26
0[Hosts]
[Guest speaker]
Li Yin, the author of AdalFlow and the founder of Sylph AI.
AdalFlow is a library to build and auto-optimize any LLM task pipeline. Before AdalFlow, Li was an AI research scientist at Meta AI and a PhD dropout from UT Arlington.
AdalFlow is named in honor of Ada Lovelace, the pioneering female mathematician who first recognized that machines could go beyond mere calculations. As a team led by a female founder, AdalFlow aims to inspire more women to pursue careers in AI.
[What you’ll hear]
01:19 Introduction to AdalFlow and SylphAI: AdalFlow aims to build and optimize LLM task pipelines and is part of SylphAI’s open-source efforts
04:42 Deepseek’s impact on Prompt Engineers: Deepseek aids in prompt optimization and enables transparent reasoning
06:44 AdalFlow’s collaborative research with UT Austin scholars excels in automated prompt generation, outperforming DsPy and Text Grade
08:39 The concept of the AI Digital Engineer product: SylphAI focuses on solving LLM application challenges to boost engineer productivity
12:37 Li expresses surprise at their latest paper trending on Hacker News’ front page, noting developers’ growing awareness of eliminating manual prompts
15:48 AdalFlow’s underlying mechanism as an automatic differentiation framework: It optimizes by building differentiable graphs and gradient-driven optimizers, using mini-batch training and backpropagation to refine prompts
18:14 AdalFlow’s origin in April 2024: The goal was to create a library for automated prompt optimization
21:28 Startups must stay focused: Despite AdalFlow trending on Hacker News, the team remains as two-person operation
26:17 Progress on finding a co-founder: Collaboration with UT Austin may naturally lead to finding a CTO
27:16 Socializing in San Francisco: After moving there, Li joined the Mission Control community and engaged actively in events
28:49 LinkedIn’s role in community-building: Attracted attention from senior executives and investors
30:19 Lessons from Connor: Learned to read the book Pitch Anything
31:18 Beware of “ChatGPT Engineers”: Caution against over-reliance on ChatGPT, especially for remote roles
39:24 Simplify code and life: Less code is better; structured code beats spaghetti code
40:20 Women still face challenges in STEM and entrepreneurship
42:55 Investors want grand visions: Founders must pitch ambitiously—no room for modesty
49:20 AdalFlow’s next steps: UT Austin students will use it for research. Future plans include automating dataset creation for components and training cheaper ML models
55:39 Advice for young people: Become a community expert, solve real problems, iterate gradually, and stay action-oriented
57:45 Release daily updates and iterate rapidly
58:39 Advice to past self: Enjoy the journey, specialize deeply, and become an expert
1:01:28 Final advice: Just do it! Never quit, work with users, and tackle real-world problems
[Mentioned articles, books, people]
1/ LLM-AutoDiff: Auto-Differentiate Any LLM Workflow (arxiv)
2/ Pitch Anything by Oren Klaff
3/ Connor Watumull, VP at Bessemer Venture Partners
- - - - - - - - - - - - - - - - - - - - - - - - - - - - -
【主播】
【嘉宾】
AdalFlow 是一个用于构建和自动优化任何 LLM 任务管道的库。在 AdalFlow 之前，Li 是 Meta AI 的人工智能研究科学家，也是得克萨斯大学阿灵顿分校的博士生。
AdalFlow 是以 Ada Lovelace 的名字命名的，她是第一位认识到机器可以超越单纯计算的女性数学先驱。作为一个由女性创始人领导的团队，AdalFlow 的目标是激励更多女性从事人工智能职业。
【你将听到】
01:19 AdalFlow 和 SylphAI 的介绍：AdalFlow 旨在构建和优化 LLM 任务管线，是 SylphAI 开源工作的一部分
04:42 Deepseek 对 Prompt 工程师的影响：Deepseek 有助于 Prompt 优化，能够以透明的方式进行推理
06:44 AdalFlow 与 UT Austin 学者的合作研究在自动 prompt 生成方面表现出色，优于 DsPy 和 Text Grade
08:39 AI Digital Engineer 产品的概念：SylphAI 致力于解决构建 LLM 应用的问题，旨在提高工程师的生产力
12:37 Li 对于最新论文登上 Hacker News 首页感到惊喜，开发者开始意识到消除手动 prompt 的重要性
15:48 AdalFlow 自动微分框架的底层机制：AdalFlow 的优化包括构建可微图和梯度驱动的优化器，使用 mini-batch 训练和反向传播来优化 prompt
18:14 AdalFlow 始于 2024 年 4 月，旨在创建一个能够自动进行 prompt 优化的库
21:28 创业公司需要保持专注：即使 AdalFlow 登上了 Hacker News 的首页，团队仍然只有两个人
26:17 寻找 co-founder 的进展：与 UT Austin 合作后，可能会自然而然地找到 CTO
27:16 在旧金山社交的经历：搬到旧金山后，Li 加入了 Mission Control 社区，并积极参与各种活动
28:49 LinkedIn 帮助建立了社区，并吸引了高级管理人员和投资者的关注
30:19 从 Connor 身上学到的：从 Connor 那里学到要读《Pitch Anything》这本书
31:18 警惕 ChatGPT 工程师：要小心那些过度依赖 ChatGPT 的工程师，特别是远程工作的
39:24 简化代码和生活：代码越少越好，结构化的代码胜过大量的 spaghetti 代码
40:20 女性在 STEM 领域和创业领域依然面临许多挑战
42:55 投资者希望听到创业者将他们的梦想和愿景说得更大，在 pitch 时不能谦虚
49:20 AdalFlow 的下一步计划：UT Austin 的学生将使用 AdalFlow 进行研究，未来探索如何自动为每个组件创建数据集，并训练更便宜的机器学习模型
55:39 给年轻人的建议：成为社区专家，解决实际问题，逐步迭代产品，并保持行动
57:45 每天发布新进展，快速迭代产品
58:39 给过去的自己的建议：享受所做的事情，专注于领域，成为专家
1:01:28 最后的建议：Just do it! 永不放弃，与用户一起工作，解决实际问题
【提到的书、资源、与人】
1/ LLM-AutoDiff: Auto-Differentiate Any LLM Workflow (arxiv)
2/ Pitch Anything by Oren Klaff
3/ Connor Watumull, Bessemer Venture Partners 副总裁
【在这里找到我们】
公众号：双成记
收听渠道：苹果｜小宇宙｜QQ音乐
海外用户：Apple Podcast｜Spotify
如果你是一名华人创业者/投资人，期待与你交流。
欢迎添加微信：jacknottim，备注“双成记 - 嘉宾”。
