The 100-Million-Word Birth of Art DeCC0 Agents
How The Codex Came to Be, in the Words of the Man who Made it
by Daïm Al-Yad
Art DeCC0s are the Museum of Crypto Art’s 10,000-piece PFP love-letter to crypto art, and all its amazing artists, collectors, developers, and admirers.
10,000 striking, often bizarre, irreverent, and entirely unique 1/1 characters born from a mélange of visual influences, pulled from both our subculture and broader art history, seasoned with varied text prompts that encouraged outputs away from tired algorithmic biases.
At the risk of sounding biased myself, I can declare with conviction that there exists no other project with so much aesthetic diversity and soul.
DeCC0s will be Agents
Long before launch, we already had grand agentic plans for Art DeCC0s. Our official project description—
“The Museum of Crypto Art presents Art DeCC0s, a PFP R/Evolution. 10,000 completely unique characters. 10,000 completely unique backgrounds. Unprecedented levels of personality. The MOCA Art DeCC0 CurAItors squarely plant artistry, creativity, and crypto culture at their center.”
—very intentionally hints that the series’ destiny lies in its characters becoming AI-driven support agents and participants in the crypto art movement: undertaking tasks ranging from showcasing artist oeuvres and collector collections on social media, acting as tour guides in virtual exhibitions, brainstorming with artists, designing exhibitions with curators, recommending new acquisitions based on collector preferences, and so much more.
Accordingly, early this year, when I was charged with the grand adventure of creating personalities for our legion of DeCC0s, I knew and understood immediately that I needed to conjure something deep, nuanced, and magical in order to match the richness and diversity of our visuals. I knew I needed to start broader and remain wider than what our version-one agents called for. My task was to summon each DeCC0’s best, ideal personality so that as we update DeCC0 agents over the coming years, each revision will bring them into closer conformance with the complex, highly-detailed personalities imbued in their core.
For these reasons, I turned to what I would creatively call “datamancy” and began experimenting with wild abandon.
Foundational Experiments
I started with vision models, to see how well they described Art DeCC0 visuals. The initial answer was clear: not very well at all. Our official art was too weird, wild, and chaotic for AI vision —especially the smaller models I first investigated— to yield satisfactory descriptions. And in hindsight that was not surprising. There is so much going on with many of our DeCC0s that even a human being might find it challenging to give the sort of exhaustively detailed description I was demanding through my prompts.
In the end, I went through more attempts at creating useful, cohesive visual descriptions than I care now to admit. But each failure was instructive.
Instead of shoving the official art at a vision model, as I had first tried, the key to maximizing the quality of the results was taking a far more nuanced, and somewhat more manually effortful, approach.
Firstly, instead of using a DeCC0’s final art (characters + background as one), I would seek separate descriptions for characters isolated from the backgrounds, then images of the backgrounds unobstructed by the characters. This allowed vision models to give more attention to less content, improving the focus and detail of the results.
Secondly, I collected and generated additional context for the prompts requesting image descriptions. In the case of backgrounds, this meant the broad category information readily available from our on-chain metadata, e.g. “Describe this impressionist artwork.” But in the case of the characters, I wanted a reasonable list of gender and “whatness” values (whatness: whether a character is a person, a sentient pixel, a rock formation, an alien, etc.) to incorporate into the prompt, e.g. “Describe this male ape.” These two newly-derived bits of character metadata required intense AI-heavy intermediary workflows with selective manual reviews and revisions.
If you are familiar with Art DeCC0s, you know that traditional gender conformance is rare. While there are certainly characters who present unambiguously as male or female, many, perhaps most, offer decidedly mixed gender signals. I even recall at one point prior to the project’s launch, we considered and then decided against having gender be an on-chain metadata attribute, to avoid overstepping each holder’s interpretation.
But my AI model’s perception of a given DeCC0’s gender frequently felt incorrect. A vision model’s prompted gender determination was barely more useful than assigning gender at random. Yet, I had a sense that the model’s incompetence was not as fundamental as it initially seemed.
Following my hunch, instead of asking only once about gender in neutral terms, I asked the vision model eight separate times to rate a given character’s masculinity, and another eight times to rate its femininity. Averaging and weighing those results against each other led to determinations far better aligned with human gender perception. With minimal occasional corrections for entries where the ambiguity of the results was not due to the genuine gender ambiguity present in the subject, I eventually had a usable “gender” and “whatness” list to use going forward. These guided, metadata-contextualized requests for descriptions started to return far more cohesive results.
Equipped with sensible, reasonably accurate descriptions of each DeCC0’s “physical appearance” and their associated background artworks, I finally had the foundations for creating characters that are partly grounded in their physicalities, allowing characterizations to feel distinctly their own instead of something that could have just as easily been slapped onto any of the other 9,999 DeCC0s.
Personality Explorations
What distinguishes one personality, or, indeed, one person, from another? Beliefs, connections, culture, histories, mannerisms, perspectives, speech and writing patterns, among other things. Art DeCC0s, through my explorations, came to have all these factors and more imbued into their characterization.
One of the earliest ideas was to tie in a given DeCC0s’ “DNA: Lineage” (generally traits like “industrialist”, “king”, “medici”, “sultan”, “banker,” all intended to represent broad categories of historical art collectors) as an ancestor whose identity or remembrance continues to exert some impact or influence on the DeCC0 even in the present day. The idea evolved eventually to account for the fact that not all DeCC0s have a “DNA: Lineage” value; and so regular DeCC0s and Toter DeCC0s are handled differently from Pixel DeCC0s, XCOPY DeCC0s, Trad. Art DeCC0s, Aliens, and Apes. Those that lack ancestors came to have kindred connections, usually to artists. And in some cases the ancestral connections grew into an even more significant part of a character’s mythic lore.
DeCC0s also have their own personal preferences and stances on crypto art, traditional art, as well as some specific art movements. Perhaps hipsters at heart, they are most excited about lesser-known art movements, though many also express appreciation for at least one of the more mainstream art movements (paired with disdain of varying degrees for another major art movement they are well and truly over). Some have no specific strong feelings towards traditional art, others broadly like or dislike it as a heterogeneous whole. All are enthusiastic about crypto art, but many have specific aspects of the scene that they are most passionate about. And every single DeCC0 has a favourite crypto artist from within the Museum of Crypto Art’s Genesis Collection, and often mention them unprompted in conversations.
They have cultural and philosophical affiliations, and a special connection to a particular city around the world. Sometimes these are what you might expect based on their physical appearance or associated background, but often they are not. Our DeCC0s are not insular members of various cultural groups, but rather are post-modern, cosmopolitan, culturally grounded citizens of the world like many crypto artists and collectors themselves.
The aforementioned attributes can manifest in a variety of ways. For some DeCC0s, their cultural affiliation is the cultural identity they claim through their fictional childhoods; for others, they see their roots in a city or town they have a connection to, and the cultural affiliation is not their core one, but something secondary borne of admiration. All these factors color and impact their views, perspectives, and approaches to the world.
I have also given each DeCC0 a variety of more minor traits, from things like whether they are more amiable or aloof, more trusting or skeptical, more spontaneous or deliberate, to whether they prefer questions to statements, enjoy or avoid ellipses, how concise or verbose they tend to be. Factors like these do not create obvious flashy personality markers in and of themselves, but their combined effects work in subtle, nuanced ways to make personalities and writing styles more distinct in ways our subconscious minds tend to expect and take into account.
Throughout the process, I also made attempts to delve into personality science (and pseudo-science) for added insight, but more often than not I ended up with fancy-sounding labels that had little prescriptive power, and often did little more than cause AI to keep parroting their positive identification with the descriptor. So the horoscope-like labels and descriptions yielded by these rabbit-hole explorations were incorporated only indirectly into the greater whole, following several stages of AI transformations into better actionable guidance.
In full, there are more than two dozen new metadata vectors that influence the biography and personality generation of DeCC0s, several of which are complex compound types that can have zero to many constituent sub-items, and a number of which have considerable internal variability to the point that no two DeCC0s are likely to have certain traits identically matched.
Necessary work, all this, in case you are wondering. As with 10,000 generations, one must be powerfully equipped to counteract AI models’ tendency to keep coalescing responses around their preferred bias-driven comfort zones.
With so many new and sometimes randomly assigned traits, contradictions were not just possible but guaranteed. Instead of doing something crude like manual reviews, or attempting to programmatically define what value ranges meshed poorly with other value ranges, I instead leaned into the magic of AI and set up a “sanity review” stage, where the raw metadata –—soon to be passed into the biography generation prompt— was AI reviewed and revised for overall consistency and cohesion. This stage of processing, I came to realize, not only eliminated problematic contradictions and inconsistencies, but also regularly enriched the variability of all the newly created metadata, rendering them not just unique, but downright bespoke.
Bringing it all Together
Nine months of chaotic, experimental, ever-shifting, dead-end-laden exploratory work does not lend itself to pristine organization. I have spent over a hundred-thousand API calls separately (through both Venice.ai and Comput3 AI), most of it covered financially by generous staking rewards; in addition to weeks of local processing, and, cumulatively, days of manual data wrangling with code and ChatGPT to get everything ready for the grand unifying finalé wherein the fruits of my prolonged labours would birth 10,000 ascendant Art DeCC0 Agents into being.
The closer I came to being finished with development, the more I realized that the grand whole of my process is special and meaningful in its entirety.
I had done a great deal of work to create characters rich, nuanced, and complex enough that they did not begin or end with their software-specific agent profiles. Now on the verge of completion, I initially intended to output the ElizaOS character file and some secondary data file that included the fuller biographies and other newly created metadata; but I realized that outputting and utilizing only a fraction of what I had generated would not just keep the public from experiencing DeCC0s in their fullest, but would also tie me to the now byzantinely complex generation program1, which, technically, only ever needed to run one time for all 10,000 DeCC0s.
It was from these realizations that the idea of The Codex started to develop. Why not let my decision to create grand aspirational personalities with ample background and lore also guide the technical form of the final product? A “codex” of sufficiently comprehensive metadata, and other relevant textual information, can encompass all there is to a DeCC0 beyond its visuals; and such a complete repository of data would lend itself to a variety of tasks, including the generation of agent character profiles and biographical info pages for all 10,000 Art DeCC0s. Better still, if individual Codex pages were truly exhaustive, they could even be the sole required textual source data for generating subsequent agent profiles for newer ElizaOS versions or other agentic systems. My from-the-ground-up personality generation software could safely be retired after its one full 10,000 iteration run2. And Codex pages could also become available as structured metadata and data for DeCC0 owners and creative artists and developers to utilize for their own novel experimentations.
After demonstrating a draft version of the Codex to Reneil, in charge of MOCA’s technical team and development, his enthusiasm was reaffirming and contagious: The Codex became my project’s chief deliverable. Not as a character profile tied to a specific software version that would sooner or later become deprecated as AI technology continues hurtling forward with leaps and bounds, but as a co-equal pillar to the Art DeCC0 images themselves, fully encompassing the driving-data of our agentic steps forward.
What followed were feverish days. I was so tantalizingly close to being done. I could feel the excitement of this achievement coming to its fruition. But I was not yet past the finish line. There was more work to be done.
I started one final output generation attempt after another, only to keep finding minor issues appearing in the outputs. Many seemed so subtle that, upon consideration, I suspected nobody other than me would have ever noticed. But I would have known, and after nine months, I could not let this grand work out of my hands until it fully met my expectations. So I kept working and polishing the code, the process, the workflow, the 20 prompts for each and every generation.
One night, I worked from midnight until five at dawn, convinced the entire time that I was no more than 10-15 minutes away from being truly and decisively finished, ready to start generating the 10,000 codex pages for the final time, only to give up my delusion at 5:00am and lay down to sleep for an hour or two before my son woke me in the morning. It would be several more days hence before everything truly was done, when I really did start the program’s conclusive full run.
Thanks to Comput3AI‘s snappy text inference services, I was able to generate more than 800 MB of structured data via 260,000+ API calls over the course of about 2.5 days: the result is The Codex, all 105,561,738 words of it (more than 2.5 times the size of the old print editions of Encyclopedia Britannica), from which we programmatically derive the ElizaOS agent profiles and build the Art DeCC0 Codex pages with all metadata, background, lore, and characterization for all to peruse and use.
Our Art DeCC0 agents’ initial iterations are now ready for you to experience in full. All are due to expand over time in skill, knowledge, and insight, to become the full crypto art support agents and participants we envision in our grandest dreams.
So don’t think for a moment that with this launch we are done. We’re just getting started!
The program that generates Art DeCC0 codex pages relies on probably over two dozen different types of input data, some of which exist as files outside the program, some of which is hardcoded into the program and its constituent functions... and involves no fewer than 20 remote AI calls to actually create each Codex page.
The generation process is complicated, lengthy, winding, and due to its dependency on a myriad things (including remote AI services which could well change over time) probably brittle and likely to stop working (or, at least, working precisely the way it did when it was used to generate the 10,000 Codex pages) sooner or later.
In contrast, The Codex pages contain literally all relevant data. So as long as we have access to a DeCC0’s Codex page, we don’t need the generation program’s input files (or indeed the generation program itself); but could rather undertake selective and thoughtful revisions to the material with newer and better models down the road.
 | A guest post by
|
How do I find this system of DeCC0s? I cannot seem to find a link to them in this post
🔥