4.a. AI model as an asset

Similar to the our definition of TrainedModel above, we can represent an AI model as an asset on the blockchain. 

4.b. Scalable

Technical scalability: In order to compete with traditional internet giants like Google, Facebook, Alibaba, Tencent, our underlying platform blockchain technology has to horizontally scalable.  If we have a AI dApp signup and use our platform for easy integration into all the existing blockchain backends, our system is able to handle tens of millions of daily active users. For this type of functional requirement, we will deploy multiple instances, abstract the core business logic into a docker and put load balancer in front of our instances to make sure every API/JSON calls went smoothly.
Business model scalability: Traditional internet company has proven to us that the best business model is to offer free service first and then offer a variety of revenue models (freemium, premium, subscription). So our underlying blockchain technology platform should have the flexibility to let application offer creative business revenue model. It is the free to use for users that causes the internet to gain more widespread adotion.

4.c. Controllable    

Constant  upgrades: Either big enterprises or small startups, even individual developers should have the flexibility to upgrade their application with new features constantly. Our platform will support both model software as well as smart contract upgrades. Even Google chrome, Facebook, Tencent WeChat, Apple iOS are subject to to different bugs or system failure, our platform and the underlying platform should be robust enough for bug fixes and automatic program call stack generation.
Continuous integration:  When developers add new functionality to their existing softwares, they do not want to brake any old features. That is the where the continuous integration solution comes in. If certain commits break the old feature(can't pass certain testcases), the code hosting platform will reject that commit and generate a report about the error. Our system as well as the underlying blockchain platform will also support this kind of good practice.
Failure recovery:  Worst case scenario: if we have network failure, system failure, underlying blockchain platform DDOS attack or any user key credential being compromised,  our platform as well as underlying blockchain platform will have a failure recovery system. By versioning different release of the AI models as well as the database state, we will be able to provide this kind of features for our end user as well as developers.

4.d. Difficult      

PoW:  PoW is the most adopted consensus algorithm for the unpermissioned public blockchain.  Pioneered by bitcoin, PoW is the most known consensus algorithm on the blockchain ecosystem. PoW is hard to break and requires more than 50% attack to destroy the system.  We will be able to safely integrates all PoW based public blockchain and finish most of our AI dApp business logic on the corresponding blockchain.
PoS:  For the unpermionsoed public blockchain, PoS/BFT-DPOS/DPOS are the new challenger. PoS is proven to be more energy friendly. Under those algorithms, users who hold tokens on the corresponding blockchain platform are able to select block producers and produce blocks. Our AI dApp platform will be able to safely communicates with  those public blockchain and implements all of our business logics. 
SIEVE:  For the permissioned private blockchain, there are numerous innovation around the consensus algorithm. One of the most interesting one is SIEVE. In short, SIEVE expands the first PBFT calculation by including theoretical execution and check stages to: 1) identify and sift through conceivable non-deterministic requests and set up the determinism of exchanges entering the PBFT 3-stage assertion protocol, and 2) enable agreement to be keep running on yield condition of validators, notwithstanding the accord on their info state offered by Classic PBFT. SIEVE is gotten from PBFT separately (motivated by thoughts portrayed in [Aublin et al., TOCS'15]) by reusing the PBFT see change convention to bring down its multifaceted nature and abstain from executing another accord convention sans preparation.By the nature of permissioned private blockchain, our system is able to securely connect those blockchain with access control implemented. 
Quantum resistance:  Post-quantum cryptography, otherwise called quantum-safe cryptography, can stand up to the assaults by quantum PCs. The improvement of such encryption innovation takes a more customary way, in view of troublesome issues in particular arithmetic fields. Through investigating and creating calculations, the post-quantum secure encryption innovation can be connected in the system, what's more, to give the most abnormal amount of information security. We will be looking forward to see exciting use case of such algorithm being put into practice on various public/private blockchain.

4.e. Maintainable

Extensibility:  The power of modern software is built on top of each other. With vibrant support of third-party libraries and packages, developer is able to develop the software at much faster speed. We will provide a set of library dependency management tools to allow developers easily call and manage the already published smart contracts or import other contracts as a library dependency.
Readability:  We will convert/parse all the transactions on the given blockchain and allow user to understand all our of systems AI dApps. User will be able to see who create the AI smart contract, the primary interactions of that smart contract, automatic smart contract abi generations, etc.  
 Explorbility:  We should be able to explore the entire blockchain. We will provide insights into the recent mined blocks on the blockchain. We will also provide insights into any transactions in any block that has already been mined and is currently attached to the blockchain network.  User can also check the history of any public address and audit the balances, transaction history etc.

4.f. Responsive

Fast response:  A great customer experience requests instantaneous feedback within couple seconds. User will get fractured if the application requires longer time to respond and the unresponsive AI dApp results in less competitive situation than traditional solutions. Our gCompute platform will try to optimize as much as we can to reduce the application response time.
Networking optimization: We will try to minimize the communication latency within the network. Network optimization ought to have the capacity to guarantee ideal use for system resources, enhance profitability and also effectiveness for the community. Network optimization takes a look at the individual workstation up to the server and the tools and connections related with it. Our system could consist of traffic shaping, redundant data elimination, data caching and data compression and streamlining of data protocols.

4.g. Concurrent                    

Synchronous execution:  If the dApp itself requires sequentially independent steps, we should support fast sequential execution as well.
Asynchronous execution:  Even though computers run in sequential mode, and uses the combination of locks & context switch to implement the pseudo asynchronous execution, modern applications all support concurrent execution by default.  With the full internet-scale AI dApp solution in mind, we will try to optimize our code for the asynchronous execution as we can.