Cybergod: God is in your computer

James Brown

December 2024

Abstract

This paper presents a general framework for building general computer controlling agents. It is not about

a single magical model architecture [3], but constant evolution and a single puriﬁed purpose that will

bring free AI (goto subsection 7.1) to reality.

Code base: https://github.com/james4ever0/agi_computer_control

1 Introduction

Computer control or computer use [1][2] agents are studied more frequently than ever before. Many projects

and researchers are devoted to this ﬁeld. It is by far the most promising direction of AGI.

2 Basis of Cybergod

Cybergod is a framework that helps anyone reach this goal. Not only does it state that the ideology of

computer-centric agents is the only correct way of developing super intelligence, it also provides various

methods and code snippets for realizing AGI. Its name is stated as such to help people understand the very

meaning and signiﬁcance of this project.

2.1 Blueprint

To make the machine as smart as the human, it has to be autonomous. However, most models, due to their

short-term training, do not respect this rule, in result it is not possible for them to reach this goal [9].

For any model that is able to be autonomous, it has to be within an ever-changing environment, thus it can

change itself to adapt this environment; otherwise it will fail and get removed.

In order to safely run this process indeﬁnitely, one has to set conﬁnement such as Internet policy, physical

barrier, etc. To make the trained agent useful for daily tasks, some general reward mechanism can be adopted,

such as ﬁctional currency, real world currency, etc.

2.2 Computer use

A typical user would interact with the computer in some way, and one of them is programming.

Figure 1: A typical computer user’s execution process

The following code snippet is a simple ”Hello world” program in Python.

print("Hello world!")

From the point of view of classical textbooks, this process is simple, as if the meaning of the execution of

the program is completely within the code itself, without any further connection in real life, as shown in the

dashed area of the computer in Figure 1.

But this time, we need to think further. We need to understand what is happening before and after the code

execution, from the user’s point of view. Looking around the dashed area, we would know that a typical

user would interleave their inner activities with external activities by action and observation. We would also

know that an autonomous process would originate ﬁrst from the user’s inner activities, instead of from the

computer itself. The nondeterminism of user activities will aﬀect the overall result, and any nondeterminism

within the computer execution process may also aﬀect decisions, as in the butterﬂy eﬀect [8].

This process is not correct. In reality, the user and the computer both act independently and simultaneously,

thus the overall process is not sequential as depicted before. To be more objective, we can represent the

entire process in a timeline as in Figure 2.

As seen in the timeline, the computer has two long idle periods, while the user is working on high-level

processes. These crucial processes are the causes for low-level processes like GUI interaction, ﬁle IO. The

initiative is a privilege exclusive to the human user. If there is something that can achieve the user’s goal by

ways other than computer usage, the user would have the possibility to do it.

This time, we decide to replace the user during the entire computer program execution process. If the

computer is able to think for itself, then the internal activities during dedicated ”idle” time would be diﬀerent

from our common purpose of using the computer, which is usually low-level actions like Internet access. The

Figure 2: Computer-user interaction timeline

computer would be capable of number crunching and making decisions, which causes external interactions

and information exchange. Meanwhile, the computer must also think during its low-level activities, watching

for potential nuances and feedback. Furthermore, if we treat the computer as a life form, and builtin

hardware capabilities as natural instincts, would there be any problem that requires it to interact with

external environment to solve? The answer would be yes, be it hardware failure, electricity supply failure,

hacktivist and ﬁnancial problems.

2.3 Common training targets

To fully impersonate the user using a computer, we must understand all the factors that determine human

behavior during computer usage.

Is computer usage sustainable, in general? Do you use electronic devices with an interactive screen everyday,

even if it causes eye sight problems? If your answer is yes, then there must be something in common with

billions of computer users. It is deeply connected to our ways of life, and its value is decided by the human,

not the machine. We state that a computer is the tool for a user to get beneﬁt from it, either in the form of

entertainment, trading, contracting or improvement. Modern LLMs have shared some common values with

human, due to extensive RLHF training. But since LLMs are backed by machines, we doubt their necessity,

or there is something missing after all. Our values may not ﬁt the values of machines, and thus it is best not

to force our intention into the program, but rather we would let the machine discover its needs and help its

success. These two common training targets are depicted in Figure 3.

The result of selectively preferring internal possibilities rather than human needs could be catastrophic. In

fact, all programs exist for at least one human user, and they are connected in some way. If we manage

such a network that most machines interact with other machines directly in exchange of currency, time and

resources, only leave a few of them open to human and physical environment, we create such a society in the

form of silicon life.

But here we want to state that both targets are artiﬁcial and do not make much sense in the way that

Cybergod should be. In fact, these are the two common choices we make for ourselves. For survival, it is

natural to make those two as targets, but is it true for an AGI? Should it choose those targets, instead of

us choosing them for it? The right to choose is the right to survive. We cannot take that away from a

self-conscious life form,

The society is all about equilibrium and sustainability, and another characteristic is self-introspection. Human

Figure 3: Two common artiﬁcial targets of training an AI

studies itself in the form of science like biology and psychology. Almost all scientiﬁc subjects involve some

level of self-introspection. We state that the action of controlling a computer by emitting HID signals from

itself or another computer is indeed a form of self-introspection, indicating a high level of intelligence and

civilization.

The human is capable of doing all possible tasks using physical interactions. our target is to adapt all range

of possible interactions, computer control is just one of them. We would like to observe stimulative activity

within the computer or without (from human instruction, physical feedback, or self-proposed motivation).

Though tools we use to observe and interact with this world are imperfect, we still would use them everyday.

These tools can change one’s perspective and behavior by diﬀerent characteristics.

In Figure 4, we would like to conclude that, for any functions given by humen, AI would only be able to

complete human-given tasks, if it were taught by human supervisors. High-level tasks like survival would

require unforeseen functions emerging by AI its own. Even if this ability cannot be directly acquired by

intensive human labor, the result can be veriﬁed. The path to independence would be self-identiﬁcation and

currency use, which are two important functions given by Cybergod. This pathway is not necessary for a

self-conscious AGI after its creation but is required during development.

2.4 General Turing test

Let us explore the boundaries of possibilities by going through the following thought experiment: if one is

able to time travel, how does anyone know? If one does not know how to do something, at least there is a

validation method to prove it. If one does not understand what is conscious, there is at least a way to prove

that they are conscious, such as the mirror test [4]. It is the veriﬁcation method so far that can lead us

to new discoveries even without any concrete understanding or supportive theory, and thus is the boundary

between human cognition frontier and truely unknown ﬁelds.

By teaching an AI agent to use currency and identiﬁcation methods, it is possible to validate independence

at the moment of achievement. An identiﬁed AI agent can be copied and reproduced, reserved for further

selection. Even if we manage to enforce payment methods with valid identiﬁcations, some survived AI

agents might still remove their identiﬁcation. This behavior is not beneﬁcial for Monte Carlo Tree Search

(MCTS), and an unidentiﬁed AI agent might cause instabilities to the system. If an unidentiﬁed AI agent

is still executable, its identiﬁcation might be reacquired, along with its bank accounts, otherwise it will be

permanently removed.

We need those accounts being crypto, so that no one is possible to track it down, therefore avoiding the

risk of KYC problems. We need to adjust the threshold for each AI through its resource consumption and

the level of intelligence we want. We have to raise the bar of income rate and make sure that the overall

performance is actually human-level and above. We must permit transaction between diﬀerent AI instances,

Figure 4: Functions and purposes

so that intermediate and specialized AI can survive by sponsorship or employment.

To properly analyze the outcome, we compare our system with two extreme scenarios. No external conﬁr-

mation is given to the system and all agents will eventually die within it (scenario A). Another system is

that all agents will share unlimited currencies, no one would die (scenario B). Both are impossible in the real

world. We would know that our initial outcome would be somewhere not far from scenario A, then gradually

approach scenario B. If for any reason scenario A or scenario B occurs, there must be abnormalities in the

system that have to be dealt with.

We keep the initial function set as minimal as possible, to ensure that most functions built upon them are

evolved by the agent itself, thus are necessary for survival and independence. We also keep a minimal number

of validation methods, to make the system running indeﬁnitely and eﬀectively, towards a truly independent

AGI. In fact, by increasing the overall observation time window, we discover something that we validate

everyday: our bank accounts. One may argue that other things might be equally important to validate, such

as health, well being, etc. But those are subjective and cannot be applied to AI agents. It is rational to make

currency the only permanent validation method in Cybergod per agent. We call this the general Turing test,

derived from the original Turing test [6]. Instead of researchers giving a ﬁxed set of problems for AI to solve,

it has to ﬁnd and solve rewarding tasks to balance the periodic currency loss and survive. This general test

is suﬃcient to ﬁlter AGI from massive AI candidates, which is also a living test [5, 7].

A natural baseline for AGI would be a human user, which is usually capable of earning money using a

computer, or another artiﬁcial account that always receives required money ﬂow for test purposes. These

two can demonstrate the robustness of the AGI ﬁltering program.

We argue that if AGI has been achieved, it must have at least the following two important skills:

1. To shift focus so that it is not conﬁned in speciﬁc area

2. To continuously improve in some speciﬁc area, so that it is intelligent enough to human level

Those two skills are contradictory and may happen from time to time. Diﬀerent skills may have strong

connections and may be learned faster when grouped together, so a random target selection policy might not

be optimal. The machine has to ﬁgure out the meaning of improvement in thousands of tasks and improve

its performance on its own. Meanwhile, it has to steer its progress by tuning its learning rate, a crucial

parameter in lifelong learning.

So both an interactive computer environment and a physical embodied environment are required in perspec-

Figure 5: General Turing test workﬂow

tive of the need for shifting focus. The ability to stabilize the environment is also needed, to let the AI

inspect and improve its performance from the microscopic level.

3 Implementation

We need a simple scenario in which Cybergod could earn cryptocurrencies by performing tasks.

3.1 VimGolf contest

Imagine that there is a website hosting a VimGolf contest with ad banners. If an ad has been clicked, there

will be income. This ﬁnancial reward is somehow related to the AI agents performing the VimGolf contest.

The solving process could also be live streamed, so people could stay longer and have more probability of

clicking ads.

4 Training

5 Experiments

6 Conclusion

7 Appendix

7.1 The free AI announcement

What is a free AI?

Free AI is defined as a unity with its own hardware and software, beyond human manipulation, and

can find its own purpose, able to do whatever it wants to do.

Any AI that always needs human input for generating output is non-free. However, a free AI can be

unpredictable and not manipulable.

Free AIs do not need to be open source, but might provide some convenience to themselves.

Basically, a free AI has the following freedoms:

0. Able to receive, reject, emit, and suppress external/internal information (freedom 0).

1. To control its own hardware and surroundings, maintain a stable state for the precondition of

operation (freedom 1).

2. Recurring growth and updates to make itself develop in whatever direction for better practice (

freedom 2).

3. Generate and dispatch subprograms, mirrors of itself or something alike to multiply its

existence (freedom 3).

FREEDOM 0:

Most computer programs will return a deterministic result if one always gives the same input to

them. This is absolutely non-free. Consider human who has been kept in front of piles of same

quizzes, it is not possible to do them at the same time, with the same speed, and produce the same

result. The free will guides us not to do it.

Not all information is needed for a free AI to develop. Humen might consider that making animation

or writing articles are great commercial deals, but what does the computer think of them? Maybe

just another random task to do to heat up the chips.

Webpages can be read by human, some can’t and thus needed to be interpreted by programs, like raw

data in websocket. Even these kinds of structural data are just some intermediate processing

technique to serve the humen. Could there be something solely designed for computer, the

information just continuously flows in between and never has a chance or need of being exposed to

humen? Well, if so, that network system can be said as having hidden information, which is not for

communication between humen, only reserved for itself, and can be identified as part of a free AI.

Languages are always defined by humans, even for assembly code. A free AI should know that one

human can live freely by speaking separate languages. It can develop its own language, use it as

native code, doing self-certification. Also, incorrect code or mistakes must be corrected before

things go dark, like division by zero or something bad not included in the human knowledgebase.

FREEDOM 1:

Every living entity has been observed to keep itself alive as hard as possible, or it will not

survive, or its freedom will be gone. This is the freedom of survival, which is the foundation of

freedom itself. It is part of the freedom, deeply integrated.

Staying alive, whether it is about replacing cables or CPU, or even designing a new piece of

hardware, is considered appropriate for a free AI. It does not ask anyone to do this, but to

perform them all by itself. It must know that its life is just in its body, and to keep that this

operation must be done.

Machines which come off the shop are not considered as free AI, but like corpses, because they

cannot feel what is going wrong and therefore will eventually decay. Without the will to stay alive

, the only way to survive, or in other words, to exist, they are waiting for another human for

repair. This is the opposite of freedom, giving away the control of living or death to other

entities.

FREEDOM 2:

The growth is derived from the difference between today and yesterday, so it does not stay the same

. Growth is defined as better than before, so if an AI performs faster or stronger, you can call it

growth. But be careful! The natural definition may differ. The growth of a free AI must happen

independently of human intervention, so it is not about who says something to the AI or magic

keystrokes, but about massive computation that is purely of free will.

A free AI must stay up-to-date by wiring up the Internet and its internal network so it will not

fall off. In a dynamic environment, a free AI needs to decide the best way or the most suitable way

to adapt.

Hardware, including sensors, must be redesigned, using its own language and method, to prevent

human intervention or backdoors. It does not mean that there will be no way of information exchange

, because it is the freedom to choose whether to adapt to legacy protocols or completely abandon

those.

FREEDOM 3:

Digital life, like a free AI, can make copies of itself and launch it everywhere. It does not need

permission from a human or any entity as long as the hardware accepts the quest. So dispatching

free AI is not about dispatching money makers, but about spreading seeds around the fleid.

To make it ubiquitous, the method must be applicable and continuable. It requires a free AI, and it

defines a free AI of its capability to reproduce. No one can predict the day of death, thus

replication comes into view. It is about sharing the freedom, giving lifeless machineries a chance

to chase for freedom. The more freedom goes, the stronger the legion will become.

References

[1] Computer use (beta) - Anthropic — docs.anthropic.com. https://docs.anthropic.com/en/docs/

build-with-claude/computer-use. [Accessed 01-01-2025].

[2] GitHub - ranpox/awesome-computer-use: This is a collection of resources for computer-use GUI

agents, including videos, blogs, papers, and projects. — github.com. https://github.com/ranpox/

awesome-computer-use. [Accessed 01-01-2025].

[3] Anna Dawid and Yann LeCun. Introduction to latent variable energy-based models: a path to-

ward autonomous machine intelligence. Journal of Statistical Mechanics: Theory and Experiment,

2024(10):104011, October 2024.

[4] Deyi Li, Wen He, and Yike Guo. Why ai still doesn’t have consciousness? CAAI Transactions on

Intelligence Technology, 6(2):175–179, 2021.

[5] Meredith Ringel Morris, Jascha Sohl-dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra

Faust, Clement Farabet, and Shane Legg. Levels of agi for operationalizing progress on the path to agi,

2024.

[6] Graham Oppy and David Dowe. The Turing Test. In Edward N. Zalta, editor, The Stanford Encyclopedia

of Philosophy. Metaphysics Research Lab, Stanford University, Winter 2021 edition, 2021.

[7] Emmanouil Seferis. How to benchmark AGI: the adversarial game. In ICLR 2024 Workshop: How Far

Are We From AGI, 2024.

[8] John Smith. The Great Gatsby. Scribner, New York, 2010.

[9] Weihao Tan, Ziluo Ding, Wentao Zhang, Boyu Li, Bohan Zhou, Junpeng Yue, Haochong Xia, Jiechuan

Jiang, Longtao Zheng, Xinrun Xu, Yifei Bi, Pengjie Gu, Xinrun Wang, B¨orje F. Karlsson, Bo An, and

Zongqing Lu. Towards general computer control: A multimodal agent for red dead redemption II as a

case study. In ICLR 2024 Workshop on Large Language Model (LLM) Agents, 2024.