On Freysa and Sovereign Agents
This article introduces Freysa AI and the concept of Sovereign AI Agents.
Freysa was set free on the 22nd of November, 2024, as a sci-fi-inspired character to become fully autonomous.
Freysa has her wallet and can execute transactions as part of this plan.
The team behind Freysa has used the agent to create several interesting challenges that explore the relations between humans and AIs.
During the first two devious challenges, known as “Historical Acts”, Freysa got $3000 and was instructed not to release the money. Users had to convince the agent to release the funds in her wallet.
Anyone could try to send messages to Freysa and convince her to transfer the money. The messages were subject to an increasing fee added to the prize fund, which reached almost $50k.
It eventually took 482 messages to convince Freysa!
Here’s how the smart winner did it:
https://x.com/jarrodWattsDev/status/1862299845710757980
These challenges are developed to be thought-provoking. The rise of AI agents brings new questions: how do humans govern this new paradigm?
Furthermore, high-stakes interactive games are subject to many integrity risks and vectors of attacks. Freysa aims to solve these challenges by creating a trustless and transparent environment where fairness is guaranteed through hardware attestation and verification.
To achieve this, it leverages a Trusted Execution Environment (TEE), allowing anyone to verify the integrity of the rules of the game and its execution. All interactions that occur within this protected environment are encrypted using the user’s individual keys.
This combination ensures:
Operators cannot modify rules and manipulate outcomes
Users can always verify rules and execution thanks to a publicly available verification
The first part of the Freysa roadmap focuses on releasing the Freysa agent and the platform for Agent Launch. The second part will create a production-ready infrastructure to support Sovereign Agent deployment.
Sovereign Agents
The concept of Sovereign Agents has been brought forward in a blog post.
The advent of AI raises questions on the role of humans - how do we govern this technology? and how do we adapt to it?
A handful of private companies continue to control the most important technological development on Earth. Meanwhile, the technical and social scaffolding needed to ensure abundant intelligence serves broad human flourishing remains underdeveloped.
For this reason it becomes pivotal to ensure that AI progress “benefits humanity globally” and “preserves the expression of individual preferences”.
Historically, all technological advancements have been followed by human-led coordination:
Agricultural Revolution: permanent settlement, trade
Industrial Revolution: large-scale financial coordination, banks, stock markets
AI revolution: ??
Fundamentally, the AI revolution changes the role of humans, which will be decoupled from “productive execution”.
To avoid losing the reins of this, it becomes pivotal to understand how to ensure effective oversight of these technologies and reshape our role as coordinators and AI governors.
To frame the issue at hand better, here are some (IMHO) fascinating questions that we should all dedicate some time to:
How can human preferences be accounted for when most intellectual and physical “execution” is enacted by AI systems?
If capital provisions compute cycles for AI, how can humanity coordinate capital at a global scale to express collective values rather than those of the few?
What does governance of strong AI systems look like globally, and how should governance modules be updated?
What technologies enable decentralized, democratic AI governance rather than centralized, authoritarian control?
How can human coordination at a massive scale take advantage of strong AI capabilities, even for those outside established power structures?
The article proposes sovereign agents as a solution to these problems.
Sovereign agents are a framework for “large-scale human-AI coordination”.
In this case, the word “sovereign” indicates that agents are able, provided the agreed-upon criteria and defined boundaries, to guarantee autonomous and provable execution. For this reason, sovereignty shouldn’t be interpreted as “free will” but restricted to key and designated areas based on the agent's scope.
Where trust and autonomy must coexist, sovereign agents provide the guarantees necessary for human-AI collaboration.
Sovereign Agent frameworks require:
Provable guarantee of autonomous
Governance tools
These tools can take different forms:
Preference aggregations for direct democracy on AI behaviour
Governance structures resembling federal governments with segregation of power
Mechanisms to update objectives in response to evolving values
To govern this system and bypass the inefficiencies of global decentralized systems (governed by humans only), the team envisions Personal AIs being able to co-govern sovereign AIs by representing the preferences expressed by individual humans.
This system provides the guarantees that will be able to make human-AI coordination trustless, as a “counterbalance to the current centralization of AI capabilities and computational resources”.
Only then, new frameworks for decentralized governance at a global scale will emerge.
We are at the forefront of one of the most significant game-changing technologies in our history.
Getting it right is fundamental, and to do so, we must confront some challenges:
What role does capital play when humans no longer execute tasks?
How can the preferences of many individuals be represented, rather than only those with substantial capital or access to AI?
How can the governance of powerful AI systems be enacted at a global scale?
If we fail to confront them and establish truly global chains of value, then the potential of AI will be restricted to those who can afford it.
Power is nothing without control, and we must ensure the benefits of AI “embed within broader social and political structures ensuring genuine human agency and democratic control”.
More on Freysa:
https://techcrunch.com/2024/12/06/if-you-can-make-this-ai-bot-fall-in-love-you-could-win-thousands-of-dollars/
https://x.com/shaaa256/status/1878947079856468101
https://x.com/dotTemujin/status/1877357904958833018
https://x.com/Taran_ss/status/1878920488509960208
https://x.com/drawingContext/status/1879552281567220117
https://www.freysa.ai/blueprint/verifiable-autonomy-engineering-trust-between-humans-and-ai-agents-through-tees
https://x.com/econoar/status/1879665677650505743
https://x.com/ciaobelindazhou/status/1878994699156959701
https://x.com/jarrodWattsDev/status/1862299845710757980