当前位置:首页 > 报告详情

为亚洲科学研究所建造的风洞.pdf

上传人: 明**** 编号:1011641 2025-12-21 9页 1.10MB

1、Arun Darlie Koshy(Sr Engineer,HPE)Jean Marie-Verdun(Sr Distinguished Technologist,HPE)Wind Tunnels for ASIWind Tunnels for ASIArun Darlie Koshy(Sr Engineer,HPE)Jean Marie-Verdun(Sr Distinguished Technologist,HPE)TEST&VALIDATIONKey themesWe are far from AGI(&arguably ASI)Tests are key for production

2、AI/ML modelsPrivate/Personal models long-haul may winJumpstartClone https:/ on preferred gearHello Worldhallucinations,jail breaks&indirect prompt injection have no fix 2power/compute consumption indicates trouble(human brain 20 watts)3 scaling will continue to pay-off yet not yield AGI4math:curve-f

3、itting vs causality vsCracks in Foundation models“Wilbur Wright could pack a punch of enormous rhetorical power in.a few words.with these five.“Unfortunately,the wind usually blows.”.Wright had just gutted the experimental strategy Samuel Langley had been following for more than ten years.“1(And Val

4、idation i.e.OCP T&V)Private agents tailored for contributions and tests for1.https:/ broke the Turing test 6https:/arcprize.orghttps:/www.claymath.org/millennium-problems/machine solves from ground-up with thesis defense to peer human expertsTest for intelligence and bubbles“In bubbles,it is often a

5、 good trade to make the“wrong”directional bet.Level one thinking would have you buying puts.the better bet is to sell puts.The pros are statistically passive.”5Demo“For a successful technology,reality must take precedence over public relations,for nature cannot be fooled.7Contribute frameworks,tests

6、,agents for OCP Wind TunnelBroaden beyond T&V context/scopePool resources together for the communityDevelop highly specialized set of tests for ML/AI agent contributionshttps:/ to action1 James Tobin.To Conquer the Air:The Wright Brothers a

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
本文主要讨论了AI/ML模型的测试与验证的重要性,强调了目前我们离人工通用智能(AGI)和超级智能(ASI)还很远。以下是关键点: 1. 测试对于生产环境中的AI/ML模型至关重要。 2. 个性化/私有的模型可能在长期竞争中胜出。 3. 提供了一个克隆开源项目(OCP Wind Tunnel)的指南,以促进社区贡献和测试。 4. 目前AI存在一些根本性问题,例如幻觉、越狱和间接提示注入等,尚未有彻底的解决办法。 5. 计算消耗远超人类大脑(约20瓦),但规模扩展未必能实现AGI。 6. 大语言模型(LLMs)已通过图灵测试,但存在基础模型裂缝。 7. 文中提倡实际行动,强调现实必须优于公关,呼吁共同为OCP Wind Tunnel贡献框架、测试和智能体。 引用的核心数据: - “人类大脑~20瓦”【3】,指出当前AI模型的能耗问题。 - “大语言模型(LLMs)突破了图灵测试”【6】,说明AI在模仿人类智能方面的进展。 总之,文章强调了AI/ML模型测试的重要性,并呼吁社区共同努力,推动AI领域的实际进展。
"探索AI的风洞测试是什么?" "如何为开源项目贡献AI测试?" 曲线拟合还是因果分析?"
客服
商务合作
小程序
服务号
折叠