Paul Nguyen

The framing that appears most consistently in both OpenLedger's own communications and in third-party analysis of the project is "data marketplace." Buyers and sellers. Supply and demand. Prices and transactions. That framing is intuitive and it is wrong in a way that has practical consequences for how the project's strengths and weaknesses get analyzed.
The structural analogy that fits much better, once you look carefully at how Datanets actually work, is a scientific pre-print server. Understanding why that analogy is more accurate, and what it predicts about the project's trajectory, is the most useful reframe I have encountered for thinking about what OpenLedger is actually building.
Pre-print servers, like arXiv in physics and mathematics, bioRxiv in biology, and SSRN in social sciences, are infrastructure for sharing research knowledge before formal peer review. Scientists submit their work to pre-print servers for several reasons: to establish priority for their ideas, to receive early feedback from colleagues, to make their work accessible before the lengthy peer review process concludes, and to build a public record of their intellectual contribution. Pre-print servers do not charge for content. They do not pay contributors. The economics are entirely separate from the knowledge-sharing function.
OpenLedger's Datanets share the most important structural characteristic of pre-print servers: knowledge flows before formal quality validation is complete, and community norms rather than institutional gatekeepers handle quality assessment. 🤔
In a pre-print server, a submitted paper is visible immediately, before peer review. Other scientists in the field read it, cite it provisionally, and form opinions about its quality based on the content and the author's reputation. Formal peer review, when it happens, either confirms the community's provisional assessment or revises it. The system works because the scientific community has developed norms around how to read pre-print work: with appropriate skepticism about unreviewed claims, with attention to the author's track record, and with awareness that the pre-print represents a snapshot of ongoing work rather than a finished product.
OpenLedger's Datanet contributors submit knowledge that becomes available to model builders before and while community validation proceeds. Datanet validators, who are domain experts in the relevant field, review contributions and provide quality assessments. The validation process is analogous to informal peer review: it filters out obvious problems, surfaces concerns about specific contributions, and establishes a community quality standard. Like pre-print peer review, it is imperfect, variable in rigor, and dependent on the depth of the reviewing community.
The pre-print server analogy predicts specific failure modes that the marketplace analogy does not anticipate and that the actual pre-print server experience has documented extensively. 💀
The first predicted failure mode is quality variability across domains. Pre-print servers in fields with large, active, quality-conscious research communities, like physics, produce pre-print corpora that are largely reliable despite limited formal review. Pre-print servers in fields with smaller or less organized communities produce much more variable output quality. OpenLedger's Datanets will show exactly the same pattern: Datanets in fields with large engaged contributor communities will develop robust quality norms. Datanets in fields where the contributor community is thin will produce variable-quality training data that reflects the absence of community quality control infrastructure.
The second predicted failure mode is usage scaling faster than quality infrastructure. The COVID-19 pandemic produced a crisis in pre-print servers when usage volume, driven by researchers urgently sharing preliminary findings, vastly outpaced the informal community review mechanisms that maintained pre-print quality in normal times. Misinformation propagated through pre-print citations faster than the community could identify and correct it. The speed advantage of pre-prints over formal publication became a liability when the volume of submitted work exceeded the community's capacity to assess it informally.
OpenLedger's Datanets will face an equivalent stress test when AI model builders begin accessing Datanet content at scale. The model training pipeline doesn't wait for quality reviews to complete. It pulls available training data and uses it. If Datanet usage grows faster than the community validation infrastructure that maintains quality, the models trained on that data will reflect whatever quality level the Datanet happened to contain at training time, including the contributions that were submitted but not yet reviewed. 🤔
The third predicted failure mode, and the one that the pre-print analogy illuminates most sharply, is the credibility problem with downstream applications. Pre-print citations in scientific papers carry an asterisk: results from unreviewed pre-prints are treated with appropriate skepticism by sophisticated readers. But unsophisticated readers, journalists, policymakers, and the general public, often treat pre-print citations as equivalent to formally peer-reviewed results. The gap between what a pre-print citation means to an expert and what it means to a non-expert has caused significant real-world harm when health claims from pre-prints were repeated in public discourse before the claims were validated.
OpenLedger's provenance records carry a similar ambiguity. A model trained on Datanet contributions has a documented provenance record. What that record does not communicate is the quality level of the Datanet at the time of training, the rigor of the community validation process, or whether the specific contributions that shaped the model's behavior most strongly were among the most or least carefully reviewed ones. An AI model with an OpenLedger provenance record is not the same as an AI model trained on high-quality, rigorously validated domain expertise. But the provenance record is legible in a way that the quality assessment behind it is not.
The pre-print server community has developed several mechanisms to address these failure modes over decades of operation. Endorsement systems where established researchers vouch for submitted work add a reputation layer on top of availability. Version control allows submitted work to be updated as community feedback arrives, so the record of a contribution reflects its current state rather than its initial state. Field-specific quality norms develop through community practice and are communicated to newcomers through explicit documentation and informal socialization. 💀
OpenLedger's Datanets need equivalent mechanisms, built for the AI training data context. Version control for contributed data, so that updated or corrected contributions are reflected in the attribution record rather than freezing the record at submission time. Endorsement mechanisms that allow distinguished domain experts to signal which contributions they consider particularly valuable. Community-developed quality norms that are explicit and documented rather than implicit and variable. These are the infrastructure investments that determine whether Datanets develop into reliable pre-print servers or remain perpetual early-stage experiments.
The pre-print server framing is the most honest and useful frame for what OpenLedger is building, and it makes the project look different from both the optimistic "data marketplace" framing and the skeptical "unproven infrastructure" framing. Pre-print servers are genuinely valuable. They changed how science works by making knowledge available faster and more broadly than the formal publication system allows. They also have well-known failure modes that took decades of community experience to partially address.
OpenLedger is attempting to compress that learning curve into the lifespan of a crypto protocol. The pre-print server history predicts exactly which problems will be hardest and in what order they will arrive. 🙏
@OpenLedger $OPEN #OpenLedger

OpenLedger's Datanet model rests on a specific chain of reasoning. Domain experts have valuable knowledge. That knowledge, when properly organized as training data, makes AI models more capable in the relevant domain. Those more capable models generate more inference demand. That demand routes attribution payments back to the contributors whose knowledge made the model better. The chain is logical and the individual links are each defensible.
The link that doesn't hold under scrutiny is the second one: that domain experts, given the opportunity to contribute their knowledge as training data, will self-select the right knowledge for AI training. This isn't a criticism of domain experts' expertise. It's a description of a specific and well-documented gap between expert self-assessment and what AI training processes actually need, and it's the gap that sits between OpenLedger's theory and what Datanets will actually produce.
Let's start with the empirical record on expert self-selection accuracy.
Research in knowledge management and organizational learning has documented consistently that subject matter experts are systematically poor judges of which aspects of their knowledge are most valuable to people learning in their domain. The knowledge that experts consider most important tends to be the formal, articulable, definitionally clear knowledge: diagnostic criteria, established procedures, documented best practices. This is the knowledge that is easiest to make explicit and the knowledge that experts believe represents their field.
The knowledge that actually differentiates expert performance from intermediate performance tends to be tacit: pattern recognition developed through exposure to many cases, intuitions about when formal criteria don't apply, judgment about how to weigh competing signals in ambiguous situations. Experienced clinicians don't just know the diagnostic criteria for a condition. They know which presentations to take seriously when the criteria aren't met, which patient histories should trigger concern before objective markers appear, and how to interpret borderline test results in context. That knowledge is what makes them valuable. It's also the knowledge they're least likely to recognize as a distinct and contributable asset.
This creates a systematic self-selection problem for Datanets. When you ask domain experts to contribute their valuable knowledge, they will contribute the knowledge they know how to articulate, which is a different category from the knowledge that most differentiates their expertise. The formal, documentable knowledge they contribute may already exist in published literature, in textbooks, in standard-of-care guidelines. The contribution that would actually improve a model's performance is harder to surface, harder to structure, and likely not what the expert thinks to submit first.
The AI training data literature has a parallel finding from a different direction. Research into what kinds of training data produce the largest improvements in model capability consistently shows that the most valuable training examples are often at the boundaries of the domain, the edge cases where the model currently performs worst, the ambiguous presentations where expert judgment diverges, the situations that don't fit neatly into standard categories. These are precisely the cases that domain experts may undervalue as training examples because they're the unusual ones, not the representative ones, and human intuition about typicality doesn't map well onto what training processes need for model improvement.
A cardiologist deciding what to contribute to a medical Datanet will likely start with clear, well-documented cases: standard presentations of common conditions with unambiguous outcomes. Those contributions are useful. The contributions that would most improve the model's performance are probably the unusual presentations, the diagnostic near-misses, the cases where the expert's judgment deviated from the initial presentation's suggestions and turned out to be right. Those cases are harder to select, harder to document, and less likely to feel like appropriate contributions from the expert's professional perspective.
OpenLedger's Datanet validation model partially addresses this through community quality assessment: validators evaluate whether submissions meet quality standards. But validation assesses whether a contribution is accurate and well-formed. It doesn't assess whether the contribution addresses a gap in the model's current knowledge or adds redundant information the model already represents well. Those are different quality dimensions, and the second one requires knowing what the model currently knows, which community validators generally don't have access to.
The self-selection problem also has a second dimension that's worth naming: domain experts don't just select the wrong content. They structure it in the wrong format.
AI training data is most valuable when it captures reasoning processes, not just conclusions. A clinical decision record that shows the diagnostic reasoning chain, including the considerations weighed, the hypotheses entertained and rejected, and the way competing signals were reconciled, is more valuable for training a reasoning-capable model than a record that shows diagnosis plus outcome. But experts documenting their knowledge for human communication typically document conclusions and rationale, not full reasoning processes. The process is implicit in the narrative. Making it explicit requires an additional translation step that most contributors won't know to perform.
The validation mechanism as described would approve a well-formatted, accurate clinical record even if it doesn't capture the reasoning process that makes clinical expertise valuable for AI training. The validator can confirm the diagnosis is correct and the format is appropriate. They can't easily determine whether the record is structured in a way that teaches the model to reason rather than to pattern-match.
The solution to the self-selection problem isn't a better validation rubric, though that would help. It's a contributor guidance system that works in the other direction: rather than asking experts to contribute what they think is valuable and then validating it, it identifies what the model currently lacks and guides experts toward contributing the specific knowledge that would address those gaps.
This is not a theoretical possibility. It's how the most sophisticated AI training data pipelines work internally at major AI labs: active learning systems that identify model uncertainty, targeted data collection that addresses specific performance gaps, contributor guidance that specifies the format and content most useful for the current training objective.
OpenLedger's Datanet model is essentially the decentralized version of this process. The gap is that the decentralized version doesn't yet include the active learning feedback loop that tells contributors what to contribute. Without that loop, contributors self-select based on their own judgment, the system validates based on accuracy and format, and the training data that accumulates may be high-quality by conventional standards while systematically missing the specific knowledge that would most improve the model.
Building that feedback loop, the mechanism that translates model performance gaps into contributor guidance, is the design work that would make the community knowledge contribution model actually produce what the Payable AI vision promises. It's also the design work that's most conspicuously absent from the current Datanet architecture description.
@OpenLedger $OPEN #OpenLedger

Perhitungan atribusi OpenLedger sering dibahas seolah-olah ini adalah solusi umum untuk masalah atribusi data pelatihan AI: sebuah sistem yang dapat melacak pengaruh data pelatihan mana pun terhadap keluaran model AI mana pun, di seluruh arsitektur model mana pun, dalam konfigurasi pelatihan mana pun. Peta jalan proyek, komunikasi komunitas, dan model ekonomi semua secara implisit mengasumsikan bahwa perhitungan atribusi akan berkembang seiring dengan ekosistem model AI, mencakup jenis model baru dan konfigurasi pelatihan baru saat mereka muncul.