註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

Eric Xu (e/Mettā)
@xleaps
polymath, polyglot, root of a ternary tree. building prev @Meta @Google @Reddit phd in classic ai; rookie pilot 🛩️; martial artist
加入 January 2007
3.5K 正在關注    34.8K 粉絲
#BuildInPublic# I am open-sourcing an AI simulation engine: SGO (Semantic Gradient Optimization) You build something. You think it's ready. But you have no idea how actual people will react, and that reaction sequence is your product's real roadmap. User research takes weeks and still misses scenarios. You can ask an LLM to role-play a buyer persona, sure, but you get back one data point shaped entirely by whatever role you made up beforehand. SGO takes a different route: simulate against census-aligned synthetic populations. NVIDIA open-sourced Nemotron-Personas-USA, a dataset of one million synthetic Americans built on top of US Census distributions. These aren't the "25-year-old tech worker" archetypes an LLM invents on the fly. They're construction workers in suburban Illinois, artisans in rural Texas, single parents in New York, each with hobbies, habits, and priorities that reflect real demographic distributions in age, education, occupation, income. Paste in whatever you want to optimize: a product landing page, a fundraising pitch, a blog post draft. One user ran his dating profile through it. SGO picks an optimization target and audience for you, then samples from the million-person pool, stratifies by segment, runs each persona through a counterfactual evaluation, and stacks up a ranked list of what to change first and why. About 30 seconds per run. Around $0.10 in API costs. Code is open-source. Live demo on HuggingFace Spaces. Also works as a standalone Skill you can drop into an inner loop with auto-research. HF Space: Things people have already run through it: resumes, business plans, app UX flows, billboard copy, logos, landing page layouts, dating profiles, and one dessert shop's name.
顯示更多
#BuildInPublic# 开源了一个 AI 模拟引擎 SGO (语义梯度优化引擎) 在 AI 世界迭代产品或者功能,目前最最缺少的就是现实世界现实用户的反馈;这些反馈意见序列实际上构成了产品的演化路径。 然而用户(真人)反馈周期较长,且不能覆盖所有的情景。当下,我们常常让 LLM "假装"某一类用户得到一个近似的反馈,但这种反馈都是一个一个的数据点,完全取决于事先规划好了的角色。 SGO 采用的思路是: 用和人口普查对齐的合成数据来模拟真人用户。NVIDIA 开源了多个主权数据集,比如对于美国,Nemotron-Personas-USA 数据集里有一百万个基于美国人口普查数据生成的合成人物。不是那种 LLM 随便编的"有着三十年经验的工程师",而是有完整背景的人——伊利诺伊郊区的建筑工人、德州农村的手工艺人、纽约的单亲妈妈等等。他们有各自的爱好、习惯、关注点。这些人的年龄、学历、职业、收入分布都跟真实人口一致。 SGO 的采样, 模拟和梯度计算框架可以让你直接从这些人里拿到反馈,周期大约 30 秒,LLM API 花费大约 $0.10。 使用方法也很简单:把要优化的东西贴进去,比如产品描述、融资 pitch、一则爆款文章等等(有一个用户甚至把他的约会 profile 放进去优化)。总之什么都行。 SGO 会很科学的帮你自动选择优化目标和目标受众。确定好以后,从这 100 万个有机数据人群中科学采样 (stratified sampling)、分类聚类、逐一询问反馈(contrafactual inquiry)、对照目标,逐一构建所谓的"语义梯度" (相当于目标对于各个变量的 Jacobian 矩阵), 以及最终的汇总反馈和迭代方向。 代码开源,目前部署在 HuggingFace Spaces 上可以直接试用。 你可以把 SGO 作为 Skill 单独使用,也可以把它放在一个内循环里,和 auto-research 联合使用。 HF Space: 希望 SGO 和 auto-resesarch 结合,帮助大家优化那些跨越数字世界和现实世界的许多场景。 PS: 现有的跑通的场景 * 简历优化 * 商业计划 * App UX 设计 * 广告牌设计 * LOGO * 网页的版式和颜色 * 约会档案 * 一个甜点屋的名字
顯示更多