Last Wednesday at three in the morning, my screen suddenly popped up an alert — the medical imaging recognition model I had trained for half a year instantly became 'unable to cook without rice' due to changes in data policies from the partner hospital. The tumor slice data that had been annotated for three months felt like a library card that was suddenly revoked, making it inaccessible. I stared at the empty data folder, feeling for the first time so strongly that doing AI in someone else's data center is like building a house on rented land.
A turning point appeared in the developer forum on Friday. Someone shared a strange project: training a model to recognize endangered bird species on Kite, with data coming from 17 reserves around the world, yet the model had never fully left any server. 'Isn't this exactly what I'm looking for?' I almost trembled as I clicked open their technical documentation.
First experience: Breaking down AI into 'kite frameworks'
I decided to start with a simple task—training a classification model to recognize common skin diseases. Traditional methods require collecting a large number of patient photos to a central server, which involves numerous obstacles such as privacy and compliance. Kite's solution is completely different:
Step one: Model cloning technique
I distributed the initial model like packaged seeds through their 'federated learning framework' to the computing nodes of three cooperative clinics. Each node received the same model architecture, but the data remained local. It felt like disassembling a kite into its frame, skin, and tail, giving them to different craftsmen to work on simultaneously.
Step two: The 'kite string' of encrypted gradients
After training began, the most exquisite part emerged. After clinic A's model learned from 100 eczema photos locally, instead of sending the data to me, it calculated the 'model gradient change'—equivalent to 'what the model learned from these photos'. This gradient was split into three parts, encrypted using homomorphic encryption, and sent back to the aggregation server through three different relay nodes.
I stared at the real-time monitoring panel, watching the encrypted gradient fragments converge like kite strings from different directions. The aggregation server reorganizes these fragments, updates the main model, and then distributes the new model parameters back. Throughout the process, no party's original data left the local area.
“We are not sharing data,” project leader Su Qing explained at the technical meeting, “we are sharing 'learning experiences'. Just like several chefs exchanging recipes across a screen, no one has to present their secret sauce to each other.”
Unexpected breakthrough: When the model began 'self-evolving'
By the fourth week of training, unexpected things happened. The node in Bangkok suddenly began returning exceptionally excellent gradients—accuracy was 15% higher than other nodes. Investigations found that local doctors had created a unique set of refined labels during annotation, and this 'local wisdom' subtly enhanced the performance of the entire model through gradient transmission.
More remarkably, the system automatically recognized this contribution, triggering a smart contract: the Bangkok node received additional training rewards, and their improvement methods were recorded in the model version log. All services using the final model would proportionally pay small copyright fees to that node.
This is the 'contribution tracing mechanism' designed by Kite. Each participating node's contribution to the model is quantified as a verifiable data fingerprint. When this model generates returns later, profits will be automatically distributed to historical contributors proportionally. I saw for the first time that AI training can continuously create a trickle of income for all participants like music copyrights.
Deep involvement: My first designed data market
Inspired by this, I attempted to build a small medical data market. In this market:
· Hospitals can publish 'data needs': what type of labeled CT images they require and how much they are willing to pay
· Institutions that own data do not need to upload it, they only need to prove they possess certain types of data (through zero-knowledge proofs)
· Training tasks are sent to data providers to run locally in encrypted containers
· Ultimately, only gradient changes and verification results go on-chain
Last week, this market facilitated its first transaction: an ophthalmology clinic needed 2,000 labeled images of diabetic retinopathy, and three medical institutions jointly completed the task. No data moved throughout the process, but knowledge began to flow. During settlement, the smart contract automatically distributed the rewards in a 5:3:2 ratio and generated an immutable cooperation record on the chain.
Professional deconstruction: Why Kite is suitable for incubating AI
Having gone through a complete development cycle, I believe this platform provides unique value for decentralized AI in three dimensions:
1. Separation architecture of computation and verification
Traditional blockchains struggle to run complex AI training, but Kite adopts a model of 'off-chain computation + on-chain verification'. The training process occurs on dedicated computing nodes, while only gradient hashes and verification proofs are stored on-chain. This ensures verifiability while allowing sufficient computational freedom.
2. Technical implementation of data sovereignty
By combining secure multi-party computation and federated learning, they achieved 'data usable but invisible'. When I was involved in a multi-center clinical trial project, I witnessed medical data from seven countries training a model together, but no patient's privacy data left their home servers. This design complies with increasingly strict data localization laws.
3. The positive cycle of contribution economy
Their token economic model is designed with multiple layers of incentives: data providers, computing nodes, model developers, and end users can all find value capture points in the ecosystem. I saw an interesting case: after 23 iterations, the initial three data providers continued to receive revenue shares—this encouraged long-termism rather than one-off transactions.
Inspiration in the late night
Now my skin disease recognition model has been deployed in eight clinics, processing about 3,000 auxiliary diagnoses weekly. The night before yesterday, I received the automatically generated monthly report: the model's accuracy improved by 7% compared to centralized training, the risk of data leakage was reduced by 100%, and the total income for participating nodes has covered 80% of their hardware costs.
I walked to the window, looking at the twinkling lights in the city, suddenly reminded of the scene from my childhood when I flew kites with my grandfather. He said that a good kite does not fly higher in stronger winds, but can fly steadily even in a gentle breeze. Today's large AI models are like heavy kites that need strong winds to take off, while what Kite is doing is breaking down the large model into countless small kites, each able to take off in a local breeze, and then using invisible strings to form an array.
Perhaps true decentralized AI is not about creating an all-knowing giant, but about nurturing a forest—each tree grows independently, but the underground mycelium network allows them to share nutrients. My medical model is just one small tree, but it is rooted in real medical soil, rather than floating on some tech giant's cloud server.
The city in the early morning still has lights awake, just like the computing nodes running on Kite around the world at this moment. They could be a clinic's server, a university lab's workstation, or even a data enthusiast's graphics card at home. Each device is learning a fragment of the world locally and then quietly weaving into a larger understanding through encrypted clues.
This is no longer a crude splicing of AI and blockchain, but a new form of intelligence—dispersed like stars, connected like constellations. What Kite provides is that piece of night sky where the stars can shine individually and converse with each other.
