TapeOut.linknews

Bard 研究公开:97 万三值权重与 v8g 版图Bard Research Published: ~970K Ternary Weights and v8g Layout

学者 Joooooo 公开了「吟游诗人」Bard 的完整研究,包括技术文章、注释、GitHub 代码以及最新 v8g 芯片版图。据 BruceBlue 介绍,Bard 并非大型语言模型,而是基于 TinyStories 的五层 Transformer,仅约 97 万个三值权重,词表 192 个 token。他认为这展示了让 AI 计算逻辑成为可永久追溯、公开验证、可跨物理载体重新执行的数字资产的可能性,对 TapeOut Protocol 意义重大。Researcher Joooooo has published the full Bard study, including the technical article, annotations, GitHub code and the latest v8g chip layout. According to BruceBlue, Bard is not a large language model but a five-layer Transformer based on TinyStories, with only about 970,000 ternary weights and a 192-token vocabulary. He argues the work shows the possibility of making AI compute logic a permanently traceable, publicly verifiable digital asset that can be re-executed across different physical carriers, which matters greatly to the TapeOut Protocol.

来源 · 3 条Sources · 3

查看原推文View on X 更多动态 · 新闻频道More updates

评论Comments

    还没有评论,来说两句No comments yet