Technologies

Trai AI. AI that runs on Tenrai's miners.

Trai AI is our research into an AI platform of Tenrai's own. Not just another chatbot: an AI that runs on the computing power of the network's miners, pays them in TRAI for useful work, and knows Tenrai inside out.

Research only: nothing here can be used yet

The Trai AI platform on this page is research: it is not part of any release, and no one can use it today. An early version already exists inside our team; everything else shows the ideas and the path we are exploring, and any detail can still change.

More than a chatbot

Trai AI is meant as one platform with four services: one talks with people, one sells computing power, one watches over the network, and one writes code.

Four services on one platform.

Trai Assistant

An assistant like the ones you know: it answers questions, writes code, reads documents, translates and understands images. On top of that, it knows Tenrai, wallets, mining and staking in depth.

Trai Compute

Trai AI's own way into Proof of Useful Work: its inference jobs run on miners' GPUs and pay them in TRAI, just like the jobs of every other customer.

Trai Network Intelligence

Reads node logs, finds network problems, measures mining performance and flags unusual behaviour. It advises operators and never makes consensus decisions.

Trai Developer

A coding assistant for Rust, blockchain development, APIs and the Tenrai SDKs, later also as a service for other developers.

A specialist in cryptography and blockchain

Trai AI is meant to have a field of its own: cryptography and blockchain. It should know post-quantum signatures, zero-knowledge proofs, consensus and DAG protocols in depth, and help developers, node operators and researchers with exactly these questions.

It does not start from zero: an early version of Trai AI was already trained and used while we developed Traidag.

Experts that work together

A mixture of experts is a model made of many specialised parts. For every token, a router switches on only the experts it needs. A large model can hold far more knowledge this way, while each answer costs about as much computing as a much smaller model.

A router sends each part of the input to the experts that know it best; the others stay idle.

First: a router of models

Specialised models are chosen per job and can already run on different GPUs. No giant model has to be trained first, which makes it the ideal start.

Later: our own mixture of experts

A true neural mixture of experts, trained for Trai AI: more capacity for the same computing, but costly to train and heavy on memory for its weights.

One customer of an open market

Proof of Useful Work is meant as an open market: any AI provider, app or company could post jobs, and miners run them. Trai AI would be one of these customers, Tenrai's own, and add steady demand next to all the others. A customer could, for example, run 10,000 text requests through the Trai AI API: a scheduler hands them to miners whose GPUs already hold the right model, the miners earn TRAI for useful work, and Tenrai settles every payment.

A job travels from the customer through the scheduler to the miners' GPUs, is checked and settled, and the payment moves in TRAI.

Checking AI is the hard part

AI answers are not always the same bit for bit, and not always simply right or wrong. Comparing two answers is not enough, so useful AI work needs its own check that resists manipulation. That is why the Tenrai Virtual Compute Machine is built to run AI with exact integer math and to record a trace of every step.

How Proof of Useful Work checks jobs

The path we are exploring

1
Our own AI system
Host a strong open-weight model ourselves, tune it for Tenrai, extend it with our own data and tools, and let a Trai AI router pick specialised models. A working platform, without training a base model from scratch.
2
AI on the miners' GPUs
Connect Trai AI to Proof of Useful Work: miners register their GPUs, receive matching jobs and get paid, while model delivery, privacy, checks and settlement grow step by step.
3
Our own mixture of experts
Develop a sparse mixture-of-experts model of our own, or train an existing one further, with our own data, GPU power and evaluations.

Agents, with a hard line at your keys

Later, Trai AI could act on its own: split a development task into steps, run tests, coordinate API calls or manage large AI jobs across many workers. With wallets it may prepare a transaction, but it never gets your private keys or the right to sign. Every payment needs your explicit approval.

More demand for real work

Trai AI would add real demand for the computing power of Tenrai's miners, next to every other customer of the open market.

Proof of Useful Work TenQ