Aditya A.

中文参考

昨晚做了一个愚蠢的实验：通过@typesafeai使用Jev（决策/分类器模型，而不是文本生成模型）来编写C++。 在每一步，我都会给它一组语法有效的下一个令牌，然后让它..选一个。没有自由形式的生成，没有幻觉的语法，纯粹是“从这个法律集中选择”。“它不是为此而建的，这就是它有趣的原因。 从根本上不生成文本的模型中获得了C++的工作部分。不要声称这是您应该如何编写代码，而且这比仅使用实际的代码模型更慢、更愚蠢。但这很好地证明了一个受足够严格约束的分类器可能会看起来像一个生成器。 这里的新颖性大于实用性，但这是一种愚蠢的实验，它比官方演示更能告诉你有关模型的实际护栏的信息。

原帖全文 / EN

Ran a dumb experiment last night: used Jev by @typesafeai (a decision/classifier model, not a text-gen model) to write C++. At every step I gave it the finite set of grammatically valid next tokens and let it just... pick one. No free-form generation, no hallucinated syntax, purely "choose from this legal set." It's not built for this, that's what made it fun. Got a working piece of C++ out of a model that fundamentally does not generate text. Not claiming this is how you should write code, and it's slower and dumber than just using an actual code model. But it's a neat proof that a classifier constrained tightly enough can stumble into looking like a generator. Novelty over utility here, but it's the kind of dumb experiment that tells you more about a model's actual guardrails than the official demos do.