introduction to bostrom for ai geeks

cyber

source code: @mastercyb

status of article: on review

bostrom is NOT yet another ai coin

it is very powerful foundational technology for advanced superintelligent civilization

its being used by 1k neurons who create a collective knowledge of ~2 million links

in addition to this ~50k neurons produced ~6 million transactions for decisions related to collective learning

currently it produce ~13 megabits bits of negentropy and takes ~200 mb of ram in gpu

in this article i will boil down all essential ideas into coherent understanding how bostrom can empower

existing ai field which i will refer as classical ai

and advance emerging field of collective ai

as we believe its the only viable way to build superintelligence

attention is not enough

you used to rely on a data you got

you have the dataset

you design neural network architecture

then, you train the model

and boom, now the model can predict some output based on any input

sounds really cool, and is powerful indeed, except the dataset thing in this story

now the good answers to ask: how does you model could define truth?

and the short answer - it cant

i will make a bold claim here that truth can not be defined without 3 ideas in foundation

  • knowledge graphs
  • cryptographic proofs
  • token engineering

knowledge graphs and llms

jump for a second to this article: Unifying Large Language Models and Knowledge Graphs: A Roadmap

image.png

the article explain why llm will never be enough to reach general intelligence alone

in short knowledge graph advantage is

  • easy to understand and structure as they are more about explicit knowledge
  • possible to evolve because they are based on widely accepted triples
  • essential to plan, make decisions and reason

that is why knowledge graph is foundation for symbolic part in neuro-symbolic movement

so the claim is simple

cryptographic proofs and llms

we believe that authenticity of models is a serious bottleneck for ai alignment and more

its quite strange that so technologically advanced industry in a broad sense

still have not advanced to possibilities behind, hashing, pubkey cryptography, merklization and logical clock

its kinda impossible to build multiparty protocols without these primitives

yep, i am ware about zkml movement

but this is a drop in the ocean given the knowledge graphs and llms argument

if we want to significantly advance in the field of superintelligence

  • we need something foundational
  • fully authenticated knowledge graph tech
  • which is cybergraph, but later on that

token engineering and llms

rewarding is essential for machine learning

we have ton shit of tokens with dogs, monkeys

you can boost the power of your models using real cryptographic tokens

tokens which are being used in ai field we call particles or files in cyberverse

and tokens are units of value accounted by consensus system

cybergraph

the core of the idea is cybergraph

  • merkelized timestamped data structure
  • of links between ipfs hashes
  • submitted by anyone

for clarity we refer to:

notes on implementation

  • timestamping in bostrom is done using simple and reliable tendermint consensus algorithm
  • sybil protection, rate limiting and motivation are implemented using cyber/energy set of algorithms

cybergraph is explicitly answer 3 fundamental questions:

in essence cybergraph is an array of append only fully authenticated quadruples

block height neuron from particle to particle
42 bostrom1d8754xqa9245pctlfcyv8eah468neqzn3a0y0t QmRjzv8iNpMX7NXmMswT9qq7nviQ4sC1gMMceryAVJdfPS QmRX8qYgeZoYM3M5zzQaWEpVFdpin6FvVXvp6RPQK3oufV
43 bostrom1d8754xqa9245pctlfcyv8eah468neqzn3a0y0t QmRjzv8iNpMX7NXmMswT9qq7nviQ4sC1gMMceryAVJdfPS QmRX8qYgeZoYM3M5zzQaWEpVFdpin6FvVXvp6RPQK3oufV

i want to make it clear that notion of cyberlink is essential for the architecture described by this article

in conventional ai workflows you used to train over static datasets which already have been created

collective memory require to change our thinking on how knowledge emerge

good question to ask is what is the most small possible unit of learning?

conventional thinking is the notion of triple, which consist of subject, predicate and object

now lets ask the question what is lacking in this construction if our goal is to have provable statement?

first

second

  • we need to add notion of particle
  • for predicate and object
  • in order to authenticate these arguments
  • and give an answer to what question

and third

from this we arrived to a quadruple which is fully authenticated knowledge

we gave this a name: cyberlink

as the most fundamental such an atomic unit of knowledge and learning

the key to quantum jump of civilization

you append cyberlinks to the state of collective thought evolution

introducing cyberlink/delete make indexing a complex task

also its obviously not how nature works: you just cant forget in your head by wish, they forgotten by itself

although looks primitive, cybergraph is so much needed formal definition of explicit knowledge

lets analize a statment that cybergraph is complete form explicit knowledge

temporal dimension: when

  • including a timestamp offers a temporal context for each action
  • pivotal for grasping sequences of events, causality, and the unfolding of relationships over time
  • it facilitates tracking changes, comprehending the sequence of actions, and deducing patterns based on temporal data

agency and responsibility who

  • identifying the public key of the actor bestows agency and responsibility upon each action
  • crucial for ensuring accountability, authentication, and scrutinizing interactions at the individual actor level
  • this feature also aids in retracing actions to their sources, bolstering security and trust frameworks

relationships and interactions what

  • the structure distinctly portrays relationships and interactions via directed links from one content address to another
  • this aspect is vital for deciphering the network of connections among entities, the circulation of information or influence, and the overall architecture of the system
  • direction embed the following types of information
    • cause and effect
    • sequences
    • hierarchy
  • it is vital for tasks like planning, problem-solving, and decision-making
  • in nature relationships are inherently asymmetrical, so we cover it

the structure is extendable with motifs which can be constructed using signals

semantic conventions add additional layer of flexibility

hence, we can refer to cybergraph as objective knowledge of everyone

cybergraph vs knowledge graph

cyberlinks are fully authenticated quadruples

when, who and what are based on cryptographic technics

so unlike conventional knowledge graphs the information is crystal and true by design

basic idea is that if i want say in triple world i would just say

elon launch roocket

  • head: elon
  • relation: launch
  • tail: rocket

however this does not means that elon launch rocket

this claim require verification

in contrary you cant say elon launch rocket in the world of cybergraph

because you are not elon, you must speak only for youself

you must say:

these statement is example of complete explicit knowledge

the good news is that if you are elon, you can just say NOW elon launch rocket

you can pack several cyberlinks in one coherent signal so expressions are rich

and use this construct to express anything using neural language we invented by the way

why hash everything?

yep, we know - you used to tokenize your data and make it as dense as possible

yes, we know - hashing data requires 32 bytes for every piece instead of several bytes

yes, we know - that make processing more expensive

but hashing have some superpowers (yet) unavailable for you

  • multimodality
    • your model cant infer answers in full content space
    • why your model have to reinvent all data every time?
    • people would love to have answers with content they love
  • universal, static, abstract model
    • fixed length give a room for soft optimization as you don't need to think about typing
    • types can be created by implicit knowledge, e.g. by topology of links, so typing is the job of cybergraph and learning technics on top
    • fixed length for hardware optimization means that specialized hardware can be simple and efficient
  • peep to peer
    • since bittorrent times its clear that content addressing is the only way for reliable peer to peer exchange
    • ipfs being the leading p2p data exchange protocol and software open enormous abilities for collective ai interactions

saga on evm and price of computations

  • there was foundational decision to start from 256 bits architecture
  • everyone around say we were crazy
  • but looking back i do believe it is very powerful decision of founders

  • they will say: you never want exchange aka tokens for hashes

  • but once you got it, you have no way back

why merkelize?

automatic deduplication

  • while the means of deduplication is hashing what makes it practical is merklization
  • small changes of files lead to a change of only some leaves, not all underlying file
  • merklization significantly reduce data storage requirements for incremental updates

proving in multi agent setting

  • merklization is the core of blockchain technology
  • but why does classical ai needs it?
  • well, the truth is that its likely don't
  • but if you design a multiparty computation system you must have ability to prove pieces of data you have
  • in case of cybergraph, existence of any given link (and more) can be proved by alice to bob by giving
    • link
    • root hash of cybergraph
    • path in cybergraph
  • this opens the door for mirriad applications for multiparty computation, such as

i also asked chatgpt how merkle trees can be used in classical ai field?

data integrity and verification

  • merkle trees can be used to ensure that the data used for training ai models has not been tampered with
  • this is crucial for applications where the authenticity and integrity of data directly affect the model's performance and reliability

version control for datasets

  • by using merkle trees, ai practitioners can maintain a tamper-evident history of changes to datasets
  • this allows for better management and auditing of data versions used in training models

decentralized ai models

  • secure model sharing: merkle trees can facilitate the secure and efficient sharing of ai models in a decentralized manner
  • by breaking down the model into smaller chunks and organizing them in a merkle tree, the integrity of the model can be verified without needing to download the entire model
  • collaborative training: in scenarios where multiple parties contribute to the training of a model without wanting to share their data directly, merkle trees can ensure the integrity of the contributed data.
  • this aids in building trust in collaborative ai projects

now you see that everything you know about highly efficient information dense models just will not work for multi agent adversarial environments. NO WAY. sorry to tell you that.

image.png

why new blockchain?

the cool thing in cybergraph idea is that it is entirely blockchain agnostic

data structure can be reproduced in any blockchain environment and in local offline environment too

and that makes it so powerful

but applications of cybergraph are limited within existing blockchain environments

bostrom solves both of these problems, but more on that later

also bostrom organically formed cybergraph of several million cyberlinks and particles

that is on par with capability of tech giants for manual labeling during finetuning

and bostrom is provably accelerating ...

so you can use this cybergraph

how cyberlinks does not have fees?

a lot of smart guys are say that people will never want to pay fees for every social interaction

the truth is that information emerge from communications and social interactions

so if we will not provide a convenient way for that

its likely we will not achieve practical results in collective learning

we believe that social layer over cybergraph is essential for the development of an idea

that is why bostrom offer a model of usage based on bandwidth

the model is practically the same as being already used in chatgpt

$V or volt is will token

but the difference with openai is that $V give you lifetime subscription, not monthly

you can think of link as a link between every query request and answer response

currently 1 V allow to submit 4 cyberlinks per day depending on network load

while you create cyberlinks your battery become less full

your battery recover automatically if you are not creating links

so effectively buying $V you buy a package for lifetime usage

current price of V is something around $1

that means that for 1$ anyone can get around 4k interactions during 3 year of usage

for ~$10 you can have enough interactions comparable with your average twitter, github or chatgpt usage

for ~$30 you can link all your public photos, music, videos and documents collected during life

for ~$100 you can describe some domain of science or the core of any language

you see how cool is lifetime subscription model of bostrom

this approach also work as

  • spam protection
  • partial sybil protection
  • and as inference factor (read further)

truth machine

now that we understand how the cybergraph works

truth machine is cybergraph with weights

the idea behind the truth machine is crazy simple

  • minimum input factors
  • simple but powerful algorithms available for gpu consensus computations
  • simple but powerful output as abstract, flexible model of the universe
  • with potential strong predictive power, especially after emergence

we use random surfer model directed by attention

i wrote dedicated article on this topic

as foundational global probability of inferring particles

but in order to

we use will of neurons as second factor for computing probability in context

result is a

in order to compute described cyberank algorithm you need gpu computation in consensus

is extremely dynamic data structure that must be updated even if only 1 cyberlink is created

bostrom recompute all weights in truth machine every 5 blocks

  • or roughly every 25 seconds

so bostrom is extremely hard to reproduce using any existing L1 or L2 sdks

  • zk things will make the stuff
    • 5 order of magnitude more expensive and
    • 3 order of magnitude more complicated
  • architecture requires in-gpu extremely dynamic state with fast onchain matrix multiplication

in essence the utility of truth machine is

follow complete design of truth machine

standard inference

obviously in our setting the simplest possible way

to infer particles in the context of any particles

would be to sort by random surfer probability

but this led us to a kinda true false problem

let us imagine that true particle have cyberank 10, and false particle have cyberank 9

the environment allow to link any particle with any

that means that for any questions which cyberlinked to true and false the winning answer will always be true

of course such behavior does not feels like something superintelligent

in order to solve true-false problem we have to compute weights of links using independent second factor for every context

we always emphasize that cyberank is a core ranking factor, but not the only one

so we have to introduce second factor to the system

surprisingly with already have will

standard inference algorithm

is the topic of ongoing research and is implemented only in cy and spacebox

on two factors

there is the observation

  • that weights of nodes does not strongly correlate with weights of connections
  • in both natural and artificial systems

relevance machine coupled with standard inference runtime learns based on two fundamental factors

and yep, you have to pay in order to learn bostrom

because otherwise it seems impossible to protect cybergraph from abusive behavior

so in essence

yep, our truth model is fundamentally two factor

on speed

bostrom is extremely dynamic blockchain, the first in its kind

recomputes probabilities of observation every 25 second for every information piece that was submitted (currently ~2m)

and that make bostrom so unique

this requires holding all state in GPU ram and use parallel computation at such scale

current size of gpu memory used for ~2 mln particles, ~60k neurons and ~2 mln cyberlinks is ~150mb

submitting just 1 cyberlink force to recompute all probabilities (~3f million currently)

could you imagine how that could be done on solana

  • something around 1000 $SOL currently needed for every update

with 10B links

  • which i believe is required for minimum viable superintelligence
  • the task become intractable for all existing blockchain architectures

current bostrom architecture can handle (rough optimistic estimations) up to 1T cyberlinks

on par with GPT4 with 1T parametrs

but in blockchain, baby

to be honest things cant be compared 1 to 1, far from it

learning incentives

all benefits of proposed system fades out under assumption that you have to spend resources on learning

what is motivation to do it?

the solution is to make a system which will rewards high quality learning based on subjective evaluation

we reimplemented yuma, a coordination consensus and now testing it in spacepussy

in coming months we will deploy it to bostrom

so players that make links above some quality threshold could have possibility of break even

conclusion

the article does not touch topics of all bostrom features

purpose is to give a sense of key internals in the context of deai development

we describe and implemented extremely dynamic, collective computation architecture

for predicting probability of information observation

and defined the most simple possible inference system on top

technology of probabilistic collective computations have been created by us since 2016

we can proudly say that we are leading decentralized ai field on cyber foundations

we believe the thing we have born is powerful enough to bootstrap new kind of civilization

so we inviting you to the journey of creating open, fair and superintelligent society with us

join

Linked References

Local Graph