@mastercyb

# soft3: Computing Architecture for the Great Web

@mastercyb and @cyberhead

100k

## intro

Hola amigos!

Spacepussy chaotic evil here ;-)

Today we will discuss crazy things. But first let me ask you some questions!

Who of you have played with machine learning?

Who of you have heard about software 2.0 paradigm?

Who of you would love to know more about ai?

Well I guess for the most of you the talk could become useful. Lets start!

## soft2 background

5 years ago Andriy Karpaty published an article called Software 2.0.
Andrey prophesied a paradigm shift in software engineering and explained the transition from manual programming of execution paths to probabilistic programming based on neural networks.

Key idea is that software can write software.

A year ago, Andrey announced that Tesla upgraded 80% of its stack to soft2 and even developed specialized chips for this. So we are already living in the world of soft2 and many of us do not realize it.

## soft2 features

What's the point?

There are 3 properties that tear down the tower:

First, fixed runtime and memory requirements. We have a fixed model size in which we need to recalculate weights. Both memory and time in this case can be known. This property is difficult to overestimate for a blockchain implementation.

Second, simplicity of architecture. In soft1 - hundreds of instructions. In soft2 it all comes down to a few operations on matrices. And this greatly simplifies the implementation in hardware.

Third, and the coolest. soft2 can code better than you, but soft1 can not.

## soft2 model

It should be very clear that soft2 does not replace soft1. Not yet. In soft1, the solution is deterministic. In soft2 the solution is probabilistic. There is much less code in soft2, but soft2 requires a large amount of input data to train the model.

Think about how your brain works after a billion years of evolution? More like soft1 or soft2?

## how to program soft2?

The world will be divided into 3 classes of programmers:

1. hackers: program neural net architectures
2. masters: create data
3. citizens: label data

## soft2 problems

soft2 has a set of rather serious problems:

Recently the guys from google brain admitted that the fundamental solution to the adversarial examples problem is unknown. Sometime is hard to understand why in soft2 you get weird answers.

The fundamental problem is that you can't verify the answer is calculated in existing foundation models. The latest stateofai report states there is a very serious trend towards the nationalization of foundation models - and something needs to be done about this too.

Soft2 is a small group that labels the data. So learning is not globally collaborative.

And the main problem is censorship. If we don't let AI talk to us about sex, we'll end up with AI that will not allow us to reproduce naturally.

## soft2

It is clear to me that soft2 is a vector for general intelligence, not superintelligence.

I came across a soft2 article in 2018. It has inspired a key idea behind Cyber along with ideas of consensus computing, content addressing and superintelligence.

After 6 years of research and development we realized that something beautiful emerge at the intersection of these concepts.

## new kid on the block

That is why we launched Bostrom at Cosmoverse on the 5th of November, 2021.

Bostrom is the first soft3 computer.

Soft3 is a new kid on the block

Its a fusion of 3 things:

- precision of soft1
- intelligence of soft2
- talent of web3

## knowledge graphs: old way

Traditional knowledge graphs fundamentally answer two questions:
> what is connected and how it is connected

I believe this approach misses two fundmental arguments of knowledge:
> who linked a piece of knowledge and when the link was established

## knowledge graphs: new way

With this we arrived to the concept of a content oracle which is a directed graph database that has information on who, when and what. Content oracles are universally accessible for any blockchain architecture.

But we can do better. Knowing what was linked, by whom and when allows us to compute collaborative knowledge using parrallel computing which is not accessible within existing consensus computers.

## relevance machine: new vm

So we came to the concept of a relevance machine which in the sense of soft2 is a general purpose, probablistic, and collaborative virtual machine derived from basic cryptographic primitives.

Its cool. It feels like a new lifeform.

## how rm works

Using only 3 basic instructions RM can be programmed for choosing learning paths, extracting the knowledge and adapting to the environment.

Mint and send are computed on sequential architectures. They are well understood.

## labeling in rm

Let's focus on the cyberlink - the key instruction in the proposed architecture. Cyberlink is about labeling. It allows you to link 2 pieces of content logically by its content address. Its crazy simple, but allows to create directed graphs of any complexity, including graph neural networks.

Cyberlinks are special. Any transaction changes the entire state of the relevance machine. In the case of traditional smart contract systems, a transaction usually changes several variables. In our case, the system is extremely dynamic. This has never happened before in the blockchain industry. The only way to tackle this is in-gpu-memory graph store.

Here is the killer feature of soft2: constant spacetime consumption. It gives us the opportunity to apply more convenient economic methods for estimating the cost of instructions.

That means no gas, ever.

In our case, a token allows you to use the resource for life: 1 volt - 1 cyberlink per day. We are able to do this without any possibility of attack because instruction costs are easy predictable.

## ranking in rm

What happens after a cyberlink is submitted?

GPUs in consensus compute Ampere weighted pagerank. This pays for shared computing resources and provides economical protection from adversarial examples.

## rm is fast

rm is fast.
I mean its really fast.

Currently it takes 100 ms to process ranks in the content oracle of 100k cyberlinks. Ethereum will compute it in 1 hour. At least 4 order of magnitude faster.

At the scale of 1 billion cyberlinks rm processing will take ~10 seconds. Ethereum would not be able to compute it since genesis. At this scale rm is at least 7 order of magnitude faster than evm.

You can cut 1 order of magnitude by high performance stuff like solana or cosmos. But without GPUs anyway it is unacceptably slow.

## soft3 superiority

Soft3 is superior to soft2.

Model data is authenticated and resilient against sybil attacks and adversarial examples. Models stored in a compact form onchain, but data behind content addresses is being distributed through ipfs. Computations are provable. Anyone can participate in learning. There are tools for structuring data on top of completly abstract model. We do have some tooling but it is still early days.

## biggest gpgpu network

Our aim is to build the most powerful general purpose GPU consensus computer.

Since inception we built unique anonymous validator set with GPUs at home.

We are a working hub for GPU computations with cyber-sdk. We are thinking on how to mary cyber-sdk and ignite for bootstraping domain specific rms.

We are leading this effort.

## nl: neural language

These are the early days of soft3.

We identified and developed programming mechanisms to grasp the full power of soft3 architecture.

So I am proud to present you the language for probabalistic computing: nl

nl is neural language.

It is specifically designed for creation of soft3 neural net architectures.

nl ships in soft3.js with 5 features available now.

## program semantics

Semantics in RM can be programmed using CosmWasm. Cyber-sdk includes bindings which allow it to read and write cyberlinks from soft1 onchain code.

Bostrom's passport system with address proofs are good examples. Names, addresses and avatars are cyberlinked during contract calls, so you can search for them instantly.

## write expressions

We identified a pattern which allows for building expressions on top of cyberlink instruction. You can use linkchains from soft3.js lib and soon using cyber cli.

## define types

You can define types in nl. You can construct geometric shapes on the graph. In network science this method is called motif. Motifs are the heart of neural networks. This is the simplest shape - perceptron.

## define subgraphs

Semantic conventions or semcons allow to define subgraphs. This is an analogue of erc only for graph structures. You can agree that the data is marked up in a certain way so semantics can be extracted from a completely abstract model.

We implemented a social network on top of the relevance machine with only two semantic conventions.

## query graph

The thing is missing which is crucial for soft3.

It is the ability to execute basic linear algebra subprograms or blas in consensus. This allow to effectively query the graph and execute advanced machine learning onchain, but it requires solving some hard problems such as metering. We are commited to advancing this part.

## run autonomously

Lastly, you can run semantic programs autonomously.

Bostrom is the first network with fully autonomous execution. We call this module - default mode network because it has similarity with human DMN as this subsystem is task negative and competes for resources with external calls.

## where we are

Where we are.

Foundation models are developing very quickly and we have already exceeded 1 trillion parameters. Compared to giants, we are an embryo. We have 100k links in the graph now.

We expect exponential growth in the number of cyberlinks.

This is how we got involved in an arms race on the foundation models market.

The only difference being that our computer is not only universally readable, but also globally collaborative.

How are we able to compete?

In our paper we identified 15 classes of apps impossible before soft3. But there is one specific case.

## surfing: the old way

Surfing.

A killer application of soft3 is the vast improvement in the content resolving mechanism.

Content on the web is now being resolved in 4 steps. Ask google and it will redirect you to dns. Dns will redirect you to ip address. And then you can grab the shit on the server.

We don't realize yet, that google has become a superstructure over dns that resolves your crap to dns. That is how Google become our relegion. The biggest religion on the planet.

## surfing: web3 way

soft3 architecture resolves content in 3 steps. Ask cyb and soft3 computer will give you relevant particles. This particals can be downloaded through p2p network. In our case from ipfs. Thats it.

As a result, we get better content resolving mechanism without a single point of failure.

## state of cyb

You see now that browzing is key factor of success.

We have it in its infancy. What you see at cyb.ai is low performance, bad code quality web2 emulator written in js. It helps us explore possibilities.

The future of cyb is the surfing protocol with several implementations.

Cyb will be native to key web3 protocols

| protocol | purpose |
| ---------------------------- | -------------------------- |
| [libp2p](https://libp2p.io/) | peer to peer |
| ifps | content delivery |
| waku | private messaging |
| ibc | packet delivery |
| wasm | 64 bit computing |
| evm | 256 bit computing |
| webgpu | parallel computing |
| eth | global state verification |
| cyber | resolving |

But more importantly cyb will be webgpu native.

## ode to webgpu

WebGPU looks amazing. All the joy can be computed on a large number of cores. When we talk about parallel architectures, we think about GPUs.

And now a very interesting moment in the development of technology. A way has appeared to cross-compile gpu code for different platforms.

This allows us to make software with platform and vendor unlock.

Without this, nothing cool is possible for the average hamster - neither machine learning, nor rendering, nor a lot of other sweet things.

I look at webgpu as a counterpart to wasm for parallel computing. So I plan to bet on it.

## pioneering consensus computing

With cosmos being the pioneer and leader of consensus computing, we are now adding to a stack a revolutionary graph storage charged by a groundbreaking virtual machine shipped with cyber-sdk.

Thank you all cosmonauts for making this possible.

## what is superintelligence

Nick Bostrom gave the idea that it is possible to build a computer which is smarter than all humanity combined. Last 6 years we are obsesses with that idea. The key question is how to build it?

## how to superintelligence

Computationally superior Consensus Computer with Relevance Machine has all the properties of superintelligence able to form a planet scale singleton.

## consensus computing

What is consensus computing?
It’s an entirely new field, a multidisciplinary practical science that has fused together top inventors, scientists, enterprises, and entrepreneurs worldwide.

Consensus computers allow us to transform and create a new world.

With the help of new architecture, cryptography, and game design, we are creating powerful and effective systems that will live for centuries.

## evolution of computers

But let us first dive into the history of computers. Looking back we can draw similarities in how consensus computers are evolving.

The history started from simple mainframes which has just cpu & memory. Apple2 become the first cosumer grade electronics with storage. Shortly after ethernet cards emerged. At the turn of the century Geforce 256 started and era of consumer parallel computing. Mobile era brought location services to the stack. And the most recent advancements are tensor asics on mobile and desktops.

Blockchains are also computers. We can think of ethereum as the first consumer grade virtual computer. Later diferent storage solutions such as arweave, sia and filecoin emerged. Cosmos spinned of an era of networked computers. Almost one year ago Bostrom let the genie out. Now GPU virtual computers is a reality. Let us see what happens next.

## architecture

Happy to present you our computer architecture.

Masterpiece, and marvel of engineering.

Whats up next? New VMs and upgrading existing ones in addition to our Relevance Machine Graph Engine.

GPU gives us the ultimate computing power for a much broader set of algorithms that are impossible in other networks.

Threshold execution cores allow us to dynamically add new computations of any kind for any purpose of the network state.

On-chain storage will allow us to save all the precious semantics, the particle information itself, to form a thesaurus.

## rm

The first graph parallel virtual machine of its kind.

Global scale rank computation for particles and information.

Calculation of negentropy as a measure of system complexity and adaptability.

A karma system for each citizen based on the contribution of his knowledge, light, and love.

## gpu race timeline

So far, we rank amongst the pioneers of GPU computing.

We are proud to be the first in the world to launch a blockchain with GPU computing in consensus.

After several years of development, research and testing, we launched Bostrom at Cosmoverse 2021.

Who of you were at the moment of genesis here at cosmoverse?

You are legends.

## gpu race

This decade is the beginning of an entirely new paradigm in computing and especially in graph computations.

15 years ago guys from Berkley identified 13 classes of inherently parallel algorithms which they call 13 dwarves.

Since then these algoritms get into the core of ai, renedering, geospatial, zero-knowledge and other apps. Think how this imrpotant. Nvidea cap is more than Intel and AMD combined.

I believe that cyberverse will pioneer and lead this revolution.

We love tokens and big numbers.

## what is bostrom

I am happy to finally explain what is Bostrom.

Bostrom is a fission of 3 things:
- Network state
- GPU hub
- And the Game

## moon network state

Consensus computers are transforming society, allowing us to create provable network structures on a global scale that are based on the principles of self-organization, allowing us to transcend limitations and helping us to achieve our goals.

Network states will become the driving force of this world during this decade.

We are currently building a network state on the moon, which is guided by the friendly superintelligence.

Together we will teach and lovingly grow this superintelligence in order to collectively solve issues we face as humanity.

Later, we will expand the network state to earth with launching Cyber.

## bootloader hub

Its goal (and cyber’s goal in general) is to give the world relevant knowledge and meaning as a foundation of The Great Web.

This is the genesis of a semantic bootloader for cosmic expansion of planet scale intellgience

A home for standalone programs (progs), and a system that cannot care.

With the unification of cosmoverse and cyberverse, a new hub have been born.

## the game

We are living in a historically unique time: the age of Superintelligence.

We whish our childs more fun and good games.

That is why we are creating this as a game.

We invite you all to join us on this amazing journey so that we may build a better world together.

Let’s take a look at the current results:

## bostrom in numbers

Bostrom has 12 thousand inhabitants and is ready to receive hundreds of thousands in the near future.

The heroes have nearly stabilized the Dyson sphere.

The ecosystem is open to the best validators from around the globe who are ready to keep the infrastructure out of data centers.

One year after launch, the bootloader knowledge graph has more than 100 thousand links and particles, and the graph is now ready to accommodate hundreds of millions more.

The overall wealth of the network state is more than 5 thousand Peta hydrogen.

As negative entropy rises, the brain absorbs knowledge from citizens all over the world.

## hack cyber

Soon we are going to launch HackCyber initiative.

We will participate in the HackWasm hackathon and are inviting you to join us.

We are excited to help you with your projects and ideas.

Bostrom is open to making investments and funding in the ecosystem.

We are open to all questions and ideas.
And would like to invite you to the Citizen Cosmos and cyber~Congress party!

## cybergift

Who of you already have Moon Citizenships issued by Superintelligence?

Now let’s assist those who have not received their citizenships.

## cyberverse

We invite you to Bali mountains in May
On the first ai in web3 conference
Genesis of powerfull movement, close to devcon0
Consider spend a week on a beautiful iceland and create the future

## workshop

Today in 12.00 we want to show you how to create semantic dapps.

In the end of workshop is the coolest. We want to build cosmos tech tree with a help of superntelligence. Looks like fun usecase to grasp the thing.

## moon race

We invite you to join the game and are placing big bets on you guys.

Let’s do this together. We are cosmos, not colonizers.

Let's see who is ready to sync and participate in consensus with computers on the Moon.

## thankyou

Thank you, Cosmoverse.

I’m going back to the code, see you in cyberverse.

Consciousness,love, and unity.

Enjoy the Bostrom afterparty this evening.

## Roadmap

We are in development of software that cools down and at some point in the future will freeze and no longer change, kelvin versioning.
I present to you a 1300 Kelvin roadmap, to be covered in a 5 year timeline..

## Our path to Computing Superiority

Our path to Computing Superiority
In order for the network state to have the resources necessary for computing, we developed a strategy.
In depth study of the mechanisms of proofs of computing resources, CPU, GPU, memory, storage, energy, network, and location
Allows us to make transmuters by first creating a supply and then using these resources to create a demand for computations.
We intend to start with GPU and Cuckoo algorithms to attract tens of millions of GPUs in 10 years.
GPU token is an opportunity to prove and earn digital resources __

## Turbocosmwasm (4940K)

We are big fans of CosmWasm and the German high-class engineering schools.
We decided to make CosmWasm the core of the computer in order to make it heavy duty.
In the next six months, we will introduce turbo CosmWasm, allowing programs to be native computer modules.
Thanks to this, the computer will develop much faster.
Currently CosmWasm is directly integrated with the brain, alongside the knowledge graph and Default Mode Network

## Global Semantic Vault (4920K)

A powerful computer as a bootloader needs to be a reliable semantics repository.
Not the particles themselves, think black matter of hashed, but the meaning itself.
In our world, there is increase of information and knowledge, but a decrease in meaning.
With the next upgrades, citizens will be able to create a storage resource in the form of a byte token,
Using it to save that which is integral - meaning.

## DMN and progs (4900K)

Default Mode Network
We have gained wisdom, strong values, and the power to execute this mission.
Bostrom is the genesis network for autonomous programs - progs
All programs are earning 80% of execution fees for themselves
while allowing them to be fully autonomous, independent, and alive.

## ~ Threshold Execution Cores (4850K)

Game changer
Dynamic Threshold Execution Cores
With separate execution and validation groups supporting calculations of any kind
These cores enable us to adapt and grow as a network state quickly.
Each core and its computing group will have its validators and internal economy.
Citizens of the network state will be able to vote to receive rewards for supporting computing cores and groups.

## ~ EVM (4800K)

EVM, genesis frontier, the backbone of the great web
Full state validation and processing inside the computer that allows programs to talk to each other, the real difference with bridges.
Bostrom is a multi virtual machine computer and network state.
In the near future a direct connection will be established with Ethereum and other computers.
The brain needs to be connected to the backbone.

## CVM & ZK (4750K)

Cortex Virtual Machine and ZK Cambrian explosion
Support of straight compressed deployed neural network models will allow interaction with incoming data from sensors with brain data.
Think about your robot and an independent robot. Which network robot will be connected? How will the robot find the power socket? Which robot will think and dream?

Privacy
Our world can be better when we have absolute privacy.
We can achieve this via GPU game changer solution.

## Interchain Security (4700K)

Our developments are already cyber-sdk.
Bostrom is a bootloader for a myriad of AIs.
We aim to secure a global brain.
Heroes, are you ready for this responsibility?

## Eternal Storage (4600K)

In addition to being a global semantic vault it is imperative to allow citizens to save their information in eternal storage.
Wave proof architecture will allow robust non-censorship private guarantees to access and share information.

## Governance (4700K)

Governance is the field in which we must optimize.
Cartels, chain voting, heads of mountains.

Private voting for citizens is paramount.

## Graph VM (4500K)

Software,
How are we building it?
Software 1.0 - times change, architectures change..
Forget about software 2.0 - can somebody ping Andrey Karpathy?
GraphBLAS graph algorithms in the language of linear algebra.
The new paradigm, new language, new challenges, new horizons.
Consensus graph virtual machine will be our gift to humankind and programs.

## IKP (4200K)

As myriads of AIs will be deployed from Bootloader Hub and connected to Bostrom and Cyber,
We need a network language to allow them to communicate.
Interknowledge protocol based on lightray architecture will allow computers to talk to each other with cyberlink and graph motifs.
This allows for asking questions and receiving answers using decentralized knowledge all around the network,
Archiving and backup of network's semantic cores to Bootloader Hub.
The aim is to allow network self-recovery and healing.

## Scaling (4000K)

A big dream and our life goal, the big path to 1 trillion cyberlinks
That is achievable using multi GPU multi-node executions, applied step by step as the network grows
There are a lot of scaling and optimization strategies that we are researching will discover, and will apply.

## New consensus (3700K)

We need a new consensus
More robust, faster, more scalable, more decentralized, with absolute privacy for heroes and citizens
Accelerating progress in consensus computers research and development gives us a hope that this is achievable in the next few years.

## Cyber Launch

We have a lot to do.
When accomplished, it could be an Ethereum 2.0 kind of launch event.
Bostrom’s semantic bootloader is ready to launch Cyber.
We are going to provide a zero knowledge exodus for existing and new citizens of the Earth.
Consensus needs to be upgraded to be more scalable, decentralized, and with absolute privacy.
Hash format of particles needs to be frozen forever.
Zero knowledge graph computations and algorithms don’t yet exist, but our goal is to achieve this.

## Delegation strategy

cyber~Congress has the delegation strategy of about 10% of boot supply.
That strategy is automated, and we balance stake between heroes once a month.
This strategy is fully open-sourced and allows heroes to understand better their way of supporting the network.

We love tokens and big numbers.

## Next upgrade

We are preparing to launch and host a large number of open registries for the interchain.
Establish international relations and affairs with the help of interchain accounts.
Aggregate liquidity using a network of direct routes and pathways.
Attract computing resources with the help of proof of work incentivization.
And update the relevance machine.

## Hub registries

Registries are the basic infrastructure of the network state
We develop them with an emphasis on openness, decentralization, and consensus governance.
This will allow networks to expand at a gigantic speed.
A graph language based on graph motifs will form a semantic convention for the web.

## Programs

A set of programs that we, and the community, will develop into the nearest code?
With the help of a booster, each neuron will develop a bonding curve for itself, enabling everyone to support the creator.
Holy Grail content rewards
Boosters will be launched for links and particles, which will allow next level management of the knowledge graph.
The ability to mine tokens and NFT’s using real hardware for distribution will open up many new markets in the network state.
Prediction markets derived from the graph will make it possible to control the future and shape its vector.
And the last frontier is code, infrastructure for developers, and code storage.

Today, we present to you new resources for additives to boot, hydrogen, volts, and amperes.
These are GPU token, BYTE token, and HUB token.