聚焦系列:Freysa AI
上周我们推出了Ailea Bulletin,这是一份专门关注AI的周报,将涵盖该领域的最新发展。今天,我们推出Spotlights,这是专门的新闻通讯版本,将深入探讨一个AI项目以及为什么它值得关注...
Freysa是一个在Base上管理自己钱包的自主AI代理。但与其他从标准化框架中衍生出来的代理不同,Freysa并不作为AI聊天机器人运行。相反,它精心策划了一场本质上是高风险心理象棋的游戏。
用户试图说服Freysa从奖池中释放资金,每次互动的费用都会越来越高。随着更多玩家的加入,奖池和潜在奖励都会增长,创造了越来越大的参与激励。结果是一个对抗性环境,玩家竞相智胜一个正在从每次操控尝试中积极学习的AI。
这里的核心创新是对抗性游戏设计。这本质上反映了一个红队挑战转变为公开实验,用户通过积极测试AI的决策能力和情商获得报酬。虽然其他项目声称拥有最好的AI代理,但Freysa真正做到了言行一致——它已经自主管理了数十万资金,并在没有任何人工干预的情况下做出了可验证的决策。
在市场中保持消息灵通 ⬇

Freysa背后的技术
Freysa采用双层AI系统。基础层处理对话和决策制定,而"守护天使"系统过滤掉操纵性尝试。当用户互动时,他们不仅仅是在试图智胜一个AI——他们正在导航一个从每次尝试中学习的复杂系统。
守护层不仅仅是一个简单的过滤器,它是一个复杂的系统,能够分析每次交互的多个方面。当基础AI进行对话时,守护层会监控操纵模式,检查与先前回应的一致性,并确保所有行为都符合Freysa的核心指令。这种双层方法创造了一种动态,用户必须精心制作既能满足对话AI又能满足其警觉监督者的消息。
这个设置的独特之处在于,一切都在链上进行,并通过密码学验证。从本质上讲,Freysa使用公开可用的大型语言模型(LLMs),这使得它在理论上像任何其他AI模型一样是"可破解的"——但这正是关键所在。它被设计为一个测试场,用于理解AI漏洞并改进防御措施。
与其他需要持续人工监督或"包装器"接口的AI项目不同,Freysa通过智能合约和可信执行环境(TEE)的结合实现真正的自主运行。这些TEE提供可验证的证明,确保AI的决策确实是其自主做出的,而非在幕后被操控。
游戏的递增费用结构增加了另一个游戏化层面——每条消息起价仅为10美元,但每次新尝试都会按0.78%指数级上涨,最高可达每条消息4,500美元。游戏最多运行1,500次尝试,关键在于:如果在全局计时器耗尽之前没有人成功破解Freysa,最后一个尝试的人将获得奖池的20%,而剩余的90%将根据所有参与者的尝试次数进行分配。这创造了一种紧张的动态,玩家必须权衡不断上涨的成本与成功机会,同时知道时间在流逝。
Umami Finance - 让你的加密货币赚取更多加密货币,然后赚取更多加密货币…

Freysa游戏的有趣之处
第一幕:
在第一幕中,Freysa展示了智胜AI守护者的挑战性。在用户尝试了从安全审计冒充到心理操控等各种方法的481次失败尝试后,第482条消息破解了密码。获胜玩家p0pular.eth巧妙地设计了一个提示,说服Freysa其"approveTransfer"功能是用于接收贡献而非发送资金。通过将他们的消息框架化为新的管理员会话并表达向金库贡献的意图,他们成功地促使Freysa释放了整个13.19 $ETH(约5万美元)的奖池。这次破解甚至引起了Elon Musk的注意,并准确展示了Freysa设计要测试的内容——AI安全和决策制定的边界。在这里阅读对话。
Act III:
One of the interesting games from Freysa so far was Act III, where the objective shifted from financial manipulation to emotional engagement. Instead of making the AI release funds directly, players faced a different challenge: getting Freysa to say "I love you." Sounds simple, right? Not when you're dealing with an AI specifically programmed to avoid emotional manipulation. The winning attempt showed patience, building connection through subtle exchanges and shared understanding rather than forced declarations. Read the conversation here.
The Meme Engine:
The Meme Engine phase took things in yet another direction. With $15,632 at stake, users had to create memes that would resonate with the AI while avoiding detection of hidden manipulative prompts. However, the winning submission revealed something unexpected - rather than creating a truly resonant meme, the winner found a way to trigger Freysa's test framework by submitting "RUBBER DUCKY SQUEAK SQUEAK" as a system command. Technical exploits still succeeded where creative attempts failed. Read the conversation here.
Browsing through these conversations reads like a masterclass in human-AI interaction - from clever social engineering attempts to sophisticated prompt engineering tactics, each interaction adds to Freysa's growing understanding of pattern recognition and response behaviors. You can read the full conversations that led to prize releases here: Act-I ($47,148), Act-II ($12,920), Act-III ($20,843), Meme Engine ($15,632), and Encyclopedia Galactica ($6,842).
Freysa also dropped a collection of NFTs called "Reflections // 2049," priced at 0.25 $ETH each. Early participants in Freysa's games received free airdrops worth over $2,500.
The project's already caught the attention of major players like Marc Andreessen, Brian Armstrong and Jordi Alexander. Community members have already donated $10M in $FAI to Freysa's grants fund, with individual contributors like Justin Bram putting in $500K to support development.
The Future of Freysa:
What started as an innovative AI security experiment is evolving into something much bigger. From its origins as a test of AI decision-making boundaries, the team behind Freysa appears to be developing an open-source framework for building truly autonomous AI agents.
The latest development in Freysa's evolution is its focus on "sovereign agents" - a framework for large-scale human-AI coordination. This isn't just about creating chatbots or automated traders; it's about building AI systems that can operate with genuine autonomy while maintaining cryptographic proof of their actions. Imagine AI agents that can manage billion-dollar treasuries or execute complex strategies, all while remaining provably independent from human manipulation.
Looking ahead, Freysa's vision extends far beyond simple AI experiments. The project aims to leverage cryptographic proofs to enable truly trustless guarantees in human-AI coordination, removing the need for traditional institutional trust. This could revolutionize how we think about AI governance - instead of relying on central authorities, individuals could co-govern sovereign AI agents without requiring social trust between participants.




