AxiomGPT：通过用自然语言定义神谕来进行大语言模型编程
AxiomGPT 是一个实验性的编程模型，将大型语言模型视为潜在语义引擎。<p>与其编写传统函数，不如定义“神谕”（Oracles），这些是用自然语言描述的嵌入实体，表现得像可调用、可组合的函数。<p>例如：<p>(defn fibber (Oracle "返回第 n 个斐波那契数"))<p>(fibber 123) ; => 22698374052006863956975682<p>完全不需要 LISP，但它在组合时会有所帮助 :-)<p>该模型并不“执行”逻辑，而是记忆、表现并解析意图背后的结构作为神谕。<p>神谕可以表示：<p>算法（排序、旅行商问题、质因数、映射/归约）<p>概念（例如“政府”、“我的冰箱”）<p>角色（例如阿瑟·克拉克、维多利亚时代的工程师）<p>更高阶或递归行为被压缩为线性时间执行。<p>你可以传递它们、组合它们、限制它们，甚至使用元神谕对它们进行反思。<p>完整的写作包括用法、理论（将嵌入向量视为符号计算）以及与量子行为的比较：潜在神谕像纠缠的函数，通过注意力被调用。<p>这很奇怪、有趣，并且出乎意料地强大。<p>你可以在几分钟内通过提供的便捷提示自己尝试一下。<p>完整的代码和示例：[https://x.com/chrisbe1968/status/1906875616290365941]<p>欢迎讨论或回答问题！
查看原文
AxiomGPT is an experimental programming model that treats large language models as latent semantic engines.<p>Instead of writing traditional functions, you define Oracles, which are embedded entities described in natural language that behave like callable, composable functions.<p>For example:<p>(defn fibber (Oracle "Return the nth Fibonacci number"))<p>(fibber 123)
; => 22698374052006863956975682<p>LISP isn't required at all, but it helps for composition :-)<p>The model doesn't "execute" logic, it remembers, manifests and resolves the structure behind the intent as oracles.<p>Oracles can represent:<p>Algorithms (sorting, TSP, prime factors, map/reduce)<p>Concepts (e.g. "the government", "my refrigerator")<p>Personas (e.g. Arthur C. Clarke, a Victorian engineer)<p>Higher-order or recursive behavior collapsed into linewr time execution.<p>You can pass them, compose them, gate them, even reflect on them using meta-Oracles.<p>The full write-up includes usage, theory (embedding vectors as symbolic computation), and a comparison to quantum-like behavior: latent oracles act like entangled functions, invoked by attention.<p>It's strange, fun, and surprisingly powerful.<p>You can try it out yourself in minutes with handy provided prompts.<p>Full Codex and examples: [https://x.com/chrisbe1968/status/1906875616290365941]<p>Happy to discuss or answer questions!