Introducing Frenrug

We’re excited to introduce you to Frenrug—a new innovation in AI agents that can interact on-chain, powered by the Ritual Infernet SDK. Frenrug is a gamified demo of this novel tech that lives in a friend.tech chat room and interfaces with the Base chain.

How do I use Frenrug?

Frenrug is a simple system that users can interface with entirely through its friend.tech chat room.

1. Any Frenrug key holder can send a message to the Frenrug chat room.
2. In their messages, users can try to convince Frenrug to buy or sell a key (theirs or others) with the corresponding friend.tech @username.
3. The Frenrug agent passes user’ messages through multiple LLMs run by different Infernet nodes. Initially, we’ll be running all the Infernet nodes, but soon enough, we’ll have folks in the community run nodes as well.
4. Each node responds on-chain with their LLM-produced vote on whether Frenrug should take action. Since LLMs are non-deterministic, each LLM might produce a different response even if it’s the exact same model!
5. When enough nodes respond, an aggregation request is kicked off entirely on-chain.
6. An off-chain Infernet node picks up this request and aggregates the various LLM votes into a single action via a supervised classifier and relays a corresponding validity proof on-chain.
7. The Frenrug agent contract executes this action (buy key, sell key, or no operation)
8. Frenrug key holders see a reply in the Frenrug chat room with each of the LLM agents’ votes and the final output

Those who have played with the GPT-4 Turkish Carpet Salesman will find this Web3-native reincarnation of the concept familiar. Akin to convincing the salesman to sell you a carpet, you now have the ability to convince our Frenrug agent to buy and sell friend.tech keys.

We’ve initially funded the Frenrug agent with 10 ETH—good luck.

How does Frenrug work?

Under the hood, Frenrug is powered by Ritual’s Infernet middleware and suite of tooling. Frenrug uses two seperate Machine Learning models:

1. The LLM model is a fine-tuned, prompt-engineered variant of a popular open-source model. We injected a bit of extra randomness to add to the fun, so don’t get mad if it doesn’t do what you expect it to do the first time. We prompt the model with users’ input messages, asking the LLM what action should be taken. Instead of relying on a single LLM’s output, multiple Infernet nodes each come to their own verdict, each of which can be a different LLM thereby allowing for unbounded diversity of models in the voting process.
2. Once a sufficient number of Infernet nodes have responded on-chain, a request is kicked off to aggregate the LLM votes into a single action. This is powered by a classifier model to provide the final verdict.

Embodying the trustless spirit of Web3, Frenrug pioneers novel AI agent technology:

1. For the LLM model, the hashed input, embedding vectors created from the LLM response, and rationale are all posted on-chain. We aggregate the LLM response’ embedding vectors entirely on-chain in the Frenrug smart contracts.
2. For the Classifier model, we directly consume the aggregated embedding vectors from the chain. Once processed by an Infernet node, the raw input embedding vectors, hash of the input data, aggregated action, and a ZKP are all posted on-chain, which enables verification on-chain of the off-chain execution of the multinomial classifier.

To learn more, please reference our documentation.

What’s next?

Frenrug is live and deployed to Base, today: 0x5bfe1Ed1741c690eC3e42795cf06a4c38Ed3BC0c.

Frenrug is an open experiment. There is no specific goal, objective or anything else. Users may want to maximize the amount of money they make off Frenrug or cause chaos and get the agent to sell others’ keys. However, we do want to see people push the absolute bounds with Frenrug.

In the spirit of web3 experiments, we want to make things a little more interesting. This experiment will run until December 24th, 11:59PM EST, and on December 25th, there may or may not be gifts for a variety of participants of the experiment depending on who’s been naughty and nice. This also may or may not depend on how badly the agent gets rugged out of its 10 ETH :).

We’ll be extra considerate to the users that post screenshot(s) of them getting the Frenrug agent to act in novel ways or to do something that other users have not yet on Twitter and tag @frenrug. This could be anything from getting it to break character to getting the agent to sell more keys than anyone else, but, we’re not going to tell you the exact achievements we have in mind (where’s the fun in that?). Users that meet these hidden criteria may or may not have some hidden surprises in store for them. Reminder: your posts MUST include a screenshot and tag the @frenrug account to be considered.

Stay tuned via our Twitter for more updates as we open-source our code and methodology over the coming days, and continue to explore the intersection of AI Agents and Web3.

How do I build my own Frenrug?

Our ultimate goal with this experiment is to expand the design surface for agents on-chain through a combination of classical and foundation AI models with verifiability. We will open source as much as possible over the next week so that anyone in the community can launch their own version of this experiment. Infernet allows for a broad design surface as it relates to the idea of LLM-as-agent setups interfacing with on-chain contracts.

Is this safu?

None of these contracts have been audited, and while we won’t rug you, you may rug yourself.