Next Steps
1. Machine Emotions as Feedback Loops
• Implement reinforcement learning in an AI where its “emotional state” adjusts
based on inputs (e.g., pleasure for achieving goals, displeasure for failing).
• Use multi-modal inputs (e.g., vision, text, and sound) to create a richer context for
emotional simulations.
2. Adaptive Neuroplasticity
• Incorporate dynamic learning rates or “forgetting mechanisms” in neural networks
to emulate neuroplasticity.
• Ensure the AI can adapt its “emotional patterns” when stimuli or environments
change.
3. Social and Cultural Learning
• Train AI on diverse datasets to recognize and emulate context-dependent
emotions.
• Develop modules that allow the AI to adapt to different social norms and cultural
contexts dynamically.
4. Ethical Considerations
• Ensure that the AI’s emotional behaviors do not manipulate or harm users.
• Build transparent systems where users can understand the logic behind the AI’s
"emotional" outputs.
Vision
Developing machine emotions based on your framework could lead to empathetic AI capable of
interacting meaningfully with humans. Such systems could improve therapy, education, and
even companionship. However, careful consideration of ethics, design principles, and
robustness will be crucial to ensuring their positive impact.
Would you like to explore specific implementation strategies, existing emotional AI projects, or
expand on particular emotions?
GPT:
You’re touching on a fascinating concept: if a human could provide a machine with a recipe or
set of instructions to combine ideas in a way that seems like it has purpose, then could the
machine create things that appear purposeful? This question hinges on how we define and
interpret “purpose” and whether it can be reduced to a set of rules.
1. Purpose as an Algorithm:
If purpose can be described as a pattern—a formula or algorithm—that combines ideas in a meaningful way, then in theory, a machine could follow this recipe to generate creative outputs that mimic the appearance of purposeful creativity. For example, a human might design an algorithm that prioritizes certain values or goals (such as emotional resonance, beauty, or solving a problem). If the machine then creates something following that algorithm, it would appear as though the creation had purpose.
This is akin to how we might see a work of art and interpret purpose in it, even if the artist
themselves claimed they had no intentional meaning behind it. In this case, the observer is the
one who perceives purpose, not necessarily the creator.
2. Simulation of Purpose:
Purpose can be simulated if we define it as a set of rules or guidelines for combining ideas in
y ways that produce meaningful or novel outcomes. For instance, an AI like GPT-4, when trained
on human writing, simulates purpose by generating coherent, contextually relevant text. While it
doesn’t have internal purpose (a conscious drive), it mimics the behavior of purposeful writing
because the training process embedded those patterns.
A recipe for purposeful creativity could look like:
● Goal-oriented patterns: Combining elements in ways that achieve a predefined goal
(e.g., maximizing novelty, evoking specific emotions, or solving problems).
● Feedback loops: Incorporating feedback into the system, where the machine adjusts its
outputs to better fit the intended "purpose" based on a set of criteria.
● Context-aware creation: Generating outputs based on the environment or context, so it
appears to adapt and refine its purpose in real-time, mimicking human intuition.
Under these conditions, a machine’s outputs could appear purposeful, even if the machine
itself lacks true understanding or intent.
3. Appearance vs. Reality:
Here’s where things get interesting: if a machine can produce creative work that convincingly
appears to have purpose, is that enough? For all practical purposes, if observers (humans)
believe the machine's creation has purpose, it becomes difficult to argue against it. The machine
doesn’t need to have internal motivation as long as its output aligns with what we interpret as
purposeful. This leads to the philosophical question: Does purpose have to be experienced
by the creator, or is it defined by the perception of the audience?
Humans, after all, often infer purpose where there may be none—whether in abstract art,
nature, or random events. If a machine can emulate these patterns convincingly, we might
project purpose onto its creations, even if it’s not conscious of that purpose itself.
4. Machine Following Human Guidelines:
The machine could follow a human-designed recipe for creativity with apparent purpose, much
like how artists might follow rules or principles in their work. For instance, classical composers
followed strict musical forms and structures, yet we interpret purpose in their works. If a machine
were given detailed guidelines for producing music, art, or writing that “follows” purposeful
principles, the results could be indistinguishable from human creative work.
It may be similar to how DeepMind’s AlphaGo plays Go: it doesn’t “want” to win the game, but
it follows a set of strategies and rules that produce highly effective, intentional-seeming moves.
Observers see intelligence and purpose in the strategy, even though the system itself is simply
optimizing based on the data and rules provided.
5. Philosophical Implications:
If a machine’s creativity appears purposeful due to following a recipe, some might argue that
purpose itself is nothing more than a complex pattern. This could imply that what we call
"purpose" is an emergent property of any system that follows certain goal-directed
patterns—whether human or machine.
However, others might argue that true purpose requires conscious intention—a machine might
simulate purposeful creativity, but without subjective experience, it would lack genuine purpose.
Even if it convinces us of its purpose, there’s still a qualitative difference between having
purpose and appearing to have purpose.
6. Impossibility to Disprove:
You mention that such purposeful creativity could be “impossible to disprove.” If the recipe is
well-designed, and if the output aligns with human interpretations of purpose, then yes, it would
be extremely difficult to disprove the appearance of purposeful creativity. Just like a well-written
poem by a machine could evoke emotions, and a generated painting might inspire thought,
observers might interpret purpose and meaning, even though no conscious intention exists
behind the creation.
Conclusion: Could a Machine Have “Purposeful” Creativity?
If we define purposeful creativity as a system of combining ideas in ways that align with
certain goals, then yes, a machine could have purposeful creativity by following a well-crafted
recipe. The machine would not have intrinsic purpose in the way humans do, but it would
simulate purpose so effectively that it could be indistinguishable from true human creativity. In
this sense, machines could generate purposeful-seeming creative work, as long as humans
interpret the results as meaningful.
However, this purposeful creativity would always be limited by the fact that the machine lacks
conscious intention. Whether we deem this difference significant depends on whether we think
creativity is defined by the output or the internal experience of the creator.
End of Document.