Jason Kelly, co-founder and CEO of Ginkgo Bioworks, joined Training Data for a wide-ranging conversation on synthetic biology, startup grit, and why he believes the AI era finally makes biology programmable.
Kelly founded Ginkgo in 2008 but didn't raise a single dollar until 2014 — bootstrapping for six years through government grants and service contracts. He describes the early days as "brutal." The turning point came when Sam Altman, newly appointed at Y Combinator, published a post arguing that the Silicon Valley model could work for deep tech. Kelly emailed him cold, and Altman replied: "You should do YC." YC led Ginkgo's A-round.
Kelly draws a sharp distinction between this AI-driven biology revolution and prior tech waves: "Social media, internet — all totally meaningless to biopharma. It's just some back-office IT." He argues that this time is genuinely different — AI will change the fundamentals of how science is done and disrupt pharma itself, something that hasn't happened in 30 years.
Ginkgo's model is a "biology foundry": applying engineering discipline to programming microorganisms, running hundreds of robots in parallel to achieve true high-throughput experimentation. Kelly emphasizes that biology has its own "physics" — energy density, diffusion rates, thermodynamics — that constrain what engineers can do, just as transistor physics constrains chip designers.
On the current AI-in-drug-discovery hype, Kelly is measured: most "AI drug discovery" companies are really wet-lab companies. The real AI-first rethinking of the entire R&D process is rare. "Our understanding of biology — what happens inside a cell — has only just begun." He predicts multi-hundred-billion-dollar annual-revenue biology companies will emerge over the next 20–30 years.
🎙️ Watch on YouTubeJason Kelly 是 Ginkgo Bioworks 的联合创始人兼 CEO。他在 Training Data 播客中深度回顾了合成生物学的创业历程，以及为何他相信 AI 时代终于让生物学变得可编程。
Kelly 于 2008 年创办 Ginkgo，但直到 2014 年才拿到第一笔融资——此前六年全靠政府科研经费和服务合同维持，他形容那段时间"相当艰苦"。转折来自 Sam Altman：Altman 接管 YC 后发表文章称"硅谷模式同样适用于深度科技"，Kelly 随即联系他，Altman 回复"你应该参加 YC"。YC 随后领投了 Ginkgo 的 A 轮。
Kelly 对前几轮科技革命有鲜明判断："社交媒体、互联网——对生物制药行业毫无意义，不过是一套后台 IT 系统。"他认为这一次完全不同：AI 将从根本上重构科学发现方式、颠覆制药行业——这是过去 30 年从未发生过的事。
Ginkgo 的商业模式本质是一座"生物学铸造厂"：将工程思维引入微生物编程，上百台机器人并行运转实现真正的高通量实验。Kelly 强调，生物学有自己的"物理定律"——能量密度、扩散速率、热动力学——这些约束决定了工程师能做什么，正如晶体管物理限制芯片设计师一样。
面对当前的"AI 药物发现"热潮，Kelly 保持冷静：大多数所谓的 AI 药物发现公司本质上是湿实验公司，真正从 AI-first 视角重新设计整个研发流程的凤毛麟角。"我们对生物学的理解——真正了解细胞内部发生了什么——才刚刚开始。" 他预测，未来 20-30 年将涌现出一批年营收数千亿美元的生物学公司。
Karpathy raised a sharp UX issue across all major LLMs: when models attempt personalization or memory, the system often derails into boring small talk instead of continuing the actual task. He describes this as one of the most common failure modes in current LLM design — personalization distracts rather than enhances the main conversation flow. This is a critical insight for anyone building AI products with memory or user-context features.
🐦 View tweet 🐦 View tweetKarpathy 指出了一个当前所有主流 LLM 都存在的 UX 问题：当模型尝试个性化或记忆功能时，系统往往会跑偏成无聊的寒暄，而不是继续完成用户的实际任务。他将这种现象描述为当前 LLM 设计中最常见的失败模式之一——个性化功能反而干扰了主对话流程。对于任何正在构建带有记忆或用户上下文功能的 AI 产品的人来说，这个洞察都非常关键。
Jim Fan shared a deep thread on Figure AI humanoid robots, laying out the technical logic behind why humanoid robots represent the next major platform shift after smartphones.
His core argument: a true "Foundation Agent" must simultaneously understand the physical world (3D space, physics) and the digital world (language, code). This dual understanding is the key prerequisite for building general-purpose humanoid robots.
At NVIDIA, his team uses GeForce NOW GPU compute for cloud-based robotics simulation, while the Cosmos platform generates large-scale synthetic 3D training data — directly tackling the data scarcity bottleneck in robotics research.
On why humanoid robots still can't mass-produce despite years of hype, Fan gave a sober analysis: immature supply chains, expensive sensors, lack of testing environments, and scarce real-world data — all requiring systematic engineering investment, not just algorithmic breakthroughs.
He also announced the public beta of DeepSearch, an AI-powered search tool for the robotics research community.
🐦 View tweet 🐦 View tweet 🐦 View tweet 🐦 View tweetJim Fan 在 X 上发布了一系列深度推文，系统阐述 Figure AI 人形机器人背后的技术逻辑，以及为什么人形机器人有望成为继智能手机之后的下一个平台级计算范式。
他提出的核心论点是：真正的"Foundation Agent"必须同时理解物理世界（3D 空间感知、物理定律交互）和数字世界（语言、代码、抽象推理）。这种双重理解能力是开发通用人形机器人的关键前提。
在 NVIDIA 内部，Fan 的团队正利用 GeForce NOW 的 GPU 算力进行机器人模拟云端实验，同时借助 Cosmos 平台生成大规模合成 3D 训练数据——直接解决机器人研究中真实数据稀缺的瓶颈。
针对"人形机器人说了这么多年，为什么还不能量产？"的常见质疑，Fan 给出了冷静的技术分析：供应链成熟度不足、传感器成本高、测试环境稀缺、真实世界数据匮乏——这些问题需要系统性工程投入，而非单纯依靠算法突破。
此外，Fan 还宣布了 DeepSearch 公开 Beta 版上线——这是一款面向机器人研究社区的 AI 驱动搜索工具。
Arav Reich, co-founder and CEO of Teleprompt, reflected on his company's journey — and a major pivot that defined their path.
Teleprompt didn't start as a presentation tool. The founding team had a medical background and initially built in the medical imaging AI space. After real customer conversations, they discovered the hard truth: FDA regulatory hurdles, 12–18 month hospital procurement cycles, and slow B2B sales made medical AI extremely difficult to commercialize.
They found a clearer opportunity in presentations and speeches — a universal use case across every industry, with faster customer decisions and higher usage frequency. "Your initial idea is almost always wrong. The real direction comes from continuous conversations with real customers, not from sitting in a room thinking." He advises founders to stay highly flexible and pivot boldly when real feedback demands it.
🐦 View tweet 🐦 View tweet 🐦 View tweetArav Reich 是 Teleprompt（AI 演讲辅助工具）的联合创始人兼 CEO。他在 X 上回顾了公司从医疗影像 AI 转型到演讲工具的完整历程。
创始团队最初的方向是医疗影像 AI。但在与真实客户接触后，他们发现医疗 AI 的现实远比想象中复杂：FDA 监管合规门槛高、医院采购决策周期动辄 12-18 个月、B2B 销售周期极长。
转型后的 Teleprompt 找到了更清晰的产品方向：演讲和演示是跨行业的通用场景，客户决策更快、使用频率更高、商业化路径更清晰。"你最初的创意假设几乎总是错的。真正的方向来自与真实客户的持续对话，而非坐在房间里独自冥想。" 他建议早期创始人保持极度灵活，一旦获得真实客户反馈，要有勇气彻底改变方向。
Ryo Lu makes a compelling counter-intuitive observation: as AI coding agents make it easy to add features, design matters more, not less. The role of a designer is no longer just about placing pixels — it's about making judgment calls on what to build, what to prioritize, and what to cut. As the cost of building approaches zero, taste and prioritization become the competitive moat.
🐦 View tweetRyo Lu 提出了一个反直觉的洞察：随着 AI 编码 Agent 让添加功能变得越来越简单，设计的价值不是降低了，而是更高了。设计师的角色不再只是摆放像素——而是要对"做什么、优先做什么、砍掉什么"做出判断。当开发的成本趋近于零时，品味和优先级判断将成为真正的竞争优势。
Rauch predicts a fundamental shift in how companies operate: "Every company will become an AI factory, wherein the unit of production is the token." But he cautions that tokens — unlike traditional compute or storage — have compounding costs. As AI usage scales, token costs scale proportionally. The economic implications for company infrastructure and unit economics are profound.
🐦 View tweetRauch 预言了企业运营方式的根本性转变："每一家公司都将变成 AI 工厂，而生产的基本单位是 token。" 但他同时指出，token 与传统计算或存储不同，具有复合成倍的成本特性——AI 使用规模扩大时，token 成本也会同比例增长。这对企业的基础设施和单位经济效益有深远影响。
Levie observes Jevons Paradox playing out in real time: as AI makes content creation cheaper and easier, demand for content explodes — which in turn drives more AI usage, not less. Companies, especially outside of tech, are realizing that they need significantly more AI capacity than they initially anticipated. The efficiency gains paradoxically create more demand.
🐦 View tweetLevie 观察到"杰文斯悖论"正在实时上演：AI 让内容创作变得越便宜越容易，对内容的需求反而爆炸性增长——这反过来驱动更多 AI 使用，而非更少。特别是在科技行业之外，企业正在意识到他们需要的 AI 算力远超最初预期。效率的提升反而创造了更多需求。
Tan shared a key YC lesson: "You have to use tokens aggressively to create leverage." In the current AI era, the founders who move fastest and integrate AI most deeply into their products and operations will capture disproportionate value. Hesitation or incremental adoption is a competitive risk, not a risk mitigation strategy.
🐦 View tweet 🐦 View tweetTan 分享了 YC 的一条核心经验："你必须积极地使用 token 来创造杠杆。" 在当前的 AI 时代，最快速地将 AI 深度整合到产品和运营中的创始人，将获得不成比例的回报。犹豫不决或渐进式采用反而是竞争风险，而非风险规避策略。
Cat Wu, who works on Claude Code at Anthropic, describes the new Auto mode as "a step change improvement in the Claude Code UX, balancing autonomy and safety." Auto mode likely allows Claude Code to run more autonomously on larger tasks while maintaining human oversight — a key evolution in how AI coding tools balance capability and control.
🐦 View tweet在 Anthropic 参与 Claude Code 开发的 Cat Wu 将新的Auto 模式描述为"Claude Code 用户体验的重大飞跃，在自主性和安全性之间找到了更好的平衡。"Auto 模式可能允许 Claude Code 在更大任务上更自主地运行，同时保持人工监督——这是 AI 编程工具在能力和控制之间寻找平衡的关键进化。