Denny,

Great exchanging with you again. We ingested all 161 million sentences of Wikipedia for our proof of concept - a context knowledge graph. I am pretty confident you asked this specific question so you can DQ this immediately, and yes, we do use an embedding model to resolve meaning, initially. However, we use a fully explainable and transparent method to move from the high dimensions of the embedding space to a 3D space we call Pixelization of Semantic Anisotropy (PSA). We use guide stars sentences to align the semantics of the 3D space, losing none of the semantic or contextual information needed for the graph in the compression from 4096 to 3D. We have submitted a paper to the ACL on our findings of meaning alignment decay by distance in the PSA field. We can even exchange coordinates to another aiidea Graph to express meaning. By ingesting sentences, the end result is a fully human readable, explainable, traceable to provenance graph of ideas. Every step and all math can be checked by a human. The original embedding model is multi-lingual and multi-modal, thus our sentence nodes (and paraphrase nodes) are the basic units of meaning tracked in the graph, regardless of language or conveyance mode.. However, we also extract the objects from the sentences and using an ensemble of approaches (all transparent and explainable to anyone that listened in grammar and algebra class) to link the objects to an ontology. While we ship with the BFO/CCO ontology as a seed, we are ontology agnostic, and can use any ontology, including WikiData. Lastly, we use ontological frontier pressure to model where the ontology is under stress (does not have coverage for the incoming objects) as an adjacent possible probe to flag the need for human intervention, or in your case, possibly new article(s).

I would argue that the aiidea Graph is more easily understood by non-programmers than Abstract. We have already demonstrated that it can work at Wikipedia scale. I’ve attached some images to help in your understanding. I am not in any way denegrating Abstract. I just think that your rigidity in a strictly function based approach is limiting. There are always N ways to get from A to B. Now that you are mostly in production, your uptake numbers will tell you if you stay in mostly technical users, or you branch out into the wider community. Your goal is about trust and transparency and so too mine. I just don’t think that AI is automatically bad. I do think that it is just the natural emergence of information under increasing pressure to reduce local entropy. Its a tool.

Doug

Douglas Clark www.aiidea.biz 707.210.4771

This message contains confidential information and is intended only for the individual named. If you are not the named addressee, you should not disseminate, distribute or copy this e-mail. Please notify the sender immediately by e-mail if you have received this e-mail by mistake and delete this e-mail from your system.





On Apr 28, 2026, at 11:30 AM, Denny Vrandečić dvrandecic@wikimedia.org wrote:

Hello Douglas, and thank you for your proposal and your answer!

In your proposal, how is the knowledge translated to text, so that it can be integrated into Wikipedia?

Cheers, Denny

On Mon, Apr 27, 2026 at 6:35 PM Douglas Clark <clarkdd@gmail.com mailto:clarkdd@gmail.com> wrote:

...

Denny and Team,

First, congratualtions on getting this coverage. Second, I want to say that I deeply respect the moral foundation behind Abstract Wikipedia and Wikifunctions. The idea of making knowledge generation contestable, transparent, and free from opaque model behavior is not just technically interesting, it’s necessary in today’s AI world.

The distinction you’re drawing between LLMs and Wikilambda is real and important. Open data, inspectable logic, and the ability to intervene at every layer of the system represent a meaningful alternative to black-box generation. Where there is still a critical gap is not in principle, but in accessibility of participation.

Wikifunctions makes every part of the system contestable, but in practice, that contestability is mediated through programming abstractions. To meaningfully engage, a contributor must understand functions, execution logic, and composability at a level that resembles software development. This creates a barrier where the right to contest exists, but the ability to exercise that right is unevenly distributed.

In other words, the system is open, but not yet broadly usable. When I proposed the aiidea Graph as an alternative years ago, it was a different approach to the same moral objective: Instead of requiring users to modify algorithms, it allows them to operate directly on ideas (sentences), objects, and their relationships, which are the natural units of human knowledge. It’s reuse of faithful recordation of information in graph form.

Rather than exposing contestability through code, aiidea exposes it through:

Direct inspection of source-backed statements Transparent linkage between ideas and underlying data The ability to challenge, replace, or extend knowledge at the sentence level (something we all know, not just some skilled users) Deterministic responses constructed from traceable subgraphs rather than generated text This preserves the core moral properties:

Transparency Contestability Auditability But shifts the interaction surface from programming to reasoning. Where Wikilambda makes algorithms defeasible, aiidea makes knowledge itself defeasible, without requiring users to understand or modify the machinery that produces it. This is a subtle but important distinction. One system opens the engine; the other lets people steer the vehicle without needing to be engineers.

I believe both approaches are aligned in intent, but optimized for different layers of participation. My concern, and the reason for reaching out, is that if contestability depends on developer-level fluency, we risk recreating a new kind of gatekeeping, even within an open system.

I would welcome the opportunity to continue this conversation. I think there is real potential for alignment in building systems that are not just open in principle, but broadly usable in practice.

Doug

Douglas Clark www.aiidea.biz http://www.aiidea.biz/ 707.210.4771

This message contains confidential information and is intended only for the individual named. If you are not the named addressee, you should not disseminate, distribute or copy this e-mail. Please notify the sender immediately by e-mail if you have received this e-mail by mistake and delete this e-mail from your system.

...

On Apr 25, 2026, at 2:22 AM, Denny Vrandečić via Abstract-Wikipedia <abstract-wikipedia@lists.wikimedia.org mailto:abstract-wikipedia@lists.wikimedia.org> wrote:

The Abstract Wikipedia/Wikifunctions project has a profoundly moral aim: to give human beings control over information in the Age of GenAI. If the problem were simply to populate minority-language Wikipedias with articles, it would be simpler just to get a Large Language Model (LLM) to translate English Wikipedia into those languages. But Wikilambda presents a stark alternative to LLMs such as Gemini, Llama or ChatGPT. LLMs rely on vast concealed datasets. Wikifunctions draws its data openly from public Wikimedia databases. LLMs generate text using opaque algorithms that even their designers struggle to control. Wikilambda makes every part of every algorithm available to anyone. In short, Wikilambda is contestable. If you ask an LLM to generate an article on a topic, the only way to contest its algorithm is to click 👍 or 👎 (Crawford and Gillespie 2016). If you are unhappy with an article generated by Abstract Wikipedia, you will be able to: change the “abstract” content in Wikidata; change the algorithms that construct or render the article in Wikifunctions; or sever the connection between the article and Abstract Wikipedia, and edit the article the old-fashioned way in Wikipedia. The role of Wikilambda in all this is to make algorithms “defeasible” (Blanton et al. 2022). Every part of every algorithm is there, and can be contested on the platform itself, even if that contestation may be culturally and politically constrained (Tkacz 2015; Ford 2022). Wikilambda is an attempt to design what Alan Blackwell (2024) calls a “moral code”: it combines More Open Representation, Access to Learning, and Creating Opportunities for Digital Expression. If nothing else, Wikilambda is a thundering critique of corporate AI hype.