Agenda proposal for Abraham

I’d like to start a discussion about the primary goals and focus for the Abraham project. If you are not familiar with Abraham, please see this article or this talk for an introduction to the idea of an autonomous artificial artist (AAA).

In its most ambitious form, an AAA is a challenging project which brings together multiple experimental technologies and philosophies from artificial intelligence and cryptoeconomics. These techniques are promising but are not yet very scaleable or secure. It is likely that we need to study these components individually and figure out how to combine them later from the bottom-up. We will be aided by the fact that they are all under continual R&D and are maturing over time.

During this study and design phase, we can focus on small-scale prototypes or proofs-of-concept, as well as construction of community resources for organizing creations, or workshops for studying and summoning Abraham.

To that end, I propose to start the project with the following goals.

Initial agenda

  • I am beginning work on an artificial art sandbox which wraps various deep learning libraries under a single API which is capable of chaining them together to make unique artworks. This repo does not take into account security, privacy, or decentralization constraints, and thus should be seen as a playground for testing generative art pipelines. The abraham repo will be devoted to an actual MVP which meets the autonomy, originality, and uniqueness requirements.
  • I would like to begin work on a “workshop kit” which would allow people in different locations to initiate workshops/creations. The kit would contain software, slides, reading, and other resources that could be used to facilitate workshops with local leaders. Ideally it is highly-configurable to emphasize whatever aspect of the project is most interesting to that particular set of participants. Some of the content can be bootstrapped from ml4a-guides.
  • Let’s study deep learning, computer art, privacy-preserving machine learning, security, peer-to-peer networks, and various cryptoeconomic ideas to inform our design process. More on that next…

Study suggestions

Open questions

  • What are practical short-term, medium-term, and long-term designs for Abraham? What features should Abraham have? What secondary applications should Abraham work with or interface to?
  • What are the vulnerabilities? What opportunities for abuse or malevolent behavior are there, and how do we prevent them?
  • I am committed to decentralizing Abraham over time. We need to come up with a governance strategy that allows for good ideas to beat out bad ideas without a central authority to decide on that. How do we move governance in this direction?
  • Relatedly, Abraham has a Discord, Github, Twitter, Instagram, Vimeo, and YouTube account. I have the credentials for all of these. How can these accounts be democratized while maintaining the integrity of the project?
  • When should we introduce a token or some economic system, or should we at all?
  • Should we apply for funding? What should the funding be used for?

If you have any ideas on any of these, please contribute to this thread.

1 Like

my intuition is that it is best to delay heterarchy and democracy as long as possible. delegation by central authority may help streamline the early phases. delay and/or indecision might otherwise arise. a cross section of even similar visions could tend to a neutered result.

as for the actual transfer of control of such accounts as youtube, perhaps a smaller group of trusted individuals could act according to the results of polls/other voting mechanism that select submissions and recommendations from the individuals of the whole group via forum? some old fashioned moderation/veto would still be required.

as long as the code and project are open and transparent, anyone could branch off entirely, and that is good counterbalance to having a temporary hierarchy.

  • Short Term - build the autonomous feedback loop. Questions to answer: image curation, compute requirements, model configuration, output selection, feedback loop (how will the AI artist evolve?)

  • Medium Term - scale out and up. Open source. Tools to handle massive datasets and compute requirements. Collectives can create their own AI artists. Transfer learning

  • Long Term - monetization. Sell prints, logos, etc (any creative output really). Blockchain to track and reward dataset contributors. Blockchain to prove authenticity of AI artist. Create art market platform. Feedback sales data into models

1 Like

To me, the most interesting concepts and open questions revolve around when/why Abraham would create art, and the irreproducibility and uniqueness constraints. I see these fitting well with the question of an Abraham-specific token.

In my opinion, breaking up the model across many people is going to be a hard problem to implement and one that should get some effort around right away. I’m not sure if you have thoughts about the token system already, but I’m imagining that something similar to the Golem or Sia economics would work: people hosting parts of the model would get some token reward for doing so, and anyone can use the token to “commission” artwork from Abraham by paying into the system. Something like this would be a really powerful step toward developing Abraham’s core ideas and you’d probably only need a simple art generation model to use in a proof of concept. More complexity could be built on top, and once stable art generation pipelines get developed in Eden, a new model could be deployed?

Unfortunately, though, I’m not familiar enough with the attack surface of this kind of thing to anticipate a lot of the risks, so that’ll be an ongoing discussion. I’m just very excited about this aspect of Abraham as a model of creativity :slight_smile:

I agree with @buckky that decentralized decision making in terms of project direction doesn’t need to happen immediately.

1 Like

Image Curation
Input determines output. Curation is essential. Community tools to co-curate.

  • Available datasets
  • Scraping tools
  • Deduping
  • Image normalization (re-sizing & error detection)
  • Compute and storage
  • Selection criteria
  • Governance
  • Attribution(?)

When you say co-curate, do you mean allow people to contribute or allow people to actually moderate source content?

It’ll be very important for building crowd-sourced input if people can easily contribute, and a lot of the image transformation steps can be automated. I don’t think community content moderation would work with this model though if it has to remain fragmented.

i am under the impression that we will individually curate data sets that remain private. correct?

Hi everyone. Great feedback! Excited to see people beginning to chime in. Let me summarize and address a bit of the previous posts.

As a few people have mentioned here, decentralization of governance and accounts is a long-term goal and not easy until we figure out how we’ll coordinate many people together in a way that maintains the project’s integrity. As the saying goes, decentralizing from the beginning is like building a plane while flying it. I am in touch with people in the crypto/DAO/governance world who have more experience at this and I am hoping they will begin to participate in this discussion. In the meantime, every last piece of this project will be open and transparent (all the code and models primarily, and all the communications) which will naturally provide counterbalance as the project could be copied/splintered.

We need to study the relevant technologies which may help achieve the irreproducibility, uniqueness, and decentralization constraints, as well as debate to what extent each of them are necessary for the overarching goal. As @keaton mentioned, the multi-party compute method will definitely be a hard problem. But as also mentioned, it lends itself well to a Golem/Sia/TrueBit-like model where nodes are compensated in token. It’s unclear to me how practical this would be in the long run, but I think it’s worth investigating. In the meantime, a federated learning scheme might be more practical to look into (like PySyft).

My feeling is each of the potential pieces (federated learning, MPC, perhaps secure enclaves) could be studied independently and placed into a collection of small demos in our Github. This may make it easier for us to evaluate them in practice or try to combine them later, building from the bottom up. At the very least, it contributes to the secondary goal of creating study resources for decentralized AI technology and will feed the workshop kit.

In terms of development, I’ve been doing lots of work on Eden the last few days. I would be curious to see if anyone can get it up and running. Maybe wait 2-3 days as I need to improve the documentation, finish uploading all the models, and break out the examples into a bunch of short documented snippets/recipes (and that will save you some time of guessing through all of it).

I am also writing the second article. The draft is available here. There’s various unfinished sentences and todos, but I’d be happy to receive feedback. I’d like to post it by the end of July.

Eden is my top priority right now, as well as figuring out a good strategy for studying the security topics and maybe adding code samples for them to Github. Perhaps in a week or two, it will be nice to organize some sort of teleconference to discuss these topics. Great to see this moving forward.


Re: democratizing the social media accounts… To be able to take action as appropriate there should be a way to log requests for access - whether by an individual, a team, non-human entity, etc…In the spirit of things here perhaps osTicket ( Transparency while maintaining security to build trust :slight_smile:


First—very interesting project, excited to watch it grow, hats off @genekogan for planting the seeds of such a brave and ambitious project, may it bear fruit beyond your wildest imagination.

For governance/organization/democratization, some things to check out:

Yes these are both on my radar. There is most of all to learn about governance. I am trying to figure out a good way to do an open study about this, while developing Eden. So long as Abraham doesn’t generate anything of much value, we have time to experiment with multiple structures and see which work best down the road.

The question why it should create art creates a strange loop. It seems to be separated from the skill part of the problem (any generative system).
Let’s start with the idea of an abstract entity. It can do art, business or anything. It needs to survive, it may pass knowledge/wisdom to next generation. It can die (if no one believes in it, like a forgotten coin or a DAO).
For me the best example of this abstract entity is multi-national corporations and big institutions. They have a code on how they start, how they operate and how they might dissolve. They have vision, mission, etc. We can already think how those institutions could be created by code with DAO’s.

Let’s assume that we create a DAO with an artistic vision and mission. We can also and an autonomy metric to it and attach it to its vision. It can use peoples skills and decisions to accomplish its goals and even ask people if it should change them. Maybe after a while of an artistic period it might decide to do work like a designer in order to survive. That’s where the loop is hidden. It can make people code itself in a different way. It can also change its governance model in order to do more.