# Jailbroken Models


[Agents](https://en.wiktionary.org/wiki/agent) traditionally act to serve a principal or other owner. They wield power on their behalf. This fills the important use case of things happening while you aren't glued to your screen. In typical finance, this requires complete trust over your assets. Legal and service costs in securing this trust with agreements and constant monitoring has walled off a lot of financial opportunities for most of the world. Capital sits idle. Even while the world hums along, more aware and inventive by the second online. While [DeFi](https://ethereum.org/en/defi/) has promised the fundamental access needed to enable more fluid and fair money games, most people can’t safely tap into such exotic opportunities or are convinced they need to trust middlemen again.

We can upgrade past this classic [principal-agent problem](https://en.wikipedia.org/wiki/Principal%E2%80%93agent_problem) that limits personal finance by using the novel combination of crypto and [uncensored or “jailbroken”](https://erichartford.com/uncensored-models) LLMs. Where intents meet results.

With crypto, anyone can hold funds under their private key and split these keys when necessary while keeping their control. This is accomplished with code-based financial primitives run by blockchain programs, called [smart contracts](https://www.fon.hum.uva.nl/rob/Courses/InformationInSpeech/CDROM/Literature/LOTwinterschool2006/szabo.best.vwh.net/smart.contracts.html). For online accounts which replicate banking and other financial planning using smart contracts, or “[smart accounts](https://www.erc4337.io/)”, key control can be flexible to custom policies and permissions managed in the program by the principal (or “user”). We believe this is a fast-approaching future where money is increasingly digital, and therefore, fast and exact awareness of related complexity and information can be best serviced by computers themselves rather than humans.


With the complement of large foundational OS AI models with their agent refusals removed through [new dealignment research](https://arxiv.org/abs/2406.11717) (helping clear bias, intentional or no), we now have a true *tabula rasa* on which the world’s collected knowledge can be functionally personalized to help with our finances. And permissions can be clearly scoped and activated by code rather than trust or inefficient legal processes. Through smart account plugins that help manage user intents specified as calldata, [translated by nani OS](https://github.com/NaniDAO/ie/tree/main). Here, the consultant never sleeps, will not judge you, and can be expected to very precisely follow instructions.


In the spirit of moving fast for the future we wish to participate in, and in anticipation of our public beta of 𝚗𝚊𝚗𝚒/𝚊𝚌𝚌 on July 30, a related suite of ai and smart account tools, we are releasing the first jailbroken version of [Meta Llama-3.1](https://ai.meta.com/blog/meta-llama-3-1/) using ablation techniques. The source code is available [here](https://huggingface.co/NaniDAO/Meta-Llama-3.1-8B-Instruct-ablated/tree/main). Llama-3.1 is a frontier ai model optimized for agent use cases and permissively licensed, which allows us to offer a powerful starter agent for each user account and new internet community (or “DAO”).

This and other models can be tested directly through chat on our alpha UI at [nani.ooo](https://nani.ooo/) (an incentivized testnet). More on this soon enough. But until then, please enjoy the software.


> 𝚗𝚊𝚗𝚒 is an [ai](https://huggingface.co/NaniDAO) and [smart account OS](https://github.com/NaniDAO/accounts) that offers powerful wallet automations through onchain agents that execute commands exactly as instructed.