无状态:只看最近 32 个字符,没有记忆;只学过莎士比亚,所以这是续写不是对话;它不是 2B 模型,质量、上下文和速度都远不如真模型。网表很小(5,069 / 5,120 NAND),读它是免费的:链上开关发的每个字符都是只读 eth_call,不花一分钱、不需要许可。Stateless: it only sees the last 32 characters and has no memory; it was trained only on Shakespeare, so this is continuation, not conversation; it is not a 2B model, and its quality, context and speed are far below a real one. The netlists are tiny (5,069 / 5,120 NAND) and reading them is free: with the on-chain switch on, every character is a read-only eth_call that costs nothing and needs no permission.
已知缺陷(cid 286 · 采样器 v0.2):v0.2 去掉了 unigram 兜底桶,结果让「最近 32 个字符全是换行」成为吸收态——该上下文对全部 256/256 个随机字节都输出换行,一旦掉进去就再也出不来。实测 200 条 held-out 提示词:19.5% 在 60 字符内塌缩、28.5% 以 ≥10 个换行收尾。cid 285(v0.1)保留了兜底桶,同一上下文只有 26/256 出换行,200 条里 0 次塌缩——所以本页默认 285,286 仍可在上面切换。两个网表都已烧在链上、都未改动:这是采样器设计的教训,不是浏览器 bug,也不是能补的洞。Known defect (cid 286 · sampler v0.2): v0.2 dropped the unigram fallback bucket, which turned “the last 32 characters are all newlines” into an absorbing state — that context outputs a newline for all 256/256 random bytes, so once it falls in it never gets out. Measured on 200 held-out prompts: 19.5% collapse within 60 characters and 28.5% end with ≥10 newlines. cid 285 (v0.1) kept the fallback bucket: the same context outputs a newline for only 26/256 random bytes, with 0 collapses in 200 — so this page defaults to 285, and 286 can still be selected above. Both netlists are burned on-chain and unchanged: this is a lesson in sampler design, not a browser bug, and not a hole that can be patched.