Siri Co-Founder Gruber Urges Builders to Target Unserved Global Needs

Humanistic AI with Tom Gruber, Co-creator of Siri, on How AI Can Augment Human Intelligence — Ai4

Ai4

Capital markets pour trillions into automating jobs, but humanistic AI can solve massive unaddressed global needs in mental health and education by building on existing foundation models.

AI-assisted summary of Humanistic AI with Tom Gruber, Co-creator of Siri, on How AI Can Augment Human Intelligence — original source. Check the original for full context.

Source published: 2026-08-22

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 22, 2026, this video features a discussion between Tom Gruber, co-founder of Siri, and Thomas Germain, senior technology journalist at The BBC. The page description outlines an overview of modern generative AI and LLMs, highlighting how these systems are built, their capabilities and risks, and their potential to augment human intelligence across fields like education, software engineering, mental health care, and knowledge management.

## TL;DR In this conversation, Tom Gruber advocates for "humanistic AI"—technology designed specifically to serve, augment, and collaborate with humans rather than fully replace human labor. He addresses the massive financial incentives driving pure automation, warning that the rapid expansion of AI creates a critical ethical crossroads for power concentration and democratic integrity. Gruber proposes cognitive augmentation, specialized FDA-style regulatory agencies, and an API-driven economy, drawing parallels to optical lenses and Volvo's safety innovations. He concludes with a call to action for developers to build vertical applications that target unserved global human needs, such as mental health and personalized education.

## Atmosphere & style The dialogue carries a thoughtful, intellectually rigorous, and urgent tone, blending deep technical perspective with high-stakes societal critique. As a veteran AI pioneer, Gruber speaks with measured authority and practical optimism, balancing serious warnings about totalitarian surveillance and market concentration with genuine excitement for human empowerment. The interaction between Germain and Gruber feels like an engaging, high-level fireside chat, offering a grounded yet visionary reflection on the ethical and economic future of artificial intelligence.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) Pure automation focuses on replacing human labor based on benchmark metrics, whereas humanistic AI is designed specifically to serve, augment, and collaborate with humans. - [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) Over $1 trillion in capital is driving AI toward job automation, creating a massive incentive gap compared to the funding directed at humanistic augmentation. - [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) Massive public demand from a critical mass of users can shift market incentives and pressure governments to prioritize human-centric AI ethics. - [02:51](https://www.youtube.com/watch?v=LLigihQXf-8&t=171s) Concentrating advanced AI surveillance and ownership in a few corporate or autocratic hands poses an irreversible threat to citizens and democratic power. - [02:51](https://www.youtube.com/watch?v=LLigihQXf-8&t=171s) Unlike previous tools, AI is an omni-purpose inference technology comparable to electricity, capable of expanding collective human capability across all fields. - [05:20](https://www.youtube.com/watch?v=LLigihQXf-8&t=320s) Humanistic AI serves as 'eyeglasses for cognition,' expanding human intellectual capacity in the same way optical lenses extend human eyesight. - [05:20](https://www.youtube.com/watch?v=LLigihQXf-8&t=320s) Market competition can elevate AI safety and humanistic design into commercial standards, similar to how consumer demand made seatbelts standard in automobiles. - [07:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=448s) AI is accelerating fundamental science, as shown by DeepMind's AlphaFold for protein folding and automated discoveries in material science. - [07:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=448s) AI regulation should be managed by a specialized expert agency modeled on the FDA or FAA, alongside laws protecting information integrity against deepfakes and synthetic fraud. - [10:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=628s) Natural language bypasses broken, complex graphical software interfaces, serving as a universal equalizer that connects non-technical and global populations to advanced intelligence. - [13:47](https://www.youtube.com/watch?v=LLigihQXf-8&t=827s) Personal AI agents will act as intermediaries across device ecosystems, directing complex real-world tasks to backend APIs and cloud-based AI models. - [13:47](https://www.youtube.com/watch?v=LLigihQXf-8&t=827s) Physical service platforms like ridesharing or food delivery will survive AI disruption by serving automated API requests, whereas ad-supported digital journalism faces existential risk. - [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) Developers should build vertical applications addressing massive unserved human needs—such as mental healthcare for one billion people and universal personalized education—rather than competing in over-funded spaces.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the transcript.

[00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) Defining Humanistic AI vs. Pure Automation

The session opens with Thomas Germain asking Tom Gruber, co-creator of Siri, to explain "humanistic AI" and how it fundamentally differs from what most technology companies are building today. Gruber traces this distinction back several decades in AI history, noting that there are two core reasons to pursue artificial intelligence. The first is automation for its own sake—creating autonomous intelligence where success is measured purely by benchmark metrics, such as beating Garry Kasparov in chess or passing the bar exam. The second path is humanistic AI, where the system's entire intent, purpose, and design are explicitly targeted at serving human needs.

Gruber highlights that while this might sound like a simple philosophical distinction, the real-world implications are huge as capital concentrates around the technology. He notes that there is currently at least $1 trillion in financial incentives driving AI toward replacing and automating human labor, whereas the funding directed specifically toward humanistic AI remains far smaller.

- **Two Historical AI Approaches:** Pure automation (benchmarks, beating human champions, labor replacement) versus humanistic AI (designed to serve, augment, and collaborate with humans). - **Incentive Disparity:** At least $1 trillion of capital is currently pushing toward job automation over humanistic augmentation.

[01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) Economic Incentives and the Ethical Crossroads

Germain and Gruber explore whether financial markets can ever support a more helpful, human-centered version of technology when so much capital rewards pure profit. Gruber points out that aligned incentives do exist when goals are universally acknowledged—such as curing cancer—but capital markets don't always naturally prioritize human benefit over the bottom line.

However, Gruber emphasizes that if a critical mass—even a billion people—demands technology that empowers humanity, that demand can shift market incentives and drive regulatory action. Because AI is on the verge of an extraordinarily rapid expansion with an uncertain endpoint, Gruber argues this moment represents humanity's last chance to make a definitive ethical choice to prioritize human benefit before optimizing for corporate balance sheets.

- **Market Pressure:** Public demand from a massive user base can reshape marketplace and government incentives. - **Critical Turning Point:** The immediate window during AI's rapid rollout is the final opportunity to prioritize human-centric ethics over pure capital efficiency.

[02:51](https://www.youtube.com/watch?v=LLigihQXf-8&t=171s) Societal Power Shifts and "Omni-Purpose" Technology

When asked why this turning point must be addressed immediately rather than fixed later, Gruber connects the technological shift to broader geopolitical and societal power dynamics. Citing historian Yuval Noah Harari, Gruber notes that once totalitarians or dictators acquire advanced AI surveillance systems, power shifts permanently away from labor and citizens toward capital and central authorities, making it nearly impossible to reclaim. Even if today's leading AI companies are run by well-meaning individuals, concentrating power over 8 billion people into a few corporate hands creates severe systemic risks.

Despite these warnings, Gruber remains optimistic. Reflecting on his time at Stanford during the dawn of the World Wide Web in 1994, he recalls how the web eliminated spatial and temporal barriers for human collaboration—acting as a massive leap forward in augmenting human intelligence. Gruber views current AI as the next, vastly larger step because it is an "omni-purpose" technology. Similar to electricity, AI can perform inference across nearly any field, creating a massive step-change in collective human capability.

- **Power Concentration:** AI surveillance and centralized ownership shift power from labor to capital and autocrats, making recovery extremely difficult once established. - **Omni-Purpose Tech:** Unlike the web (primarily communication), AI is omni-purpose inference technology, comparable in scope to electricity.

[05:20](https://www.youtube.com/watch?v=LLigihQXf-8&t=320s) Eyeglasses for Cognition and the Competition Model

Moving into practical applications, Gruber explains that humanistic AI operates through collaboration and augmentation, such as collaborative software agents that assist humans in completing tasks. To illustrate cognitive augmentation, he draws a parallel to eyeglasses. Glasses represent a perfected optical engineering solution that corrects vision, while specialized optics like microscopes and telescopes afford humans "superhuman vision."

Gruber suggests that modern AI will act as "eyeglasses for cognition," extending human mental capacity. To prevent monopoly control over these "cognitive lenses," competition and market choices are vital. He compares the situation to the automotive industry: when Detroit automakers initially resisted adding seatbelts in favor of larger muscle cars, Volvo entered the market to demonstrate that safety was a feature consumers would actively pay for. Over time, features like seatbelts and airbags transitioned from premium safety selling points to universal standards. Gruber predicts a similar path for AI safety and humanistic design.

- **Cognitive Augmentation:** AI acts as "eyeglasses for the mind," expanding human intellectual capacity much like optics expand eyesight. - **The Volvo Analogy:** Just as Volvo commercialized car safety features into industry standards, consumer demand will turn AI safety and human-centric design into viable, profitable products.

[07:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=448s) Accelerating Scientific Discovery and Government Guardrails

Gruber points out that humanistic AI is already demonstrating immense value in scientific discovery. He highlights DeepMind’s Nobel Prize-winning AlphaFold system for protein structure prediction, as well as its work in material science, where AI discovered tens of thousands of novel materials. Because materials science laid the foundation for semiconductors, fiber optics, and lithium-ion batteries, using AI to accelerate scientific breakthroughs serves as a prime example of humanistic technology that yields massive economic and societal dividends.

Germain then asks how governments should intervene beyond blocking corporate mergers. Gruber outlines two immediate responsibilities for policymakers:

``` +-----------------------------------------------------------------+ | GOVERNMENT REGULATORY FRAMEWORK | +-----------------------------------------------------------------+ | | | 1. Specialized Expert Agency | | * Modeled on FDA (drugs) & FAA (aviation) | | * Establishes clear safety standards for models | | * Protects open-source innovation while stopping capture | | | | 2. Information Integrity & Truth Protection | | * Enforces laws against impersonation, fraud, & deepfakes | | * Prevents digital noise ("slop") from replacing facts | | * Safeguards democratic discourse | | | +-----------------------------------------------------------------+ ```

- **Scientific Breakthroughs:** DeepMind’s AlphaFold (protein folding) and material science discoveries prove AI can accelerate fundamental science. - **Specialized Regulation:** Regulatory responsibility should move away from individual politicians and into a dedicated expert agency, mirroring aviation or pharmaceutical governance. - **Democratic Preservation:** Protecting truth and preventing synthetic fraud are essential to maintaining a functioning democratic society.

[10:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=628s) Language as the Universal Interface and Siri’s Legacy

Germain shifts the discussion to Gruber’s background as a co-creator of Siri, asking whether voice and conversational language represent the ultimate interface for computing. Gruber clarifies that voice is simply a modality, whereas human language itself is the true, underlying interface. Traditional graphical software interfaces remain overly complex and broken for most non-technical users, who routinely struggle with basic IT troubleshooting. Language bypasses this friction entirely, enabling anyone to simply state their intent to an agent.

Gruber highlights how language interfaces break down global access barriers. By pairing high-speed telecommunications with conversational AI, a villager in rural India who speaks neither English nor Hindi can converse naturally with the world's most advanced AI models in their local dialect. Reflecting on Siri's evolution, Gruber recalls that the original standalone Siri app featured multi-turn conversations for e-commerce tasks like ordering flowers or booking hotel rooms. When Steve Jobs acquired Siri for Apple, he presented the founders with a choice: build a standalone business, or join Apple to build a product everyone uses.

- **Language as Interface:** Natural language eliminates complex user interfaces, enabling non-technical users to control powerful systems effortlessly. - **Global Reach:** Combining global telecommunications with spoken AI allows rural populations to access advanced intelligence in native local languages. - **Steve Jobs’ Pitch:** Jobs convinced the Siri team to join Apple by offering them the chance to create a universal product used by everyone rather than a niche business.

[13:47](https://www.youtube.com/watch?v=LLigihQXf-8&t=827s) Intermediary Agents and the Future of Digital Services

As Apple rolls out deeper system integrations, Gruber foresees conversational assistants fulfilling their original vision: acting as a frictionless intermediary across a user's entire ecosystem of devices (phones, Macs, home speakers). These local agents will communicate with massive cloud-based models to solve complex computational or scientific problems seamlessly in the background.

Addressing concerns raised by journalists like The Verge's Nilay Patel regarding market concentration—where a single assistant acts as the sole layer between users and third-party services—Germain asks if assistants will displace companies like Spotify, Uber, or Grubhub. Gruber argues that real-world service platforms won't be erased because their underlying value lies in physical logistics (delivering food or providing rides), not just their app interface. While search engine business models and digital journalism face severe disruption from direct conversational answers, physical service providers will simply transition to servicing automated requests submitted via AI APIs.

- **Device Intermediary:** The long-term goal is a personal assistant that manages hardware ecosystems and routes complex requests to cloud AI. - **API-Driven Economy:** Services like Grubhub and Uber will survive because their core value lies in real-world physical operations, adapting to receive orders via AI APIs rather than manual app taps. - **Media Disruption:** Unlike physical services, ad-supported digital journalism faces existential disruption as AI shifts search away from web page traffic.

[16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) Building Humanistic AI: Focus on Unserved Human Needs at Scale

As their conversation draws to a close, moderator Thomas Germain transitions from broader market incentives and government regulations—factors outside an individual's direct control—to actionable advice for the audience. Directing his focus to the developers, founders, and engineers present in the room who are actively building AI tools, Germain asks Tom Gruber what specific actions they should take today to bring the vision of humanistic AI to light.

Gruber responds with a direct call to action, urging builders to focus on vertical applications that address massive, unserved human needs rather than competing in already crowded arenas. He explicitly advises against taking on over-resourced domains like "curing cancer," noting that plenty of well-funded teams are already tackling those problems. Instead, he highlights under-addressed global challenges where AI can act as a massive force multiplier:

* **Mental Healthcare:** Target mental healthcare for the **one billion people** worldwide who, according to the World Health Organization, can never be realistically served by human healthcare workers alone. * **Global Education:** Create a personalized AI tutor for every child on the planet to level the playing field of knowledge work globally.

Gruber emphasizes that solving these critical problems does not require "a trillion dollars." Developers can leverage the immensely expensive baseline engineering provided by foundation model companies and immediately build high-impact vertical applications on top of those existing platforms. He concludes with a sense of urgency, warning builders to solve these hard problems now, "cuz in three or four years you may not have the chance."

---

## Notable quotes > [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) At least $1 trillion in financial incentives driving AI toward replacing and automating human labor, whereas the funding directed specifically toward humanistic AI remains far smaller. — Tom Gruber

> [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) This moment represents humanity's last chance to make a definitive ethical choice to prioritize human benefit before optimizing for corporate balance sheets. — Tom Gruber

> [02:51](https://www.youtube.com/watch?v=LLigihQXf-8&t=171s) Once totalitarians or dictators acquire advanced AI surveillance systems, power shifts permanently away from labor and citizens toward capital and central authorities, making it nearly impossible to reclaim. — Tom Gruber

> [05:20](https://www.youtube.com/watch?v=LLigihQXf-8&t=320s) Eyeglasses for cognition — Tom Gruber

> [10:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=628s) Language bypasses this friction entirely, enabling anyone to simply state their intent to an agent. — Tom Gruber

> [13:47](https://www.youtube.com/watch?v=LLigihQXf-8&t=827s) Services like Grubhub and Uber will survive because their core value lies in real-world physical operations, adapting to receive orders via AI APIs rather than manual app taps. — Tom Gruber

> [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) Target mental healthcare for the one billion people worldwide who, according to the World Health Organization, can never be realistically served by human healthcare workers alone. — Tom Gruber

> [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) Cuz in three or four years you may not have the chance. — Tom Gruber

## Mentioned Tom Gruber, Thomas Germain, Siri, Ai4, BBC, Garry Kasparov, Yuval Noah Harari, World Wide Web, Stanford, Volvo, Detroit, DeepMind, AlphaFold, FDA, FAA, Apple, Steve Jobs, The Verge, Nilay Patel, Spotify, Uber, Grubhub, World Health Organization

## Video description Tom Gruber - Co-Founder, Siri Thomas Germain - Senior Technology Journalist, The BBC

Almost overnight, the world has come to see that the new technology of large language models and generative AI could change the nature of knowledge work. Trained on the cultural legacy of humanity, these systems can now do many tasks involving language, images, and information that are the expertise of educated professionals. In this overview of modern AI, which is updated with fresh insights every time it is given, Tom explains how these systems are created, what they are great at, what to be careful with, and the exciting applications that they enable. He brings to bear decades of experience in designing AI systems and advising AI companies. Applications include education, creative services, software engineering, knowledge management, mental health care, and augmented cognition. He offers insights into the potential risks to individuals and society from ungoverned applications, and a positive vision for the future of truly personal AI.

[Source](https://www.youtube.com/watch?v=LLigihQXf-8)

中文简报

Siri联合创始人Gruber敦促开发者聚焦未被满足的全球需求资本市场正向岗位自动化投入数万亿美元，但人文AI可以通过在现有基础模型上进行构建，解决心理健康和教育领域巨大的未被满足的全球需求。在本次对话中，Tom Gruber提倡“人文AI”（humanistic AI）——即专门用于服务、增强并与人类协作，而非完全取代人类劳动的技术。他探讨了推动纯自动化的巨大资金激励，警告称AI的迅速扩张在权力集中和民主完整性方面造成了关键的伦理十字路口。Gruber提出了认知增强、仿照FDA设立的专门监管机构以及API驱动型经济，并将其与光学镜片及沃尔沃（Volvo）的安全创新相提并论。他最后呼吁开发者构建垂直应用，瞄准心理健康和个性化教育等未被满足的全球人类需求。纯自动化侧重于基于基准指标取代人类劳动，而人文AI则专门设计用于服务、增强并与人类协作。 (0:04) 超过1万亿美元的资本正推动AI走向工作岗位自动化，与投向人文增强的资金相比，造成了巨大的激励差距。 (0:04) 达到关键规模的用户群体所产生的巨大公众需求，能够改变市场激励，并促使政府优先考虑以人为本的AI伦理。 (1:21) 将先进的AI监控和所有权集中在少数企业或专制者手中，会对公民和民主权力构成不可逆转的威胁。 (2:51) 与以往的工具不同，AI是一种堪比电力的通用推理技术，能够拓展所有领域的人类集体能力。 (2:51)

Reading summary

Siri co-creator Tom Gruber argues that artificial intelligence must be deliberately steered toward augmenting human capability rather than pursuing pure labor automation, warning that humanity faces a closing window to establish ethical guardrails before market forces and political power permanently consolidate. Aligning Capital Incentives with Cognitive Augmentation (0:04) Gruber distinguishes between two historical paths in AI: pure automation designed to replace workers and satisfy benchmark metrics, and humanistic AI designed to expand human capability. While at least $1 trillion in financial incentives currently accelerates labor displacement, Gruber maintains that large-scale consumer demand can redirect commercial incentives, comparing the transition to Volvo establishing consumer safety features like seatbelts as profitable industry standards. Under the humanistic model, collaborative software agents act as 'eyeglasses for cognition,' enhancing human intellectual capacity in the same way optical engineering extended human vision. However, Gruber emphasizes that preserving competitive markets is essential to ensure these cognitive tools do not become proprietary monopolies controlled by a handful of platform operators. Regulatory Safeguards Against Centralized Power (2:51) Because AI functions as an omni-purpose inference technology comparable to electricity, Gruber cautions that concentrated control over advanced systems creates profound societal and geopolitical risks. Citing historian Yuval Noah Harari, he warns that if surveillance and generative tools are monopolized by authoritarian regimes or centralized corporations, power permanently shifts away from labor and citizens, making democratic recovery extraordinarily difficult once established. To counter this trajectory, Gruber proposes removing regulatory oversight from partisan politics and establishing dedicated expert agencies modeled after the FDA and FAA. Such agencies would enforce rigorous safety standards for foundational models while preserving open-source innovation. Additionally, he highlights the urgent necessity of legal frameworks to combat synthetic fraud, impersonation, and deepfakes to safeguard democratic discourse. Natural Language as a Frictionless Intermediary (10:28) Reflecting on Siri's origins and acquisition by Apple, Gruber identifies natural language—rather than voice alone—as the universal interface capable of dismantling traditional software complexity. Conversational models enable non-technical populations, such as rural communities communicating in local dialects, to utilize advanced computational systems across global telecommunication networks without needing technical literacy. As AI assistants integrate more deeply across personal hardware ecosystems, Gruber anticipates they will act as intermediaries dispatching automated API requests. While this structural shift threatens advertising-supported digital media and web search, he predicts that real-world service platforms like logistics and transportation providers will endure because their economic value resides in physical execution rather than app interface monopolies. Targeting Neglected Global Needs at Scale (16:41) Gruber points to scientific discovery platforms—such as DeepMind's AlphaFold for protein structure prediction and AI-driven material science breakthroughs—as existing proof that humanistic AI generates profound economic and societal dividends. Rather than duplicating heavily funded commercial efforts, such as oncology research, he urges software developers and founders to leverage foundation models to tackle under-resourced global crises. Specifically, Gruber calls for building dedicated vertical applications in global education, providing personalized AI tutors for every child, and expanding mental healthcare access to the one billion individuals worldwide identified by the World Health Organization who cannot realistically be served by human clinicians alone. Because foundation models shoulder the baseline engineering costs, impactful applications can be deployed rapidly before commercial dynamics harden over the next three to four years. Gruber's thesis hinges on the assumption that mass consumer demand and specialized regulatory bodies can overcome the estimated $1 trillion financial momentum favoring automation over augmentation. Whether non-profit vertical deployments in mental health and education can resist commercial capture remains an open challenge for developers.

精炼阅读摘要

Siri联合创始人Tom Gruber指出，当前AI发展正处于伦理分水岭，资本市场投入至少1万亿美元用于纯劳动力替代，因此业界必须转向以增强人类能力为目标的“人本主义AI”（Humanistic AI）。人本主义AI与纯自动化路线的根本分歧 (0:04) Gruber回顾了AI数十年来的发展历史，将其归纳为两条截然不同的路径：第一条是以基准测试和击败人类冠军（如击败国际象棋特级大师卡斯帕罗夫或通过司法考试）为导向的纯自动化路线，其核心目的在于替代人类劳动；第二条则是人本主义AI，其系统设计、初衷与运作全面服务于人类实际需求，强调人机协作与智能增强。Gruber指出，这两者的哲学分歧在现实中带来了巨大的资本分配失衡，当前全球至少有1万亿美元的财务激励正在流向劳动力自动化，而明确投入人本主义AI的资金规模则相对匮乏得多。权力过度集中与全用途技术的治理窗口 (2:51) Gruber引用历史学者尤瓦尔·赫拉利（Yuval Noah Harari）的观点警告称，AI监视系统一旦落入集权者手中，权力将永久性地从劳工和公众转移至资本所有者与中央权力机构，这种失衡日后极难逆转。尽管领先的AI企业管理者初衷良好，将涉及全球80亿人的技术控制权集中在少数商业巨头手中仍具有严重的系统性风险。Gruber将AI定义为一种类似于电力的“全用途”（omni-purpose）推理技术，其影响范围远超1994年打破时空协作壁垒的万维网；他认为，若能形成包含数亿至数十亿用户的需求规模，公众力量便可倒逼市场激励与监管模式发生转变，当前AI快速扩张且终点未明的阶段，是确立以人为本伦理选择的最后窗口期。认知眼镜隐喻与安全驱动的市场机制 (5:20) 在实际应用模式上，Gruber将人本主义AI比作“认知领域的眼镜”（eyeglasses for cognition）。如同光学眼镜矫正视力、显微镜赋予人类“超人视力”一样，AI能够作为认知增强工具拓展智力边界。为防范认知透镜被单一垄断者掌控，市场必须保留充分的竞争与选择权。Gruber以汽车工业发展史为对照：底特律车企最初为了销售大马力汽车而抗拒配备安全带，但沃尔沃（Volvo）通过将安全做成卖点并商业化，最终促使安全带与安全气囊从高端特性转变为行业通用标准。他预测，随着消费者对安全与人本设计的要求提高，类似的商业竞争路径将在AI领域重演。科学加速成果与专业化政府监管框架 (7:28) Gruber指出人本主义AI在基础科学中已经展现了巨大价值，例如DeepMind荣获诺贝尔奖的AlphaFold蛋白质结构预测技术，以及发现了数万种新材料的AI系统。材料科学历来是半导体、光纤与锂电池的技术基石，AI加速底层科学突破能够带来显著的社会与经济红利。针对政策制定，Gruber向政府提出两项核心职能：一是设立类似于美国食品药品监督管理局（FDA）或联邦航空管理局（FAA）的专业专家监管机构，取代非专业的政治操作，制定模型安全标准并保护开源创新以防行业垄断；二是严格打击虚假信息、伪造与深度伪造（deepfakes），防止数字垃圾内容侵害事实基准，以守护民主对话机制。自然语言接口与服务生态的演进 (10:28) 关于人机交互，Gruber强调语音仅是一种模态，人类自然语言本身才是真正的通用计算接口。传统图形界面的复杂性一直困扰非技术用户，而语言使得用户仅凭陈述意图即可调用代理系统；配合高速电信网络，这甚至能让讲本土方言的印度偏远村民平等使用先进大模型。针对市场对AI代理切断第三方服务商流量的担忧，Gruber认为物理服务型平台（如Uber、Grubhub）的核心资产是现实物流运作能力，这类企业不会消亡，而是会转向接入AI API继续提供配送服务；相比之下，高度依赖搜索流量与广告模式的数字新闻媒体则面临着更致命的商业冲击。 Gruber向开发者呼吁，不必耗费巨资重复研究“攻克癌症”等已有重资投入的项目，而应直接利用大模型底座，聚焦填补全球未被满足的刚性需求，例如为世界卫生组织指出的全球10亿缺乏医护支持的心理健康人群提供辅助，以及为所有儿童开发个性化AI导师。他警告称，技术快速固化的时间窗口仅剩三到四年，行业必须尽快行动。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台上。页面简介指出，这是 Siri 联合创始人汤姆·格鲁伯（Tom Gruber）针对现代大语言模型与生成式 AI 发展现状带来的深度概述，探讨了 AI 的构建原理、潜在风险以及在教育、医疗和认知增强等领域的赋能前景。 ## TL;DR 访谈围绕“人文主义 AI”与“纯自动化 AI”的分野展开，汤姆·格鲁伯指出当前绝大多数资本都在追逐替代人类劳动的自动化，而真正的关键在于构建服务人类需求、增强人力的协同系统。他回顾了从 1994 年 Web 浪潮到 Siri 创办的历程，强调语言界面对消除数字鸿沟的意义，并警示了极权监控与认知混乱风险。最后，他建议开发者抓住当下的技术窗口期，利用底层大模型去解决全民心理健康和个性化教育等长期被忽视的重大社会需求。 ## Atmosphere & style 整个对话气氛睿智而充满深度的理性思考，既有科技先驱对技术演进的深刻洞察，又洋溢着人文主义 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) 人文主义 AI（Humanistic AI）旨在服务人类需求并通过协作增强人类能力，与追求纯粹自动化以替代人类劳动的路线有着本质区别。 - [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) 当前全球有上万亿美元的资本正在推动替代劳动的自动化 AI，导致权力加速从劳动者向资本方转移。 - [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) 如果极权者或少数巨头垄断了现代 AI 监控系统，公众将极难夺回权力，因此现在是人类做出伦理选择的关键时刻。 - [04:07](https://www.youtube.com/watch?v=LLigihQXf-8&t=247s) AI 是一种类似电力的“全用途技术”，能够为人类认知能力戴上“近视眼镜”，其安全性与人文属性将成为未来的市场核心竞争商品。 - [04:07](https://www.youtube.com/watch?v=LLigihQXf-8&t=247s) AI 在加速科学发现方面展现出巨大的人文价值，例如预测蛋白质结构及研发新材料。 - [08:34](https://www.youtube.com/watch?v=LLigihQXf-8&t=514s) 政府应设立独立的专业监管机构制定统一安全标准，并通过法律打击诈骗与冒充，维护理性的社会认知与“思想市场”。 - [10:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=628s) 自然语言是无需学习的通用交互界面，能大幅降低使用门槛，结合语音能力可穿透至偏远地区，消除数字鸿沟。 - [12:45](https://www.youtube.com/watch?v=LLigihQXf-8&t=765s) Siri 初始愿景是作为通用访问接口；未来设备端 Agent 将成为用户与云端大模型及复杂技术之间的中介。 - [12:45](https://www.youtube.com/watch?v=LLigihQXf-8&t=765s) AI 对实体服务（如物流送餐）的解耦冲击有限，未来世界将演变为由 AI 驱动的 API 生态。 - [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) 创业者和开发者应利用现有的底层大模型，专注于未被满足的大规模垂直领域需求（如全民心理健康和全民个性化教育）。 ## Detailed walkthrough 以下是根据视频访谈前半部分（ Part 1 of 2）整理的详细、逐字级还原的深度记录： ### [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) 人文主义人工智能（Humanistic AI）与纯自动化 AI 的本质区别访谈由 BBC 科技资深记者托马斯·杰曼（Thomas Germain）主持，受邀嘉宾为 Siri 联合创始人汤姆·格鲁伯（Tom Gruber）。托马斯首先邀请汤姆解释什么是“人文主义 AI”（Humanistic AI），以及它与目前绝大多数科技公司正在构建的 AI 有何不同。汤姆指出，人工智能领域几十年来一直存在这种根本性的路线分歧： * **追求纯粹自动化的 AI（Automation of Intelligence）：** 构建 AI 的目的仅仅是为了实现智力本身的自动化。这类系统通过在各种 Benchmark 基准测试中获胜、在棋盘上击败卡斯帕罗夫（Gaspard/Kasparov）或通过法律资格考试来证明其优越性。其核心动力在于用技术直接取代人类劳动。 * **人文主义 AI（Humanistic AI）：** 其设计理念、立项初衷和整体目标，都是为了服务于某种具体的人类需求。它是通过协同与增强人类能力来实现价值，而非纯粹的替代。汤姆感叹，这种理念上的差异在今天显得尤为关键。目前全球至少有**上万亿美元的资本激励**正源源不断地砸向“取代人类劳动”的自动化 AI 方向。虽然指向人文主义 AI 的资金比例尚不明确，但他幽默地调侃：“希望我们‘人类队’（Team H）能分到更多资金！” ``` 人工智能的两种发展路线 ┌───────────────────┐ │ 人工智能 (AI) │ └─────────┬─────────┘ │ ┌─────────────────┴─────────────────┐ ▼ ▼ ┌───────────────────┐ ┌───────────────────┐ │ 纯自动化 AI │ │ 人文主义 AI │ │ (Automation) │ │ (Humanistic AI) │ ├───────────────────┤ ├───────────────────┤ │• 目的: 替代人类劳动│ │• 目的: 服务人类需求│ │• 评估: 基准测试/胜负│ │• 手段: 协作与认知增强│ │• 驱动: 万亿资本激励│ │• 愿景: 赋能个人与群体│ └───────────────────┘ └───────────────────┘ ``` --- ### [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) 资本激励、资本与劳力的博弈，以及为何现在是关键转折点托马斯接话指出，商业资本的驱动力往往是让企业赚更多钱，而非无条件地帮助用户。汤姆回应道，向善的激励其实一直都在——比如“是否应该攻克癌症”，全人类都会一致同意。问题的核心在于：**当前风险资本追求的目标与人类的长远利益是否对齐？** 汤姆强调，哪怕是在现有的政府体制和资本主义框架下，如果几十亿大众能够真正意识到“AI 应当是用来赋能人类的”，这股认知力量就能撼动市场与政府的决策。我们正处于一个爆发式扩张的前夜，而未来充满了巨大的不确定性： * **资本与劳力的转移：** 纵观全球，技术正在将权力迅速从“劳动者”（Labor）手中转移到“资本方”（Capital）手中。 * **极权与监控的风险：** 汤姆引用了著名历史学家尤瓦尔·赫拉利（Yuval Harari）的警示——一旦极权者掌握了现代监控技术与 AI 系统，公众将极难再把权力夺回；哪怕权力集中在少数动机良好的巨头公司手里，影响着 80 亿人，同样难以逆转。 * **最后的抉择时刻：** 目前行业处于一种“大家表面上都宣称为了全人类利益”的假象中（比如成立 B Corp 公司，一边喊着为了人类，一边再要 1000 亿美元资金）。汤姆警告，**现在可能是人类做出强有力伦理选择的最后机会**——必须优先去做对人类有益的事，至于利润和企业财务底线（Bottom line），完全可以稍后整理清理。 --- ### [04:07](https://www.youtube.com/watch?v=LLigihQXf-8&t=247s) 从 1994 Web 浪潮到全用途（Omni-purpose）AI 的认知“眼镜” 托马斯追问，过去每个技术时代（如搜索引擎索引全网）人们都声称能带来变革，为什么汤姆对这次的 AI 技术格外乐观？汤姆回忆起他在斯坦福大学工作时经历的 1994 年 Web 互联网浪潮。当时令他无比兴奋的是，互联网消除了人类协作的时空阻隔，那是**人类集体智慧增强（Augmentation）的上一次巨大飞跃**。而这一次的 AI 变革要远大于互联网： * **全用途技术（Omni-purpose Tech）：** 汤姆更正了自己口误提到的“无所不知”（Omniscient），强调 AI 是一种像**电力**一样的“全用途技术”，能够应用于推理、决策等几乎所有领域，引发集体能力的阶跃。 * **认知界的“近视眼镜”：** 汤姆以眼镜做类比——人类在光学上早已解决了眼镜和手术问题，甚至能通过显微镜和望远镜获得“超人类”的视觉。AI 的本质，就是**为人类的认知能力戴上“近视眼镜”**。 * **竞争与安全诉求：** 汤姆进一步补充，正如早期的底特律汽车厂商排斥安全带，直到沃尔沃（Volvo）证明“安全性是消费者愿意付费的卖点”，AI 的安全性与人文属性也将成为市场竞争中的核心商品。此外，AI 在加速科学发现方面的表现已展现出极高的人文价值。例如 DeepMind 的 AlphaFold 预测蛋白质结构（斩获诺贝尔奖），以及利用 AI 探索出数万种全新材料——材料科学正是半导体、光纤和锂电池的基石。 --- ### [08:34](https://www.youtube.com/watch?v=LLigihQXf-8&t=514s) 监管机构与“思想市场”：政府在防止 AI 混乱中的责任面对许多初创公司在壮大后抛弃初心、向市场妥协的现象，托马斯询问政府除了反垄断（如禁止 Anthropic 收购 OpenAI）之外，还能做些什么？汤姆提出了两点具体建议： 1. **设立专业的独立监管机构：** 就像对待药品（FDA）或航空安全一样，不能只靠个别国会议员去拍脑袋决定。行业需要一个汇聚专业人士的监管机构，制定统一的安全标准，让包括开源社区在内的所有玩家都能在公平、安全的框架下竞争。 2. **维护社会认知的基本理性（Keep Society Sane）：** 民主制度赖以生存的基础是理性的公众讨论。如果未来人们分辨不出眼前的文字/图像是真是假，或者屏幕对面和你对话的是人还是机器，整个社会集体将会陷入发疯状态。政府必须通过针对诈骗、冒充（Impersonation）的法律，保持极其轻量但有效的干预，打造一个**“思想的市场”（Marketplace of ideas），而不是“垃圾内容的市场”（Marketplace of slop）**。 --- ### [10:28](https://www.youtube.com/watch?v=LLigihQXf-8&t=628s) 语音与语言界面：从 Siri 的诞生到消除数字鸿沟作为 Siri 的联合创始人，汤姆聊到了语音与自然语言作为交互界面的深远意义。托马斯吐槽自己经常对手机发脾气，询问语音是否是计算形态的终极答案。汤姆指出： * **自然语言才是核心：** 语音只是载体，真正的核心是**语言（Language）**。软件界面在各处都极其笨拙（极具共鸣的是，许多技术人员经常收到家里长辈因软件太难用而打来的求助电话）。语言是全人类通用、无需专门学习的界面。 * **从繁琐点击到直接意图：** 自然语言交互让所有人都能享有少数“技术高手”才能掌握的电脑能力，极大降低了使用门槛。目前 Siri 每天被使用数十亿次，即便随后的各类 Chatbot 兴起，其核心范式依然没有脱离 Siri 开创的框架。 * **赋能偏远地区（Emancipation）：** 结合电信基础设施，AI 的语言能力可以穿透至偏远乡村（如印度乡村）。一个既不懂英语也不懂印地语的村民，能够直接用当地方言与全球最顶尖的 AI 进行语音实时对答。这种技术的普及具有真正的“解放”意义。 --- ### [12:45](https://www.youtube.com/watch?v=LLigihQXf-8&t=765s) Siri 的原始愿景、苹果的整合与生态系统解耦（Disintermediation）谈及苹果对 Siri 的更新以及 Siri 离最初的“终极蓝图”还有多远时，汤姆分享了 Siri 早期创业的历史细节： * **通用访问界面（Universal Access）：** Siri 创办初期的宏大愿景，是为整个移动和计算生态系统打造一个通用接口。在被苹果收购前，Siri App 已经具备多轮对话能力，甚至是一个电子竞技/电商应用（可以直接订花、预订酒店等）。 * **史蒂夫·乔布斯（Steve Jobs）的邀请：** 当时乔布斯找到他们团队并说：“你们可以继续自己做生意赚钱，也可以加入我，做出一个全世界每个人都在用的产品。”团队随即选择加入苹果。 * **解决中间层摩擦：** 汤姆指出，随着架构演进，设备端的小型 Agent 可以作为用户与复杂技术之间的中介。这个中介既能处理手机本地任务，也能调用云端巨型 AI 模型去解决诸如雅可比猜想（Jacobian conjectures）等高难度数学难题。针对《The Verge》主编尼莱·帕特尔（Nilay Patel）提到的“终极目标是让计算机替你使用它自己”，托马斯提出担忧：如果 Siri 或 ChatGPT 成了唯一的中介，Uber、Grubhub 等服务商是否会被脱钩（Disintermediated）？汤姆解释道： * **实体服务的不可替代性：** 真正的 AI 解耦冲击主要发生在新闻出版与传统搜索（因为新闻依赖广告和付费墙，AI 总结会导致流量流失）。但像 Grubhub 或 Amazon 这样的实体服务，其核心壁垒是线下的物流、送餐与供应链，而非仅仅是一个 App 界面。 * **AI API 时代：** 未来世界将演变为由 AI 驱动的 API 生态。虽然软件界面的“入口地产”（Real estate）价值可能会缩水，但提供扎实物理服务的公司依然不可或缺。最后，汤姆也幽默地表示希望未来能迎来免费新闻业的复兴，托马斯则笑着回应：“但愿他们还能想办法给我们发工资！” --- <ElicitationsGroup message="想要进一步探讨本访谈的内容？可以尝试以下方向："> <Elicitation label="生成汤姆·格鲁伯的人文主义 AI 核心观点逻辑图" query="请帮我总结汤姆·格鲁伯在访谈中关于人文主义 AI 的核心观点与逻辑架构，并用清晰的结构展现。"/> <Elicitation label="对比 Siri 初始愿景与现代大模型 Agent 的异同" query="请结合访谈内容，深入分析 Siri 早期创业时的 Agent 愿景与现代基于大语言模型的 AI Agent 有何异同。"/> </ElicitationsGroup> ### [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) 践行人文主义 AI：给创业者与开发者的建议在对谈接近尾声之际，主持人 Thomas Germain 总结道，要让“人文主义 AI”（Humanistic AI）真正落地生根，最终起到决定性作用的莫过于市场激励机制——无论这种激励是由政府法规引导建立，还是由广大大众的强烈需求所驱动。然而，对于单打独斗的普通个体而言，要凭一己之力改变庞大的市场机制无疑极其困难。面对现场众多专注于开发 AI 产品的技术人员与创业者，主持人向 Siri 联合创办人 Tom Gruber 提出了一个关键问题：作为这些工具的缔造者，大家现在具体应该怎么做，才能将人文主义 AI 的宏伟愿景变为现实？ Tom Gruber 给出的核心建议是：开发者和创业者应当将目光投向那些“尚未被满足、且尚未被主流巨头覆盖的大规模人类需求”，去打造专注于特定垂直领域的应用（Vertical Apps）。他特别强调，大家没有必要盲目地挤进那些已经被资本和精英高度关注的赛道，例如去攻克“治愈癌症”这样的宏大课题，因为已经有无数顶尖团队和资源在为此努力；相反，创业者应该去关注那些规模巨大但长期缺乏人手与服务的社会痛点。针对具体的创业方向与社会价值，Tom Gruber 提出了两个极具潜力的垂直领域范例，并阐明了利用底层大模型进行上层创新的可行性： - **全民心理健康护理（Mental Health Care）：** 根据世界卫生组织（WHO）的数据，全球有超过 10 亿人存在心理健康需求，而仅凭人类专业的心理医生和咨询师数量，是永远无法完全覆盖这一需求的。AI 能够在此提供大规模、低成本的辅助与护理服务。 - **全民个性化教育（Education for All）：** 为地球上的每一个孩子配备专属的 AI 导师（A tutor for every kid on the planet）。通过这种方式，可以从根本上抹平教育资源的鸿沟，拉平全球知识劳动的竞争起跑线。 - **低成本上层应用创新：** 攻克这些社会级难题并不需要耗费上万亿美元的巨额投入。开发者完全可以站在基础大模型公司（Foundation Model Companies）昂贵且坚实的成果之上，利用现有的底层能力去快速解决上层的实际痛点。 - **抓紧当下窗口期：** 这种技术与市场的红利期非常紧迫，创业者必须“现在就去解决这些难题”（Solve them now），因为在三到四年之后，这种阶段性的市场切入机遇可能就不复存在了。在探讨完对创业者的寄语后，主持人 Thomas Germain 对 Tom Gruber 带来的深刻洞见与精彩分享表达了诚挚的感谢，对谈在现场观众的热烈掌声中圆满落幕。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=LLigihQXf-8&t=4s) 希望我们‘人类队’（Team H）能分到更多资金！ — 汤姆·格鲁伯 > [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) 当前风险资本追求的目标与人类的长远利益是否对齐？ — 汤姆·格鲁伯 > [01:21](https://www.youtube.com/watch?v=LLigihQXf-8&t=81s) 现在可能是人类做出强有力伦理选择的最后机会 — 汤姆·格鲁伯 > [04:07](https://www.youtube.com/watch?v=LLigihQXf-8&t=247s) 人类在光学上早已解决了眼镜和手术问题，甚至能通过显微镜和望远镜获得‘超人类’的视觉。AI 的本质，就是为人类的认知能力戴上‘近视眼镜’。 — 汤姆·格鲁伯 > [08:34](https://www.youtube.com/watch?v=LLigihQXf-8&t=514s) 打造一个‘思想的市场’（Marketplace of ideas），而不是‘垃圾内容的市场’（Marketplace of slop）。 — 汤姆·格鲁伯 > [12:45](https://www.youtube.com/watch?v=LLigihQXf-8&t=765s) 你们可以继续自己做生意赚钱，也可以加入我，做出一个全世界每个人都在用的产品。 — 史蒂夫·乔布斯 > [12:45](https://www.youtube.com/watch?v=LLigihQXf-8&t=765s) 终极目标是让计算机替你使用它自己 — 尼莱·帕特尔 > [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) 开发人员和创业者应当将目光投向那些‘尚未被满足、且尚未被主流巨头覆盖的大规模人类需求’ — 汤姆·格鲁伯 > [16:41](https://www.youtube.com/watch?v=LLigihQXf-8&t=1001s) 为地球上的每一个孩子配备专属的 AI 导师（A tutor for every kid on the planet）。 — 汤姆·格鲁伯 ## Mentioned Tom Gruber, Thomas Germain, Siri, BBC, Gaspard/Kasparov, Yuval Harari, Volvo, DeepMind, AlphaFold, FDA, Anthropic, OpenAI, Steve Jobs, Nilay Patel, The Verge, Uber, Grubhub, Amazon, WHO ## Video description Tom Gruber - Co-Founder, Siri Thomas Germain - Senior Technology Journalist, The BBC Almost overnight, the world has come to see that the new technology of large language models and generative AI could change the nature of knowledge work. Trained on the cultural legacy of humanity, these systems can now do many tasks involving language, images, and information that are the expertise of educated professionals. In this overview of modern AI, which is updated with fresh insights every time it is given, Tom explains how these systems are created, what they are great at, what to be careful with, and the exciting applications that they enable. He brings to bear decades of experience in designing AI systems and advising AI companies. Applications include education, creative services, software engineering, knowledge management, mental health care, and augmented cognition. He offers insights into the potential risks to individuals and society from ungoverned applications, and a positive vision for the future of truly personal AI. [Source](https://www.youtube.com/watch?v=LLigihQXf-8)

Source quotes

[0:04] 汤姆·格鲁伯: 希望我们‘人类队’（Team H）能分到更多资金！ [1:21] 汤姆·格鲁伯: 当前风险资本追求的目标与人类的长远利益是否对齐？ [1:21] 汤姆·格鲁伯: 现在可能是人类做出强有力伦理选择的最后机会 [4:07] 汤姆·格鲁伯: 人类在光学上早已解决了眼镜和手术问题，甚至能通过显微镜和望远镜获得‘超人类’的视觉。AI 的本质，就是为人类的认知能力戴上‘近视眼镜’。 [8:34] 汤姆·格鲁伯: 打造一个‘思想的市场’（Marketplace of ideas），而不是‘垃圾内容的市场’（Marketplace of slop）。 [12:45] 史蒂夫·乔布斯: 你们可以继续自己做生意赚钱，也可以加入我，做出一个全世界每个人都在用的产品。 [12:45] 尼莱·帕特尔: 终极目标是让计算机替你使用它自己 [16:41] 汤姆·格鲁伯: 开发人员和创业者应当将目光投向那些‘尚未被满足、且尚未被主流巨头覆盖的大规模人类需求’ [16:41] 汤姆·格鲁伯: 为地球上的每一个孩子配备专属的 AI 导师（A tutor for every kid on the planet）。

Source quotes

[0:04] Tom Gruber: At least $1 trillion in financial incentives driving AI toward replacing and automating human labor, whereas the funding directed specifically toward humanistic AI remains far smaller. [1:21] Tom Gruber: This moment represents humanity's last chance to make a definitive ethical choice to prioritize human benefit before optimizing for corporate balance sheets. [2:51] Tom Gruber: Once totalitarians or dictators acquire advanced AI surveillance systems, power shifts permanently away from labor and citizens toward capital and central authorities, making it nearly impossible to reclaim. [5:20] Tom Gruber: Eyeglasses for cognition [10:28] Tom Gruber: Language bypasses this friction entirely, enabling anyone to simply state their intent to an agent. [13:47] Tom Gruber: Services like Grubhub and Uber will survive because their core value lies in real-world physical operations, adapting to receive orders via AI APIs rather than manual app taps. [16:41] Tom Gruber: Target mental healthcare for the one billion people worldwide who, according to the World Health Organization, can never be realistically served by human healthcare workers alone. [16:41] Tom Gruber: Cuz in three or four years you may not have the chance.

Original transcript

[0:04] All right. Thanks everyone for having us. Uh Ben, I could give us an intro, but I think you'd do a [0:09] better job. Can you explain what humanistic AI is and how it differs from what most of the companies [0:15] out there are building right now? Yeah, sure. It's a distinction in AI that's been in the history for [0:21] decades. There are two reasons to do AI. One is to create automation of intelligence for its own [0:26] sake. And so, you know when you're winning when you like beat Kasparov in chess or beat, you know, [0:33] champion of this or that. Um and you know, it's winning busy in benchmarks, you know, [0:37] passing the law. That's AI for its own sake, automation. There's another side which the AI [0:44] community has also been working on. It's called humanistic AI. And that's AI whose design, whose [0:48] intent, the whole purpose of building it is to serve some human need. And it's you think it's a [0:55] what's the big deal, right? Except that years later, the distinction is starting to matter. [1:02] There's about at least $1 trillion of incentives aimed at AI for just automating human labor. [1:10] And and not I don't know how much is pointing at humanistic AI, but let's hope we can [1:14] team H can get some more. Yeah. [laughter] Well, I think that's, you know, last time we talked [1:21] this is where the conversation went. That intent of I think is a big part of it, right? There's [1:24] a lot of money that is pushing companies towards building tools that will help them make more money [1:30] as opposed to necessarily help users depending on what the situation is. Are the incentives there [1:36] to build that more uh I don't I don't know gentle or helpful version of this technology? [1:43] Oh, there's incentives to do good things for humanity. I mean, you know, that's not a hard [1:48] problem. If someone says, "Oh, should we cure cancer?" Everyone says, "Yes." That's the kind of [1:52] thing. The question is if the are the incentives that the capital is pursuing aligned correctly? [1:57] Maybe not is is you know according to my rules of the game, but that doesn't matter. [2:01] The good news is that even um even with the system we have a government right now and the system of [2:06] capitalism here right now, if a billion people get it that AI could actually empower humanity, [2:14] they can move things. They can move incentives. They could be marketplace, it could be government, [2:19] it could be anything, but I think that's the key. People have to understand we are at this point [2:24] where we're about to go through this very very rapid expansion with uncertainty on the other end. [2:30] And so now is the last chance, possibly the last chance we have, to make a strong ethical choice, [2:36] which is to do the stuff that's good for humans first. We can clean up and do the stuff that's [2:41] good for the bottom line later. Right? But let's let's solve these problems we've never been able [2:46] to solve before for human benefit and do it and put our energy and intelligence into that now. [2:51] And why is why is this turning point? Like why do we have to do that first before we [2:57] go in the business direction? Couldn't you do the human stuff later as well? Like what what's [3:02] How does this play out? I think we go there, but it's interesting cuz it's really a bigger [3:06] picture than technology. It's a tech conference, right? But we're suddenly in a situation where [3:11] we're on the cusp of totalitarian overtaken several other countries. We have basically [3:15] every kind of thing is moving power from labor to capital all over the world. Um the Yuval Harari, [3:22] for instance, the famous historian has written eloquently about how once a dictator gets [3:28] um surveillance technology and other kind of AI technologies, it's hard to ever take it back. [3:33] Okay? It's And if the if a small number of companies, even if they're good people, [3:38] own all of that power that's affecting 8 billion people, it's kind of hard to take that back, too. [3:44] But right now we have this kind of kumbaya world where everyone's saying, "Yeah, we're building it [3:48] for everybody. It's a Well, it's a B corporation. It's for the benefit of humanity. Let's not forget [3:52] Give me another 100 billion dollars. Thank you very much. Anyway, but we're at that point and [3:56] I think I'm actually optimistic that we'll do the right thing. I think we will. Because people are [3:59] going to see what you can do with AI in the near future and they're going to go, "Yeah, [4:03] let's do that." Why are you more optimistic about this technology than others? We could say the same [4:07] thing about all the different eras of technology. Oh, finally we can do XYZ. We can we can [4:12] index the entire internet. Now with all these things are possible. Why is this make you more [4:17] hopeful than previous technologies? I was a big fan of the of web when it happened. I was [4:22] working at Stanford at the time. It was pretty exciting to think we could now eliminate space [4:27] and time barriers for human collaboration. That was pretty good. And but remember notice [4:32] the sentence. It wasn't the machines took over. It was like we could make people smarter collectively [4:36] by we can enable them to work together and so on. So it was it was a ma- last giant step in [4:41] augmentation for human intelligence in 1994. We have now that the next step [4:47] which is much bigger actually because it's omniscient. I mean not not omniscient, sorry. [4:52] Big mistake. Scratch. It's not omniscient. It's omni-purpose omni-purpose tech. That is [4:59] it can be be used for a lots and lots of different things. Like electricity is omni-purpose, right? [5:03] And yeah, the internet's kind of that way too, but it's mainly communication. But this is anything. [5:07] So all kinds of inference, right? So now we're talking about the the the really big step change [5:13] in human capability, collective human capability with this stuff. Makes a lot of sense. So [5:20] let's get a little more practical here. What does this look like? Like what makes a tool humanistic [5:25] as opposed to more business oriented if that's the other end of this. Is it the interface? Like [5:30] what how when we're building it? Like what are we doing differently? Well, that's exactly it. Starts [5:33] from when you're building it. So the main point is the definition I gave out there is just it's [5:37] AI designed to meet human needs and it's generally done through collaboration or augmentation. [5:44] And so all this business of agents, they're collaboration agents. They're collaborating with [5:47] people to do things. And if the human need is is to get a job done, to do something that people pay [5:52] you to do, fine. That's that's a human need. And and and it could be a localized human need. Um and [5:57] then the augmentation thing is a little bit more uh far-reaching, especially when you're talking [6:02] about cognitive or augmentation. But we both wear glasses. I just didn't have mine on. Glasses [6:07] are awesome augmentation technology. I mean, we have this technology for optics is a solved [6:12] engineering problem. It's amazing. We can and e and we can you know, we can cut our eyes I mean, [6:17] we can do anything we want to make our vision like Oh, but let's not forget, we can have superhuman [6:22] vision already with microscopes and telescopes, right? So, we're kind of like that with AI now. [6:29] AI is about to give us the equivalent of eyeglasses for our cognition. And can you [6:34] imagine if there's only one company that made eyeglasses in the world or one government [6:38] that controlled it? It's kind of interesting. But it's so competition is a big part of it. [6:43] Oh, uh thanks for that. Yes, thank you. I needed that underhanded pitch cuz uh competition is going [6:49] to Look, look what happened with the car industry, right? Um there used to be Detroit was like, "No, [6:54] no, you can't make me put seat belts in cars. We would just want to make them bigger and faster, [6:57] more muscle cars, you know." And then eventually Volvo came along and said, "You Americans are just [7:03] missing the point. Safety is something people will pay for." And they made a whole brand over it, [7:08] right? And then and then years later they thought about airbags again. And after a while airbags [7:11] became a thing you paid a lot for. So, you know, things like AI safety or humanistic uses of it [7:18] are something people really want and they'll pay for it. Um also, I think there's there's [7:23] a little bit of government incentive and kind of and philanthropic incentive can go on here. If you [7:28] If you If all the money only goes to how do we cut labor force, then you know, we'll get what we paid [7:33] for. Uh but there's a lot of incentive that can be done, for instance, accelerating science. I I [7:40] a small amount of intelligence at the right place, well, a small amount of bit strong intelligence. [7:45] So, right now we you probably heard of the AlphaFold thing that got a Nobel Prize for this [7:50] protein structure prediction thing that DeepMind did. They also did a thing with material science. [7:55] They dis- had the AI discover tens of thousands of new materials. Now, materials is what created [8:02] the semiconductor, it created optical fibers for the network, it's what created lithium for [8:07] lithium batteries, lithium ion. So, materials are really the important thing that drive tech and we [8:13] and the AI is already grinding away making new discoveries in that case. That's humanistic AI. [8:19] Um and it the the benefit could be economic. But you see that companies like Google, they are [8:24] investing, they're putting money into that part of it. They're not just putting it all into how [8:27] do we get rid of jobs? So, part of the incentive can be economic, right? But I think we watched, [8:34] you know, I've been reporting on technology for 10 years now. We've watched a lot of companies [8:38] that started out really driven by ideals. You know, as they move towards, uh you know, [8:45] bigger and bigger market power, shift away from some of that. I think and some people would argue [8:50] they've abandoned a lot of those ideals. Uh if the government is going to step in and [8:54] create additional incentives, I mean, I suppose it could just be there's going to be competition, so [9:00] safety can be built into the model and the people can get what they actually desire. But if the [9:04] government is going to do something aside from, you know, saying Anthropic you can't buy Open AI, [9:11] what is that going to look like? Like what would you ask them to do? Well, I'd ask them first of [9:15] all, actually enforce some guardrails. We need to establish the agency that everybody in the [9:19] industry is asking for. Please just listen and do what we ask for. Regulate the industry in a [9:25] sen- not not a regulatory capture, in a sensible way with an agency full of experts like we do [9:29] with drugs and aviation and everything else that's dangerous but powerful. And not just leave it in [9:35] the hands of individual congressmen or whatever to make decisions, right? So, we make an agency [9:39] that helps produce a process that everybody can play equally on, including open source, [9:45] to say this is a safe model. So, that's one thing. That's low-hanging fruit. The second thing, [9:49] we have a government has a moral responsibility to kind of keep help keep society sane, especially [9:54] democratic governments, because democracy depends on there being a kind of informed conversation, [9:59] right? And if you cannot tell whether the thing you're looking at every day is real or not, [10:04] or that thing you're talking to is a person or not, you we're going to go crazy as a collective [10:10] thing, and that's coming really fast, right? Governments can do things about that. There are [10:14] laws for fraud. There's a law for impersonation and so on. Um there's a lot of ways that [10:19] governments can do very lightweight but sensible things to create a kind of a marketplace of ideas, [10:25] not a marketplace of slop, right? So, let's shift gears slightly. You're one of these guys [10:30] that did something that was so successful that no matter how successful you are at everything else, [10:35] that's always going to be the first thing in your bio, right? But I think there's an interesting [10:40] uh thing to talk about here. So, you're one of the co-creators of Siri. Do you think that voice [10:43] is the ultimate form of computing? Is that how we want to Like I'm constantly getting angry [10:48] at my phone. If I could talk to it, would it be better? Is that going to make it more humanistic? [10:53] I I think again, humanistic is where the the point of it is was to help us use our mobile devices, [10:58] use our devices. So, it's help us get the benefit of the computers. And so, uh the voice was [11:04] is a modality in service of this It's just too hard to use those little things to get real thing [11:09] real work done. Um however, it's I think it's voice and it's actually language. The bigger thing [11:14] is language. So, why is language a good interface to AI? Because we all know how to use it. Right? [11:22] I mean, I don't know how many of you are all most of you technical. How many times do people in your [11:26] family call you about the most idiotic IT problems at home because they're so hard to use? Right? [11:31] I mean, software interfaces are broken everywhere. But if you can talk to whatever the AI agent it [11:38] is, this this flavor this year, and you say, "I just want to do that." And it just does it. [11:44] You are now empowering all of us to get the benefit of the a few of us who have mastered [11:48] all the clicking and all that. Um I think that's really the big payoff of the whole the voice [11:53] paradigm that Siri introduced. Um it's I'm very proud of the fact that it's used a billion times [11:58] a day. It's still the one that people use a lot. And And also, you know, the paradigm hasn't moved. [12:05] The other chatbots are the same paradigm, right? And it's because you can't beat human language. I [12:11] mean, if you want to get outreach, you go you got you need two things. You need telecommunications [12:15] and and language. So, telecommunications and they did their own job with the internet and [12:19] dark fiber and all this stuff. All of a sudden now, I can reach all the way into rural India [12:25] into a village that doesn't speak any English or Hin- Hindi even with a- the best AI in the world. [12:33] Because the AI can speak voice all the way to the rural Indian and come back. And that's an [12:39] amazingly emancipating result of all this. So, you used the word emancipation. I I felt a little [12:45] of that with the new update to Siri's. I go, "Finally, it can do more than set a timer." You [12:51] know, obviously does a lot more than that to begin with. How close are we to like your original like, [12:56] you know, blue sky vision of what Siri could be? And this is a major step in that direction. What's [13:02] missing? What What do you hope it could bridge into next? Well, the full vision was, you know, [13:06] create create an interface to the whole ecosystem of technology, especially mobile and and computers [13:14] that anyone can use. So, it's basically the idea of they should call it universal access at Apple. [13:17] That That was the goal. And AI was the best way of doing it and and the conversational interface [13:22] was the best way of achieving it. Now, at the time, we could when we built the app, we actually [13:26] had multi-turn conversations and we could do a lot of use cases. It was kind of an e-commerce [13:31] app back then. You could like order flowers and book hotels and all that kind of stuff. And when [13:35] Steve Jobs found us, he says, "Look, you guys are doing good stuff. You can build a business [13:39] and make money or you can join me and make a product that everyone uses." So, we we said, [13:43] "Okay." We we signed up. Sounds good. Yeah. And so, what what Apple has done now is they [13:47] figured out that they've gotten rid of a bit of a lot of the friction, but not all of it. [13:51] And then I had that I think this current uh these haven't really played with that much in September. [13:55] September, they'll be probably the original vision of like just do the stuff on your phone. Right. Uh [14:01] but you know, once you have that little agent thing intermediating between you and all that [14:06] other tech. Remember, it could be your phone, your Mac, your speaker at home, you know, whatever. Um [14:13] that little intermediary can do a lot of other things. Like talk to [14:17] big honking AI models somewhere else and do hard problems like solve, you know, Jacobian [14:23] conjectures and things like that you know, while you're on your little device. So, [14:28] uh I think it was Nilay Patel at The Verge talked about how the ultimate goal here is a computer [14:32] that uses itself for you. Uh and we talked about market concentration. If you've got one assistant [14:40] that's controlling everything, right? Like why would I go to Uber or to Grubhub directly [14:47] when I could just ask Siri to do it for me. At that point, we're disintermediating Grubhub [14:53] and Uber from their customers. Right now, there's a lot of things I wish Siri would do. It'll add a [14:58] song to my Apple Music playlist. It won't do that for Spotify. I think because neither of them want [15:02] it to. Uh how do we solve that problem? Does it just get to a point where like Siri or ChatGPT [15:08] or whatever it is is so big that you don't have a choice, you have to play ball with them? Or is [15:12] there a way that this is also empowering for the smaller players? Yeah, I don't think this is this [15:17] is as disin- Like the the real disintermediation with AI was it really via via search and the other [15:23] the intelligent now the conversational search. Yeah. Because it would put you out of business, [15:28] journalism. Right? Because you won't be here next year. You won't be. [15:32] I mean, but journalism depends on advertising and subscription basically the paywalls and so on. [15:39] And that's going to go away because of all this, right? So but actually the agents like Siri [15:43] don't do that. So that Grubhub business doesn't survive because it has a cool mobile interface. [15:50] It survives because it does an important service that you can't do yourself. [15:53] Right? It's It's like you know, I call the they bring the bring the food. It's It's really cool, [15:58] right? All right. That service is still there even if it was invoked over an API via an AI assistant. [16:06] Right? So the world will come to be AI APIs driven by AIs and it and only in the real [16:13] estate of the the little box you type into will be a slightly less valuable, right? [16:19] But who did the hard I mean, who's built these companies that really have you know, long-lasting [16:24] most like Amazon. These are physical hard things. You don't think your way to displace that company. [16:32] Right? So I'm hoping that's I'm hoping that's the outcome we see. I also hope we [16:34] get a renaissance of maybe free journalism or something. Hey, that'd be nice. Yeah, [16:39] but we're still going to pay your salary though somehow. Yeah, somehow. We'll [16:41] We'll talk about figuring that out. Yeah. Uh so we're getting close to running out of time here. [16:46] But we've talked about what really is going to bring humanistic AI to fruition if it happens [16:52] is market incentives more than anything else whether that's the government creating [16:56] them or the people demanding them. Uh that's not something individuals have a ton of control over. [17:03] But there's a lot of people in this room who are building these tools. What would you ask [17:07] them to do to bring this vision to light? Well, I'd say either yourself or other people [17:14] start on the vertical apps for the unserved human needs at scale that other people are not doing. [17:20] Don't go after curing cancer, people are working on that. Go after mental health care for the [17:24] billion people the World Health Organization says can't be served ever by humans, right? [17:29] Um, go after education, a tutor for every kid on the planet so we can level the playing field of [17:35] of knowledge work on the planet. Go after these problems. These don't require a trillion dollars. [17:40] Right? These can be so You can build on top of the hard work that the foundation model [17:44] companies are doing, expensive hard work, and start solving these hard and solve them now [17:50] cuz in three or four years you may not have the chance. Well, Tom, this is a [17:53] great conversation. I appreciate the time. Thanks very much, man. Thanks, everyone. Yeah.