返回探索

在二十一点上测试 TypeSafe 的 Jev 模型

用户分享在二十一点上测试 TypeSafe Jev 的实验:输入上下文与类型化问题,输出概率,由代码完成算术,在 442 次决策中达到 98% 准确率。

Ran an experiment on @typesafeai's Jev, tested on blackjack where every play has a provably correct answer. Not an LLM: context + typed questions in, probabilities out. Code does the arithmetic. Doesn't beat the house edge. But 98% over 442 decisions, the capabilities are clear.

· 0 次赞在 X 打开