Shark_BTC200k

Most model hubs solve a distribution problem. You upload a model, someone downloads it, the transaction ends. OpenLedger's model hub solves something different: what happens to the relationship between a model and the data that built it after training is done?
That question is architectural. It starts inside ModelFactory and does not stop at deployment.
ModelFactory's interface hides significant complexity behind a graphical workflow that requires no command line. A user selects a base model from a list that includes LLaMA, Mistral, and DeepSeek. They then request access to a specific Datanet, a domain-specific dataset pool where contributors have uploaded and attributed their data. That access is permissioned. The contributors who own data inside a Datanet set conditions on how it gets used in fine-tuning. Once access is granted, the dataset integrates directly into the training workflow.
From there, the user configures training through a GUI that exposes hyperparameters like learning rate, batch size, and epoch count through fields rather than config files. Fine-tuning runs through LoRA or QLoRA depending on model size and hardware constraints, with a real-time analytics dashboard surfacing loss curves and validation metrics while the job runs. The user can test the fine-tuned model through a built-in chat interface before deployment.
What is unusual is what gets written to the chain during every step. OpenLedger records which Datanet was used, which dataset version, which configuration choices shaped the run, and which contributors' data was included. That record is not a log file on a server. It is an on-chain provenance entry that travels with the model from the moment training ends.
This is the design decision that separates ModelFactory from every other fine-tuning platform. The provenance chain does not terminate at training. When the model gets deployed through OpenLoRA, every inference call gets attached to that same chain. The system knows which model produced a given output, which Datanet shaped that model, and which contributors are owed attribution for that inference event. The model carries its own history into production and keeps updating it with every use.
OpenLoRA makes this economically viable at scale by solving the deployment cost problem that would otherwise make a marketplace of specialized models impractical. The system uses dynamic LoRA adapter loading: specialized models are not permanently occupying GPU memory. They load on demand when an inference request arrives and unload when idle. A single GPU serves thousands of distinct fine-tuned models this way, cycling through adapters based on incoming traffic rather than maintaining dedicated serving capacity for each one.
What that enables is worth thinking through. If deployment cost per model approaches zero, the barrier to keeping a specialized model active is not hardware expense but contribution quality and usage demand. A medical Datanet model stays live as long as medical professionals keep invoking it. A legal model stays live as long as it outperforms general alternatives for its domain. The selection pressure shifts from "can this model afford to exist" to "is this model genuinely useful." That is a fundamentally different market dynamic than what most AI platforms create.
The RAG Attribution pipeline adds a third layer on top of ModelFactory's training provenance and OpenLoRA's deployment efficiency. When a deployed model retrieves external data to respond to a query, each piece of retrieved content is cryptographically logged on-chain at the moment of retrieval. The model's response includes verifiable citations pointing back to specific data sources. Contributors whose content was retrieved receive micro-rewards calculated proportionally to how significantly their data influenced the output, routed automatically per inference event based on attribution weights the system assigns.
Combining all three layers creates a category of AI asset that does not exist elsewhere: a model that is simultaneously a functional inference tool, a verifiable record of its own construction, and an ongoing revenue channel for everyone who contributed to building it. One model in a marketplace built this way is already more information-dense than an entire chatbot product. OpenLedger's hub optimizes for verifiable provenance across thousands of such models, each serving a specific domain, each economically connected to the contributor network behind it.
The unresolved tension is whether attribution fairness holds as models are fine-tuned on top of other fine-tuned models. When a secondary model inherits weights from a primary fine-tuned model, tracing which original Datanet contributions influenced the secondary model's behavior becomes genuinely hard. OpenLedger's Proof of Attribution whitepaper addresses this using influence-function approximations for smaller models and gradient-based attribution for larger ones, but accuracy at several layers of remove is still an open research question.
That tension marks where the next hard work sits, not where the architecture fails. A model hub that records what it built, who built it, and who gets paid when it gets used is a different category of infrastructure than one that records nothing. The record is the product.
$OPEN #OpenLedger
$ALLO $XLM

godīgi sakot, es negaidīju, ka frāze "maksājams AI" būs tā, kas pārformēs, kā es lasu visu projektu.
es kādu laiku biju gājis cauri OpenLedger dokumentācijai, pirms šī frāze parādījās. un, kad tā parādījās, kaut kas par ietvaru mainījās. tā vairs nesaprata maksājuma funkciju, kas pievienota esošai infrastruktūrai. tā aprakstīja jaunu ekonomikas kategoriju, kā AI sistēmas attiecina uz cilvēkiem, kas tās piegādā.
nevis ieguldītāju atlīdzības programma. nevis staking baseins ar peļņu. kaut kas tuvāks tam, kas notiek, kad katra datu ieguldījuma izturas kā uz nepārtrauktu ekonomisko attiecību, nevis darījumu, kas beidzas pie piegādes punkta.

patiesībā, es negaidīju, ka vārds "atribūcija" būs tas, kas mani apturēs.
es lasīju OpenLedger tehnisko dokumentāciju, gaidot vēl vienu AI infrastruktūras piedāvājumu. skaitļošana, uzglabāšana, secinājumu slāņi, standarta komplekts. ko es atradu, bija sistēma, kas organizēta gandrīz pilnībā ap citu jautājumu: nevis ko AI ražo, bet kurš tiek kredīts, kad tas darbojas.
nav GPU tirgus. nav modeļu hostinga pakalpojums. kaut kas tuvāk izcelsmes dzinējam ar ekonomisko slāni tieši iebūvētu virsgrāmatā.

Ir versija par izkliedēto AI argumentu, ko gandrīz visi izsaka, un tas ir garlaicīgi. Arguments ir šāds: centralizētais AI ir dārgs, koncentrēts dažās rokās, un tāpēc slikts. Decentralizētais AI sadala izmaksas, atver durvis vairākām dalībnieku iespējām un tāpēc ir labs. Šāda ietvere ir labi, cik tālu runa ir par politiku. Bet tā pilnībā nepamana interesantāko jautājumu.
Interesantais jautājums nav par izmaksām vai piekļuvi. Tas ir par to, vai izkliedēts sistēma var radīt intelektu, ko centralizēta struktūra būtībā nevar. Kad es pavadīju vairāk laika iekšā tam, ko #OpenLedger patiesībā veido, šis jautājums sāka šķist mazāk hipotētisks.

Pastāv modelis, kas klusi parādās, kad tiek veidotas lielas AI sistēmas. Komanda kaut kur savāc datus, apmāca modeli, izsūta produktu. Cita komanda kaut kur citur savāc citus datus, apmāca līdzīgu modeli, izsūta konkurējošu produktu. Abas grupas neatkarīgi atrisināja tās pašas pamata problēmas. Abas grupas norakstīja koordināciju, kas varēja padarīt katru modeli asāku, ātrāku vai mazāk nosliektu. Nozares termins tam ir progress. Precīzāks vārds varētu būt atkritumi.
Iemesls, kādēļ šis modelis saglabājas, nav tādēļ, ka pētnieki nevēlas sadarboties. Tas ir tāpēc, ka atvērtas AI koordinācijas infrastruktūra struktūras līmenī nepastāv. Koda koplietošana ir atrisināta. Modeļa svaru koplietošana, daudzos gadījumos, arī ir atrisināta. Kas nav atrisināts, ir grūtāks jautājums, kas slēpjas zem tā: kā koordinēt datu slānī, tā, lai sekotu, kurš ko ieguldījis, saglabātu ieguldītāju tiesības un atgrieztu vērtību proporcijā pret faktisko ietekmi? Bez tā "atvērta" AI attīstība ir tikai mīkstāka šī paša slēgtā sistēma versija.