## Executive Summary
- The next paradigm shift in computation are AI and AGI
- We want to explore how we can get training and using these models can be distributed, decentralised and incentivized
- We want to explore how Logos Network State can benefit from "baby AGI", LLMs and generative AI

## Introduction
We find ourselves before the breakthrough of the next paradigm shift in computational technology. Digital neural networks are approaching the capacity to integrate the internet and thus the entirety of registered human knowledge, culture and intelligence. Foundational AI, Large Language Models and Generative AI are terms used to describe specific applications of these neural networks, and serve as a jargon to differentiate between the uses.

As a member of the open source community, freedom movement and cypherpunk adept, we see a few threats and opportunities. After researching and deep diving into this matter, the contours of a new project for Logos Innovation Lab are beginning to take form.

**Threats:**
1. Corporate AI
GPU power is fungible. The real value is in the training data.
Corporations like Meta, Reddit, Quora, even linkedIn etc, have a huge competitive advantage in this: in some cases they have decades of people's conversations, ideas, opinions. We already see Reddit for instance disabling access to this data, and making it really expensive to use their api. In general we could say, the ones who have the best training data will be the next Google.

2. Nation state actors rewriting history / future

**Opportunities:**
1. Freedom
2. Emancipation through technology
3. As a defensive weapon

**Logos Network State specific opportunities:**
1. Tools for objectively allocating resources and guide the creation and sustainability of Logos
2. Creation of AI daemons who can do specific tasks
3. Generative AI to create a shared cultural expression
4. In the long run: a virtual environment for Logos citizens to work and live

## References
"The Penultimate Truth" by Philip K. Dick
"The Matrix"
"The Congress"
"Alice In Wonderland"

## Resources
Dreamlook.ai ML researchers are really interested in working with us on this
Lukso also really interested to collab (genAI for designers)
bittensor
PSE!!!! <3 <3 <3 Again PSE is already working on this, i am trying to get in touch with 852dev.xyz
Talked to catsnacks, will ping Andy after the 3rd (holidays)
https://docs.ezkl.xyz/

https://medium.com/@danieldkang

## How it could work / what the project outlines
Training an AI model:
- take a foundational model
- start training it using a specific technique (GAN, diff, etc)
- have traceability, ownership of training data (who is putting in what) (kurate)
- integrate an economic model (bittensor) (for GPU use)
- create a client / api
- make an interface to use it

## Cathie:
> I think you can go off the 3 scenarios of ZKML
> 1. private data, public model
> - applications: biometric authentication, privacy related
> - status: the farthest from production, since client-side ZKP is still not possible for ML operations
> - future: need to wait for better hardware, better proving systems, but ultimately probably the most appealing use case of ZKML
> 2. public data, private model
> - applications: proprietary model as a service
> - status: only can hide model weights but not model architecture right now
> - future: need something like functional commitment where we can "hash the ZKP circuit" as well
> 3. public data, public model
> - applications: ML "rollup", make offchain computation provable -> e.g. verifiable AIGC NFT, verifiable algo trading
> - status: can achieve reasonable models now as long as we have servers powerful enough to prove
> - future: we need some nice applications with good UIUX to push this
>
> Use case No. 3 in my opinion is the most promising right now. We should be able to get a few production-grade applications in the next few months.
>
> As for the direction of research in PSE, we are deeply interested in proof recursion via folding, and the ZKML effort right now focuses on Nova and RNN which could bring powerful models onto the blockchain soon.
>
> I also have a playlist on youtube compiling all the talks I have given on the topic if that is something that interests them:
> https://www.youtube.com/playlist?list=PL3C_huBgySWsdOs5QN5qa5LyWQb4Y1I3I
----
UNPROCESSED RANTING
----

I understand / believe that what builds a nation is a shared culture, a shared expression and understanding of emotional storytelling. It would be so cool if Logos can have a collectively trained generative AI model, which our community we can use to express / give form to an emotional / cultural consensus., but maybe also to create "daemons" or "agents" who can help analyse and give advice on resource allocation etc.

In the conversation I had with dreamlook.ai today, with this machine learning researcher startup dude, we talked about how we could use kurate to create a generative AI model that is built by contributors. The cool thing is that as a practical outcome, you could have the following scenario:

Imagine the model version is 0.45. We sort of know what it is capable of and what it can do. Now I train from this checkpoint on, and I teach it things like let's say a collection of apparel and accessories. Or a new vocal artist. A graphic / photograpic style.

The community tests it out, and if they like it, my contribution gets merged. Because we use Kurate-like stuff, my contribution is anonymous, yet, it can be rewarded, because I can prove I did it so I can collect money. Or tokens, or rep or whatever symbol of value these ppl wanna have. So we have a continuously growing model, with for instance new fashion, new music or voice capabilities, while having an economic model for sustainability. Im other words, the people using the AI model also control the resources the AI model has to grow, and as such have control over it.

(We need to protect ourselves from corporate AI, the only way I see it working is by having a distributed, curated and human controlled alternative.)

The sales pitch would be that this is the best model / community / economy / network state to be in if you are a creator. A fashion designer, a musician / producer, photo generator,...).

dreamlook reply:
Interesting question... I would say approximations are possible, yes. Let's say you take a base model like Stable Diffusion. Then you finetune it on a bunch of images. This gives you a delta of the weights (imagine this as a huge vector). Then later you generate an image from this finetuned model. As you do the forward pass through the network you get the activations which you can then compare to the diff from before (you could e.g. compute a cosine similarity between the two vectors). This gives you a scalar number which says something about how much the change impacted the new results. You could also compute the same against the vanilla base model (Stable Diffusion).

The other (probably simpler) option is to just compare the generated image to the training set using CLIP distance. This gives sort of a similarity measure between a set of training images and a generated result.

mplur.eth
A model that is not controlled by corporations
A model where no human data slaves were used
Etc

As for the training data - we could use something like IPFS or swarm + blockchain or Rollup, meaning we hash the data before it gets added to the data set, the hash is stored in the Rollup by the contributor, registering their key as the origin.

If we can “prove” or validate to a degree of certainty which training data lead to the output, we can add accounting.
Also thinking about making the economy grow with the ability of the model, using a bonding curve mechanic. This would create controlled growth translated in value for the contributors.

The next big thing to solve is then distributed gpu
But bitcoin seems to have solved that 🙂

bittensor also seems to have solved some of these problems?
https://bittensor.com/


## Principles:
As with everything we do, this project must adhere to these basic principles:

I. Liberty
We believe in the sovereignty of individuals. As a platform that stands for the cause of personal liberty, we aim to maximize social, political, and economic freedoms. This includes being coercion-resistant.

II. Censorship resistance
We enable free flow of information. No content is under surveillance. We abide by the cryptoeconomic design principle of censorship resistance. Even stronger, Status is an agnostic platform for information.

III. Security
We don't compromise on security when building features. We use state-of-the-art technologies, and research new security methods and technologies to make strong security guarantees.

IV. Privacy
Privacy is the power to selectively reveal oneself to the world. For us, it's essential to protect privacy in both communications and transactions, as well as being a pseudo-anonymous platform. Additionally, we strive to provide the right of total anonymity.

V. Transparency
We strive for complete openness and symmetry of information within the organization, and have no border between our core contributors and our community. We are frank about our shortcomings, especially when making short-term tradeoffs in service of our long-term goals.

VI. Openness
The software we create is a public good. It is made available via a free and open source license, for anyone to share, modify and benefit from. We believe in permission-less participation.

VII. Decentralization
We minimize centralization across both the software and the organization itself. In other words, we maximize the number of physical computers composing the network, and maximize the number of individuals who have control over the system(s) we are building.

VIII. Inclusivity
We believe in fair and widespread access to our software, with an emphasis on ease-of-use. This also extends to social inclusivity, permissionless participation, interoperability, and investing in educational efforts.

IX. Continuance
We create software incentivized to continue to exist and improve, without the stewardship of a single entity or any of the current team members.

X. Resourcefulness
We are relentlessly resourceful. As we grow and have ready access to capital, it is our obligation to token holders to fight bureaucracy and inefficiencies within the organization. This means solving problems in the most effective way possible at lower economic costs (in terms of capital, time and resources).

* getting a community of fashion designers up
* educating them on how to use Kurate to create designer persona's
* making them create fashion and then train loras
* developers hack a way to get proof of training when their training is used they get paid
* ppl buy nft's of fashion stuff
* train a lora on themselves
* can generate instagram pics of themselves with the fashion on them