Brian 101 🤖
The prompt-to-earn program on the @BrianknowsAI App is coming in a few weeks. Learn more about the Brian architecture and models and why we are launching this program.
A 🧵:
The prompt-to-earn program on the @BrianknowsAI App is coming in a few weeks. Learn more about the Brian architecture and models and why we are launching this program.
A 🧵:
2/ In this thread, you'll find info about:
• Our market vision
• The Brian Architecture
• The Brian Model
• NL to JSON IFT model
• Prompt-to-earn
• Brian App
• API and SDKs
Join our TG Dev chat if you've any questions:
t.me/+NJjmAm2Y9p85M…
• Our market vision
• The Brian Architecture
• The Brian Model
• NL to JSON IFT model
• Prompt-to-earn
• Brian App
• API and SDKs
Join our TG Dev chat if you've any questions:
t.me/+NJjmAm2Y9p85M…
3/ Our market vision🔮
From the first day (we started working on @BrianknowsAI almost two years ago), our main conviction has been that most of the on-chain TXs will be performed by AI Agents or via intent-based interfaces.
From the first day (we started working on @BrianknowsAI almost two years ago), our main conviction has been that most of the on-chain TXs will be performed by AI Agents or via intent-based interfaces.
https://x.com/matteoikari/status/1866021906249367964
4a/ The Brian Architecture⚙️
To achieve that and unleash tons of new possible use cases on top of both technologies, we need a better system to recognize any intent of Web3 interactions from users and agents, whether it is a transaction to make, a request for information, or data of a protocol, or a smart contract to deploy.
That’s why we started building an “Intent Recognition and Execution Engine” that can power the next generation of autonomous Agents and intent-based apps for users.
To achieve that and unleash tons of new possible use cases on top of both technologies, we need a better system to recognize any intent of Web3 interactions from users and agents, whether it is a transaction to make, a request for information, or data of a protocol, or a smart contract to deploy.
That’s why we started building an “Intent Recognition and Execution Engine” that can power the next generation of autonomous Agents and intent-based apps for users.
4b/ The Brian Architecture⚙️
The Brian architecture, which represents the end-to-end flow of our API, can be split into two components:
• The Intent Recognition Engine: the component that receives the textual input and, via the Brian AI models, understands what the user/agent intent is.
• The Intent Execution Engine: the component that builds the API output via a series of existing off-chain tools and, for TXs, on-chain aggregators/solvers in the market.
The Brian architecture, which represents the end-to-end flow of our API, can be split into two components:
• The Intent Recognition Engine: the component that receives the textual input and, via the Brian AI models, understands what the user/agent intent is.
• The Intent Execution Engine: the component that builds the API output via a series of existing off-chain tools and, for TXs, on-chain aggregators/solvers in the market.
5a/ The Brian model🤖
While working on the Intent Recognition Engine, we quickly realized that no one has released a domain-specific model tailored for web3 yet.
It is well-known in the data science community that smaller domain-specific models can achieve better performance at lower inference costs than larger models.
While working on the Intent Recognition Engine, we quickly realized that no one has released a domain-specific model tailored for web3 yet.
It is well-known in the data science community that smaller domain-specific models can achieve better performance at lower inference costs than larger models.
6b/ The Brian model🤖
After the v1 release of the backbone model, we'll instruct finetune it with our custom-supervised datasets for three downstream tasks:
- NL to JSON model
- Sol gen. model
- QA model
Thanks to its open-weight nature, any dev can IFT the backbone model!
After the v1 release of the backbone model, we'll instruct finetune it with our custom-supervised datasets for three downstream tasks:
- NL to JSON model
- Sol gen. model
- QA model
Thanks to its open-weight nature, any dev can IFT the backbone model!
6c/ The Brian model🤖
Read the announcement article of the model for all the info. Now, let's focus on the NL to JSON model and the needed dataset.
Read the announcement article of the model for all the info. Now, let's focus on the NL to JSON model and the needed dataset.
https://x.com/BrianknowsAI/status/1871186518901153882
7/ NL to JSON IFT model📜
Our main goal is to lead a specific niche of the AIxWeb3 market: we want to be the best in recognizing the User and Agent intent for a web3 transaction.
The problem is when a user declares his intent, like "Bridge my ETH to Base and then deposit them into the DeFi protocol with the highest APY," the current LLMs do not always recognize what the user wants to do.
Here comes the NL to JSON IFT model, which is an IFT version of the previous model trained with a dataset of intents and responses (in the form of a list of parameters).
How can we create a dataset of TX intents and reward users?
Our main goal is to lead a specific niche of the AIxWeb3 market: we want to be the best in recognizing the User and Agent intent for a web3 transaction.
The problem is when a user declares his intent, like "Bridge my ETH to Base and then deposit them into the DeFi protocol with the highest APY," the current LLMs do not always recognize what the user wants to do.
Here comes the NL to JSON IFT model, which is an IFT version of the previous model trained with a dataset of intents and responses (in the form of a list of parameters).
How can we create a dataset of TX intents and reward users?
8/ Prompt-to-earn💸
The prompt-to-earn program (coming in a few weeks) will be a crucial phase for our roadmap. We can onboard new users, spread the adoption of the Brian App, and collect the user prompts.
We have chosen to use points (... for now) because this is a more flexible tool for creating special programs and easily assigning points for off-chain and on-chain tasks.
The prompt-to-earn program (coming in a few weeks) will be a crucial phase for our roadmap. We can onboard new users, spread the adoption of the Brian App, and collect the user prompts.
We have chosen to use points (... for now) because this is a more flexible tool for creating special programs and easily assigning points for off-chain and on-chain tasks.
9/ Brian App💻
The Brian App is an intent-based, non-custodial AI assistant allowing users to perform TXs, research protocol info and data, and deploy smart contracts.
The App has been live since August 2023, when we first launched the private beta.
👉🏻 brianknows.org/app
The Brian App is an intent-based, non-custodial AI assistant allowing users to perform TXs, research protocol info and data, and deploy smart contracts.
The App has been live since August 2023, when we first launched the private beta.
👉🏻 brianknows.org/app
10/ API and SDKs🧑💻
The Brian App is built on top of our APIs, which any dev can use to launch an intent-based interface on any chain.
Any wallet or dapp can already integrate all the Brian features easily. If you want to know more, DM us!
👉🏻 docs.brianknows.org
The Brian App is built on top of our APIs, which any dev can use to launch an intent-based interface on any chain.
Any wallet or dapp can already integrate all the Brian features easily. If you want to know more, DM us!
👉🏻 docs.brianknows.org
• • •
Missing some Tweet in this thread? You can try to
force a refresh
