Dr. Fei-Fei Li Pleit Voor Mensgerichte AI En Buiten-Cortex Modellen

李飞飞博士倡导以人为本的AI与皮层外模型 只有将人类自主性置于核心，并将视野从单纯的语言模型拓展至空间智能，人工智能才能真正增强人类的直觉与创造力。 李飞飞博士与Andrew Huberman博士回顾了人工智能从ImageNet和计算机视觉演进至Transformer以及World Labs 3D空间智能的历程。李博士指出了当前AI的关键缺陷，强调人类智能依赖于生物直觉、情绪状态以及无法量化的内在体验，而这些是互联网数据训练出的模型所无法复制的。他们探讨了AI在医疗保健、软体机器人和K-12教育中的实际应用，主张采用苏格拉底式提示法和能够保留认知努力的AI导师。最终，李博士呼吁赋予公众能动性、建立强有力的伦理防护栏，并在极端的末日论与乌托邦主义之间寻求平衡的中间地带，以确保人工智能提升人类尊严。 视觉是生物演化与人工智能发展的核心驱动力；正如5.4亿年前寒武纪大爆发由感光细胞催生一样，人类大脑约一半的皮层活动专注于处理视觉信息，奠定了现代AI架构的神经科学基础。 (3:46) 李飞飞博士团队于2012年推出的 ImageNet 数据集，结合卷积神经网络（CNN，如 AlexNet）与 GPU 算力，开启了现代深度学习革命，使机器图像分类准确率在数年内超越人类基准。 (12:11) 生物学习机制与当前 AI 存在本质差异：儿童仅需接触极少数样本即可掌握抽象概念，而主流 AI 模型依然极度依赖海量互联网数据的统计模式匹配。 (23:36) 虽然 AlphaGo 的“第37手”展示了机器在规则框架内的创新，但人类真正的创造力与直觉源于未被上传或数字化的内在主观经验与生理信号。 (39:55) 在医疗与科学探索等复杂前沿领域，人机共生协作（如达芬奇手术机器人辅助）远胜于完全自主的机器系统，技术的开发与应用应始终以增强人类自主性（Agency）、动力与尊严为核心。 (58:14)

Dr. Fei-Fei Li argues that artificial intelligence should be built and governed to augment rather than replace human agency, emphasizing that while modern models excel at statistical pattern synthesis across vast datasets, they remain fundamentally distinct from biological cognition, emotion, and lived experience. Vision as the Catalyst for Biological and Artificial Intelligence (3:46) Dr. Li highlights a parallel between evolutionary biology and computing: roughly 540 million years ago, the evolution of primitive photoreceptors in trilobites triggered the Cambrian Explosion, and vision now occupies roughly half of the human brain's cortical processing. In AI, early neural networks drew directly from the hierarchical visual processing discovered in mammals by David Hubel and Torsten Wiesel. However, algorithmic architecture stalled until the development of massive training datasets. Observing that human children categorize tens of thousands of objects by age six through continuous visual immersion, Dr. Li spearheaded ImageNet in the mid-2000s, curating over 15 million images. This precipitated the 2012 computing watershed, where the convergence of ImageNet's structured data, convolutional neural networks (AlexNet), and GPU acceleration caused error rates to plunge into the mid-teens. By 2015–2016, machine accuracy on the benchmark surpassed human visual classification baselines, which Andrej Karpathy had established at approximately 4%. Statistical Pattern Synthesis Versus Biological Cognition (23:36) Despite rapid advancements across language models, speech analysis, and video generation tools like Sora, Dr. Li clarifies that modern AI remains grounded in statistical probability over scraped internet data rather than genuine biological comprehension. A child acquires object constancy and identifies a obscured animal after seeing only three to ten examples, whereas contemporary AI models require billions of data points to adjust parameter weights. Similarly, video generators synthesize physically plausible movement from ingested footage rather than an internal grasp of anatomical or physical laws. Furthermore, Dr. Li and Andrew Huberman note that internal subjective states—such as nostalgia, personalized memories, or the origins of creative artistic impulses—remain inaccessible to AI because they are not digitized or uploaded to the web. While systems can recombine known information in novel ways (exemplified by AlphaGo's unconventional Move 37 within a closed mathematical system), machine outputs like sympathy or urgency reflect objective functions and probability constraints rather than genuine empathy, pain, or visceral human intuition. Embodied AI, Spatial Intelligence, and Collaborative Healthcare (58:14) Looking beyond screen-bound language interfaces, Dr. Li identifies spatial and physical intelligence as the next foundational horizon in AI, which led her to co-found World Labs in 2024 to generate interactive 3D and 4D worlds from text prompts, sketches, and images. True embodiment bridges digital models with the physical world, offering tools to train autonomous robots, assist architects, and relieve healthcare workers through tasks like hospital logistics and elder care. In complex scientific and clinical settings, Dr. Li and Huberman advocate for collaborative human-AI systems over full machine autonomy. In dense, multi-variable medical scenarios—such as robotic liver surgeries using the Da Vinci platform—anatomical variance and limited procedure volumes mean training data is too sparse for safe autonomous execution. Instead, AI serves best by synthesizing cross-disciplinary literature, assisting human surgeons, or powering neural prosthetics that restore motor-impaired patients' vocal tones and expressions. Safeguarding Agency, Educational Guidance, and Multi-Stakeholder Governance (80:37) Dr. Li warns against extreme public rhetoric that alternates between existential doomerism and naive utopianism, arguing both treat the public as passive observers. True human flourishing hinges on a triad of motivation, agency, and dignity. In education, the primary danger is not that students will use AI, but that passive consumption will bypass the cognitive friction required for neural maturation, or conversely, that schools will issue wholesale bans out of cheating fears rather than teaching iterative, Socratic prompting. To manage deployment responsibly, Dr. Li rejects allowing a handful of corporate executives to unilaterally dictate safety rules, citing misaligned market incentives. She calls for multi-stakeholder governance modeled on biomedical Institutional Review Boards, sector-specific federal regulation (such as the FDA for clinical AI), and direct institutional support for parents and K–12 teachers, whom she identifies as the most overlooked stewards of the coming generation. A central unresolved question remains how society will bridge the gap between technical capability and institutional readiness. While spatial models and robotic companions progress rapidly in research labs, deploying them safely depends on creating robust legal guardrails, supporting overwhelmed educators, and ensuring that emerging tools empower human agency rather than encourage passive cognitive dependence.

斯坦福大学计算机科学家李飞飞（Dr. Fei-Fei Li）与神经生物学家安德鲁·休伯曼（Andrew Huberman）指出，当前人工智能的本质是基于海量互联网多模态数据训练的统计模式识别，其核心价值在于赋能并增强人类的主体性与尊严，而非取代人类。他们强调，AI的未来在于跨越纯语言界面走向物理与空间智能，并呼吁在科研医疗、教育启蒙与制度治理中建立人机协同体系，避免技术被两极化的舆论所扭曲。 视觉演化与数据驱动的技术跃迁 (3:46) 李飞飞回顾，在生物演化史上，约5.4亿年前寒武纪三叶虫感光细胞的诞生催生了物种大爆发，如今人类大脑皮层约一半区域均参与视觉信息处理。现代人工神经网络的逻辑同样源于视觉神经科学，直接继承了大卫·休伯尔（David Hubel）和托斯滕·维泽尔（Torsten Wiesel）发现的哺乳动物视皮层层级特征提取机制。然而，早期的机器学习受限于小样本算法；直到2006年李飞飞从儿童依靠海量日常视觉刺激认知世界的过程中获得启发，主导创建了涵盖逾1500万张标注图像的ImageNet数据集，彻底打破了长期制约AI发展的训练数据瓶颈。 2012年，海量结构化数据（ImageNet）、成熟算法（卷积神经网络AlexNet）与图形处理器（GPU）硬件算力三者交汇，使计算机图像分类错误率大幅下挫，并在2015至2016年间超越了人类基准测试表现（斯坦福研究生卡帕斯测得的人类基准错误率约为4%）。这一技术范式随后外溢至听觉与自然语言处理；2016至2017年Transformer架构的诞生，结合算力与全网文本的扩展，在经历约五年沉淀后促成了2022年底ChatGPT的爆发，推动AI技术迈入跨模态生成的新周期。 统计模式匹配与人类认知的本质鸿沟 (23:36) 尽管诸如OpenAI在2024年初发布的Sora等视频生成模型能够输出逼真的动态画面，但李飞飞与休伯曼澄清，其底层逻辑依然是海量视频序列的统计概率拟合，模型并不理解生物肌肉组织或底层物理定律。这与人类大脑的学习机制存在根本背离：儿童在成长过程中往往只需接触三到十只猫，即可建立强大的对象恒常性并在部分遮挡时迅速推断；而现代模型必须消化数以百万计的图像与视频数据，通过庞大参数空间的激活来输出概率上最合理的结果。 在高级认知与情感体验上，二者划定了明确的界限。AlphaGo在击败李世石时下出的“第37手”（Move 37）展现了机器在预设闭环数学规则下的重组与深度搜索创造力，但它无法复制毕加索式的范式突破，更触及不到未被上传至网络的个体内心体验。人类的情感、怀旧与私人记忆具有独特的第一人称主观性；当聊天机器人输出安慰性语言时，本质上只是“目标函数”（objective functions）约束下的统计对齐，并不具备基于痛苦记忆与深层关怀而产生的真正共情。 具身智能、空间世界模型与人机协作 (58:14) 在实际应用场景中，李飞飞与休伯曼倡导人机协作而非激进的完全自动化。在生物医药领域，AI能够打破学科壁垒，实时综合文献并挖掘复杂的非线性规律（休伯曼举例AI曾比专科医生更快查明其低血压诱发的眩晕病因，并在两小时内通过补液缓解），其变革性堪比150年前电力的普及。而在临床手术中，如李飞飞父亲在斯坦福接受的达芬奇（Da Vinci）机器人微创肝脏手术虽减少了十倍失血量，但由于人体器官高度变异且全球手术样本稀疏，完全自主的算法在缺乏数据密度时远逊于优秀人类外科医生与机器的协同。 为了赋予AI理解现实物理规律的能力，技术必须从平面屏幕迈向空间智能与具身实体。加州大学旧金山分校（UCSF）张复伦（Edward Chang）团队通过脑机接口解码运动皮层电信号，驱动数字化身恢复了闭锁综合征患者的说话语调与微表情。李飞飞于2024年初联合创立的World Labs，则致力于开发能够根据文本、图像或草图生成3D与4D交互环境的空间智能基础模型，为机器人训练、建筑设计与医疗护理构建虚拟演练场；在机器人形态上，她推崇类似《超能陆战队》中大白（Baymax）的柔性设计，以安全承担养老护理与高危作业任务。 捍卫人类主体性、治理护栏与教育变革 (80:37) 针对当前媒体中流行的“极端毁灭论”与科技精英推崇的“盲目乌托邦主义”，李飞飞主张采取务实的人本主义中道。她于2018年离开谷歌并联合创立斯坦福人本AI研究院（HAI），呼吁建立涵盖工程专业伦理规范、类似机构审查委员会（IRB）的研究审查机制以及特定行业（如FDA对医疗AI）的政府监管体系，坚决反对由少数商业科技领袖垄断规则制定权。技术的部署前提必须兼顾社会道德与法律约束，如同汽车制造商具备技术可行性，却绝不可随意编写解除刹车指令的代码。 在教育与社会发展层面，两位学者认为核心在于捍卫个体的“主体性”（Agency）、内在动力与尊严。正如苏格拉底通过精准追问探求真理，结构化提示词（Prompting）本质上是一种批判性思维技能，应尽早纳入基础教育课程。李飞飞强调，过度的被动算法消费会剥夺大脑发育所必需的认知摩擦，但因防范作弊而全面禁止AI工具同样是因噎废食；当前公共叙事最大的失误，是将K-12及高校教师和家长边缘化。硅谷与政策制定者必须停止居高临下的说教，为教育一线的看护者提供充足资源与指引，使AI成为赋能学习者的互动导师。 本期对话确立的核心结论是：AI是人类多模态行为数据的数字化镜像，其算力极限取决于物理世界数据的供给与建模深度，而人类独特的价值在于尚未被数字化的深层情感、直觉与原创思想。当前未决的核心挑战在于，社会各界能否在技术指数级爆发与商业资本垄断的推力下，及时建立兼具伦理审查与法律约束的制度护栏，并在教育中帮助看护者跨越信息鸿沟，确保AI的发展终究服务于扩展人类潜能而非削弱人类自主决断的权力。

[23:36] Dr. Fei-Fei Li: As neuroscientists, I think we depart from human brain because that child who learns about what you say kitty cat would not have the chance to download the internet of images of cat... yet they're able to identify that tail... through a different kind of learning pathway. [33:43] Dr. Fei-Fei Li: Internet is not some random thing. Internet is the biggest collection of human behavior in multimodal forms... That typing is a sensing mechanism that captured everything from teenager chitchat all the way to deep scientific articles. [39:55] Dr. Fei-Fei Li: That thought has never been captured... it's not on the internet, therefore AI has not seen it. So that is where humans still remain so unique. [39:55] Dr. Fei-Fei Li: Move 37 has symbolized AI's creativity... [AlphaGo] made a move that the human masters of Go never thought about. [45:48] Dr. Fei-Fei Li: Agency is so important for humanity. You know that boils down to motivation, agency, and dignity at every individual level. [101:15] Dr. Fei-Fei Li: Humanity should have the agency to decide how we imagine this... our society should be collectively, proactively imagining. One thing I worry about in this AI rhetoric is that the public is put in a position of being reactive.

[0:00] I think the biggest thing humanity never [0:02] learns is the older generation lamenting [0:05] about the future generation as if the [0:08] future generation doesn't know anything. [0:10] They're rude. They're they're they're [0:12] forgetting the past. But if you look at [0:14] arc of history, of humanity, by and [0:17] large, we advance for the better. Now, [0:20] I'm not denying the atrocities. I'm not [0:22] denying the setbacks. I'm not denying [0:25] this. But fundamentally I'm a optimist [0:29] in humanity. I look at kids, they're [0:32] curious. Of course, they get massively [0:34] entertained by this technology, but they [0:36] also are starting to use it. What I [0:40] worry about are teachers and some [0:42] parents because I think our society [0:45] today and especially Silicon Valley are [0:48] not doing them a service. We're [0:50] forgetting about them. [0:52] Hey everyone. To celebrate the launch of [0:54] my new book entitled Protocols, I'm [0:57] pleased to share that I'll be hosting [0:58] three live events very soon. The first [1:01] live event is in New York City at Radio [1:03] City Music Hall on September 17th. The [1:06] second event is in Los Angeles at the [1:08] Dolby Theater on October 8th. And the [1:10] third live event is in San Francisco at [1:12] the Masonic on October 28th. At each of [1:15] these events, I'll be discussing topics [1:16] from the book and my favorite part, [1:18] taking questions directly from you, the [1:21] audience. To get tickets, you can go to [1:23] hubermanlab.com/events [1:25] and use the code protocols to get early [1:27] access. Again, that's [1:28] hubermanlab.com/events [1:31] and use the code protocols to get early [1:33] access to tickets. Welcome to the [1:35] Hubberman Lab podcast where we discuss [1:37] science [music] and science-based tools [1:39] for everyday life. [1:42] [music] [1:44] I'm Andrew Huberman and I'm a professor [1:46] of neurobiology and opthalmology at [1:48] Stanford School of Medicine. My guest [1:50] today is Dr. Fay Lee, a computer [1:53] scientist and professor at Stanford and [1:55] one of the pioneers and luminaries of [1:57] artificial intelligence and computer [1:59] vision. As you all know, millions of [2:01] people use AI chat bots to look up [2:03] information every single day. And of [2:05] course, many people are concerned about [2:07] AI, where it's going, and how it might [2:09] replace certain human jobs or degrade [2:12] our experience of life in one way or [2:14] another. Today we discuss from a [2:15] neuroscience perspective what [2:17] intelligence really is and the ways that [2:19] AI can and is being used for good [2:22] meaning to truly enhance learning health [2:25] and to enrich rather than diminish the [2:27] human experience. We start off by [2:28] talking about how human brains of all [2:30] ages learn new information. What rules [2:33] the brain follows in that process and [2:35] how AI because it is based on the [2:37] content of the internet both resembles [2:39] and falls short of what human brains can [2:42] learn. and we discuss exciting uses of [2:44] AI and robotics in medicine. To be [2:47] clear, FFE acknowledges and addresses [2:49] the many valid concerns about AI. But as [2:52] the director of the Stanford Institute [2:54] for Human- Centered Artificial [2:55] Intelligence, her goal is to make sure [2:57] that humans and humanity at large are [2:59] represented in where AI goes next. As [3:02] you'll soon hear, Dr. Fa Lee is an [3:04] extraordinary scientist and educator. [3:06] She has been called the godmother of AI [3:09] for her ushering in of AI technologies, [3:11] but also for her insistence that the [3:13] ethics and benevolent uses of AI stay [3:16] central to AI and robotics. So whether [3:19] you are young or old, today's [3:21] conversation will inform and empower you [3:23] to understand and use AI in ways that [3:25] truly benefit you and enrich your life. [3:28] Before we begin, I'd like to emphasize [3:29] that this podcast is separate from my [3:31] teaching and research roles at Stanford. [3:33] It is however part of my desire and [3:35] effort to bring zero cost to consumer [3:36] information about science and science [3:38] related tools to the general public. In [3:40] keeping with that theme, today's episode [3:42] does include sponsors. And now for my [3:45] discussion with Dr. Fay Lee. Dr. Fay [3:48] Lee, welcome. [3:49] >> Thank you. I'm excited to be here, [3:51] Andrew. [3:51] >> Yeah, this is a long time coming. And [3:53] yes, [3:53] >> you are a luminary in this AI field, but [3:56] I also consider you a neuroscientist and [3:58] computer scientist, and we share a [4:00] common path through vision science. And [4:02] so I'd like [4:03] >> and fellow colleagues [4:04] >> and fellow colleagues at Stanford. So [4:06] I'd like to start in vision. What is so [4:10] special about vision and seeing and [4:14] light as it pertains to AI and where [4:17] it's all going? Because I think for most [4:19] people those probably sound like very [4:21] divorced themes but actually that's [4:24] where it all starts. [4:25] >> Yeah. I see vision as a cornerstone of [4:28] intelligence in almost two parallel way. [4:32] One is what evolution has taught us. You [4:35] know what's the evolution of vision and [4:37] animal intelligence and human [4:39] intelligence. The other one is computer [4:41] vision and AI what that relationship is. [4:45] So I'll go into each evolution. I always [4:49] say that 540 million years ago animals [4:53] saw the first light. These are simple [4:56] sea ocean animals, trilobytes and and [4:59] the the cousins. And before that there [5:02] was very little sensing. Uh around that [5:05] same time tactile and haptics was [5:08] starting also to emerge in animal bodies [5:11] but but there was no hearing. There's no [5:14] you know smelling there's no but there's [5:17] absolutely no nervous system. But the [5:20] first photoreceptive [5:22] cells created a evolutionary force that [5:27] propelled animals to evolve because [5:31] sensing the external world changes your [5:34] self-perception changes the way your [5:37] relationship with the external world. To [5:39] put it simply, if you seek you can see [5:42] food, it changes your your life, right? [5:45] from a evolution point of view and you [5:48] become someone else's food and also [5:50] you're actively seeking food. You're [5:51] actively seeking mates and and and all [5:55] that. So really because of sensing and [5:59] perception evolution took a incredibly [6:04] accelerated pace in terms of uh animal [6:08] speciation. Fossil studies have told us [6:11] that 10 million years after the first uh [6:15] light for animals [6:18] was what we call the the big ban of [6:20] evolution or Cambrian explosion of [6:23] animal speciation. And fast forward I [6:27] think vision has always played a huge [6:29] role in not only in the early evolution [6:32] of animals but as well as um advanced [6:36] intelligence and how that emerged. You [6:39] and I are both vision student and and [6:42] scientists. It is estimated half of the [6:45] cortical AC activities in human brain is [6:48] involved in visual function. [6:51] Children were first visual before they [6:55] were verbal in development. So vision [6:58] really to this day plays a central role [7:02] in both the evolution of animal [7:06] intelligence as well as in the daily [7:08] life of human human life. Now in [7:12] parallel, vision as a uh as a discipline [7:16] or as a area of uh artificial [7:19] intelligence was really played a pivotal [7:23] role in what we see as this modern AI [7:27] moment in a couple of ways. First of all [7:32] is the the uh algorithms the neuronet [7:36] network algorithms. Neural network [7:39] algorithms were first computer [7:42] scientists start dabbling that in the [7:46] early 1950s. [7:48] And [snorts] Andrew, you might remember [7:51] what's happening on the neuros side in [7:54] the early 1950s [7:57] is that neuroscientists like Hubo and [7:59] Viso were starting to record visual [8:02] cells in malian brain and starting to [8:06] realize there is a hierarchical [8:08] structure of nervous cells that stack [8:12] against each other and pass [8:14] neuroinformation [8:15] across these hierarchy. And it goes from [8:19] you know collecting light from retina [8:22] all the way to recognizing there is a [8:25] shape in front of you. And that very [8:29] neuro architecture that we see in [8:31] mamalian brain is also part of the [8:34] inspiration of neuronet network [8:37] algorithm. Now today's neuronet network [8:40] algorithm runs on hundreds of billions [8:43] and even trillion of parameters. It has [8:45] the complexity that departs from what we [8:49] recorded in the mamalio uh brain or the [8:52] visual pathway but the origin is very [8:56] close to each other about half a century [8:59] ago um a little more than half a century [9:01] ago. That's one aspect of uh vision's [9:05] contribution to AI. There is another [9:08] aspect of vision's contribution to AI [9:11] that is also pivotal which is through [9:14] big data is that that comes closer to my [9:18] own work is that AI around the century [9:23] was a field of machine learning a lot of [9:26] different labs different research [9:28] scientists were were trying out [9:30] different algorithms and it's not just [9:33] neuronet network there are other methods [9:36] jargon words like Beijian methods, [9:39] support vector machine methods. It [9:41] doesn't matter what these methods are, [9:42] but it's a explorative phase that we're [9:46] trying to get these algorithms to work [9:48] so that we can empower the machine to [9:50] read or to see. A group of us computer [9:54] vision scientists [9:56] were struggling with these algorithms [10:00] and uh I was a very young faculty [10:04] um first year faculty 2006 at Princeton [10:09] and my students and I are looking at [10:11] these algorithms and how little data [10:15] were fed into these algorithms to learn. [10:19] So I turned to cognitive neuroscience [10:22] literature per namely vision literature [10:25] and started to study how much humans [10:28] learn, how much humans can see and the [10:31] numbers were incredible. Humans were by [10:34] age six can learn tens of thousands of [10:37] different object categories and the [10:40] exposure to visual world is also [10:43] massive. Right? babies can see the mo [10:46] most of the time the moment they're [10:48] born. So they're inundated with this big [10:51] data. So we conjectured that the lack of [10:55] data was a huge part of the reason [10:58] that's the lack of progress in AI. So we [11:01] took a departure from everybody else who [11:04] are really focusing only on algorithm [11:08] and said that we need data. we need data [11:12] to drive these algorithms. So long story [11:16] short, we led this um image that project [11:20] that collected the first ever internet [11:24] scale large data set for the field of [11:28] artificial intelligence, but really [11:30] through the field of vision because [11:32] imageet is a collection of 15 million [11:36] images. And the goal of imageet was to [11:39] drive machines to recognize everyday [11:42] objects, you know, microphones, cups, [11:45] chairs. And that work converged with the [11:50] advances in neuronet network algorithm [11:53] as well as in GPU computing. And by 2012 [11:59] that work uh that the convergence of the [12:02] three elements of modern AI became the [12:06] defining moment [12:09] of what um what modern AI is. I recall [12:13] somewhere around 2012 it seems there was [12:17] this debate at this vision course at [12:20] Cold Spring Harbor that was held every [12:21] other summer like could a computer learn [12:24] to recognize specific faces as well as [12:26] humans. Now I think most people would [12:28] say computers are actually much better [12:30] at it than humans are even though you [12:33] have these super super recognizer people [12:35] who are exceptional at this. [12:37] >> Could you tell us how is it that this [12:40] technology went from [12:42] a state basically where it would confuse [12:44] you and maybe a a a cousin or or even [12:47] someone that looks somewhat like you [12:49] could [12:50] >> or to the point where uh to the point [12:53] where now it is exquisitely precise. [12:56] >> How do we get here? I want to definitely [12:59] double triple click on the convergence [13:03] of this technology. I think around the [13:06] second decade of 21st century. So like [13:10] you said around 2012 [13:13] the the huge convergence was the [13:16] capability of GPU computing which [13:21] basically accelerated or parallelized [13:24] computing so that you can have more [13:27] flops going through algorithms right you [13:30] need that speed then you also have a um [13:35] after many decades of research neuronet [13:38] network algorithm them is getting more [13:40] mature. Um you know starting as we said [13:44] 1950s people start to um create these [13:48] very simple algorithm that behaves [13:51] similarly to neurons but much simpler. [13:53] Neurons as you know are very complex but [13:56] here the idea is that you have one unit [13:59] of node that takes some some input and [14:03] outputs another input and within it it's [14:05] just a function a very simple function. [14:08] So you stack them together. That's what [14:10] neuronet network is. But by by the time [14:13] it's in the um after you know around 20 [14:18] uh 2010ish the maturity of these [14:22] algorithms have have gotten to a level [14:26] that it's it's becoming really good. But [14:29] also last but not the least the [14:32] recognition of big data. Internet [14:34] definitely fueled that. It made data [14:37] more available. But the reckoning moment [14:40] of wow big data needs to be part of that [14:43] equation. We need to use big data to [14:46] drive these algorithm to learn these [14:48] patterns. So this convergence of these [14:50] three things really set off um the the [14:55] the [14:56] revolution of AI. The specific moment is [14:59] also worth mentioning because you [15:01] mentioned face recognition is this image [15:04] net challenge. My lab put forward that [15:07] starting 2010 after we collected this [15:10] humongous data set, we at that point GPU [15:14] was not yet mature uh and and uh we put [15:19] out a uh public challenge for the [15:22] research community uh for for multiple [15:25] years in a row and invited people to [15:27] solve this major computer vision problem [15:30] called object recognition. The task was [15:33] very easy. We have a data set of a [15:36] thousand different categories of objects [15:40] and this data set is more than a million [15:43] images large. It's what we call the [15:45] testing data set and uh the task for the [15:49] algorithm is I'll show you a picture. [15:52] You have to name the the main objects [15:55] inside and if you guess right you're you [15:58] you get a point. If you guess wrong you [16:00] don't get a point. So that image that [16:03] challenge uh we later a couple of years [16:07] later benchmarked human performance by a [16:10] very smart graduate student at Stanford [16:14] and that was roughly 4%. So random [16:18] chance will be one over a thousand [16:21] >> right? So 4% for humans is not that bad. [16:24] The first few years machines were not as [16:28] good as humans. The turning point was [16:31] 2012 the convergence of neuronet network [16:35] image net data set and GPU even that [16:38] year even though the error rate was was [16:42] cut um to oh by the way the human [16:45] performance error rate was 4%. Sorry I I [16:48] need to correct that the error rate was [16:51] cut down to to the teens. It wasn't [16:55] where human performance was. So this is [16:58] looking at images and and assigning a a [17:00] a [17:01] >> one out of a thousand labels. [17:02] >> Got it. [17:03] >> Yeah. But 2012 was so momentous that [17:06] year because the error rate from [17:09] previous algorithm dropped a lot by this [17:14] neuronet network algorithm. And we know [17:16] in the research community when something [17:19] this drastic happens it it means a [17:22] inflection point. But it still took [17:25] another [17:28] three years I remember by 2012 2016 for [17:32] the algorithm to beat humans in in [17:36] naming a thousand objects. [17:38] >> Could I ask you where this 4% error is [17:41] coming from in this very smart graduate [17:43] student? Is it that they don't recognize [17:45] the objects or it's a recognition [17:47] against time pressure? like they have to [17:49] they're being fed images fast enough [17:51] that occasionally they do an incorrect [17:53] assignment. [17:54] >> I don't think the time pressure was the [17:56] main issue even though for a graduate [17:58] student to do this I don't think they [18:00] want to do this forever. Um but I think [18:03] you know the the human brain as you know [18:07] has limited memory whether it's [18:10] long-term or short-term right so [18:12] retaining the patterns of a thousand [18:15] object classes even if some classes [18:18] you're you're familiar is is not that [18:22] easy [18:23] >> you know so so I think there is the [18:26] confusion and and also for example [18:29] different species of dogs gets really [18:31] close. [18:32] >> Mhm. [18:32] >> And that that's a challenge. [18:34] >> I'd like to take a quick break and [18:36] acknowledge our sponsor, Lingo. Lingo is [18:39] an everyday wearable that tracks your [18:40] glucose 24/7. Glucose drives a lot of [18:44] key processes that support energy, body [18:46] composition, and long-term health. When [18:48] glucose is constantly spiking and [18:50] crashing, that's where we can start to [18:52] see metabolic dysfunction. And over [18:54] time, that can even progress to [18:55] pre-diabetes. Right now, about 115 [18:58] million adults in the US have [19:00] pre-diabetes. Most don't know it, and a [19:02] higher percentage of men have it than [19:04] women do. Often, there aren't clear [19:06] symptoms of pre-diabetes early on, so [19:08] people don't tend to look into it. But [19:10] the fact is that metabolic health is [19:12] shaping how your body functions every [19:14] day, whether you feel it or not. [19:16] Tracking your glucose with Lingo can [19:18] help you see how food, activity, and [19:20] stress impact your glucose throughout [19:22] the day. I personally have used Lingo [19:24] and it's been an invaluable tool for [19:26] improving my metabolic health. If you [19:28] would like to try Lingo, Hubberman Lab [19:30] listeners in the US and UK can save 10% [19:32] on a four-week plan. Just visit [19:34] hellolingo.com/huberman [19:36] for more information. Terms and [19:38] conditions apply. Again, that's [19:40] hellingo.com/huberman. [19:43] Today's episode is also brought to us by [19:45] Wealthfront. In today's financial [19:47] landscape of constant market shifts and [19:48] chaotic news, it's easy to feel [19:50] uncertain about how to save and invest [19:52] your money. Wealthfront is the solution [19:54] that helps you take control of your [19:55] money while managing risk. For nearly a [19:58] decade, I've trusted Wealthfront to [20:00] navigate this volatility. With the [20:01] Wealthfront cash account, I can earn [20:03] 3.3% annual percentage yield or APY on [20:07] my cash from program banks. And I know [20:09] my money is growing until I'm ready to [20:10] spend it or invest it. One of the [20:13] features I love about Wealthfront is [20:14] that I have access to instant no fee [20:17] withdrawals to eligible accounts 24/7. [20:20] That means I can move my money where I [20:22] need it without waiting. And when I'm [20:24] ready to transition from saving to [20:25] investing, Wealthfront lets me [20:27] seamlessly transfer my funds into one of [20:29] their expert-built portfolios. For a [20:31] limited time, Wealthfront is offering [20:32] the Huberman Lab audience an exclusive [20:35] 75% APY boost over the base rate for 3 [20:38] months, meaning you can get up to 4.05% [20:40] 05% variable APY on up to $150,000 in [20:45] deposits. Over 1 million people already [20:47] trust Wealthfront to save more, earn [20:49] more, and build long-term wealth with [20:51] confidence. If you'd like to try [20:52] Wealthfront, you can go to [20:54] wealthfront.com/huberman [20:56] to receive the boost offer and start [20:58] earning 4.05% variable APY today. That's [21:02] wealthfront.com/huberman [21:04] to get started. This is a paid [21:06] testimonial of Wealthfront. Client [21:08] experiences will vary. Wealthfront [21:10] brokerage is not a bank. The base APY is [21:12] as of January 30th, 2026 and subject to [21:16] change. For more information, please see [21:18] the episode description. [21:20] I can see the rationale for doing this [21:22] in the vision domain. But has a similar [21:25] thing been explored with hearing with [21:28] sounds? I mean, it's, you know, as [21:29] humans, we we are amazing at recognizing [21:32] speech inflection, emotional tone, [21:34] things like that. But if I had to [21:36] discriminate, you know, even 15 [21:37] different sound frequencies, I can tell [21:39] you as a non-m musician, um, it would be [21:42] very difficult for me. [21:43] >> Absolutely. I think that what you see is [21:46] the floodgate got open and every sub [21:50] area of AI whether it's speech [21:53] recognition sound recognition uh natural [21:56] language processing which more than [21:59] recognition uh vision all areas got [22:02] really a boost in terms of the [22:04] technology we have colleagues at uh [22:06] Stanford who are studying whale sound [22:09] right uh whale songs using machine [22:12] learning and AI AI now and speech [22:15] recognition is another area that did so [22:18] well in the early days of this AI [22:20] revolution and of course the technology [22:23] continues to um advance by the time the [22:27] transformer paper was uh published [22:30] around 2016 2017 it quickly showed that [22:36] it is even more powerful than the early [22:38] imageet AlexNet algorithm there it was [22:43] not the field of computer vision that [22:45] made the next big uh progress. It's the [22:48] field of natural language processing. So [22:50] because the recipe hasn't changed now we [22:54] have a even more powerful neuronet [22:55] network algorithm called transformer but [22:58] we have even more data on the internet [23:01] from at least more readily available [23:04] data on the internet in the form of [23:06] texts and now we have more powerful [23:09] GPUs. So companies like Open AI and [23:13] Google quickly rallied beyond this this [23:18] very important technology and um it [23:21] still took [23:23] about 5 years from [23:27] 2017 to 2022 to get to the chat GPT [23:32] moment in natural language. But that's [23:34] yet another step forward. So I think for [23:38] people who are not computer scientists [23:40] nor neuroscientists, the um natural [23:44] human uh experience will perhaps [23:47] resonate with them and and maybe I can [23:48] just frame my question through that [23:50] lens. So when a child learns that [23:53] there's something called a kitty cat, [23:54] they go, "Oh, cat." Then they usually [23:56] drop the kitty part. They may say kitty [23:57] and then they learn cat. [23:59] >> And if they have enough interactions [24:00] with a cat, they'll realize what a cat [24:03] is. Even if they see it from the side, [24:05] from the back, and eventually if they [24:07] see a tail that looks a little bit like [24:09] a cat and it's, you know, behind some [24:11] books, you say, "What is that?" They're [24:13] very likely to say cat. Even if they've [24:15] also seen foxes and other animals with [24:18] tails, just based on their experience, [24:20] they're making a probability judgment. [24:21] And that's essentially what uh AI can [24:24] do. That's essentially what machine [24:26] learning can do. Mhm. [24:27] >> But it seems to me that there's a key [24:29] moment that had to happen in the [24:31] progression of, you know, from [24:33] calculators to the AI we have now to be [24:37] able to see an image of a tail and make [24:41] the reasonable assumption that it's most [24:43] likely a cat if it's indoors or [24:44] something like that because foxes [24:46] generally aren't indoors. This sort of [24:47] thing. So, at what point did machine [24:50] learning and AI gain the ability to do [24:53] kind of contextual learning and come up [24:54] with the most likely assignment of what [24:57] something is? Because it's one thing to [24:59] show apples and bananas and oranges, [25:00] they're all fruit. Okay, you could [25:01] distinguish them. You could distinguish [25:03] those from cars and trucks, etc. But [25:05] this object constancy piece [25:07] >> that if something is moving, you're only [25:08] getting a partial image. This isn't what [25:10] most people think of in terms of [25:12] intelligence, but it's part of what [25:14] makes our brains and the brains of other [25:16] animals, but especially our brains so [25:18] remarkable [25:19] >> and why we consider ourselves [25:22] >> probably the smartest species on earth [25:23] and if not the smartest and certainly [25:25] the best at technology development. [25:27] >> Yeah. [25:27] >> So when did AI achieve this and how was [25:30] that scripted into these computers to [25:33] allow them to do that? So let's just [25:35] take the problem very you you have [25:37] described it so well this problem of [25:39] seeing a glimpse of a cat tail and being [25:43] able to recognize cat right or or or a [25:46] sign of high likelihood there is a cat. [25:50] The interesting thing is Andrew [25:52] generations of [25:55] machine learning computer scientists [25:58] have tried this problem. So before today [26:03] that machines can reliably do it there [26:06] were different algorithm you know you [26:08] can imagine a common sense way of [26:11] thinking about this is oh maybe we [26:13] should recognize all the furniture to [26:15] know it's a indoor so it's unlikely to [26:18] be a fox. So though there are rules like [26:21] that that it was built into uh previous [26:24] generations of algorithms, there are [26:27] also rules like well let's only instead [26:31] of guess it's a cat, let's only guess [26:34] one out of the 10 potential animals, you [26:36] know, cat being one of them. That limits [26:38] the the the the search or guess uh space [26:42] and that would help. So many ideas were [26:45] tried. [26:47] So when was the moment it became much [26:50] more reliable is this current era when [26:55] the huge data that these algorithms have [26:58] learned let's take Gemini or GPD uh have [27:02] learned really created the capability in [27:07] the machines uh uh learned space so much [27:12] knowledge so much pattern that when [27:16] presented with this more or less maybe a [27:19] newish [27:21] photo of a cat's tail sticking outside [27:24] of a bookshelf. That pattern activated [27:28] the learned what we call learned weights [27:31] or learned parameters that put put the [27:35] machine's um assessment or or guess of [27:39] this this object closer to what it has [27:42] seen which is likely to be a cattail or [27:46] or just tail because there's just so [27:49] much data. Got it. This is where Andrew [27:52] as neuroscientists I think we depart [27:55] from human brain because that child who [27:59] learns about what you say kitty cat [28:01] would not have the chance to download [28:05] the internet of images of cat. They [28:08] likely have seen three cats, 10 cats at [28:11] most, [28:13] but yet they're able to identify that [28:17] tail as a cattail instead of a fox tail [28:20] through a different kind of learning [28:22] pathway. These are the mysteries we [28:24] haven't fully solved. But I I do want to [28:28] point out that departure between today's [28:31] AI algorithm that is learned with the [28:34] humongous amount of data versus how uh [28:38] humans have evolved. If we continue to [28:42] um ascend the kind of hierarchy from [28:45] simple object recognition to what you [28:48] and I would call higher order brain [28:51] functions like moving more towards what [28:53] most people they hear the word [28:54] intelligence and they just think oh it [28:55] must be some higher order thing [28:56] creativity imagination. Let's go to um a [29:00] middle step and then a and then a much [29:02] further step out. So staying with the [29:04] cat example, if a computer or a child [29:06] learns to recognize a cat through the [29:11] tail, the whole thing, whatever, and [29:12] they've seen a cat move, it's a very new [29:15] world at that point for that brain, that [29:17] child or that computer [29:19] >> because now [29:21] they know that the cat generally moves [29:23] in the direction of its head, not its [29:24] tail. These are simple simple learning [29:26] rules, right? It might go after mice, [29:28] but it might run from dogs. Maybe yes, [29:30] maybe no, and on and on. And so it seems [29:33] that the next layer up in terms of quote [29:36] unquote intelligence is to assign [29:38] likelihoods of direction to move, [29:41] directions not to move, other objects [29:44] that that object is likely to interact [29:46] with. This all sounds very basic to [29:48] people, but like this is how brains [29:49] learn and this is how machines learn. So [29:51] when was the next sort of big inflection [29:53] in terms of like giving a computer AI um [29:58] a picture of a cat and saying um uh [30:02] animate this cat for me, make it move [30:04] like a cat without giving it any [30:06] specific instructions about how to move [30:08] its limbs etc. But I would imagine that [30:11] was a pretty quick but a but a [30:13] remarkably important transformation in [30:16] this whole thing that we call AI because [30:17] that's what a brain does. [30:19] >> Yeah. So it's it's really funny you [30:21] asked this and you put it beautifully. I [30:23] never thought it to put it in this way [30:25] for a uh public audience but that moment [30:29] came when video become part of the [30:31] training data. So see again I'm going [30:34] back to the training data. So around [30:37] 2023 very shortly after uh chatbt moment [30:42] multiple research teams start to put [30:45] video into the training data. Of course, [30:48] I'm not going to get into the nuance [30:49] stuff, the algorithm. There's a little [30:52] bit of uh changes and variations. So, [30:55] remember tw January 2024, Sora [30:59] was released and that's where people see [31:03] a video can be generated literally what [31:06] you just said. People can then type and [31:08] say [31:10] a cat running towards a mouse and then a [31:13] a a few second clip would be generated [31:17] and there would be a cat moving its leg [31:20] in a plausible way running towards the [31:22] mouse. At that time there were still [31:24] mistakes still to even today it's not [31:27] perfect but things gotten have gotten a [31:29] lot better but that opened the floodgate [31:32] of video generation as you described it. [31:36] So what happened there? What happened [31:38] there is actually not as revolutionary [31:42] as you might think because the bottom [31:46] line is it's still data. As a scientist, [31:49] I can tell you there are all kinds of [31:52] algorithm tweaks and changes and [31:54] improvements and and all that. But [31:57] overall, if you zoom out, it's still [31:59] part of this great neuronet network era, [32:03] right? But what happened is that we're [32:05] now able to um process video data in a [32:10] way again some clever engineering [32:13] tokenize it whatever you call it. And [32:16] now we can generate these short clips of [32:20] videos which is frames put together that [32:25] look like plausible cat movement. Now [32:28] you might ask does the algorithm know [32:31] the muscle structure of a cat's legs so [32:36] that when the algorithm shows that the [32:39] cat is mo moving in a plausible way with [32:42] the paws you know in a sequence I would [32:45] say the algorithm doesn't but what it [32:48] does have is so many video especially [32:51] cat on the internet so many videos of [32:54] cat so it learned what it should look [32:57] like so in a way humans do that most of [33:00] us without education would not know the [33:03] how muscles move in cats I still don't [33:07] know you know our colleagues in medical [33:09] school might know but we have just got [33:12] so used to seeing cats moving this ways [33:14] that we have a plausible idea of how [33:18] cats move so that is similar that's how [33:22] similar AI is the it's the statistics [33:26] It's the large amount of data that [33:29] showed you what is the plausible um [33:32] generation of cat movements. [33:34] >> Yeah. So when people have heard almost [33:36] certainly that the brain is a prediction [33:38] machine, it's a learning machine, this [33:40] is exactly what Yes. you're referring [33:42] to. [33:42] >> Yeah. [33:44] >> Let's go to a really far out there [33:46] aspect of brain function that we know [33:48] exists in humans, which is [33:50] >> thoughts [33:52] >> and creativity. M now there are probably [33:57] rules for thoughts and creativity. [34:00] They're a little bit harder to tack down [34:02] than um examples from the visual system. [34:05] Like if it's a tail and it's indoors is [34:06] likely a cat. This kind of thing, but [34:08] they're there. The rules are there. [34:10] >> If you use apple as an example, [34:12] >> we could have gone from lowle seeing an [34:16] apple [34:16] >> to midle seeing apple always drop, not [34:20] fly off. at highest level what is the [34:24] equation that governs the Apple's [34:26] movement [34:27] >> right so that's ascending to like a [34:29] higher order more reductionist analysis [34:32] >> what do you think about the idea that [34:34] while AI is indeed intelligent it can do [34:37] things that brains can do maybe even [34:39] >> well certainly things that individual [34:41] human brains can't do we know this by [34:43] virtue of beating humans at chess and [34:45] this sort of thing [34:47] the idea right now as I understand it is [34:49] that AI is trained on the internet, [34:52] >> images, discussions, videos, songs, but [34:57] that's not all of human cognition, [35:00] right? So, are there aspects of AI that [35:03] are whether or not it's chat or it's [35:05] claude or even the most powerful not yet [35:08] released machine learning and and AI [35:11] tools that don't have access to features [35:14] of human brain function yet because [35:18] they've never been uploaded to the [35:20] internet, at least not in a way that the [35:22] AI can pull out. So, for instance, you [35:25] know, you could put a symphony there and [35:27] it follows certain rules of music and [35:29] mathematics and sound like that that [35:30] makes sense, but you have thoughts all [35:33] day long and I have thoughts all day [35:34] long that don't quite mesh with language [35:37] in a way that I can just type them out [35:39] on the internet. Stay with me here. I [35:41] know this is a long question, but I feel [35:42] like this is the one thing you are [35:44] perfectly poised to answer, and I've [35:46] been waiting to ask you this for a year [35:48] and a half since I saw you in Utah. In [35:50] the world of art, we have this thing [35:52] called abstraction, right? And [35:54] occasionally somebody will come up with [35:55] a painting or a drawing that it doesn't [35:57] look like anything specific. This [35:59] happens in music, too, where you just [36:01] feel something like there's like a [36:02] fundamental rule or an emotion [36:04] associated with it. Like they've tapped [36:06] into some aspect of brain function, but [36:08] you can't say what it is. [36:11] I feel like this is the sort of thing [36:13] that is complicated for AI or for me to [36:15] understand how AI could do because you [36:17] can put that piece of art into AI and [36:20] say, you know, what fundamental feature [36:22] of human uh experience does this reveal [36:25] and it only has access to what's on the [36:27] internet. So, h how can you capture a a [36:30] complex constellation of feelings and [36:33] experience with AI? That seems to be the [36:35] gap for me. And I'm sure we'll get there [36:37] with AI, but I'm not seeing from [36:41] neuroscience to AI in any kind of direct [36:43] way. The same way we could ratchet [36:44] through visual motion, sadness, [36:47] happiness. You could pull out a lot of [36:48] things, but it's hard to get to these [36:50] higher order abstract representations [36:52] that can't be spoken or written down or [36:54] drawn. If I just say, give me your [36:56] example of whatever nostalgia for your [37:00] childhood home. You could write about [37:02] it, but those are just words. It's not I [37:06] can't understand your experience at a [37:08] first person level. [37:09] >> Totally. Andrew, I I know you put a lot [37:12] of thoughts into this question and I [37:15] think it's a very important question and [37:18] let's let's peel this one step at a [37:21] time. First of all, TLDDR short answer [37:24] is I agree with you that we do have to [37:28] be very careful recognizing what AI can [37:31] do, is likely to do, not conjecturing [37:35] over a 100 years or or whatever. I [37:38] recognize what you just said are these [37:41] extremely nuanced personalized [37:45] hard to characterize or not even [37:48] captured human cognitive behaviors and [37:53] because they were not captured they then [37:56] they were not uploaded on the internet [37:59] and we don't have today's AI doesn't [38:02] have a way to do that. So when you call [38:06] internet [38:08] which is the source of AI's data, let's [38:11] be very clear what is internet. [38:14] Internet is not some random thing. [38:17] Internet is the biggest collection of [38:20] human behavior [38:22] in multimodal forms. Let's break it down [38:26] further. Internet has the world's [38:29] population typing on it for many many at [38:33] this point multiple decades. That typing [38:36] is a sensing mechanism that captured [38:40] everything from teenager chit chitchat [38:43] all the way to deep scientific articles [38:46] who dig got digitized and get uploaded. [38:50] Right? So that capturing human language [38:53] is what internet is super good at. [38:57] Then internet captures [39:00] images. How? Because we now have digital [39:04] cameras. That's so prevalent in [39:06] smartphones and digital cameras. So that [39:09] humans love taking photos from, you [39:13] know, the cat in your house to selfies [39:16] to beautiful, you know, BBC captured [39:19] photos. Those also got uploaded in our [39:23] digital sphere. On top of that, there's [39:26] videos. Videos now has sound, has [39:30] movements that also got uploaded to our [39:34] digital sphere. On top of that, there's [39:37] music. We're not even getting into the [39:39] legal discussion of copyrights, but [39:41] let's just table that aside. I'm just [39:44] talking about the forms of data. The [39:46] speeches and and singing and music and [39:49] orchestra that also got uploaded into [39:52] the digital sphere. So now we have [39:55] created this humongous library of human [40:00] knowledge in words, human behavior in [40:05] videos, human expressions or even [40:08] nature's whatever in in sound and now AI [40:12] gets trained on that. That is why it's [40:15] so powerful. This is why especially in [40:18] the words front that AI can recognize [40:22] patterns can can synthesize patterns [40:25] because so much of this is already [40:28] there. But the thing that you just [40:31] talked about that when let's say Picasso [40:35] had that incredibly profound thought [40:39] about that particular way of expressing [40:43] that that portrait of the of the young [40:46] woman that thought has never been [40:50] captured. [40:51] In fact, as neuroscientists, if I ask [40:54] you which brain area did that thought [40:56] come from, you don't know, right? Is it [40:59] Broa? Is it V1? Is it motor? Is it [41:04] preffrontal? We don't know. Maybe it's [41:06] diffused everywhere because that thought [41:10] is so personalized, so special. You can [41:12] call it creativity, you can call it [41:14] emotion, you can call it whatever you [41:16] want. You can call it cat 231 whatever [41:19] name you can give it that thought is not [41:23] captured therefore it's not on the [41:26] internet therefore AI has not seen it [41:30] so that is where humans still remain so [41:34] unique but we also need to give credit [41:37] to AI because AI has learned so many [41:40] things it can combine information in [41:44] highly creative way did you Remember [41:47] move 37? [41:48] >> This is Alph Go. Right. [41:50] >> Right. Move 37 has symbolized AI's [41:53] creativity. I think it's both true but [41:57] can be taken out of context because that [42:00] was a game when Alph Go was plain Lisa [42:04] doll and in I think it's a third game [42:07] out of the five games that Alph Go as a [42:11] computer algorithm made a move that the [42:14] human masters of Go never thought about [42:19] and that is an incredible move right [42:22] because it really humans collect [42:24] collectively these are the masters never [42:26] thought about it. But if you really go [42:29] deep into what AI did there [snorts] it [42:33] was because first of all go is a highly [42:36] mathematical game. It it has very clear [42:39] mathematical objective very clear [42:41] mathematical rules in terms of move. So [42:44] when AI having a bigger compute and um [42:49] ways to retain how many moves it can uh [42:52] it can remember it was able to do things [42:55] that human brains don't typically do. So [42:59] is that called creativity? I think it is [43:02] but we do have to recognize that's a [43:06] special kind of creativity. I was [43:08] talking to a incredible mathematician of [43:11] our time and I was asking him about the [43:15] unsolved problem of mathematics and how [43:17] AI can contribute to that and he was [43:20] very positive. He said there are many [43:22] problems in today's mathematics. [43:25] As hard as they are, even as say a field [43:30] mentalist, I probably have forgotten [43:34] there are known methods in math that can [43:38] solve these problem because I have a [43:40] human brain. I don't remember I I don't [43:43] know all of math's, you know, solutions [43:46] in the past hundreds of years. even if I [43:49] were a a field medalist. So AI can help [43:53] us to solve these problems. But as a [43:56] mathematician, he was also telling me he [43:59] said, I don't know if AI can solve all [44:02] of math problems because some of these [44:05] math problems require solutions that [44:07] have not been invented [44:10] that will push creativity to a whole [44:12] different level. And this is where you [44:15] you know I'm we should be curious is it [44:18] going to be a human creativity or AI [44:22] would go through its iterations of uh of [44:25] improvement and get to a point of [44:27] creativity that humans don't have or is [44:29] it a combined creativity. My current [44:32] conjecture is hybrid is that humans [44:36] working alongside AI would help us to [44:39] solve these problems whose solutions [44:41] have yet to be invented. And then what [44:44] you said especially you touched on [44:47] emotion is even more personalized. This [44:50] is not necessarily logic. This is not [44:52] necessarily deduct deductive reasoning. [44:55] This is maybe Andrew you look at this [44:58] cup and say it's a great cup. What if it [45:01] evoked an emotion in me, a childhood [45:06] moment that a gray cup might mean [45:09] something that only me and my best [45:11] friend share? That is a completely [45:15] inaccessible [45:17] piece of information in my brain that is [45:20] never uploaded on the internet and no [45:23] matter how mighty AI is today cannot [45:28] access that. So that my reaction to this [45:31] cup and potentially what I would do with [45:34] it because of that piece of memory can [45:37] be completely different. You can call it [45:40] creativity. You can call it expression. [45:42] You can call it storytelling. You can [45:44] call it in many ways. But that's where [45:47] it AI cannot access. [45:49] >> I feel like at some point in the not too [45:51] distant future uh computers will have [45:54] access to our brain activity in [45:57] non-invasive ways. M [45:58] >> so you know like I might even imagine in [46:02] 5 10 years I'm wearing something on my [46:04] head right now you can't see it [46:05] >> it's a very very fine hairet makes it [46:08] sound like it whatever like some [46:09] electrodes that are just there on the [46:10] outside of my skull not bothering me [46:12] sensing my activity inside the brain [46:15] maybe also sensing my heart rate [46:17] autonomic activity how alert I am and [46:20] comparing that yes to what I'm saying [46:21] and what I'm doing this is all totally [46:23] within reach and it's going to happen [46:25] you and I both know this and it's [46:27] probably already starting to scare [46:28] people, but let's let's let's keep it [46:29] benevolent, right? There's this world [46:31] where a computer that I own and I'm not [46:34] worried about data getting out or [46:35] anything like that. We've can manage [46:37] that problem is sensing all these [46:40] aspects of me and is picking up on the [46:42] fact that yes, what I say might be [46:44] important, but there are aspects of my [46:47] internal state and brain activity that [46:49] I'm not even aware of. [46:50] >> Yeah. and I can decide to collaborate [46:53] with this aspect of me and say, let's [46:56] let's come up with a really interesting [46:59] uh picture that I've never seen before, [47:02] but comes from some experience of mine [47:04] that's important based on whatever like [47:08] and and it could reveal that to me [47:10] because it has access to my [47:11] >> to unconscious features of my brain [47:14] activity. I think this is very likely to [47:16] happen in in the not too distant future. [47:18] And perhaps if people thought about it [47:20] within the bubble of their own [47:21] experience, like this isn't immediately [47:23] going to the internet or it's not going [47:24] to be used against them, you're actually [47:27] learning about yourself, [47:28] >> of course, [47:29] >> and and I feel most people have an [47:30] inherent interest in what's going on for [47:32] them also with other people, thank [47:34] goodness. But [47:35] >> they're I think like amazing. Like I [47:37] would love to know why [47:39] >> I trip up in certain ways and don't have [47:42] the best day or why some days I have the [47:44] best day or where ideas come from in me. [47:47] What states I could, you know, kind of [47:48] elaborate on, but I'm not going to know [47:51] how to do that except okay, one cup of [47:53] coffee good, one and a half a little [47:54] better, two is too much. If I like right [47:57] now, if you think about how primitively [47:58] we go about this, it's kind of crazy. [48:01] It's crazy. And everyone has a different [48:03] method and we all try and get this right [48:05] and then you've aged enough by the time [48:06] you get it right that then you have to [48:07] update it. And like we're probably not [48:09] getting the most out of our biology and [48:11] our brains at all right now. [48:12] >> No, we're not. And this is why I keep [48:15] saying this is why it bothers me when [48:18] people talk about AI. Some people make [48:20] it sound like it's replacing humanity. [48:24] But what we really what you describe is [48:27] about enhancing and augmenting humanity. [48:30] Right. This is where it doesn't even [48:32] have to go as sci-fi as a smart hairet [48:36] uh accessing your brain waves. Just AI [48:39] learning your patterns of writing [48:43] >> can already help you to be you know a [48:47] better communicator, a more effective [48:50] communicator, a more efficient [48:52] communicator and that is an empowering [48:55] capability that we could unleash in [48:58] today's AI. I think one of the most [49:00] important thing Andrew that as a [49:03] neuroscientist [49:04] and also faculty we know is agency is so [49:08] important for humanity. You know that [49:11] boils down to motivation, agency and [49:15] dignity at every individual level. And I [49:18] think we need to recognize that we need [49:21] to think about AI as a tool that helps [49:24] us in our agency. It does it should not [49:28] take away our agency and people who lead [49:32] in today's AI should not try to talk [49:35] like that this this work will take away [49:40] agency from people. [49:42] >> Yeah. I think people who are very [49:43] familiar with the technology whether [49:45] it's computers or it's biology or any [49:48] technology cars for that matter [49:52] we they become such nerds of that thing [49:55] that we forget that [49:57] >> it can be scary to people [49:58] >> and that the languaging around it is [50:00] essential it is. [50:01] >> And I remember a time in the early 90s [50:03] I'm sure you remember this too when [50:05] genetic testing was viewed as this thing [50:07] like would you want to have it? Would [50:09] you want to do a blood test? Because oh [50:10] my goodness, you might see something [50:12] that could really scare you. And that [50:14] discussion is happening now around, you [50:16] know, self-elected MRIs and things like [50:18] that. None of which people have to do. [50:21] >> But I come from the stance like more [50:22] information is better. But I've come to [50:23] understand that not everyone feels that [50:25] way. Some people don't want to know. [50:27] They don't want to know. [50:28] >> Yeah. But they should have the choice. [50:30] In the meantime, we should have enough [50:32] public education and communication to [50:35] let people know the pros and cons, but [50:37] not to deny them the choice and also not [50:40] to take away, you know, um, and and and [50:43] [clears throat] [50:44] say, well, since you don't understand [50:46] this, let me decide for you what's good. [50:49] That is not good, you know, and and the [50:52] rhetoric around AI right now is getting [50:54] really skewed because people who know [50:58] what this is tend to talk down at the [51:01] public. It tend to talk whether the [51:04] motivation is a positive one or negative [51:06] one. it there there's a rhetoric of you [51:10] guys don't know what this is and I will [51:14] tell you allow will make you whether [51:17] happy safe whatever it is and I will [51:19] decide for you these are not healthy and [51:22] not helpful. Yeah, I agree. And I think, [51:25] you know, one of the reasons for [51:26] starting this podcast was to showcase [51:28] the scientists and physicians who really [51:32] have a benevolence about them and they [51:33] have no interest in dumbing things down, [51:35] but they do have an interest in people [51:37] understanding things and many people [51:39] would feel that, you know, health [51:40] information is among the more important [51:42] things to understand. Absolutely. [51:44] >> Well, thankfully you're um you're [51:46] breaking the mold of of the, you know, [51:48] the phenotype you just described. um and [51:50] and there are a few others but you've [51:52] really uh you've been doing this at at [51:55] the highest levels really encouraging [51:56] people to think about the collaboration [51:58] that is AI the the agency that exists [52:01] and whether to use it or not to use it [52:02] and so forth [52:03] >> one of the agency I do think it's [52:06] important for individual humans whether [52:08] you're a student a teacher doctor a [52:10] policy maker is learn about this not [52:13] necessarily learn about how to code I [52:16] don't think that's it's necessary depend [52:18] on your job right So for example, if [52:20] you're artist or if you're a teacher or [52:23] doctor, you don't necessarily need to [52:25] code, but learn about what this l uh [52:27] this technology is, learn about how you [52:30] can use it yourself to empower yourself, [52:34] your [snorts] learning or your work or [52:36] your expression. By learning, [52:40] one feels more in control. By learning, [52:43] you're less scared of trying. And by [52:46] learning, you retain that agency. and [52:49] that dignity because at the end of the [52:51] day, no matter how advanced technology [52:54] is or medicine is, as humans, we want [52:58] that benevolence that helps us to live [53:00] better, keep our dignity, and and make [53:03] our community better. [53:05] >> I'd like to take a quick break and [53:07] acknowledge our sponsor, AG1. I'm [53:10] excited to share that AG1 has just [53:11] launched their newest formulation, AG1 [53:14] Pro. AG1 Pro takes the clinically backed [53:16] AG1 formula, which is a blend of [53:18] vitamins, minerals, probiotics, and [53:20] adaptogens, and adds three important new [53:22] ingredients. Creatine monohydrate, [53:24] calcium HMBB, and zinc carnosine. Each [53:27] serving has 5 grams of creatine [53:29] monohydrate to support muscle strength [53:30] and performance, as well as brain [53:32] health. Calcium HMBB to support muscle [53:34] recovery and reduce muscle breakdown, [53:36] and zinc carnosine to support and [53:38] improve the lining of your gut. All [53:40] three of these ingredients have [53:41] compelling science to support them, and [53:43] therefore, I love seeing them added to [53:44] the existing AG1 formula. As most of you [53:47] know, I've been taking AG1 every day for [53:49] nearly 14 years now. I started taking it [53:52] long before I even knew what a podcast [53:53] was. It's a great product, and it's now [53:55] made even better with the new AG1 Pro [53:57] formula. If you would like to try AG1 [54:00] Pro, you can go to drinkag1.com/huberman [54:03] to get a special offer. AG1 is giving [54:06] away a free bottle of Omega-3 co-enzyme [54:08] Q10 with your first subscription. Again, [54:10] go to drinkag1.com/huberman [54:13] to get a free bottle of omega-3 [54:15] co-enzyme Q10 with your first AG1 [54:17] subscription. Today's episode is also [54:20] brought to us by Element. Element is an [54:22] electrolyte drink that has everything [54:24] you need and nothing you don't. That [54:26] means the electrolytes, sodium, [54:27] magnesium, and potassium, all in the [54:29] correct ratios, but no sugar. Proper [54:32] hydration is critical for brain and body [54:34] function. Even a slight degree of [54:36] dehydration can diminish your cognitive [54:38] and physical performance. It's also [54:40] important that you get adequate [54:41] electrolytes. The electrolytes, sodium, [54:43] magnesium, and potassium are vital for [54:45] the functioning of all cells in your [54:47] body, especially your neurons or your [54:49] nerve cells. Drinking element makes it [54:51] very easy to ensure that you're getting [54:53] adequate hydration and adequate [54:55] electrolytes. My days tend to start [54:57] really fast, meaning I have to jump [54:59] right into work or right into exercise. [55:01] So, to make sure that I'm hydrated and I [55:03] have sufficient electrolytes when I [55:04] first wake up in the morning, I drink 16 [55:06] to 32 ounces of water with an element [55:08] packet dissolved in it. I also drink [55:10] Element dissolved in water during any [55:12] kind of physical exercise that I'm [55:14] doing, especially on hot days when I'm [55:15] sweating a lot and losing water and [55:17] electrolytes. Element has a bunch of [55:19] great tasting flavors. In fact, I love [55:21] them all. I love the watermelon, the [55:22] raspberry, the citrus, and I really love [55:24] the lemonade flavor. So, if you'd like [55:26] to try Element, you can go to [55:28] drinkelement.com/huberman [55:30] to claim a free element sample pack with [55:32] any purchase. Again, that's [55:34] drinkelement.com/huberman [55:36] to claim a free sample pack. The idea [55:38] that technologies can be connectors as [55:40] opposed to separators, I think, has to [55:42] sit at the center of the discussion. [55:43] Yes. And we all know who they are that [55:45] they're they're several of them. But the [55:47] big names in this field, you know, they [55:48] they are also in a developmental process [55:51] where they're learning how to be public [55:52] facing and it happens very fast. Like, [55:55] you know, the the microscope is on them [55:56] and the cameras are on them and and so [55:58] every every subtle dysfunction is [56:01] magnified. So I like to think that they [56:05] will mature quickly enough to realize [56:09] that and I think they are that some are [56:11] that the public needs to hear the [56:15] correct the true message but in a way [56:18] that makes them understand. That's the [56:20] the kind of dirty secret of medicine and [56:23] academia that you break this mold. I [56:25] like to think I break this mold is that [56:27] there's a power in not sharing how [56:30] things work. Yep. [56:32] >> But it doesn't serve anybody well at the [56:34] end of the day. Like you pull back the [56:36] veil and let people in and people feel [56:37] safer. [56:38] >> Yeah. There's a power in in not sharing. [56:41] There's also a power to say just trust [56:44] me I will tell you and the neither as [56:47] educators that is we don't go to our [56:50] lectures and say just trust me you know [56:53] 2 plus 2 equals four. We actually say [56:56] here's how you break it down and learn [56:58] about it so next time you can do it [56:59] yourself. Right. I also think that [57:03] especially you are your podcast is so [57:06] important as part of public [57:08] communication education of knowledge. I [57:11] also think that we need to hear voices [57:13] of different different background, [57:17] right? So because there are plenty of [57:19] scholars, technologists, builders, uh [57:23] thinkers out there who have been dealing [57:26] with AI, using AI, thinking hard about [57:30] how to use AI to empower people, and [57:33] these voices are so important. Well, [57:35] certainly I'll take names of people to [57:37] to host in addition to you, but since uh [57:40] you're here, I'm going to go next to [57:42] something that I think most everybody [57:43] would agree would be a wonderful thing [57:46] if it existed and it's already starting [57:48] to happen, which is the use of AI to [57:50] augment health discovery, treatment of [57:53] disease, and so on. So, using the [57:56] AlphaGo example from before, and people [57:59] surely still remember the cat example, [58:01] those just follow certain rules. Alph Go [58:03] is very complicated set of rules, but if [58:05] you learn them, there's a constrained [58:07] set of rules. [58:08] >> With the cat, it seems unconstrained, [58:11] like infinite possibilities, but it's [58:13] constrained enough that machines and [58:14] humans can learn it really well. [58:17] >> When you start getting into medicine, [58:20] >> there are rules of medicine. There are [58:22] rules of science. You have a question, [58:23] you pose a hypothesis, you test the [58:25] hypothesis, you try and rule out your [58:26] hypo and so on like the the scientific [58:29] method. And in medicine, every field has [58:31] its methods. We observe, we observe [58:34] disease, we observe who recovers, we [58:36] have a case report, we do a randomized [58:38] control trial. So there are rules and [58:39] the internet knows these rules. So LLMs [58:42] can be used to mine health information [58:43] very well because there are constrained [58:45] rules. But I think you and I both know [58:49] because I also consider you a biologist [58:50] that the rules of biology are still [58:53] revealing themselves to us. Which is not [58:55] to say that the dermatologists, [58:56] neurosurgeons, and oncologists don't [58:58] know what they're doing, but they're [59:00] doing what they're doing within a [59:01] constrained set of rules that they [59:03] learned. And even if they continue to [59:04] learn and update them, [59:06] >> it's every month it seems now that a [59:08] discovery comes out that violates the [59:10] rule. Like I learned that action [59:12] potentials are unitary. They always look [59:14] the same. You either fire or not. [59:16] >> But there was a paper not but 12 years [59:18] ago that showed that the shape of an [59:20] action potential can vary quite a lot. [59:22] It was published in Nature. Mhm. [59:24] >> Everyone saw it and then no one wanted [59:26] to deal with it. It's just too much. It [59:28] changes the rule. [59:29] >> Neurons are supposed to be either graded [59:31] or all are one. And the all I mean it's [59:33] in every single textbook. So now if I [59:35] take a bunch of neural activity and I [59:38] give it the rule, oh well you know [59:39] action potentials can be big, they can [59:41] be small in the same neuron. It [59:43] completely confuses everything we [59:44] understand about neuroscience [59:46] >> and it just our understanding of the [59:48] brain just breaks down to zero. Yeah. [59:50] But if you gave AI the rule that it [59:51] could be, you know, a hundred different [59:53] shapes of this signal, well, AI could [59:56] probably do a lot more than even the [59:59] very very best graduate student at dare [60:01] I say Stanford or to be fair MIT or [60:04] Caltech. I don't think it can do it and [60:07] it can do it like in the duration of [60:09] this question, which admittedly is a bit [60:10] long. So, I'd like to get your thoughts [60:12] on how is it that humans in health care, [60:15] the general public and AI can [60:17] collaborate to help solve disease and [60:20] ideally come up with new rules for [60:22] discovery so that we can finally [60:24] understand our biology at a level that [60:26] can really change the course of humanity [60:29] for the better. [60:30] >> Yeah. No, Andrew, this is probably [60:32] perhaps you touch one of the most [60:34] exciting usage of AI, which is [60:37] scientific discovery. And in the case of [60:39] biio medicine, you know, scientific [60:41] discovery directly connects to human [60:44] health and diseases, I think we're we're [60:47] ready for complete re rewriting of how [60:51] scientific discovery can be done because [60:53] for ages, I don't even know how long, it [60:56] relies on smart humans retaining what [61:00] they have learned from other smart [61:02] humans and and and doing things at the [61:04] speed of our own muscles, I guess, you [61:07] know. Most likely of course there's like [61:11] super colliders and and all that but by [61:14] and large the the ways of doing [61:17] scientific discovery [61:19] human brain or scientists brain are the [61:23] only central character in this process. [61:27] Now we have a new tool whose brain that [61:32] can retain humongous amount of [61:35] information can help us synthesize [61:37] knowledge can go across disciplines in [61:40] ways that you and I cannot go. So for [61:43] example we happen to be both in the [61:45] vision neuroscience AI domain. [61:50] I know nothing about you know oactory [61:53] zero like I don't even know how to spell [61:55] most of probably the these words in that [61:58] our colleagues know right so it's so [62:01] hard for our brain but now we have a [62:03] tool that can break open so so I think [62:06] that [62:07] >> we need to change we need to use this [62:10] tool we absolutely I I was just thinking [62:14] 150 or I don't know exactly when years [62:17] ago we electricity [62:20] changed everything in in in our life, [62:23] right? I'm sure that's a moment we were [62:27] thinking about how [62:29] the changes, the opportunities, the [62:32] scary moment. I think we have to come to [62:35] reckon that scientific discovery is one [62:38] of the most exciting opportunity for AI [62:43] and for health, right? How information [62:45] can be synthesized, how information can [62:48] be presented not only to clinicians but [62:51] also to patients and how patients can [62:54] participate in that process from [62:57] diagnosis to treatment is also there is [63:01] just so much we can do now. [snorts] [63:03] >> Yeah. I mean AI I won't say AI is better [63:06] than all doctors but AI was able to [63:09] disambiguate vertigo from low blood [63:12] pressure for me a few months back and [63:16] one of the people who got it wrong is a [63:18] ENT who works on the vestibular system [63:21] >> what information did you provide just [63:23] your subjective [63:24] >> my subjective experience over a day or [63:26] two [63:27] >> okay good [63:27] >> um turns out it was a medication that a [63:29] doctor had prescribed me that I had a [63:31] like a mild but adverse event and it's a [63:34] weird thing to step and feel like the [63:35] whole world's dropping down and then [63:37] kind of spinning and I thought my [63:39] goodness like feels like vertigo but I [63:41] remember dizzy and lightheaded or [63:42] different. So I started like looking [63:44] into that and then and um sure enough it [63:46] was a it was a blood pressure issue. It [63:48] brought brought my blood pressure excuse [63:50] me down too low [63:52] >> and but I consult we know some smart [63:54] doctors um none of these were at [63:56] Stanford. I will say that this is the [63:57] truth. But it was [63:58] >> we should just be intellectually honest. [64:00] >> But it's just remarkable. And when I ran [64:02] it back to them, they were like, "That's [64:03] really incredible." You know, had you [64:04] not been on the phone with me and in my [64:06] clinic, I would have been able to do [64:07] some additional testing to be fair. But [64:10] this was zero cost. It took a morning to [64:13] know if I drank some uh electrolytes at [64:16] what I would have thought would be [64:18] excessive level that by two hours later, [64:21] I would be fine. Now, of course, there's [64:23] the possibility of a placebo effect [64:25] here, but two hours later, I was fine. [64:27] >> And so, it's also very consoling to the [64:29] patient [64:30] >> to have this. And so, it's not to say [64:32] don't go to a doctor, but it it's [64:34] incredible. I mean, this exists now. [64:36] >> Doctor can use this tooling. By the way, [64:38] I have a very interesting example. You [64:40] know that we have to reschedule this uh [64:42] our conversation because my father was [64:45] going through a surgery right at [64:47] Stanford uh with an incredible surgeon. [64:50] But the surgery was done by a robot, the [64:53] Davinci robot system because it was a [64:56] liver surgery and the surgeon, [64:59] incredible surgeon was driving the [65:00] robot. So it was a deep human machine [65:04] collaboration. After the surgery, I [65:07] asked the surgeon, I said, "Do you [65:09] imagine if say you've done a million, [65:12] which is impossible for a surgeon, but [65:14] human surgeon, but let's collect all of [65:17] human surgeons uh for for this liver, [65:20] this type of liver surgery data. Can we [65:24] possibly train a automatic AI to do [65:28] this?" The answer was not clear. So we [65:33] went a little bit down the rabbit hole [65:34] because liver is a very complicated [65:37] organ. It's extremely vascular. It has a [65:40] lot of vessels and everybody's liver is [65:43] very different. So given the reality of [65:47] how many patients undergo liver surgery [65:51] per year, even if you aggregate um the [65:54] world's liver patient um surgeries, you [65:57] might not have enough data to train [66:00] these algorithm. So this speaks of a [66:03] very important fact that um AI learns [66:07] from patterns. When the patterns are not [66:09] abundant, [66:11] then we have to be careful. We have to [66:13] know how to use AI or how not to use AI. [66:16] You know in this case that having a [66:19] human collaborating with the robot is [66:21] way better than a underlearned robot [66:24] doing the surgery by itself. But the [66:27] same issue might be true for surgeons [66:29] because how many surgeries a surgeon can [66:31] get trained on. So these are [66:33] opportunities that humans and AI can [66:36] totally collaborate with and might [66:39] reveal the best result. Right now the [66:42] future remains to be seen. Can we create [66:45] a artificial simulation of a liver that [66:47] we can now train infinite possibility? [66:51] These are all incredibly open scientific [66:54] possibilities that is waiting ahead of [66:57] us. But then there are [67:01] uh situations like your situation where [67:03] the vertigo versus low blood pressure [67:06] probably have been reported so many [67:08] times that in the database there's [67:12] enough of that that AI has learned that. [67:15] So we can then now take advantage of [67:18] that for people who don't have immediate [67:20] access to doctors. [67:22] >> Amazing. Is your father's surgery went [67:24] okay? [67:24] >> It did. It actually [67:27] lost [67:28] >> 10x less blood than a typical surgery [67:33] >> uh thanks to the laparoscopic capability [67:36] of a robot surgery. [67:39] >> I'd like to talk a little bit about some [67:41] features that we think are uniquely [67:43] human that may or may not be. You'll [67:45] tell me. These are genuine questions, [67:46] not loaded questions. And then I'd also [67:49] like to get educated on how AI is [67:53] structured to allow these things to [67:55] happen. For instance, intuition. We all [67:59] like to think of intuition as this like [68:01] mystical very like it certainly is [68:04] powerful, but this thing that like we [68:05] own that no one can take from us that [68:07] can't be mimicked kind of thing. But I [68:09] could also break intuition down to be [68:11] well, it's my experience over time. It's [68:14] a data set [68:16] coupled to some bodily and brain [68:18] sensations and some prediction cues like [68:21] the last time I felt this this happened. [68:23] The last two times I felt that things [68:25] didn't work out that way so I'm going to [68:26] go this ways. I mean that you could [68:28] assign these rules to a computer. But [68:31] there are other aspects of our deeper [68:33] self if I can refer to them that way. [68:35] Like we don't know where intuition is [68:36] mapped in the body could do an imaging [68:38] experiment but you're not going to [68:39] collect all the neurons and hormones and [68:42] everything simultaneously. who don't [68:43] really have like a location or even a [68:45] network to to point to like things like [68:48] creativity, intuition, premonition, the [68:51] idea that you know you really sense [68:53] something is coming on but it hasn't [68:56] happened yet. [68:57] What sorts of rules can AI get that [69:00] could give it these sorts of [69:02] capabilities? And here I'm want to talk [69:04] about it in the context if you will of [69:07] energy. So whatever this thing is, it's [69:09] like mitochondria driving cells more [69:12] around one thing versus another, the [69:14] same way fear or happiness would, right? [69:16] We were just talking about energy. But [69:17] within AI systems, and I'm not a [69:19] computer scientist, within AI systems [69:21] and GPUs, can we actually allocate more [69:25] energetic flow through particular [69:27] learning rules? So we could tell maybe [69:30] someday you know based on everything you [69:32] know about [69:35] my sister who I love you know what is [69:38] your intuition about how our uh brother [69:41] sister relationship will evolve over [69:43] time and what is your sense about what [69:46] would be great for us to do perhaps for [69:48] our birthdays this year that's different [69:50] than before giving and it only has [69:52] access to the internet can it actually [69:54] become sort of mindlike or mindbody like [69:57] and come up with a sort of sense of what [70:00] might actually be worthwhile or does it [70:02] just need more and more prompts like [70:04] it's just going to keep asking me [70:06] questions so I'm actually doing the [70:07] work. [70:07] >> Such a interesting question Andrew. So [70:10] um I do want to separate intuition from [70:14] creativity for the sake of argument here [70:17] and maybe we'll come back to merging. So [70:20] let's talk about this intuition of given [70:24] my sibling love [70:26] what's going to happen right is it [70:28] really intuition so today when you go to [70:31] a AI chatbot you're going to prompt you [70:35] know I'm a Stanford professor and a um a [70:39] um neuroscientist [70:41] um give me this information that is [70:44] already called context I don't know if [70:46] you call it intuition but because you [70:48] gave that piece of information. The AI's [70:52] answer for you is already going to be [70:54] different if I type that I'm a 14 year [70:57] old teenager, you know, loving race [71:00] cars. Even if we ask the same question, [71:04] it'll have customized answer. That is a [71:08] mathematical [71:09] I wouldn't call it energy. I want to be [71:12] that is just a mathematical [71:14] uh fact of how these um these algorithms [71:19] takes these context and tailor the the [71:22] the outputs and it's called context. [71:25] It's not that deep in the in computer [71:28] science. That's one type of intuition [71:32] that is fairly shallow because you [71:35] already are able to use language to [71:38] describe it or you can say I'll upload [71:40] an image that that also is is already [71:44] expressable and then AI gets it. The [71:47] deeper intuition you just said is like [71:50] you don't even know where they come [71:51] from, right? Like is it because I smell [71:54] something? Is it hormones? Is it you [71:56] know the the mixture of mood? Is it my [71:59] breakfast? [72:00] That intuition, [72:03] what would AI do with it? That is what I [72:06] would say is inaccessible. There's no [72:10] sensory apparatus yet that can glean [72:13] that data and feed it to [72:17] not only AI cannot even feed it to, you [72:20] know, for example, sometimes as a couple [72:23] you might have moment that you're just [72:25] rubbing each other in the wrong way. [72:27] >> Never. No, I'm just kidding. Yeah, of [72:29] course. [72:29] >> If you're really familiar with each [72:32] other, you kind kind of can sense it, [72:33] but you can't quite tell. Maybe you just [72:36] leave quietly, leave that person alone. [72:38] So that means whatever that intuition [72:41] that person has, they could not even [72:44] express it in words or or a gesture to [72:48] give it to another person to use as a [72:52] piece of information. So when you cannot [72:55] even access that [72:58] neither a human a different human nor a [73:02] machine [73:04] can can do anything about it because [73:06] there's no access to that highly [73:09] individualized intuition. [73:12] There's no technology that can do that [73:14] till you say we put brainwave collectors [73:17] or you know skin conductance [73:21] sensors. I mean by the time we do those [73:24] maybe they become accessible. So we have [73:26] to recognize. So so what I'm trying to [73:29] say here is it's not what's not very [73:32] deep is is the data accessible [73:35] you know either through language or [73:37] through picture or through imaging or [73:40] through brain waves whatever it is it [73:44] needs to be an accessible piece of [73:46] information. If it's accessible [73:50] then [73:52] if we have collected enough of that you [73:54] can train machines with or if a machine [73:57] is well trained it can like you said in [74:00] a private way forget about privacy uh uh [74:04] uh breach but in a private way the [74:07] machine can probably you use it. What [74:09] I'm trying to do, Andrew, here is not to [74:12] make it sound mystical, [74:14] >> but try to give it a scientific [74:18] process to describe if it were to [74:22] happen, how would that happen? [74:23] >> Yeah. Because um pattern recognition [74:26] based on big data sets and rules get us [74:29] a long way is what I'm hearing. And we [74:32] earlier we were talking about where [74:33] doctors fail and robots and machines [74:36] perhaps do better or they collaborate to [74:39] do better than either one alone. You [74:41] know I as a neuroscientist you spend a [74:44] lot of time looking at cells at some [74:47] point in your career. And it's amazing [74:48] how like the electrophysiologists for [74:51] decades if not longer you develop an [74:54] intuition. I'm not really a [74:55] physiologist, but I learned to recognize [74:57] cells based on like kind of these things [74:58] that were not written up in any papers. [75:00] But like if there was kind of a like a [75:02] like a straighter edge along this thing [75:04] and it had a certain shape and [75:05] roundness, like I tell you right now, [75:07] that's a transient offpha cell in the [75:09] retina. Eventually, we we developed [75:11] genetic labels to reveal that that was [75:13] true in every case. But then you also [75:15] saw some that didn't fit the rule. [75:17] Machines can learn that, computers can [75:19] learn that. And with all that [75:20] information from all those papers, now [75:22] we have a pretty good parts list of the [75:24] retina. [75:25] >> Cool. That works. And then you can apply [75:27] rules like they fire this way, they fire [75:28] that way. Okay, I'm good with all of [75:30] that. What I think I was trying to get [75:32] to with intuition, and I probably didn't [75:34] give the best example, is like what are [75:37] some internal states of humans that are [75:39] really hard to imagine machines could [75:42] recapitulate, but perhaps they can like [75:46] motivation. Do machines, do robots get [75:49] motivated? We have rules of motivation. [75:51] Like when I'm really motivated to do [75:53] something, we call that urgency, a state [75:54] of urgency. And I might move faster to [75:56] do it. Less activation energy. You say, [75:58] "Let's go." I stand up a little bit [75:59] faster. Machines could like go quicker [76:01] in a certain direction. But can you say, [76:04] "Hey, I want you to seek this out, but [76:07] with a heightened level of urgency, or [76:10] are they just constrained by the [76:11] mathematical rules they can work with?" [76:13] >> So you could build this in the [76:14] mathematics. So certain things whether [76:18] you call it motivation or in machine [76:21] learning world we call them objective [76:23] functions you can build certain things [76:26] into math for example now you go to say [76:29] JBT it has different mode like think [76:32] deeper mode or or like give me a quick [76:35] answer mode if you don't know how this [76:38] works you're like oh this is interesting [76:39] one has more urgency that gives me a [76:42] quicker answer the other one has to go [76:45] deeper into the search, right? And and [76:48] take longer to give me the answer. So, [76:51] as a human, if you anthropomorph [76:54] anthrop morph morphalize it too much, [76:59] you might call it urg urgency or [77:02] motivation. But the truth is this is [77:04] just a different kind of um objective [77:06] for the uh algorithm. You can say, well, [77:09] the the one that think quicker has a [77:12] time limit or token limit. the one that [77:14] thinks slower can activate a different [77:17] part of the model that would take [77:19] longer. So it become actually [77:22] mathematically very dry and not that [77:24] deep. But for a human you can call that [77:27] motivation or urgency. But let's go [77:30] deeper because you're asking something [77:32] deeper than than that, right? Is that [77:36] there are cognitive states that humans [77:40] you truly just whether it's motivation [77:43] or urgency or fear or love that is very [77:48] hard to access and express. And do [77:52] machines have it today? No. Let's make [77:55] it very clear. we tend to imagine that [77:58] the machines feel or or they're not they [78:01] don't have that data they don't have [78:04] that mathematical objective function so [78:07] they can say when the machine says I'm [78:10] sorry you're so sick today it's very [78:14] different from how your friend says it [78:16] to you because the machine said that [78:18] because it has learned through pattern [78:20] when someone tells it I'm sick you [78:22] should say I'm sorry you're sick instead [78:24] of I'm so glad you're sick because that [78:27] data exists. Whereas your friend who [78:30] hears that, they genuinely want your [78:32] well-being. They love you. They want [78:35] they don't want to see you suffer. They [78:38] have that empathetic [78:40] feel of, "Oh, wow. If you're in pain, [78:43] I've experienced pain." So that's it's [78:45] not mirror neuron, but it's at least a [78:47] memory of what pain means. The machine [78:50] doesn't have any of that. So we do need [78:53] to make sure we differentiate [78:55] uh that. So a lot of what drives human, [78:59] what ticks human, what triggers human is [79:03] doesn't exist in today's machine. We [79:05] operate fundamentally different from [79:08] today's AI and we have to recognize that [79:11] respect that and this is where public [79:13] communication is so important. We cannot [79:17] confuse the public about this. [79:20] I'd like to take a quick break to [79:21] acknowledge one of our sponsors, David. [79:24] David makes protein bars unlike any [79:26] other. Their newest bar, the Bronze Bar, [79:29] has 20 gram of protein, only 150 [79:31] calories, and zero gram of sugar. I have [79:34] to say, these are the best tasting [79:36] protein bars I've ever had, and I've [79:38] tried a lot of protein bars over the [79:40] years. These New David bars have a [79:42] marshmallow base, and they're covered in [79:44] chocolate coating, and they're [79:46] absolutely incredible. I of course eat [79:48] regular whole foods. I eat meat, [79:49] chicken, fish, eggs, fruits, vegetables, [79:52] etc. But I also make it a point to eat [79:54] one or two David bars per day as a [79:55] snack, which makes it easy to hit my [79:57] protein goal of one gram of protein per [79:59] pound of body weight. And that allows me [80:01] to take in the protein I need without [80:03] consuming excess calories. I love all [80:05] the David Bronze Bar flavors, including [80:08] cookie dough, caramel chocolate, double [80:10] chocolate, peanut butter chocolate. They [80:12] all actually taste like candy bars. [80:13] Again, they're amazing. But again, they [80:15] have no sugar and they have 20 grams of [80:17] protein with just 150 calories. If you'd [80:20] like to try David, you can go to [80:22] davidprotein.com/huberman. [80:24] Right now, David is offering a deal [80:26] where if you buy four cartons, you get [80:28] the fifth carton for free. You can also [80:29] find David on Amazon or in stores such [80:32] as Target, Walmart, and Kroger. Again, [80:34] to get the fifth carton for free, go to [80:36] davidproin.com/huberman. [80:39] I feel like people assume there's an [80:42] emotion, a person or whatever inside of [80:44] the AI chatbot because we're so language [80:48] oriented. It's talking to us. It's [80:50] writing things to me. And we do that [80:52] more now than we did 30 years ago. Yeah. [80:54] >> Certainly, we've gotten very accustomed [80:55] to receiving communications in fairly [80:58] deprived language. Texts are not like [81:01] extensive pros. Language has changed. [81:04] Modes of communication have changed. [81:06] more deprived as opposed to more [81:09] enriched. [81:10] >> Yeah. [81:10] >> But at some point soon, I'm guessing [81:12] faces are going to start to enter the [81:14] picture. [81:14] >> Uh no pun intended. Um like how far off [81:17] are we from? Like if you or I were to [81:20] text the other person, oh uh see you on [81:23] campus for coffee next week at this [81:24] time. [81:25] >> How soon is it that that text is going [81:27] to be actually a photo or video like [81:30] image of you just talking to me telling [81:32] me that? I mean, this would be trivial [81:33] to do nowadays. [81:34] >> The technology is there. Mhm. [81:36] >> But we have to now look zoom out a [81:41] little bit and look think about the [81:43] social parameters, the legal [81:45] implications. I mean, humans are capable [81:48] of doing a lot of things with our tools, [81:50] but we don't do all of them. For [81:52] example, today any car manufacturer can [81:56] say every Friday the brake doesn't work. [82:00] This is a trivial technology. There's a [82:03] clock in the car's computer and it just [82:05] turns off the brake every Friday. But we [82:08] don't do that because it has deeply bad [82:12] implications to our human society. [82:15] That's where rules comes in, laws come [82:17] in, social norm comes in, morality comes [82:20] in and I think this is where we exit the [82:25] pure technical discussion of AI and need [82:28] to enter the social discussion of AI. [82:31] Mhm. Well, let's do that because one [82:33] thing that I know about biologists or [82:35] technologists is they like to go fast [82:39] cuz it it's exciting. It's the next [82:40] edge, right? I remember long ago I had a [82:43] friend he was studying viruses and ways [82:46] of putting uh these weren't infectious [82:48] disease viruses. These were viral [82:50] vectors for getting genes expressed as [82:52] experimental tools in animals. But there [82:54] came the opportunity to actually put the [82:56] rabies virus, a modified rabies virus [82:59] into Drosophila, into fruit flies. [83:02] >> Oh my god. [83:02] >> Now, that's fine and good in my opinion [83:05] if you are absolutely certain, 100% [83:08] certainty that that is a nonfunctional [83:11] version of the rabies virus because you [83:13] can put other cargo in there and do all [83:14] sorts of important experiments on, [83:16] believe it or not, disease and things [83:17] like that. [83:18] >> But if there's just one fruitly that [83:22] somehow an escaper and you get the [83:24] actual rabies virus. [83:26] >> There's the potential it mates with [83:27] another and then they eventually find [83:29] the others. I don't know if this would [83:30] be a dominant or recessive situation, [83:32] but [83:33] >> now you have fruit flies with rabies and [83:36] those things move really fast. So, [83:37] there's a reason why you don't do that [83:39] experiment. [83:40] >> But it was exciting for them to think [83:42] about and then they got denied, right? [83:45] For good reason. I was grateful, right? [83:47] Go to any biology department, you're [83:48] going to see some fruit flies flying [83:49] around. They love vinegar, by the way. [83:51] you know, so they're coming to your [83:52] salad. But the point here is that [83:55] technologists love to go fast. They love [83:58] sensing that next edge of things. So how [84:00] is it that [84:01] >> between government, the general public, [84:04] technologists, and now I'm just leaving [84:07] out biology here and medicine. How is it [84:11] that that conversation can occur in a [84:13] way that's going to satisfy each of [84:15] those groups enough, not hold us back? [84:18] Because we're also supposedly in an AI [84:19] race right now. So that that warrants [84:22] going faster, not slower. How do you [84:24] think about this? [84:26] >> I mean, Andrew, this is why I returned [84:28] from Google um eight years ago back to [84:33] Stanford and started the human center AI [84:35] institute. These are profound societal [84:38] questions we had to face. And back in [84:41] 2018, there was no Chad GPT. But as a AI [84:44] scientist, I knew that this is only [84:47] going to accelerate. This is why I went [84:50] to my colleagues and university [84:52] leadership and say let's put a framework [84:54] but it's not just my framework or [84:56] Stanford's framework. The entire society [85:00] in every way need to wake up to the [85:04] social implication as we have done this [85:06] in h human history whether it was cars [85:10] or airplanes or or or biotech is that [85:15] it's multi-dimensional with [85:17] multistakeholders right there is the [85:21] professional norm for example [85:24] you guys as biologists don't sneak into [85:27] the lab and and try to put rabies into [85:30] drosophilas or fruit flies because [85:32] that's a professional norm and your [85:33] ethical training. There is industry uh [85:37] rules for example IRBs every uh human [85:42] subject experiment today on university [85:44] campuses are subject to the IRB [85:47] regulatory framework so that we can look [85:50] at this and then there are uh laws and [85:54] regulatory laws depending on if it's [85:57] applied to humans versus [86:00] uh crops or or you know so AI has to go [86:04] through the same right we need to have [86:06] our professional norms we need to have [86:08] education computer scientists are not [86:12] educated in ethics and societal studies [86:16] you know they're starting to I mean this [86:18] is why a number of universities [86:20] including Stanford are feverishly [86:23] putting that part of curriculum into our [86:26] education now that those are the norms [86:29] and education but we also should work [86:32] with the government and different kind [86:33] of Governments and society have [86:35] different kind of norms and traditions [86:38] and heritage and look at where the [86:40] regulatory measure should apply AI for [86:44] example crossing biology FDA I think [86:47] that's a very important area to look at [86:49] how AI uh should be used to help but [86:54] also guard rail to harm uh so that we [86:56] can avoid harm. What I would not like to [87:00] see is one person or or a few people [87:05] coming from industry and telling [87:08] everybody what to do. I think that would [87:10] be dangerous because market forces are [87:14] different from uh societal norms and [87:18] culture and heritage are different from [87:21] uh education and ethics and and these [87:24] are multistakeholder [87:27] problems to solve together. [87:30] >> I love that answer and it's something [87:31] that's very very timely right now. Um, [87:34] this aspect of our conversation is [87:36] surely going to expand over time, but [87:38] you bullseyed it. I'd like to get your [87:40] thoughts on how the human brain is being [87:43] shaped on machines and how machines are [87:46] being shaped by our understanding of the [87:48] human brain. So, first question first. [87:51] Many people, parents and kids are [87:53] thinking, oh, like my kid is never going [87:54] to learn anything now. They're just [87:56] going to look everything up on a [87:57] chatbot. But if you look back in the [88:00] history of learning, similar arguments [88:02] were made about calculators um and [88:04] computers and the typewriter and on and [88:07] on. However, it is an interesting [88:09] question that this hardware that we have [88:11] in our heads evolved to process physical [88:14] things in the world, light, sound, it [88:16] smells, etc. And then it got this really [88:18] cool piece up front, the prefrontal [88:21] cortex that can learn learning rules and [88:23] can update those learning rules. So like [88:24] if anything we were gifted with a a [88:27] learning tolearn machine and updating [88:29] learning. So that's how kids can adjust [88:31] and use LLMs. So I as a generation that [88:34] grew up with the personal computer [88:36] showed up. Granted I grew up in Palo [88:38] Alto. It was like here's Pong and [88:40] there's the Apple 2e and like we had and [88:44] I think oh cool like the brain can [88:46] mature around technology collaborate [88:49] with technology in a way that I think my [88:51] life has been greatly enriched by it. [88:53] But I think the smartphone and perhaps [88:55] the camera smartphone combination as [88:57] people like Jonathan hate have pointed [88:59] out [89:01] have created a situation where most [89:03] people like they love these technologies [89:05] for the ease and convenience. [89:07] >> But we're all a little bit more aware [89:09] now or a lot more aware that we're [89:11] giving up something too. Yeah. [89:13] >> And that they're traps that people in [89:15] particular young people can fall down. [89:16] >> Yeah. So what is the very optimistic meh [89:22] and very pessimistic view in your in [89:24] your mind if three if three flavors [89:26] actually exist there of how young brains [89:29] can be enriched are unaffected or can be [89:32] uh harmed by AI as it exists now. Let's [89:36] just kind of stay with what we've got. [89:38] Great question, Andrew. And the answer [89:40] almost fall out of our previous [89:42] conversations because you use the word [89:44] motivation and I was using the word [89:47] agency. The absolute [89:51] bad outcome is that our young [89:54] generation, [89:58] their agency [90:00] and human level motivation of learning [90:04] and living is taken away by tools. [90:08] So doom scrolling, [90:11] passive watching of shorts, all this are [90:16] not helping [90:18] agency, human agency. Learning [90:21] fundamentally respecting the hardware [90:23] you're talking about takes time, takes [90:27] effort, sometimes takes some pain. That [90:29] is just how our brain is. It doesn't [90:31] matter how transistors move, our neurons [90:34] move in certain ways, our chemistry, our [90:37] hormones move in certain way. So for [90:40] young generation, no matter how the [90:43] society will be different, jobs will be [90:45] different, our human body needs to go [90:48] through a deeply developmental phase [90:52] where learning needs to happen. And that [90:55] agency of learning that motivation of [90:58] learning cannot be taken away by anybody [91:01] should not be taken away by humans nor [91:03] should it be taken away by machines. [91:06] That would be my concern which is that [91:09] if AI is not used right the agency and [91:13] motivation is taken away then we are [91:15] left with generations or generations to [91:18] come who have not properly developed the [91:21] brick. The other kind of danger is in [91:25] the name of agency and and uh and [91:28] motivation the tools are denied to our [91:31] students because we're worried you cheat [91:33] or worry you only got your answer from [91:35] Chad GBT. That is very bad as well [91:38] because with the proper agency, proper [91:41] motivation, proper ways of using this [91:44] tool, we can go a lot deeper with AI [91:48] than we have ever learned. I I was just [91:50] thinking about I was a premed student [91:53] for for a while. Man, organic chemistry [91:56] was hard, you know. I remembered trying [91:59] to learn the the molecules, their [92:02] orientations, [92:03] but the TA hours are too short or it [92:07] overlaps with my other class and my [92:09] professors only have certain number of [92:11] office hours. It was just a struggle to [92:14] learn that. Right? If today I were to [92:17] have a AI companion, I would ask so many [92:21] questions about organic chemistry [92:23] because I know what where I'm stuck, [92:26] right? I have the motivation to learn. I [92:28] just need to uh guidance. That would be [92:31] such a powerful tool for me to learn. So [92:34] that we should not deny students from. [92:37] So both things worry me is either [92:40] denying the tool or taking away agency [92:44] and motivation. Of course, the flip side [92:47] is is great is let's find a way to keep [92:51] our children and students motivation and [92:54] agency. Let's find a way to give them [92:56] the access and the right way of using [92:58] these tools. Then this generation, this [93:01] coming generation and many generations [93:03] to come will be way smarter than us [93:06] because they are superpowered. [93:08] >> I love that answer. Um I have great [93:10] faith in neuroplasticity and the younger [93:13] generations too. Yeah, even our own I [93:15] know we're old [laughter] but [93:17] >> not so let's give ourselves some credit [93:19] plasticity does exist throughout the [93:21] lifan [93:21] >> even our own neurop plasticity right [93:24] like I I find AI a great tool for my [93:28] learning [93:28] >> I mean for me it's been a remarkable [93:31] discovery of what it can do [93:33] >> but I I tend to approach it from the [93:35] position of consumer if I know nothing [93:37] about something and from the position of [93:39] creator if I have some [93:42] >> uh knowledge set [93:43] >> inside of whatever it is I'm asking. [93:46] >> Well, I actually have another thing [93:47] because Stefer undergrad taught me [93:49] something last year and I realized [93:52] before Chad GPT sometimes I got lazy. I [93:56] if I have a question I ask the person I [93:59] think is smart next to me. Now I realize [94:02] I should not ask lazy questions because [94:05] it's so much easier to get information [94:09] before you spend somebody else's time to [94:12] ask something that's that's that's too [94:14] lazy. And AI is forcing me not to be too [94:16] lazy. [94:17] >> How essential is the specificity of the [94:20] prompt to getting the best information [94:23] out of AI? [94:23] >> Prompting is very important. [94:25] >> And that's a skill, right? [94:26] >> That is a skill. This is why public [94:28] education is so important. This is why [94:31] education is so important. I would love [94:33] to see our schools K12 [94:36] teaching prompting. I here's a quiz. Who [94:40] is humanity's best prompter? [94:43] >> I'm going to flunk this quiz. [94:45] >> Socrates if he were alive [94:47] >> because that is the method of prompting. [94:51] Right? Think about it. What is Socrates [94:54] method is prompting and seeking truth by [94:58] asking questions. [94:59] And we should go back and teaching kids [95:02] that [95:02] >> and taking a walk while you have those [95:04] discussions. [95:05] >> Yes. [95:05] >> Which actually is a a good transition [95:08] perhaps to this notion of embodied AI. [95:11] You know, it's a world apart to attach a [95:14] face speaking to hearing words. Uh my [95:17] good childhood friend um who I hope [95:20] you'll meet soon because you both would [95:21] benefit from the conversation so much [95:23] and I just want to be a fly on the wall. [95:24] Um Dr. Dr. Eddie Changeng, chair of [95:26] neurosurgery, bioengineer, and he [95:28] studies speech and language. He and [95:30] others have figured out the [95:30] transformation of neural activity to [95:33] control of the larynx and ferings. And [95:35] he's brought people essentially out of [95:38] lockedin syndrome so they can speak. [95:40] >> Wow. [95:40] >> For the first time in 10 years, he has [95:43] this patient who was sadly paralyzed and [95:45] he could speak through a computer. He [95:47] has others, many examples of these in [95:49] fact. But the incredible thing is when [95:51] he started putting an iPad next to this [95:53] person who is uh one woman in particular [95:56] who's wheelchair bound, they had a video [95:58] of her at her wedding. So they knew her [96:00] voice. They knew her emotive patterns. [96:03] They knew a bit about how she moved her [96:05] body as well. And she now speaks through [96:08] an iPad next to her frozen real face. M [96:13] >> but she can interact with the world and [96:15] it can interact with her in a completely [96:17] different level of depth [96:20] >> than if it were just a microphone. The [96:22] sort of Stephen Hawking thing [96:24] >> and it's constantly being updated [96:25] through machine learning. [96:27] >> What and now also paying attention to [96:29] the people she's speaking to and their [96:31] responses. I mean this is [96:33] >> this is embodiment. Yes, [96:34] >> it's on a 2D flat 2D screen [96:36] >> admittedly, but this is like a [96:41] exponential leap over just robot sound [96:44] or even accurate sound alone. [96:46] >> It's not just embodiment of people, it's [96:48] embodiment also embodied AI goes into [96:51] robotics. Right? The next frontier of AI [96:54] as I have been saying is beyond language [96:57] because again humans develop first [97:00] preverbally. [97:02] evolution took, [97:04] you know, 500 million years without [97:07] verbal communication [97:09] and uh and also the world would in the [97:14] right version would be a lot better [97:16] place with robots helping humans. [97:19] >> Could you give me some examples? I love [97:20] this idea, but again, I'm I realize I'm [97:23] probably a little too deep into the [97:25] technology rabbit hole and it's probably [97:27] scaring some people. So, robots, we've [97:29] got self-driving cars. Actually, the [97:30] Whimo always stops for me and my puppy. [97:34] My beautiful little six-month old puppy. [97:37] How could you not stop when he wants to [97:38] cross the street? A lot of people won't [97:40] stop. They'll almost run us over in the [97:42] morning. The Whimo is very respectful. [97:44] >> The Whimo has to learn the rules, right? [97:47] >> Exactly. Exactly. So, there's [97:48] benevolence there that doesn't always [97:50] exist in humans, but um where do you [97:53] think this is going to show up first? [97:55] And what's it going to look like if we [97:57] zoom out 12 months from now? 12 months [98:00] is a little bit too fast for robotics. [98:02] >> Two years. Three years. [98:03] >> I would say if we zoom out 30 years. [98:06] >> 30 years. [98:07] >> Okay. I'm not saying that's the first [98:08] time it robots hit the street. We [98:10] already have robotic uh cars. I'm just [98:13] saying it takes longer for especially a [98:17] hardware also involved technology to [98:20] manifest. But I would say hopefully in [98:24] you and my lifetime, I would love to see [98:28] robots being part of [98:31] our society, helping us. For example, [98:34] I'm a single grown-up child taking care [98:37] of two very advanced aged and very sick [98:41] parents and they happen not to speak [98:44] English either. The amount of work I do [98:48] is incredible, right? So I would love to [98:51] have have help. It doesn't take away [98:55] family's responsibility. It doesn't take [98:57] away love. It doesn't take away the [98:59] necessary communication. [99:01] But the physical labor would really [99:05] certain part I would love to get help. [99:07] We live in the state of California. What [99:10] is the one thing we all experience? [99:13] >> Traffic. [laughter] [99:15] >> High taxes. Certain part of certain part [99:18] of California doesn't have traffic but [99:20] wildfires. [99:22] >> Oh wow. Yes. [99:22] >> Right. Who is fighting these wildfires? [99:25] Putting humans in danger of rescue [99:29] natural disaster is not a great idea. [99:33] Right? So my family and my parents [99:36] happen to have have enough means. But I [99:39] was just thinking a elderly living [99:41] alone. [99:43] How do they go get grocery? How do they [99:45] go get medicine? Now, there might be [99:49] some shipping we are starting to see, [99:53] but what if they want to go, you know, [99:55] um um for a walk or want to go to a [99:58] park? So, there are just so many things [100:00] that uh Oh, by the way, you're in the [100:03] school of medicine. [100:06] We don't have an excess of caretakers. [100:09] We have a shortage of caretakers. Our [100:13] nurses are deeply fatigued and [100:16] overworked. I was literally in the [100:18] hospital with my dad for the past month [100:21] and just watching the amount of work [100:23] nurses do. We know that on a given shift [100:26] nurses walk miles to fetch things, get [100:30] medicine. There's just so many. Can you [100:33] imagine robots helping, right? Like so [100:36] there are just so many ways that our [100:39] society can be structured and can [100:41] benefit from uh help. [100:45] >> Oh I I love these examples that you know [100:47] so many spring to mind based on what you [100:49] described. You know crossing guards. [100:51] Yeah. You imagine with video that [100:53] somebody who's you homebound because of [100:55] age or illness could navigate to the [100:57] store and pick things off the shelf. Um [100:59] it doesn't have to be so disconnected [101:02] that they just program and it comes [101:03] back. that could be an option, too. [101:05] >> Yeah, I think that we have to revise our [101:09] notions of what this picture looks like [101:12] because I think there are a couple [101:13] things about robots and computers that [101:15] scare people. Um, one is that is their [101:17] physical hardness, [101:19] >> right? And so the way we share space [101:21] with them is very different than the way [101:23] we share space with other things. [101:25] >> Of course, I'm not thinking, oh, like [101:26] you you cuddle with a with a robot, [101:27] although some people might think that. [101:29] That's not my mindset. But I am thinking [101:30] like, okay, if I had a robot that could [101:32] fold clothes, vacuum, water the plants, [101:36] and feed my fish. Although I like to [101:37] feed my fish myself. I really enjoy it. [101:40] I love seeing them eat. I love being [101:41] tactile, you know, in in literally in [101:44] touch with them. They'll eat from my [101:45] hand. [101:45] >> Does your puppy like your fish? [101:46] >> Uh, he does. He has his own fish tank. I [101:48] just got him some tropical fish. Yeah. [101:50] Right in front of his little [101:51] >> He's taking care of them. [101:52] >> Well, he looks at them. He's not [101:53] equipped to take care of them yet, I [101:55] don't think. [laughter] Unfortunately, [101:56] there's not enough prefrontal cortex in [101:58] him. So, and he's a he's a kind, but [102:00] he's a bulldog mut. They're not the [102:02] smartest breed. [102:03] >> They only have a few learning rules, but [102:05] they're very kind. [102:06] >> Um, but if you want a dog that can take [102:08] care of a fish tank, you probably need [102:10] like a West Highland Terrier or [102:12] something like that. Different, more [102:13] prefrontal cortex. [102:14] >> Take him. [102:14] >> But the idea here is [102:16] >> if one robot is doing one thing and [102:18] another robot is doing another, it feels [102:20] like a lot of hardware in my life. And I [102:22] think that's kind of how people feel. [102:24] >> But you could imagine a multimorphic [102:27] robot. Do you know Baymax? [102:29] >> I don't. [102:30] >> Disney's robot probably 10 years ago, 15 [102:33] years ago. It's a Google the image. The [102:37] This is the white medicine robot. A [102:39] healthcare robot that is very not [102:42] fluffy. It's very spongy like it's it's [102:45] feels like a big balloon. [102:47] >> Mhm. Yeah. So, you might like that. [102:48] Yeah. More more contours. [102:50] >> Yeah. Yeah. [102:50] >> And um more multitasking from the same [102:53] robot. feels like a a world that I could [102:56] adjust to more quickly than the idea of [102:59] my world filled with robots. [103:00] >> Yes. Again, Andrew, I think as we [103:03] imagine the future and we talk about how [103:06] we imagine the future, I keep coming [103:09] back to the word agency. Humanity should [103:12] have the agency to decide how we imagine [103:15] this. It cannot just be a company or or [103:19] I don't know an investor decide that the [103:22] the world should be filled with metal [103:25] like robots, right? like our society [103:28] should be collectively [103:30] proactively [103:32] imagining and and one thing I worry in [103:35] this AI rhetoric is that the public is [103:38] put in a position of being reactive [103:42] >> when it feels some people are just [103:45] deciding [103:46] >> and the multistakeholders are not [103:48] participating in this [103:52] designing the future together [103:54] >> like with your example of your father's [103:55] surgery [103:56] to cross the the the robot with the [103:59] physician, right? [104:01] >> If we cross a problem where there's a [104:04] vulnerability with a robot that clearly [104:07] makes things better, [104:09] >> the picture changes Yeah. [104:11] >> in the right direction. So, I'm thinking [104:12] of a few examples off the top of my head [104:14] like um I think most people would agree [104:17] that if their kids could walk themselves [104:19] to school and home, it would be great, [104:20] but you worry about safety. But if a [104:22] robot was really a good guardian of your [104:24] kid to the point where they could alert [104:26] the authorities or maybe even protect [104:28] physically protect your child, [104:30] >> that would be awesome. Give them more [104:31] agency in the world. [104:33] >> You think about um some of the darker [104:36] but nonetheless unfortunately real [104:38] predatory behavior online. [104:40] >> Parents can only oversee their kids [104:42] behavior so much. Kids are only aware of [104:44] so much that's happening. But you could [104:46] imagine uh kind of an avatar in there [104:48] with you that's really advocating for [104:50] you that can spot things and keep [104:53] predators at bay. [104:54] >> Here you go. That's a great startup [104:56] idea. [104:56] >> Like that would be cool. But here's [104:57] what's missing I think from the picture [105:00] >> for me. I remember seeing this [105:03] incredible guy. I know people some say [105:05] he was kind of prickly but this [105:06] incredible guy walking around downtown [105:08] PaloAlto when I was a posttock and when [105:10] I was a kid growing up working at the [105:11] Paltoy and Sport World and that was [105:14] Steve Jobs. no shoes, kind of look like [105:16] a hippie. [105:18] Yes, he shouted at people at work, and [105:21] you know, probably HR wouldn't look too [105:23] kindly upon him nowadays, but he [105:25] understood that these things we call [105:27] computers needed to have rounded edges. [105:29] >> Yes, [105:29] >> they needed to fit kind of seamlessly in [105:31] our pocket. They needed to have Bob [105:34] Dylan on the landing page or whatever so [105:37] that it softened the relationship to [105:39] technology. Some people would say, well, [105:40] it went too far. It was a Trojan horse. [105:42] But I don't think so. somebody who [105:44] really understands human nature to allow [105:47] these like what are clearly going to be [105:49] benevolent collaborations between robots [105:51] and humans to happen because as you've [105:55] pointed out and with total respect to [105:57] the technologists that have built AI and [105:59] the scientists that do amazing science, [106:03] there's a hardness to either the way [106:05] they're being presented or what they're [106:08] capable of sharing that is a real [106:10] separator. [106:11] >> Yes. And I'm not a therapist, but if I [106:13] could like wrap my arms around him, I'd [106:14] be like, "Listen, guys, you're the [106:15] smartest people in the room, guys and [106:16] gals. To be fair, you're the smartest [106:19] people in the room." But people don't [106:21] like you because they don't understand [106:23] you and they're maybe you need a [106:26] collaborator to help you share your [106:30] vision in a way that isn't going to [106:31] allow the press because the media is [106:32] guilty of of building this chasm because [106:34] it's like these technologists, they're [106:36] coming for us. I I think that's I think [106:38] that's a total trick of media, too. [106:40] that's just to put money in their [106:41] pocket. Like there's a lot going on [106:43] right now. So who's the Steve Jobs or [106:46] the Stacy Stacy whoever it's I mean [106:48] could be a man could be a woman. Someone [106:49] who really understands human beings. [106:51] >> Many of them there are many of us you [106:54] know I mean Stanford started human [106:56] center AI institute. Well, there's you. [106:59] There's you. [106:59] >> Okay. But there are many. [107:00] >> Yeah. [107:01] >> There are plenty of entrepreneurs who [107:03] are doing incredible uh startups on AI [107:07] for drug discovery, AI for health care, [107:10] AI for aging, AI for mental health. [107:13] These people care about AI, right? There [107:16] are many designers and and product [107:18] managers who are trying to I I do think [107:22] the megaphone is too much focused on [107:25] people pumping their chest and talking [107:28] about tech in a certain particular way. [107:32] So, you know, even this podcast is is [107:35] making a positive difference, I hope, is [107:38] to to put that human angle, the the the [107:41] rounded human angle, human perspective, [107:44] human human future into these [107:47] conversations. I don't feel despair, [107:50] Andrew. I'm an educator. I'm a builder. [107:53] I'm a technologist. I see many people [107:56] around me, including my entire startup. [108:00] They they these brilliant young [108:02] technologists could join any startup or [108:05] company they want but they come to world [108:08] labs because they want to empower people [108:10] right so I see many people but I don't [108:14] think there's enough you're right I [108:16] don't think the com the public discourse [108:19] is um is balanced right now and there is [108:24] too much extreme rhetoric either in [108:28] terms of extreme doomerism and [108:33] lack of safety like it's just freaking [108:35] people out or extreme utopian as if [108:39] technology can do no run and then that's [108:42] disingenuous people would say well okay [108:44] you're the halves of course you say that [108:47] so I think we should come to the middle [108:49] and talk about what this technology is [108:52] how to use it how we can collectively [108:56] have that agency to guide the future [108:59] Yeah. One thing that was pointed out to [109:01] me by one of my podcast colleagues that [109:03] that was should have been obvious but [109:05] wasn't and and clearly this is something [109:07] that you you uh for lack of a better [109:10] word you embody among many other things [109:12] is people don't really want to hear [109:15] stories about machines. But people love [109:18] hearing that some person cured their [109:20] dog's cancer or their child that was [109:23] experiencing [109:26] >> crazy symptoms. They had no clue. The [109:28] doctors had no clue and their fingertips [109:31] AI [109:32] >> solved the problem. [109:34] >> These are the stories that really need [109:35] amplification because I think that they [109:38] we can relate to them and and they're [109:40] beautiful stories. They're incredible [109:41] stories, but they're not getting nearly [109:43] as much attention as the other stuff. [109:45] Yeah. [109:45] >> And that's a it's a challenge. [109:47] >> Yeah. [109:47] >> You know, traditional media doesn't [109:49] really care about the long arc of [109:53] things. they are on a like a 12 to 24 [109:55] hour cycle but other names perhaps of [109:58] like people who are really trying to [109:59] like talk about the benevolent use of AI [110:02] these collaborations that you know we [110:04] should be aware of [110:05] >> Stephan for HI's newsletter our website [110:08] our seminars we uh promote a lot of [110:11] those work [110:12] >> I would love to learn more about your [110:13] startup um because you don't pick [110:15] projects haphazardly so what is the what [110:18] is the project what's the goal [110:20] >> so my startup uh co-founded with um a [110:24] couple of other co-founders is called uh [110:27] World Labs. [110:28] >> We co-ounded it at the beginning of uh [110:31] 2024. [110:32] It really is for for me a kind of my [110:36] life's work. You know, we both come from [110:39] vision and the recognition of uh there's [110:43] more beyond language intelligence is [110:46] what really motivated me to to think [110:48] hard about what's the next chapter of AI [110:52] frontier and uh we recognize that [110:55] unlocking spatial and physical [110:57] intelligence is really the next chapter [111:00] that it's not excluding languages of [111:03] course the language technology is [111:05] incredible is where we can um uh devote [111:09] more time to build um models or build [111:14] eventually products that can help [111:17] unlocking capabilities in spatial [111:19] intelligence like generating 3D 4D [111:24] worlds that are um deeply useful for [111:28] creators for robot training for uh [111:31] architecture design design or to en [111:35] enable those interactive environments. [111:38] Whether you're talking about healthcare [111:40] usage or education usage or robotics [111:43] usage or industry usage, these [111:46] capabilities goes beyond language [111:49] >> per se. And uh so World Labs was founded [111:53] based on that premise. We are still a [111:56] young company. We're very much a um a [111:59] model focused company where we're [112:01] building these this foundation model and [112:04] we're started by a lot of PhDs [112:08] but now we we we we're starting to build [112:11] products and uh so it's a it's still the [112:14] beginning. It's very exciting and as a [112:17] technologist I feel deep in my heart I'm [112:19] a builder [112:21] >> you know it's maybe it's because also [112:23] I'm an immigrant so that that [112:27] rolling your sleeves up and just get in [112:30] with the young generation that's so [112:32] incredibly smart and just build [112:34] something from scratch is just so [112:36] exciting. [112:37] >> I recall a time not but what 15 20 years [112:40] ago when there were cars driving around [112:43] >> Oh yeah. taking images [112:44] >> still still driving around [112:45] >> still driving around taking images but I [112:47] imagine that there and there are [112:49] certainly aerial views as well but you [112:51] could imagine little tiny drones like [112:54] the type that could fly through a neuron [112:56] and just kind of look at everything or [112:58] um so to speak or drones picking up [113:00] information about every nook and cranny [113:02] of the fjords in Norway has that been [113:04] done to sort of map the the [113:06] three-dimensional world [113:07] >> first of all let's not make it sound [113:09] scary that drones are [clears throat] [113:11] getting into people's homes and [113:13] properties. I think that the ability to [113:15] capture imageries of the world is really [113:18] rapidly advanced, right? Like our cell [113:21] phones are incredible sensors. They're [113:24] not drones, but people take a lot of [113:26] photos. And of course, our camera [113:28] technology has improved. What World Labs [113:31] is doing is not just taking real world [113:34] images. It's we allow people to imagine [113:37] what's in their mind's eye. As long as [113:39] you could type a sentence or show a [113:43] picture or a sketch of what you imagine, [113:46] we try to turn that into worlds [113:49] >> and environments. Um why is it useful? [113:53] Because uh entertainment industry would [113:55] use it, design industry would use it, [113:58] robotics industry uh very much would use [114:01] it for training environments and and and [114:04] all that. So the combination of [114:07] capturing what's in the real world as [114:09] well as capturing what's in your [114:11] imagined world is the new frontier. [114:14] >> If you don't mind, I'd like to just take [114:16] a couple of more minutes and talk about [114:18] this uh moving from imagination to [114:21] >> something. Because this is Los Angeles, [114:23] it occurred to me that a lot of people [114:25] write scripts [114:26] >> and then they try and get them their [114:28] movie made. Mhm. [114:29] >> But with AI, in theory, you could take a [114:32] script and give it to AI and it could [114:33] make the movie in theory, right? Going [114:36] from words to pictures to video. Um, and [114:39] you could maybe edit it a little bit [114:40] here and there where it needed help, of [114:42] course. Has that been done? Has a a [114:44] successful movie been made start to [114:47] finish using AI? [114:49] >> So, this is a very nuance topic. This is [114:51] where we also get into people's [114:55] weariness of AI and creativity when if [115:00] not careful it might sound like we're [115:03] taking away from storytellers and [115:06] creators job. Right? So, so let's [115:09] separate this job conversation from the [115:12] technology conversation a little bit [115:14] even though they're entangled. [115:16] technology has advanced enough that [115:18] taking scripts and generating shots, [115:22] video shots is is getting really good. [115:25] We have seen short movies even almost [115:28] feature length films being assembled by [115:31] AI AI tools. We have and and there are [115:36] many companies US companies, Asian [115:39] companies creating technology. [115:41] But what remains deeply human and that [115:46] is important is [115:48] every part of storytelling and story [115:52] creation. [115:53] There are humans behind it with their [115:57] unique, [115:59] emotion, story, [116:02] technique, [116:03] how they see the world, how they move [116:06] the cameras, how they characterize pe [116:10] characters. A lot of that is what [116:14] Hollywood and and novel writers is [116:18] about. So how do we meet the human need [116:23] and human [116:25] desire of storytelling with modern tools [116:29] is actually a a challenge because there [116:32] is a fear very much coming from [116:35] Hollywood that AI is taking over and [116:40] storytellers and actors and [116:42] screenwriters the jobs are being [116:44] impacted and I think it is but how is it [116:48] being impacted, what are we doing about [116:50] it? Who is working in a in a [116:54] constructive way? You know, this is not [116:56] my industry per se, but I would love to [116:59] see much more nuanced work [117:02] >> in in this and also nuanced public [117:04] discussion about that. But I do think [117:08] just like healthcare, we were talking [117:10] about how AI can rapidly change and [117:14] disrupt the old ways of doing [117:17] healthcare. I think AI is absolutely [117:19] changing the way we're doing um [117:22] storytelling. So one story speaking of [117:24] which I have a co-founder whose name is [117:27] Ben and Ben and I met with Ben Affleck. [117:32] So I was joking Ben meeting Ben who is [117:35] also thinking very avanguard about using [117:39] AI tools about film making right. So [117:43] having conversations between [117:46] technologists [117:47] and storytellers or movie makers at this [117:50] moment is critical. [117:52] >> Yeah. I feel like in every example of [117:54] technology, there's some crossover point [117:57] that when somebody who's truly an [117:59] insider embraces a technology [118:03] >> and then [118:04] >> it just kind of takes off like [118:06] >> you know uh Steven Spielberg or [118:08] something like that or these probably [118:10] aren't the best examples but like the [118:11] the Steve Jobs Wnjak crossover kind of a [118:14] designer [118:16] technology curious guy and and a real [118:19] forgive me to the Jobs film but a real [118:21] computer scientist. right? That merge [118:23] these collaborations are really key like [118:25] so you need an insider and an outsider [118:27] to do it right because you have to [118:29] understand both cultures and how to [118:31] include [118:32] >> the industry the the people [118:33] >> so I really hope right cuz world labs [118:36] works with VFX industry as well it's so [118:40] important for me that our customers and [118:42] users feel empowered [118:44] >> it's not that technology should be [118:46] taking their jobs away technology should [118:48] be making their jobs better [118:50] superpowering their creativity And [118:52] that's how I I see this technology and [118:55] that's how I would like to work with the [118:57] users and customers. [118:59] >> It's wild to think that, you know, when [119:01] I was a kid on California Avenue in Palo [119:03] Alto, there was this store, Keeblin [119:04] Shucket, and it was just a photograph [119:06] store and camera store. Yes. [119:08] >> You go in there, you got your film [119:09] developed, and there were all these guys [119:10] behind the counter, and they tell you [119:12] all you could rent a longdistance lens [119:14] and this kind of thing. Um, none of that [119:16] exists anymore. or everything went [119:18] digital, you know, but there are still [119:21] camera stores. So, industries can morph. [119:24] They don't always get obliterated. [119:26] >> Yeah, it morphs. People also get [119:28] reskilled, upskilled. You know, we are [119:31] working with a lot of creators who are [119:33] using AI tools because they see where [119:36] technology is going and they want to [119:38] reskill and upskill themselves. So, I [119:41] think moments of change is moment of [119:44] both opportunity and loss. We [119:46] [clears throat] need to really be [119:48] thoughtful about that. [119:50] >> My last question is about the young [119:53] generation. How do they feel about AI? [119:57] Because there is this [119:58] >> How young are you talking about? [119:59] >> I'm talking about kids between the age [120:00] of uh seven and 20. [120:04] >> Okay, that that's literally my kids. [120:07] >> Yeah. So, I might have asked that [120:08] question for a reason. You know, how do [120:10] they feel about it? Are they excited by [120:12] it? Because there is this phenomenon [120:14] where like computers come along and you [120:17] know your handwriting teacher is getting [120:18] nervous that people aren't just typing. [120:20] Now they're all writing with their [120:21] fingertips and no one's going to know [120:23] how to write and we wrote for there's [120:24] these these stories have been around for [120:26] a long time about how we're just going [120:28] to dissolve into a puddle of our own [120:29] neurons if we don't uh embrace the past [120:33] as much as the future. And I like to [120:36] think some of both is what's important. [120:38] But how do the kids feel? What do they [120:40] think? This is actually my pet project [120:43] as a educator and technologist. [120:46] Everywhere I go, I try to talk to to [120:49] students, parents, and teachers because [120:52] I think that is the most forgotten [120:55] population, [120:56] our policy makers and our technologist [121:00] and our investors. They don't talk about [121:02] teachers, parents, and students. They [121:04] all have opinions and they all have [121:06] kids, but they don't talk about it. I [121:09] always have hope for kids. Maybe because [121:11] I'm an educator because I think the [121:14] biggest thing humanity never learns is [121:17] the older generation lamenting about the [121:19] future generation as if the future [121:22] generation doesn't know anything. [121:24] They're rude. They're they're they're [121:25] forgetting the past. But if you look at [121:28] arc of history of humanity, by and [121:31] large, we advance for the better. Now, [121:34] I'm not denying the atrocities. I'm not [121:36] denying the setbacks. I'm not denying [121:39] this but you know humanity there [121:43] fundamentally I'm a optimist in humanity [121:47] right so so that's where I come from so [121:49] if you're a total pessimist maybe we're [121:51] already on the wrong footing but I look [121:54] at kids they're curious that's why [121:58] they're kids they're curious they of [122:00] course they get massively entertained by [122:02] this technology but they also are [122:04] starting to use it what I worry worry [122:08] about our teachers and some parents [122:10] because I think our society today and [122:13] especially Silicon Valley are not doing [122:16] them a service. We're forgetting about [122:18] them. We are lecturing them. We are [122:22] berating them. We are looking down at [122:25] them. They are the most important people [122:28] in our society. We should be talking to [122:30] them. We should be uplifting them. We [122:32] should be supporting them. We should be [122:34] providing resources to them. K12 teacher [122:37] or K16 teachers, they share the most [122:42] important critical burden of our [122:45] society. I'll tell you a real story. [122:48] November [122:50] 2022, Chad GPT came out. Obviously, I'm [122:52] an insider in terms of technology, but [122:55] the first thing I did was emailing the [122:58] principal of the elementary school my [123:00] kid was in and said, "I would like to [123:03] come and guest lecture for your students [123:05] and teachers." It's not because I'm so [123:08] special. It's because I want them in [123:11] real time to know what's happening. [123:13] Because nobody, nobody in Silicon [123:15] Valley, no investors, multi-billion [123:18] dollar investment firms or multi-million [123:20] dollar, multi-t trillion dollar [123:22] companies. When chap GBT came out, the [123:25] first thing is what about our teachers [123:27] in the neighborhood? Nobody think like [123:29] that. [123:31] But we need to we need to be talking to [123:33] teachers. We need to show teachers. Of [123:35] course, they're going to ask the [123:37] question about what if kids cheat. It's [123:40] okay. They ask those questions. Let's [123:43] just show them. Let's work with them and [123:46] empower them to come up with ways to [123:49] deal with that. They are smart, too. [123:51] They are eager to change. They're just [123:53] forgotten. [123:55] So, I have hope for kids, but in order [123:58] not to have a blind hope. I think we [124:02] should all remember our teachers and [124:04] help our teachers and parents so that we [124:06] can help our kids. [124:08] I absolutely love that answer and I know [124:11] that sentiment is shared by many many [124:13] people listening. Um, God bless the [124:15] teachers and they need help, support and [124:18] information because now they they turn [124:20] on not yours but most podcasts they're [124:23] just scared. [124:25] They're so scared they hear these [124:27] doomerism. [124:29] They hear the dooms say or they say, [124:31] "Oh, don't worry. It's utopian." Neither [124:34] of these messages can help our teachers [124:37] and if they're not helped, our kids are [124:39] not helped. [124:40] >> Couldn't agree more. [124:41] >> Yeah, couldn't agree more. Fifi, thank [124:44] you so much for taking the time out of [124:46] your incredibly busy schedule. I'm so [124:49] glad to hear your father's okay. And [124:51] that is also part of your schedule, [124:53] taking care of your parents, kids, and [124:54] all the rest to come educate us on this [124:57] thing that's not just important, it's a [125:00] major wedge of where we're at and where [125:02] we're headed. And I I share great [125:04] optimism with caution even more so on [125:07] the basis of what you shared today. And [125:08] also thank you for teaching us more [125:11] neuroscience uh as we went along uh [125:13] because these machines are informed by [125:16] the brain and the brain is informed by [125:17] these machines and this is the world [125:19] we're living in and uh I have great [125:22] optimism in no small part thanks to the [125:24] fact that you exist in this world and [125:26] thank you for taking the time to come [125:28] here to share. I I know many people are [125:30] very grateful. So thank you. [125:32] >> Thank you. Andrew and I really [125:33] appreciated this conversation. It's a [125:35] civilizational moment. [125:37] >> Thank you for joining me for today's [125:39] discussion with Dr. Feay Lee. To learn [125:41] more about her work, please see the [125:43] links in the show note caption. If [125:45] you're learning from and or enjoying [125:46] this podcast, please subscribe to our [125:48] YouTube channel. That's a terrific [125:49] zerocost way to support us. In addition, [125:52] please follow the podcast by clicking [125:54] the follow button on both Spotify and [125:55] Apple. And on both Spotify and Apple, [125:57] you can leave us up to a fivestar [125:59] review. and you can now leave us [126:01] comments at both Spotify and Apple. [126:03] Please also check out the sponsors [126:04] mentioned at the beginning and [126:05] throughout today's episode. That's the [126:07] best way to support this podcast. If you [126:10] have questions for me or comments about [126:11] the podcast or guests or topics that [126:13] you'd like me to consider for the [126:14] Huberman Lab podcast, please put those [126:16] in the comment section on YouTube. I do [126:18] read all the comments. For those of you [126:20] that haven't heard, I have a new book [126:21] coming out. It's my very first book. [126:24] It's entitled Protocols: An Operating [126:26] Manual for the Human Body. This is a [126:27] book that I've been working on for more [126:29] than five years and that's based on more [126:30] than 30 years of research and [126:33] experience. And it covers protocols for [126:35] everything from sleep to exercise to [126:38] stress control protocols related to [126:40] focus and motivation. And of course, I [126:42] provide the scientific substantiation [126:45] for the protocols that are included. The [126:47] book is now available by pre-sale at [126:49] protocolsbook.com. [126:51] There you can find links to various [126:52] vendors. You can pick the one that you [126:54] like best. Again, the book is called [126:56] Protocols, an operating manual for the [126:58] human body. And if you're not already [127:00] following me on social media, I am [127:02] Huberman Lab on all social media [127:04] platforms. So that's Instagram, X, [127:06] Threads, Facebook, and LinkedIn. And on [127:08] all those platforms, I discuss science [127:10] and science related tools, some of which [127:12] overlaps with the content of the [127:13] Hubberman Lab podcast, but much of which [127:15] is distinct from the information on the [127:17] Hubberman Lab podcast. Again, it's [127:19] Huberman Lab on all social media [127:21] platforms. And if you haven't already [127:22] subscribed to our neural network [127:24] newsletter, the neural network [127:25] newsletter is a zerorost monthly [127:27] newsletter that includes podcast [127:29] summaries as well as what we call [127:30] protocols in the form of one to [127:32] three-page PDFs that cover everything [127:34] from how to optimize your sleep, how to [127:36] optimize dopamine, deliberate cold [127:37] exposure. We have a foundational fitness [127:40] protocol that covers cardiovascular [127:41] training and resistance training. All of [127:43] that is available completely zero cost. [127:45] You simply go to hubermanlab.com, go to [127:48] the menu tab in the top right corner, [127:49] scroll down to newsletter, and enter [127:51] your email. And I should emphasize that [127:53] we do not share your email with anybody. [127:55] Thank you once again for joining me for [127:57] today's discussion with Dr. Fay Lee. And [128:00] last, but certainly not least, [music] [128:02] thank you for your interest in science.