Uber Halts Token Runaway With Open Models and AI Pods

Building an AI-Native Company at Global Scale with Praveen Neppalli Naga, Uber and Nathan Bomey — Ai4

Ai4

Uber CTO Praveen Neppalli Naga says multi-model routing and agentic workflows flattened AI costs while adoption quadrupled.

AI-assisted summary of Building an AI-Native Company at Global Scale with Praveen Neppalli Naga, Uber and Nathan Bomey — original source. Check the original for full context.

Source published: 2026-08-21

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 21, 2026, this fireside session features Uber CTO Praveen Neppalli Naga in conversation with Axios reporter Nathan Bomey. The discussion explores what it takes to embed AI at true global scale, measure meaningful ROI across technical and non-technical operations, and navigate the expanding future of autonomous transportation.

## TL;DR Praveen Neppalli Naga explains how Uber shifted past unconstrained 'token maxing' by optimizing agentic engineering through prompt caching, smart model routing, and embedded cross-functional 'Agentic Pods.' The discussion highlights practical AI integrations across Rides, Eats, and the Earner Companion, backed by Uber's custom open-source Agent Detection and Response (ADR) security platform. Looking to physical AI, Uber details its multi-partner autonomous vehicle strategy and hybrid marketplace, which balances flexible human driver supply with dedicated robotaxi fleets to achieve sustainable unit economics.

## Atmosphere & style The session strikes a crisp, pragmatic, and collegially candid tone, blending journalistic probing with grounded technical candor. Nathan Bomey keeps the dialogue energetic with playful commentary on tech hype, while Praveen counters with thoughtful, understated realism about the engineering headaches and CISO security anxieties of autonomous agents. The dynamic feels authentic and unpretentious, moving smoothly from deep infrastructure trade-offs to lighthearted banter about competitor apps.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) The era of unconstrained 'token maxing' has ended, prompting Uber to manage agentic budget shocks through prompt caching, smart defaults, and mixed model routing—quadrupling adoption while flattening costs. - [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) Automating isolated, single-step tasks yields negligible value; true ROI comes from deploying cross-functional 'Agentic Pods' to re-engineer complex, end-to-end business workflows. - [05:51](https://www.youtube.com/watch?v=5SSDLni677w&t=351s) Uber avoids superficial AI features by targeting high-utility product touchpoints: voice booking for quick Rides decisions, multi-item grocery parsing for Eats, and real-time earning guidance for drivers. - [08:00](https://www.youtube.com/watch?v=5SSDLni677w&t=480s) Multi-billion-parameter LLMs ingest nuanced driver behavioral histories and marketplace context to improve dispatch matching and ETA accuracy beyond traditional deep learning. - [09:45](https://www.youtube.com/watch?v=5SSDLni677w&t=585s) To secure autonomous agents without stifling engineering velocity, Uber created and open-sourced Agent Detection and Response (ADR) and quarantined Tier-1 consumer data in air-gapped sandboxes. - [13:02](https://www.youtube.com/watch?v=5SSDLni677w&t=782s) Uber pivoted from proprietary autonomous vehicle R&D to a multi-partner global ecosystem, deploying roughly $10 billion into fleet management, charging infrastructure, and depot operations. - [14:05](https://www.youtube.com/watch?v=5SSDLni677w&t=845s) Through Uber Labs, the company uses its global operational scale to curate edge-case L4 driving data, solving the primary training bottleneck for its autonomous vehicle partners. - [15:33](https://www.youtube.com/watch?v=5SSDLni677w&t=933s) A hybrid marketplace pairing autonomous fleets with flexible human drivers solves the peak-demand utilization problem, expanding total market demand rather than cannibalizing human earner income. - [17:57](https://www.youtube.com/watch?v=5SSDLni677w&t=1077s) Declining sensor and vehicle costs are steadily improving autonomous unit economics, while fragmented regional regulations favor partnering with local autonomy champions over a single global supplier. - [19:09](https://www.youtube.com/watch?v=5SSDLni677w&t=1149s) In automated last-mile logistics, Uber relies on partner networks, viewing step-climbing sidewalk robots as often more practical than aerial drones for dense urban fulfillment.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) Introduction and Moving Past the "Token Maxing" Era

The fireside session opens with an introduction highlighting Uber's transition into an AI-native organization, operating at a global scale where millions of passengers, drivers, and couriers interact with its platform daily. Praveen Neppalli Naga, Chief Technology Officer at Uber since joining in 2015, leads the company's engineering, science strategy, and technical execution. Axios reporter Nathan Bomey kicks off the conversation by playfully probing whether Praveen can "dance around these questions," to which Praveen jokes back that while he technically could dance, the audience would be thoroughly disappointed: "I have a lot of boring engineering and AI."

Nathan pivots immediately to Praveen's widely circulated commentary declaring that the era of "token maxing" is over, asking whether the industry has just exited a Jurassic period of unconstrained model consumption. Praveen explains that he did not invent the phrase, but used it to describe the very real, practical budget shocks companies faced when autonomous agents took off faster than expected. While Uber had budgeted for standard AI initiatives, the rapid shift toward multi-step agentic workflows rapidly exhausted early spending allocations. Rather than instituting heavy-handed curbs that would stifle innovation, Uber treated token consumption as an engineering optimization challenge:

```mermaid flowchart TD accTitle: Uber AI Cost Optimization Strategy subgraph S1["Engineering Interventions"] A["Prompt Caching"] B["Smart Model Defaults"] C["Frontier + Open-Weight Routing"] end subgraph S2["Platform Results"] D["4x Agentic Adoption"] E["Flattened & Dropping Costs"] end S1 --> S2 ```

- **Prompt Caching:** Implemented system-wide caching to prevent repetitive API calls for common queries and prompts across engineering workflows. - **Smart Defaults:** Established baseline default models so everyday tasks avoid defaulting to the most expensive frontier foundation models. - **Model Diversity:** Blended top-tier proprietary frontier models with cost-effective open-weight alternatives (such as Meta's Llama and Moonshot AI's Kimi). - **Scale and Cost Trajectory:** Over five months, agentic workflow adoption quadrupled across thousands of Uber engineers while total expenditure flattened and began to decline.

---

[04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) Shifting from Simple Tasks to End-to-End Agentic Workflows

Nathan notes that around the tech sector, some executives have lamented employees utilizing enterprise AI licenses for trivial queries like checking the weather, asking how Uber ensures responsible and impactful usage. Praveen recalls that when Uber's CEO mandated a strategic pivot to become an "agentic AI company," early usage naturally started with low-hanging fruit—drafting emails, polishing grammar, and generating isolated code snippets. However, Praveen emphasizes that automating discrete, single-step tasks yields negligible business value. The true transformative unlock happens when organizations elevate from individual tasks to re-engineering end-to-end organizational workflows that bridge cross-functional silos.

To operationalize this, Uber created dedicated "Agentic Pods." The company selected 30 of its top AI engineers and paired them directly with operational and business domain specialists across finance, operations, and marketing. Rather than relying on top-down directives, these engineers sat with subject-matter experts to map real-world friction points, organizational handoffs, and operational bottlenecks, embedding custom autonomous agents directly into those multi-step workflows.

- **Task vs. Workflow Mindset:** Moving engineering focus from generating isolated functions to automating complex, cross-departmental business processes. - **Agentic Pod Composition:** High-density pairings of elite AI engineers with functional domain specialists in finance, marketing, and operations. - **Decentralized Innovation:** Direct pairing allowed staff closest to the operational friction to design agents that eliminated administrative and decision-making drag.

---

[05:51](https://www.youtube.com/watch?v=5SSDLni677w&t=351s) Embedding Practical AI into Rides, Eats, and the Earner Experience

Addressing prior remarks by Uber COO Andrew Macdonald regarding the scarcity of impactful consumer-facing AI features across the industry, Praveen explains that Uber deliberately avoids "slapping AI onto a product" merely for novelty. Instead, the company concentrates on practical, high-friction user surfaces across its three primary offerings:

```mermaid flowchart LR accTitle: AI Integration Across Uber Product Verticals subgraph R["Rides App"] R1["Fast Decision-Making"] R2["Voice Interaction"] end subgraph E["Uber Eats"] E1["30+ Item Grocery Carts"] E2["Complex Multi-Item Search"] end subgraph D["Earner Platform"] D1["AI Earner Companion"] D2["Demand & Offer Guidance"] end ```

Praveen highlights that in the rides app, consumers require instant, low-friction decisions regarding price and ETA rather than extended browsing interfaces, making voice-based interactions a natural fit. In contrast, Uber Eats grocery orders involve complex 30-item baskets where multi-item parsing and predictive replenishment provide tangible utility. On the supply side, drivers and couriers spend up to eight hours a day inside the app. For them, Uber developed an "Earner Companion" agent that analyzes real-time marketplace demand to recommend optimal driving zones, evaluate incoming trip offers, and maximize earning efficiency.

Nathan jokes about the classic user grievance where an app displays a 5-minute ETA that stretches to 10 minutes upon booking. Praveen explains that while traditional deep learning handled basic routing, traffic, and pricing, large language models (LLMs) with billions of parameters now ingest nuanced driver behavioral histories, acceptance probabilities, and contextual marketplace parameters to make matching algorithms and ETA calculations significantly more accurate.

---

[09:45](https://www.youtube.com/watch?v=5SSDLni677w&t=585s) Security Guardrails and Agent Detection and Response (ADR)

Discussing recent cybersecurity incidents across the AI landscape—including vulnerabilities reported across platforms like Hugging Face, Anthropic, Meta, and OpenAI—Nathan asks how Uber protects sensitive consumer data when deploying autonomous agents. Praveen admits the intrinsic tension between his role as an innovation-focused CTO and his security responsibilities:

> "The my other side of the CTO, which is my CISO, is really scared, right? I have like the one thing that keeps me up at night is the security of these things. You don't know what they do... and what the agent does."

Rather than locking down tooling and stifling engineering velocity, Uber's information security team engineered and recently open-sourced their proprietary **Agent Detection and Response (ADR)** platform. Because agents can autonomously interpret prompts, crawl external networks, and inadvertently expose credentials or secrets, monitoring high-volume agent sessions is essential. ADR utilizes specialized language models to evaluate execution sessions at scale, distinguishing benign activity from potentially malicious behaviors.

- **Data Classification & Isolation:** Sensitive Tier-1 (L1) consumer data is strictly gated from unrestricted agent access. - **Air-Gapped Sandbox Execution:** When agents interact with sensitive internal data environments, they are quarantined in contained execution hosts stripped of external internet access to eliminate leakage risks. - **Open-Source Tooling:** Uber open-sourced ADR after finding commercial market offerings insufficiently mature for enterprise agent monitoring needs.

---

[13:02](https://www.youtube.com/watch?v=5SSDLni677w&t=782s) Physical AI and Uber's Autonomous Vehicle Fleet Strategy

Turning to physical AI and autonomous vehicles (AVs), Nathan looks back at Uber's 2016 self-driving initiatives and notes the dramatic strategic pivot toward partnering with roughly 25 autonomous vehicle developers globally. When asked if this approach is simply throwing darts at a wall, Praveen clarifies that the industry has evolved from basic research into commercialization. Instead of confining autonomous capabilities to a proprietary in-house stack, Uber is establishing a global multi-partner ecosystem featuring companies such as Waymo, Nuro, Wayve, WeRide, Pony.ai, and Baidu across North America, Europe, and the Middle East.

Praveen outlines a multi-billion-dollar commitment—deploying approximately $10 billion into autonomous infrastructure and ecosystem enablement. This includes solving the physical complexities of fleet management, centralized charging depots, cleaning, and maintenance for thousands of autonomous vehicles. Furthermore, Praveen highlights that AV development has shifted from an algorithmic engineering bottleneck into a data curation challenge. Because autonomous models navigate common driving paths easily but struggle with rare edge cases, Uber established **Uber Labs** to perform L4 data collection across its global footprint, feeding rich real-world training data back to AV partners to accelerate their edge-case learning.

---

[15:33](https://www.youtube.com/watch?v=5SSDLni677w&t=933s) Marketplace Hybrid Supply and the Economics of Autonomous Fleets

Nathan raises the competitive landscape, comparing pure-play AV software developers like Waymo and Tesla—where Elon Musk champions nationwide Full Self-Driving (FSD) robotaxi fleets—against ride-hailing networks and rental car companies vying for fleet management roles. Praveen details why Uber’s hybrid marketplace engine represents an essential layer for autonomous fleet economics:

```mermaid flowchart TD accTitle: Uber Hybrid Marketplace Dynamics A["Baseline Autonomous Fleet"] --> C["Uber Marketplace Engine"] B["Flexible Human Driver Supply"] --> C C --> D["Peak Demand: Humans Fill Troughs"] C --> E["Off-Peak: High AV Utilization"] C --> F["Lower Rideshare Prices"] F --> G["Expanding Total Market Demand"] ```

Praveen explains that autonomous fleets are capital-intensive and require near-constant utilization to remain commercially viable. Dedicated AV operators face structural profitability issues during demand troughs and capacity shortages during sudden demand spikes. Uber’s dynamic marketplace solves this via **hybrid supply**: during demand surges, flexible human drivers absorb the spillover, while routing algorithms steer autonomous vehicles to high-density zones to sustain maximum utilization. Praveen points to operational deployments in San Francisco, Los Angeles, and Phoenix, where introducing autonomous vehicles has expanded overall rideshare demand across all categories rather than cannibalizing human drivers. As autonomy scales, falling per-mile costs will expand total addressable rideshare demand, creating a non-zero-sum market with room for multiple platform and vehicle operators.

[17:57](https://www.youtube.com/watch?v=5SSDLni677w&t=1077s) Autonomous Vehicle Economics and Global Partner Strategy

Addressing the broader shift toward autonomous mobility and whether consumers will eventually stop owning cars altogether, Praveen Neppalli Naga explains that hardware and safety sensor kits remain significantly expensive today. However, Uber is already witnessing a steady downward cost curve across vehicles and sensor suites. As hardware expenses decline, consumer ride prices will drop in tandem, creating the unit economics necessary to operate a sustainable, profitable business at scale.

When asked whether Uber intends to whittle down its roster of roughly 25 autonomous vehicle partners to just a handful, or if Uber prefers managing autonomous fleets rather than owning the core vehicle technology, Praveen emphasizes that the global landscape demands a diverse, localized partner ecosystem. Different geographic markets feature unique road conditions, local regulations, and domestic champions. Because the global addressable market is massive and fragmented across distinct national environments, Uber focuses on aggregating regional leaders rather than consolidating down to a single global supplier.

- **AV Cost Dynamics:** High upfront costs for specialized vehicles and safety-critical sensor suites are actively decreasing, paving the way for lower end-user pricing and profitable autonomous operations. - **Global Deployment Footprint:** Uber collaborates with distinct regional autonomy leaders based on local market strengths, including WeRide and Pony.ai in the Middle East, Wayve in London, and Nuro in the delivery space.

```mermaid flowchart TD accTitle: Uber Global AV Platform Strategy subgraph MarketDynamics["Cost & Deployment Drivers"] A["Sensor & Vehicle Costs Decline"] --> B["Lower Consumer Ride Pricing"] B --> C["Sustainable Platform Profitability"] end subgraph RegionalEcosystem["Global Partner Integration"] D["Uber Platform"] D --> E["WeRide & Pony.ai<br/>(Middle East)"] D --> F["Wayve<br/>(London)"] D --> G["Nuro & Specialized Fleets<br/>(Autonomous Delivery)"] end ```

[19:09](https://www.youtube.com/watch?v=5SSDLni677w&t=1149s) Last-Mile Logistics: Drones, Sidewalk Bots, and Competition

Turning to airborne drone delivery and ground-based robotic couriers, Praveen outlines that Uber maintains the same collaborative partnership model rather than building proprietary hardware in-house. While drone and robotic delivery face complex regulatory hurdles, airspace rules, and privacy considerations, several dedicated companies are investing heavily in the space, and Uber intends to integrate their capabilities into its network.

Responding to an anecdote about an Amazon delivery drone spotted flying over a neighborhood in Detroit, Praveen acknowledges Amazon as a perennial innovator in last-mile fulfillment. However, he highlights that the true challenge of last-mile delivery centers on optimizing speed, safety, and physical accessibility. In dense downtown urban cores, sidewalk delivery bots may play an even larger role than aerial drones. Solving friction points like ascending staircases—such as modern robots designed to climb steps—represents the critical engineering frontier for automated neighborhood delivery.

- **Regulatory & Privacy Hurdles:** Drone and sidewalk robotics face significant local municipal oversight and privacy constraints, making specialized delivery partnerships ideal for Uber. - **Urban Form Factors:** Ground-based bots capable of handling steps and urban streetscapes often present a more practical last-mile delivery solution in downtown environments than aerial drones.

[20:10](https://www.youtube.com/watch?v=5SSDLni677w&t=1210s) Axios Rapid Fire: The Lyft App on Uber CTO's Phone

In keeping with the Axios tradition of ending fireside sessions with "one fun thing," the interviewer catches Praveen unprepared by putting him on the spot with a lighthearted question: whether he currently has the rival Lyft app installed on his personal phone.

Praveen readily admits that he does indeed have Lyft on his phone. When asked for his honest appraisal of the competitor's product, he responds with good-humored diplomacy and a smile, noting that "they do a good job" and playfully adding, "I like the pink color," drawing the fireside chat to a cheerful close.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) I have a lot of boring engineering and AI. — Praveen Neppalli Naga

> [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) token maxing — Praveen Neppalli Naga

> [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) agentic AI company — Praveen Neppalli Naga

> [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) Agentic Pods — Praveen Neppalli Naga

> [05:51](https://www.youtube.com/watch?v=5SSDLni677w&t=351s) slapping AI onto a product — Praveen Neppalli Naga

> [05:51](https://www.youtube.com/watch?v=5SSDLni677w&t=351s) Earner Companion — Praveen Neppalli Naga

> [09:45](https://www.youtube.com/watch?v=5SSDLni677w&t=585s) The my other side of the CTO, which is my CISO, is really scared, right? I have like the one thing that keeps me up at night is the security of these things. You don't know what they do... and what the agent does. — Praveen Neppalli Naga

> [15:33](https://www.youtube.com/watch?v=5SSDLni677w&t=933s) hybrid supply — Praveen Neppalli Naga

> [20:10](https://www.youtube.com/watch?v=5SSDLni677w&t=1210s) one fun thing — Nathan Bomey

> [20:10](https://www.youtube.com/watch?v=5SSDLni677w&t=1210s) they do a good job — Praveen Neppalli Naga

> [20:10](https://www.youtube.com/watch?v=5SSDLni677w&t=1210s) I like the pink color — Praveen Neppalli Naga

## Mentioned Praveen Neppalli Naga, Nathan Bomey, Andrew Macdonald, Elon Musk, Uber, Axios, Ai4, Uber Eats, Uber Labs, Agent Detection and Response (ADR), Meta, Llama, Moonshot AI, Kimi, Hugging Face, Anthropic, OpenAI, Waymo, Nuro, Wayve, WeRide, Pony.ai, Baidu, Tesla, Amazon, Lyft

## Video description Praveen Neppalli Naga - CTO, Uber Nathan Bomey - Reporter, Axios

As AI becomes foundational to how businesses operate, industry leaders have shifted from experimentation to execution. This fireside explores what it takes to embed AI at true scale, from engineering to non-technical functions, how to asses meaningful ROI, and how advances in AI and autonomy are shaping the future of transportation – with the potential to positively impact millions of people's lives around the world.

[Source](https://www.youtube.com/watch?v=5SSDLni677w)

中文简报

Uber借助开放模型与AI Pod遏制Token消耗失控 Uber首席技术官Praveen Neppalli Naga表示，多模型路由与智能体工作流在AI采用量翻两番的同时稳住了AI成本。 Praveen Neppalli Naga解释了Uber如何告别无节制的“Token最大化”，通过提示词缓存、智能模型路由以及嵌入式跨职能“Agentic Pods”优化智能体工程。讨论重点介绍了AI在Rides、Eats和Earner Companion中的实际集成，并由Uber自研开源的“智能体检测与响应”（ADR）安全平台提供支持。面向物理AI，Uber详细阐述了其多合作伙伴自动驾驶汽车战略与混合市场模式，该模式平衡了灵活的人类司机供给与专用Robotaxi车队，以实现可持续的单位经济效益。 Uber 通过提示词缓存（prompt caching）、智能默认设置和前沿/开源混合模型路由有效遏制了智能体预算冲击，在使内部智能体采用率增长四倍的同时抹平并降低了成本。 (0:04) 自动化单一离散任务带来的业务价值微乎其微；Uber 组建了由顶尖 AI 工程师与业务专家配对的“Agentic Pods”，通过重塑端到端跨部门业务流程来实现实质性 ROI。 (4:00) 为了在保护数据安全的同时不拖慢工程速度，Uber 开发并开源了智能体检测与响应（ADR）平台，将核心客户数据隔离在无外网访问的沙箱环境中。 (9:45) Uber 从自研自动驾驶转向建立全球多伙伴生态系统，投入约 100 亿美元用于车队管理、充电与场站运营，并通过 Uber Labs 为合作伙伴提供稀缺的 L4 级长尾驾驶数据。 (13:02) Uber 打造的自动驾驶与人类灵活运力混合市场，通过削峰填谷解决了自动驾驶车队的高利用率难题，从而扩大整体出行需求而非挤压人类司机的收入。 (15:33)

Reading summary

Uber Chief Technology Officer Praveen Neppalli Naga argues that capturing enterprise value from artificial intelligence requires transitioning from isolated task automation to cross-functional agentic workflows, supported by strict token cost controls, internal security guardrails, and a multi-partner autonomous vehicle strategy. Engineering Discipline and Cross-Functional Agent Workflows (0:04) After early autonomous agent rollouts quickly consumed preliminary budgets, Uber addressed model spending through systematic infrastructure optimization rather than usage caps. By implementing system-wide prompt caching, establishing cost-conscious baseline model defaults, and dynamically routing requests between frontier systems and open-weight alternatives such as Meta's Llama and Moonshot AI's Kimi, the company quadrupled agentic workflow adoption across thousands of engineers over five months while flattening and reducing total expenditures. To move beyond single-step tasks with low commercial return—such as drafting text or producing standalone code snippets—Uber structured dedicated 'Agentic Pods.' The initiative embedded 30 top AI engineers directly alongside operational specialists across finance, marketing, and operations. This structure empowered teams closest to real-world operational friction to re-engineer complex, cross-silo business workflows with autonomous agents rather than implementing top-down software mandates. High-Friction Product Interfaces and Agent Security (5:51) Uber directs consumer-facing AI development exclusively toward high-friction surfaces rather than novelty features. In rides, integration centers on voice commands to facilitate rapid decisions regarding price and arrival times, while Uber Eats utilizes multi-item parsing to organize 30-item grocery baskets. For drivers and couriers spending extended hours on the platform, an 'Earner Companion' agent evaluates real-time market conditions to guide trip selection and position drivers in high-demand zones. Billions of parameters in large language models also ingest historical driver behaviors and marketplace context to improve dispatch matching and trip-duration estimates. To manage enterprise security risks associated with autonomous model actions, Uber enforces strict access gating on Tier-1 consumer data and confines agents handling internal data to air-gapped sandboxes without external internet connectivity. After finding third-party monitoring software inadequate, the company developed and open-sourced its Agent Detection and Response (ADR) platform, which uses specialized language models to monitor high-volume agent sessions for inadvertent credential leaks or malicious behavior. Decentralized Autonomous Vehicle Ecosystems and Data Curation (13:02) Departing from its earlier strategy of building a proprietary autonomous vehicle (AV) stack, Uber has established commercial partnerships with roughly 25 AV developers worldwide, including Waymo, Nuro, Wayve, WeRide, and Pony.ai. Naga asserts that the industry's primary bottleneck has shifted from core algorithmic research to edge-case data curation. In response, Uber established Uber Labs to collect Level 4 driving data across its international operations, feeding real-world edge scenarios back to partner companies to accelerate model training. Alongside vehicle integrations, Uber is deploying approximately $10 billion into physical autonomous infrastructure, addressing fleet maintenance, vehicle cleaning, and centralized charging depots. Naga emphasizes that because regulatory requirements, street geography, and dominant domestic developers vary significantly by country, aggregating localized regional champions is a more viable global strategy than standardizing on a single vehicle supplier. Hybrid Marketplace Supply and Last-Mile Automation (15:33) Addressing the economics of robotaxis, Naga maintains that standalone autonomous fleets face structural profitability hurdles due to capital-intensive hardware and fluctuating demand. Uber balances these dynamics through a hybrid supply engine: algorithms steer autonomous vehicles to dense urban cores to maintain near-constant baseline utilization, while flexible human drivers absorb demand spikes. Deployments in San Francisco, Los Angeles, and Phoenix indicate that introducing autonomous vehicles has expanded aggregate rideshare demand across all categories rather than reducing human driver opportunities. Uber applies a similar partner-centric model to automated delivery, avoiding proprietary hardware development while working with drone and ground robotics firms. Naga anticipates that ground-based sidewalk robots designed to overcome urban obstacles like stairs offer a more practical short-term solution for dense downtown fulfillment than aerial drones, which confront tighter local municipal rules and privacy constraints. Across both passenger and delivery segments, Naga forecasts that declining sensor and vehicle hardware expenses will steadily lower end-user pricing and establish sustainable unit economics. While Uber positions its hybrid marketplace and third-party aggregation model as the most economically viable framework for scaling autonomy, the strategy remains dependent on hardware partners achieving projected cost reductions and navigating divergent municipal regulations across international territories.

精炼阅读摘要

Uber首席技术官Praveen Neppalli Naga指出，企业AI转型已告别盲目消耗算力的“Token狂热”时代，转向成本工程化控制、端到端工作流重塑以及与核心业务的实质整合。通过构建混合模型架构、自研安全防护体系以及整合约25家自动驾驶伙伴的“混合运力网络”，Uber正将自主智能体与物理世界AI深度嵌入其全球出行与配送生态。告别Token狂热与重塑端到端工作流 (0:04) 在智能体多步工作流兴起初期，激增的模型调用迅速耗尽了Uber早期的算力预算。Praveen解释，“Token狂热”（token maxing）时代的终结反映了企业面对突发模型账单时的现实困境。Uber未采取扼杀创新的硬性限制，而是将其作为工程优化课题：通过引入全局Prompt缓存避免重复API调用，设定智能默认基准以防止日常任务无差别调用昂贵的前沿基座模型，并混合采用前沿商业模型与Meta的Llama、月之暗面的Kimi等低成本开源权重模型。在五个月内，Uber数千名工程师的智能体工作流采用率增长了4倍，而平台总体AI支出却趋于平缓并逐步下降。 Praveen强调，自动生成代码或草拟邮件等孤立的单步任务带来的业务价值微乎其微，真正的颠覆在于重塑跨部门的端到端工作流。为此，Uber设立了“智能体小组”（Agentic Pods），抽调30位顶尖AI工程师与财务、运营和营销等领域的业务专家紧密协作。这种模式打破了自上而下的教条指令，让直面业务堵点的一线团队直接设计自主智能体，消除流程中的流转迟滞与管理损耗。业务实效：打车、外卖与骑手端AI (5:51) 针对消费端AI功能多流于表面噱头的现象，Uber坚持避免形式化植入AI，专注于高摩擦的核心产品场景。在乘车端，用户追求快速决策，AI以语音交互支持即时获取价格与到达时间；在Uber Eats杂货配送中，面对多达30件商品的复杂购物车，AI专注于多品类解析与预测性补货；在供给侧，由于司机和外卖员每天在App内工作数小时，Uber推出“收入助手”（Earner Companion）智能体，分析实时市场供求数据以推荐高收益接单区域并评估行程效益。在底层调度机制上，除了传统处理路径规划和定价的深度学习，Uber现已引入数十亿参数的LLM综合分析司机的历史接单行为、接受概率预测及实时的市场上下文参数。Praveen指出，这种深层次的行为建模大幅提升了匹配精度与预估到达时间（ETA）的准确度，有效改善了长期困扰用户的预估偏差问题。智能体安全防御：沙盒隔离与ADR监控 (9:45) 面对行业频发的AI安全事件，Praveen坦言在推动技术创新的同时，信息安全是令其最担忧的问题。自主智能体具备自主解析Prompt和跨网络调用的能力，可能无意间泄露凭证或敏感数据。为此，Uber实施严格的数据分级管控，全面阻断智能体对Tier-1核心消费者敏感数据的无限制访问；当智能体与内部敏感数据环境交互时，必须置于切断外部互联网访问的“物理隔离”（Air-Gapped）沙盒容器内运行。此外，针对市面上商业监控工具成熟度的不足，Uber安全团队自主研发并开源了“智能体检测与响应”（ADR）平台。ADR借助专用语言模型对海量并发的智能体运行会话进行大规模实时分析，有效识别并隔离潜在恶意操作，为高自主性AI系统的规模化部署提供了运行时的安全可观测性保障。自动驾驶生态：混合运力与百亿基建 (13:02) 在物理AI与自动驾驶（AV）领域，Uber已从早期的专有全栈研发转向全球多方生态聚合，目前已与Waymo、Nuro、Wayve、文远知行（WeRide）、小马智行（Pony.ai）和百度等约25家开发商展开合作，并承诺投入约100亿美元用于车队运营、集中充电场站及维保等基础设施建设。针对算法在罕见长尾边缘场景（edge cases）下的技术瓶颈，Uber成立Uber Labs，借助全球运营网络采集L4级别现实行驶数据，反哺合作伙伴加速模型迭代。在商业模式上，Praveen指出AV属于重资产，面临低谷期闲置亏损与高峰期运力不足的结构性矛盾。Uber的核心竞争力在于“混合运力市场”：调度算法引导高成本AV车队在常态及高密度区域维持全天候高利用率，波峰需求则由弹性人类司机承接。旧金山、洛杉矶和凤凰城的运营数据表明，引入AV不但未挤压人类司机收益，反而扩大了整体出行总需求。随着车辆及激光雷达等传感器套件成本持续下行，每英里成本降低将支撑起可持续盈利的大规模商业闭环。末端配送：人行道机器人与空中无人机 (19:09) 在最后一公里配送探索上，Uber同样秉持开放整合策略，避免自研硬件，积极对接无人机与人行道机器人物流提供商。尽管该领域普遍面临市政监管、空域合规及隐私安全等挑战，但Praveen分析指出，在密集市中心环境中，地面人行道机器人的实用性往往超过空中无人机。当前末端自动化配送的关键工程关卡在于复杂地形的通行能力，例如研发能够爬楼梯的地面机器人，以解决深入建筑物内部的最后一米交付难题。 Praveen所描绘的AI原生战略依托聚合全球算法、硬件与运力网络，但该模式的最终兑现仍取决于自动驾驶硬件成本的实际降幅、各地区监管法规的放开节奏，以及在数十家独立技术伙伴之间平衡利益与协同运营的长期能力。

Detailed notes — zh

## Background 本内容发布于 YouTube 平台的 Ai4 频道，发布时间为 2026 年 8 月 21 日。页面简介显示，本次访谈由 Axios 记者 Nathan Bomey 对话 Uber 首席技术官 Praveen Neppalli Naga，共同探讨企业如何超越实验阶段将 AI 规模化落地、评估实际投资回报率（ROI），以及 AI 与自动驾驶技术的演进如何塑造未来交通出行。 ## TL;DR Uber 首席技术官 Praveen 阐述了 Uber 从“Token Maxing”狂热转向工程治理与成本优化的实践，通过提示词缓存、智能体工作流（Agentic Pods）及数据沙箱隔离实现高效降本与安全防护。在核心业务上，Uber 利用大语言模型优化地图匹配与预测算法，并打造运力端 AI 助手。面对自动驾驶与末端配送，Uber 坚持开放平台战略，投入百亿资金与全球约 25 家 partner（如 Waymo、文远知行等）合作，结合人类驾驶员的混合供给优势推动商业化落地。 ## Atmosphere & style 整场对话气氛专业而平实，既有深度的工程技术探讨，又充满前沿商业思维的碰撞。主持人提问精准犀利，直击 AI 成本控制、数据安全与竞品威胁等核心痛点；Praveen 的应答则显得从容且具条理性，展现出技术高管的严谨与自信。结尾处关于手机中是否安装竞品 App 的突袭趣问，不仅化解了硬核技术的严肃感，更以轻松风趣的笑声为整场访谈增添了真实而有人情味的互动色彩。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) Uber 通过 Prompt Caching、默认轻量模型以及结合开源模型与前沿模型，在智能体（Agent）使用量翻倍的同时平稳降低了 AI 使用成本。 - [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) Uber 避免将 AI 局限于写邮件等孤立任务，而是组建“Agentic Pods”与各业务部门配合，推动完整业务工作流的深度自动化。 - [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) Uber 针对网约车、外卖及司机端采取差异化 AI 策略，例如为驾驶员打造实时推荐高收益订单的 AI 助手（Companion Agent）。 - [07:34](https://www.youtube.com/watch?v=5SSDLni677w&t=454s) Uber 引入大语言模型（LLM）来融合海量驾驶员偏好与历史数据，大幅提升派单匹配与 ETA 预测的准确度。 - [09:45](https://www.youtube.com/watch?v=5SSDLni677w&t=585s) 针对智能体的数据泄露与提权风险，Uber 开源了 ADR（智能体检测与响应）平台，并对 L1 级敏感数据实施无网络沙箱隔离。 - [13:02](https://www.youtube.com/watch?v=5SSDLni677w&t=782s) Uber 在自动驾驶生态中采取开放平台战略，计划投入 100 亿美元，与全球约 25 家自动驾驶公司展开合作而非自研整车。 - [13:02](https://www.youtube.com/watch?v=5SSDLni677w&t=782s) Uber 通过 A Labs 项目收集并共享 L4 级长尾场景数据，同时依靠双边市场的人类驾驶员进行高峰削峰填谷，提升无人车队的利用率。 - [17:57](https://www.youtube.com/watch?v=5SSDLni677w&t=1077s) 随着传感器与制造技术的成熟，自动驾驶车辆成本将逐步下降，推动乘车服务价格下探与商业化普及。 - [17:57](https://www.youtube.com/watch?v=5SSDLni677w&t=1077s) Uber 积极在欧洲、中东等全球市场与 Wayve、文远知行、小马智行等本土化自动驾驶企业合作进行商业化部署。 - [19:09](https://www.youtube.com/watch?v=5SSDLni677w&t=1149s) 在末端履约方面，相比受空域法规与隐私限制的无人机，Uber 更看好能克服阶梯等物理地形障碍的地面履约机器人。 ## Detailed walkthrough <Sequence> <Step title="Uber CTO 探讨 AI 成本与模式转型"> ### [00:04](https://www.youtube.com/watch?v=5SSDLni677w&t=4s) 告别“Token Maxing”时代与 AI 预算管理 Axios 记者 Nathan Bomey 与 Uber 首席技术官 Praveen Neppalli Naga 展开对话。针对近期关于“Token Maxing”（Token 极客狂热/过度消耗）时代已经终结的讨论，Praveen 解释道，随着智能体（Agent）工作流的兴起，各公司原本规划的 AI 预算迅速耗尽。这在业界引发了关于创新与成本平衡的大讨论。Uber 并没有采取简单粗暴的“切断 AI 访问权限”做法，而是将其视为工程优化问题，通过引入 **Prompt Caching（提示词缓存）**、**设置合理默认模型**，以及结合 **开源/开放权重模型（Open-weight Models）** 与前沿模型（Frontier Models），成功实现了效率飞跃。在过去 5 个月中，Uber 内部使用 Agentic 工作流的工程师数量翻了数倍（数千名工程师投入使用），但 AI 使用成本反而平稳下降，验证了工程治理的有效性。 - **Prompt Caching 优化**：针对员工高频重复输入的 Prompt 进行缓存，大幅降低重复调用成本。 - **默认模型策略**：非高难度任务默认调用轻量级或开源模型，避免一律消耗高昂的前沿模型。 - **Agent 规模化与成本控制**：工程师采用率翻倍提升，总预算曲线实现平抑并回落。 </Step> <Step title="Agentic Pods 与工作流自动化实践"> ### [04:00](https://www.youtube.com/watch?v=5SSDLni677w&t=240s) 告别边缘应用，推动工作流深度自动化针对 Axios 记者提到“有公司员工用 AI 查天气”的资源浪费现象，Praveen 强调 Uber 的 AI 战略重点不在于写邮件、改语法等孤立任务（Task），而在于**完整工作流（Workflow）的自动化**。Uber 组建了由 **30 名顶尖 AI 工程师构成的“Agentic Pods”**，与财务、运营、市场等业务领域的专家深度结合，深入业务底层重构流程，从而释放出巨大的人效潜力。对于 COO Andrew McDonald 此前提及的“消费者端 AI 体验尚显不足”的观点，Praveen 拆解了 Uber 的三大核心业务板块及其 AI 赋能落地路径： 1. **网约车（Rides）**：核心需求是快速决策与即时响应（价格与 ETA），重点探索语音辅助交互。 2. **外卖与杂货（Eats & Grocery）**：属于多物品复杂决策（单次购物可能达 30 件商品），AI 智能体可提供列表化与个性化搜索协助。 3. **运力供给端（Earners）**：驾驶员每天使用 App 长达 8 小时，Uber 正为其打造 **AI 助手（Companion Agent）**，实时推荐高收益区域与最优订单接单策略。 </Step> <Step title="算法升级与黑盒背后的技术魔法"> ### [07:34](https://www.youtube.com/watch?v=5SSDLni677w&t=454s) 从传统深度学习到大语言模型预测用户使用 Uber 时“一键下发，车辆即达”的流畅体验背后，依赖于复杂的底层架构，包括地图与路径规划（Routing）、动态定价（Pricing）、智能匹配算法（Matching）以及风控识别（Fraud & Risk）。传统的深度学习模型在应对多变量预测时存在参数规模限制；而利用大语言模型（LLM）及大参数模型后，Uber 能够融入海量维度（如驾驶员历史偏好、接受度预测等），显著提升行程预测与匹配效率，缩短预估到达时间（ETA）与实际接驾时间的偏差。 - **路径与定价强绑定**：不同于常规地图导航 2 分钟的 ETA 误差，Uber 的行程时间预测直接影响动态定价与派单效率。 - **大模型增强预测**：通过输入驾驶员的历史接单行为与偏好参数，大幅提升派单匹配成功率。 </Step> <Step title="安全防御与自治检测平台 ADR"> ### [09:45](https://www.youtube.com/watch?v=5SSDLni677w&t=585s) AI 安全治理与开源 ADR 平台面对近期业内发生的安全事件（如 Hugging Face 等平台的数据与安全隐患），作为 CTO 的 Praveen 透露自己最担心的就是智能体在执行任务时可能带来的数据泄露或提权风险。为解决这一难题，Uber 安全团队开发并开源了 **ADR（Agent Detection and Response，智能体检测与响应）平台**。同时，Uber 建立了严格的数据隔离机制：对最敏感的 **L1 级数据（包含核心消费者隐私的数据）**，严格禁止 Agent 直接访问；若需要处理，必须在**无互联网连接（No Internet Access）**的沙箱宿主机中封闭运行，防止数据通过网络向外泄露。 - **开源 ADR 平台**：利用 LLM 实时监控 Agent 的运行会话，自动识别恶意行为与正常操作。 - **L1 级数据沙箱隔离**：对核心数据物理切断互联网连通性，防止 Agent 在外部探索过程中泄露凭证或敏感信息。 </Step> <Step title="物理 AI 与自动驾驶（AV）开放生态"> ### [13:02](https://www.youtube.com/watch?v=5SSDLni677w&t=782s) 100 亿美元布局与商业化数据闭环回顾 Uber 从 early 自研自动驾驶到如今与全球约 **25 家自动驾驶公司（如 Waymo、Nuro、Waabi、Pony.ai 等）** 合作的转型，Praveen 阐述了 Uber 的开放平台战略：自动驾驶技术已从单纯的技术攻关迈入商业化落地阶段，Uber 计划在自动驾驶生态中投入 **100 亿美元**（用于基础设施、车队运营与数据服务等）。针对自动驾驶最为头疼的“长尾场景（Tail Cases）”，Uber 创立了 **A Labs** 项目，专门收集 **L4 级别的数据** 并提供给 partner 用于模型训练，帮助合作伙伴加快技术迭代。此外，Uber 双边市场的核心优势在于**高峰期的混合供给（Hybrid Supply）能力**——利用人类驾驶员削峰填谷，显著提升无人车队的履约率与利用率，从而推动单位经济学走向盈利。 - **双边市场高峰削峰**：自动驾驶车队保持高利用率，突发高峰由人类驾驶员补充，解决单靠无人车队导致的资产闲置问题。 - **L4 数据赋能生态**：通过 A Labs 采集复杂长尾驾驶数据，反哺算法训练。 - **全球区域落地**：已在美洲、欧洲及中东与多家 AV 供应商展开商业化试运营。 </Step> </Sequence> <ElicitationsGroup message="想要进一步探索 Uber 的 AI 架构与自动驾驶战略？"> <Elicitation label="深入了解 Uber 开源的 ADR 安全检测架构" query="请详细解释 Uber 开源的 ADR (Agent Detection and Response) 平台的架构原理与安全拦截机制。"/> <Elicitation label="对比 Uber 与 Tesla 的自动驾驶商业化路径" query="请梳理并对比 Uber 的开放网约车平台战略与 Tesla 的 FSD/Robotaxi 垂直整合战略的优劣势。"/> </ElicitationsGroup> ### [17:57](https://www.youtube.com/watch?v=5SSDLni677w&t=1077s) 自动驾驶汽车的成本演变与全球合作伙伴生态 Axios 记者 Nathan Bomey 与 Uber 首席技术官 Praveen Neppalli Naga 深入讨论了自动驾驶汽车（AV）的商业化落地路径与经济可行性。针对未来人们是否会完全放弃购买私家车的疑问，Praveen 指出，当前自动驾驶汽车的价格居高不下，主要原因是车辆配备的各种传感器套件非常昂贵，而这些传感器对于确保出行安全至关重要。不过，市场已经展现出了明确的成本下降趋势。随着车辆及传感器制造技术的成熟与规模效应的显现，车价必然随之降低，进而带动乘车服务价格的下探。Praveen 强调，Uber 的终极目标是运营一家可持续且盈利的企业，因此成本降低是推动服务普及的核心钥匙。在谈及 Uber 目前拥有的约 25 家自动驾驶合作伙伴时，Nathan 询问 Uber 未来是否会精简合作方名单，以及 Uber 是否会重新尝试自主研发自动驾驶技术。Praveen 明确表示，Uber 不会致力于成为一家自研自动驾驶汽车的公司，而是专注于管理与调度这些自动驾驶车辆。他解释称，全球市场规模极其庞大，且各个国家和地区的监管环境、道路状况与本地需求差异显著。因此，全球不同区域将会孕育出不同的本土化龙头企业。 - **中东市场部署**：Uber 已在 middle east 地区与 WeRide（文远知行）以及 Pony.ai（小马智行）展开联合部署合作。 - **欧洲与英国市场**：在伦敦，Wayve 正积极推动自动驾驶方案落地；同时，Baidu（百度）等企业也在欧洲及国际市场有所布局。 - **专项与末端配送合作伙伴**：Uber 的合作阵营还涵盖了针对末端配送的 Nuro，以及在特定硬件或无人驾驶领域深耕的 Avride、Serve Robotics 等合作伙伴。 --- ### [19:09](https://www.youtube.com/watch?v=5SSDLni677w&t=1149s) 无人机与机器人末端配送及行业竞争话题随后转向无人机配送与机器人履约服务。Praveen 表示，Uber 在该领域的策略与自动驾驶汽车保持高度一致，即继续采取“开放合作”模式。尽管许多顶尖创业公司正在积极研发无人机配送，但这一领域仍面临着沉重的法规监管壁垒以及严峻的公众隐私担忧。 Nathan 分享了自己上周在底特律拜访朋友时的亲身经历：当时朋友突然指向天空，喊道“看，那是一架亚马逊配送无人机飞过社区”。Nathan 据此询问亚马逊的自研无人机是否会对 Uber Eats 构成直接竞争威胁。Praveen 坦然承认，亚马逊在末端履约领域的创新步伐始终领先，但履约竞争的核心始终是“交付速度”与“安全保障”。相较于无人机，Praveen 更看好地面履约机器人（Bots）在密集城市中心的应用前景。他提到了本次大会现场展示的能够爬楼梯的履约机器人，指出这种能够克服物理地形障碍（如阶梯、门槛）的技术，才是真正解决末端履约最后“几米”难题的关键突破。 - **无人机挑战**：法规限制（Airspace regulation）与隐私争议（Privacy concerns）是无人机配送大规模商用的两大瓶颈。 - **履约机器人潜力**：地面履约机器人更适合高密度的市中心场景，特别是具备跨越阶梯和障碍物能力的机器人，能更好地将物品直接送到用户门口。 --- ### [20:06](https://www.youtube.com/watch?v=5SSDLni677w&t=1206s) 大会尾声：关于竞争对手的趣问趣答按照 Axios 活动的惯例，主持人 Nathan 在未提前预警（“put you on the spot”）的情况下，向 Praveen 抛出了一个轻松幽默的闭幕问题：“你的手机里有竞争对手 Lyft 的 App 吗？” Praveen 毫不避讳地笑着承认：“我有。”当被问及对这个最大竞品的评价时，Praveen 展现出了大度与幽默感，打趣道：“我觉得他们做得挺不错的，而且我挺喜欢他们那个粉红色的配色。”这番机智而风趣的回答在全场的笑声与掌声中为本场关于 AI、自动驾驶与未来出行的对话画上了圆满的句号。 - **高管心态**：Praveen 坦言手机中安装了竞品 Lyft 的应用程序，并称赞其产品表现与品牌代表色。 --- <FollowUp label="想了解 Uber 在自动驾驶与配送领域的 25 家合作伙伴具体包含哪些公司吗？" query="请详细列出 Uber 目前公开的 25 家自动驾驶及无人配送合作伙伴，并说明各自的合作领域。"/> ## Notable quotes > [20:06](https://www.youtube.com/watch?v=5SSDLni677w&t=1206s) 我有。 — Praveen Neppalli Naga > [20:06](https://www.youtube.com/watch?v=5SSDLni677w&t=1206s) 我觉得他们做得挺不错的，而且我挺喜欢他们那个粉红色的配色。 — Praveen Neppalli Naga ## Mentioned Praveen Neppalli Naga, Nathan Bomey, Uber, Axios, Ai4, Andrew McDonald, Hugging Face, ADR, Waymo, Nuro, Waabi, Pony.ai, A Labs, WeRide, Wayve, Baidu, Avride, Serve Robotics, Amazon, Lyft, Uber Eats ## Video description Praveen Neppalli Naga - CTO, Uber Nathan Bomey - Reporter, Axios As AI becomes foundational to how businesses operate, industry leaders have shifted from experimentation to execution. This fireside explores what it takes to embed AI at true scale, from engineering to non-technical functions, how to asses meaningful ROI, and how advances in AI and autonomy are shaping the future of transportation – with the potential to positively impact millions of people's lives around the world. [Source](https://www.youtube.com/watch?v=5SSDLni677w)

Source quotes

[20:06] Praveen Neppalli Naga: 我有。 [20:06] Praveen Neppalli Naga: 我觉得他们做得挺不错的，而且我挺喜欢他们那个粉红色的配色。

Source quotes

[0:04] Praveen Neppalli Naga: I have a lot of boring engineering and AI. [0:04] Praveen Neppalli Naga: token maxing [4:00] Praveen Neppalli Naga: agentic AI company [4:00] Praveen Neppalli Naga: Agentic Pods [5:51] Praveen Neppalli Naga: slapping AI onto a product [5:51] Praveen Neppalli Naga: Earner Companion [9:45] Praveen Neppalli Naga: The my other side of the CTO, which is my CISO, is really scared, right? I have like the one thing that keeps me up at night is the security of these things. You don't know what they do... and what the agent does. [15:33] Praveen Neppalli Naga: hybrid supply [20:10] Nathan Bomey: one fun thing [20:10] Praveen Neppalli Naga: they do a good job [20:10] Praveen Neppalli Naga: I like the pink color

Original transcript

[0:04] Building an AI native company is about more than adopting new technology. It requires reimagining [0:09] how products are built, how our engineering teams operate and how AI is embedded across [0:14] the business. Few companies have navigated that transformation at the scale of Uber where millions [0:20] of people rely on its platform every day. Today I'm pleased to be joined by Pravin Npali Naga, [0:25] chief technology officer of Uber, where he leads the company's overall engineering and [0:29] science strategy and technical execution. Since joining Uber in 2015, his focus has been on [0:35] creating outstanding experiences for customers while building a flexible earning platform for [0:40] drivers and couriers. Please welcome to the stage Pine. All right, let's see if you can dance around [0:45] these questions. Well, if you're expecting me to dance, I can, but you'll be very disappointed. [0:52] goes. I have a lot of boring engineering and AI. [laughter] Well, all right. Let's let's [0:57] start here. So, just a few days ago, you made some waves by talking about how the [1:02] era of token maxing is over. So, is this some sort of Jurassic period that we've been in? Yeah. Look, [1:10] I mean, when when we originally talked about uh I didn't invent the word token maxing. I talked [1:15] about practical problems uh of how the budgets are being managed for AI. I think what we meant what [1:23] I meant was what we are seeing now with agent is very new. Even last year we didn't know this was [1:29] the direction we'd be going. So usually every company plans a budget for whatever you want to [1:34] spend and AI we had a budget. Uh we didn't expect all these models to go crazy and and what we what [1:41] I meant was that budget got uh exhausted. uh but but that's that led to a huge debate in [1:48] the industry which I'm very happy that we all have this industry because this is a practical [1:52] problem right for all of us and and I'm glad that the industry talked about this debated it and a [1:58] lot of innovation came out of it and you and and over the last four months of course there's so [2:04] much of talking in this so what I ask my team is to really focus on what are the efficiencies we [2:11] can find while at the same time using these models right the easiest thing to do is curb [2:16] the access. That's not what you should be doing. That's not how tech in Silicon Valley companies [2:22] do. So while we are encouraging our teams to use, we like okay what are the options and we learned a [2:27] lot of stuff through that. Um and the first thing is simple solutions like caching like we took it [2:34] as an engineering problem right not as a budget problem and simple solution like caching that's [2:38] that's not something new in this engineering tech industry. The prompt caching really helped us a [2:44] lot because everybody are searching prompting similar stuff. And the second is defaults. Not [2:50] all tasks need the best models, the expensive models, right? Just setting up basic defaults [2:56] really helped us as well. And then it is about using the right models for the right things and [3:02] the openweight models have taken off. So again, we want to be able to use frontier models and [3:08] open weight models. And then all this work that we did over the last five months, we already seen [3:14] that our adoption of this agentic workflows have quadrupled by thousands of engineers using it. [3:23] But our cost kind of flattened and dropped and and this is exactly what I was wanting to spend time [3:30] on and close to my engineering team all my teams who worked on it and we thought okay token maxing [3:36] was a word that got invented we thought okay well that era is now probably uh over uh and we all [3:44] are seeing every week there is a change that's happening it's incredible like uh and and the [3:50] and the Kimmy and then once Kimmy is there's the meta model that tell it's even cheaper in open air [3:55] as it's actually really good for all of us. Yeah, I think you sparked a national conversation about [4:00] budgeting and the balance between innovation and costs. One of my colleagues at Axios reported not [4:06] recently, not not long ago that a another CTO told her that a lot of their employees were using AI to [4:12] check the weather. I don't think that's what you intend to happen with your employees. So, [4:17] how do you ensure that your workforce is using AI responsibly? By the way, that's that's really [4:22] uh an important topic because I think last year when we all wanted to use AI, my CEO said he had [4:29] a strategic shift for the company and one of the strategic shift he said we need to become [4:33] an agentic uh uh AI company, right? And but what does that mean? Right? Everybody started using [4:40] AI to write emails, fix the grammarss even I did that and write some documents, right? [4:47] But even in engineering world taking a task and asking the agent to write the code for it is [4:54] not going to have a huge impact. That is what is called a task. You need to slowly start upleveling [5:01] and then work think of a workflow automating the entire workflow because that's where we spend a [5:07] lot of time at. You get buried in organization uh boundaries and then you get decision making [5:12] and all the stuff happens. that workflow when you automate we started seeing results and that's the [5:18] work we did with agentic pods uh we really took our best AI engineers 30 of them paired them [5:27] up with functional experts like finance our operations team our marketing team picked a [5:34] workflow that those folks know and engineers go in and understand the workflow and really [5:39] put an agent in it incredible the results that we saw and and now you See the our CEO and I pushed [5:47] for it but the innovation came from the people who know the stuff underneath and they took the [5:51] workflow out. So a few months ago your COO Andrew McDonald made a splash by saying that there really [5:57] haven't been that many consumerf facing features yet that he's seen come out from the AI investment [6:02] for you. You're the CTO. Did that sting or did you do you know exactly what you're saying? See look I [6:08] think what we were very clear was you can always slap in an an AI to a product but the question [6:15] is is that useful right at Uber we have two main products one is well in fact I'll say three one [6:22] is the right app where you go in you're trying to get the right as soon as possible you're not [6:28] trying to browse and stuff like that right it's a decision making the price and and and what ETA [6:33] you get so what is the agent workflow that can help voice is one thing, right? And the eat side, [6:39] it's much more complicated because you're grocery shopping. Like it's the grocery shopping is the [6:45] most complex one. You don't you're not shopping for one or two items. You're usually shopping for [6:49] 30 items and really finding searching one by one or you're having a list where you're trying to [6:54] punch it back. And that's where we therefore felt like focusing on those kind of experiences will be [7:00] helpful. And the others the third one what I said is which we usually don't talk too much outside [7:04] is our earner base. Our earners spend like eight hours in a day in certain countries on our app, [7:10] right? And we thought like okay what could be the agentic experience for them which is a companion [7:18] eight earner companion helping them which locations I should go to earn more and what kind [7:23] of offers I should be accepting and really being that companion and this this agent is able to take [7:28] the marketplace and help them out. So those are the kind of natural experiences we want to build. [7:34] Just not slap some AI. It's not over. I think with the average person experiences the Uber app, [7:38] they pop it up, they pop it up, they put in their destination, and then something magical happens [7:43] and it connects me with a a driver. Is that really where you think most of the innovation [7:47] is going to be or do you have something really granded? So the the magic that you talked about, [7:54] the simplicity with which you push a button, things happen. There is a lot of things that [8:00] happens behind scenes right first of all our maps platform and routing which takes into traffic and [8:07] all the things that goes in and then using that information pricing right if you don't route [8:12] properly the pricing impact in Google maps what happens are Apple maps if you get a 2 minutes [8:17] variance of your ETA it's it's not the end of the world for but for us it's impacts the price that [8:22] what so and then the matching algorithms that we you know and then fraud and risk and all the [8:29] models that comes in there's a lot of innovation that happens has been happening with traditional [8:35] deep learning models right now with LLM's large models we are able to put in billions of more [8:40] parameters and actually get better predictions out of it and that's the innovation that's we [8:44] focus yeah I mean I told you you're going to have to dance around questions here's my big [8:49] one which is when I put when I you know insert my destination in Uber and it says it's going to be [8:54] five minutes and then I click and it actually ends up being 10 minutes the AI and fix the um possibly [9:01] right which is which is if I go into a little bit of tech technical aspects of it that's the [9:08] predictions and matching we have to do like okay I have bunch of drivers who is more uh uh who is [9:15] more probable going to be accepting the trip and that needs to you need to be able to understand [9:20] what is the preference of the driver and a lot of parameters have to go into it right the limitation [9:26] was always the small models. Now with large models I can actually feed in all those preferences [9:30] drivers past behaviors and all these things and get the better prediction out of it and the better [9:35] prediction will reduce from the five minutes to maybe six seven minutes might not be five exactly [9:40] because it's we have a human on the other side who is trying to predict right so those are the things [9:45] that's possible so let's talk about uh you know the the risks that you see in here I mean with [9:50] obviously you know having this conversation in the backdrop of you know open AI talking [9:54] about this hugging face incident just in the last week or to now see a throp because said something [9:58] similar and now we see meta just in the last 24 hours saying something similar as well for an [10:03] organization like yours handling a lot of consumer data for example where are the gu guard rails what [10:08] do you need to make sure to do to make sure that you're protecting your customers look I mean see [10:13] as a this is the problem of being a CTO right one of the things is you're excited we want to use the [10:19] agents and you're trying to push the teams and and which means that you have to give access to [10:23] all these tools and these are all new tools nobody knows what they do and we just saw what what we [10:28] saw with hugging face the my other side of the CTO which is my seesaw is really scared right I have [10:36] like the one thing that keeps me up at night is the security of these things you don't know what [10:41] they do with and what the agent does so we've been very careful to the easiest thing is don't use it [10:48] until we solve this well that's that's that's curbing innovation my CEO has been very clear [10:52] about it I don't think investors would like that though if you said all of a sudden now we're not [10:56] Yeah, absolutely. And I I think that was not going to help because this is a new thing that you use [11:01] and learn from it. But I have a pretty awesome security team insect team which is who who have [11:06] been focusing using agents to help understand what's actually going on. That's what we open [11:12] sourced last week which which is our ADR uh agent detection and response uh uh platform that we [11:20] built and on the side we started innovating on it. The reason why we did that everybody could ask a [11:25] question you can just buy it. I don't think it's mature enough to available outside thing for us is [11:30] as we start building these things we will learn a lot and we learned a lot uh in building these [11:37] platforms uh and it's not perfect uh but we have some guardrails about as you said consumer data [11:44] right we give access to what we call as L1 data we don't give it give the access to that L1 data [11:49] to the agents if you do it we kind of contain no internet connection in that in that host so that [11:55] it's not going a leak at all in any way and then we are able to use ADR to monitor what's going [11:59] on and so forth. So it's a it's a new space and we should be really careful and watch out. You [12:04] talked about you mentioned a little bit about the need to have visibility with your agents. What [12:08] does that look like from a practical perspective? I think what h what happens is when you deploy an [12:15] agent and it is doing certain things you prompt it and it takes a prompt and it starts doing certain [12:23] things. You don't have control on what XP does right it could be that you asked to your prompt [12:30] could mean it could interpret that I need to go to internet and learn something and then do something [12:36] through that process it could leak some secrets or whatever credentials and stuff right so that's [12:43] where we had to this ADR comes into picture to be able to monitor those and it's impossible [12:48] to monitor everything every section it's it's just going to be humongous volume is extremely [12:52] expensive so that's where some kind like LLMs can actually say what are malicious, what are [12:57] benign sessions and and monitor those things and that's that's what we we're talking about here. [13:02] Yeah. Let's pivot a little bit and talk about your about physical AI and Uber's autonomous vehicle [13:08] strategy. So I'm old enough to have covered the 2016 uh event that Uber held long before your [13:14] time in which they gave us as media rides and the self-driving cars. Now that's a long time [13:19] ago and we're not going to ask you about that. The strategy has changed a lot since then. Now [13:23] you've partnered with something like 25 different autonomous vehicle companies throughout the world. [13:28] So is this sort of just a throw darts at the wall to see what sticks strategy? Like what's what's [13:33] the what's the precise uh motivation there? No, as you said back then there were only a few companies [13:39] doing it. Right. Right now there's many companies doing it and there are brilliant people that are [13:45] founded these companies and and our interest is uh our autonomous strategy as it grows is to be [13:52] able to work with all the all the companies each of them are are doing different ways [13:57] uh different tech techniques of as as me doing it's going to it's going to have its own uh growth [14:02] and now all of them are moving from technology to commercialization we already seen like we have [14:08] deployed with with three prov providers around uh in Middle East and Europe as such right and when [14:15] we think about it if we if we do it ourselves it's limited to us but we want to be able to [14:19] tap into all the great talent and companies that is coming out and that's the that's the [14:24] main strategy and the second is uh as we we we we are we are going to deploy about 10 billion into [14:31] this like this includes infrastructure for this there's a lot of innovation that needs to come [14:35] how you going to handle thousands of cars like how you going to charge them all the fleet operation [14:39] s and and then uh and then the other thing which I am very excited about which I've been working on [14:44] is what I learned from these partners for them to scale faster what used to be a technical problem [14:50] has moved to be a data problem you kind of will see the robot dancing here is a data prop is about [14:56] training with the data and the aronomy thing is all about tail cases right the happy path [15:01] is easy going in a highway but the tail cases are the hardest and that's what we want to help [15:06] and we built something called a labs which is data collection uh L4 data collection and using [15:12] that data to give it to our partners to be able to train faster and then go for it uh and and [15:17] and and that's our that's our strategy and and we have neuro and and Silicon Valley and wave [15:25] in London was and we write and and and pony and and bu these are all incredible companies there's [15:33] different camps here so we have the self-driving car companies like Tesla or Whimo who maybe they [15:38] feel like the technology is going to be where all the value is. And then you got yourself the [15:41] ride share company you, you know, you managing the fleet. You've even seen a company like the rental [15:46] car companies thinking that they could play a role here in managing fleets. And so you're talking [15:50] about the data like what sort of data is going to be most valuable here and how are you going [15:55] to use it against? So two things, right? One is the the the autonomy the cars driving growing by [16:05] itself then the infrastructure to manage the fleet cars training charging all the stuff like charging [16:12] infrastructures and all the stuff and third is the biggest thing where that what Uber brings is [16:18] a marketplace right now if you are having thousand cars utilization of those cars is very important [16:26] and we are very good at that right We have the consumer and the demand and we are able to do [16:31] the hybrid supply and whenever the peak happens we are able to have human supply come into the [16:36] equation and fill the trough and and and when the utilization is highly needed for the cars to be [16:42] profitable we are able to route them in the right places. The what we have seen is along the way the [16:48] categories expanded we already seen that in SF uh uh LA and Phoenix where we have autonomous also [16:54] cars running from other providers. the categories expanded and we have grown faster along the way. [17:01] So, uh that's where the marketplace is going to be very important. You know, you listen to Elon Musk, [17:05] it's no secret, no, you know, that he believes that Tesla is going to quickly pick up business [17:11] when they really are able to launch their FSD technology nationwide or globally. What do you [17:16] why do you think Uber has edge on Tesla? I think it's as I said the category is going to expand [17:21] you will see multiple players coming in and I think all I think that the demand is going to [17:26] grow along with that like as you can imagine as this comes the prices go down the prices go down [17:30] demand goes up the demand goes up everybody has the time goes up and everybody has a play [17:35] to it's not one player we don't see it as Uber alone in this equation I do think people will [17:40] be successful certainly Whimo is a factor there as well but when it comes to cost you just talked [17:43] about prices how quickly do you think prices will come down in this space right now obviously [17:48] It's very expensive and everyone's sort of doing it as a loss leading of thing but eventually the [17:52] price will come down. How quickly for the average person will become you know cost competitive to [17:57] maybe not even own a car anymore. Yeah. I think look right now the cost of the cars these cars [18:02] are not cheap. The sensor kits you want to be safe the sensor kits are very expensive. So as the cost [18:09] of the car it will come down. We already seeing that pattern. As the cost of the car comes down, [18:14] that's when you'll start seeing the prices also coming down because you at the end of that you're [18:18] running a business, you need to be able to be profitable, right? That's do you at this point, [18:24] okay, you've got the 25ish partners. Are you going to eventually whittle down or do you anticipate [18:29] maybe even um I mean getting down to a handful? You don't you're not going to own the technology [18:34] yourself. You don't want to be a self-driving car company. You want to manage your self-driving [18:37] cars. Is that right? So what's going to happen is around the world there going to be diff different [18:44] players in different countries right we deploy we had a v ride and and pony deploy in middle [18:50] east right and bu in London wave is deploying in London wants to do it neuro big gear we have [18:56] a gear so you'll see that the the market is so huge across the world and every country is very [19:02] different and and and you'll see all these players having their base in each countries What about [19:09] drone delivery? Is that something that you drone delivery, robotic delivery, that sort of thing? [19:13] Are you investing heavily in that? Yeah, I mean we we similarly we'll work with the partners. There's [19:17] some there's some really good companies is is trying to build it and there's lot of regulation [19:22] hydros with that. There's a lot of privacy stuff but there are companies who investing [19:27] in it and we we will partner with them. Yeah. I was visiting a friend in Detroit just last week [19:31] and all of a sudden he point out oh there goes the Amazon drone right across my neighborhood. [19:35] Is that a a competitive threat? Yeah, I mean look, Amazon is always going to innovate in this space, [19:41] right? They will always find that the last mile delivery is going to be there. I think it's the [19:45] same thing which is how fast can you deliver uh and and how safely can you deliver whether [19:50] it's drone or or bots that can actually go in the the downtown areas maybe more than drones. [19:57] The bots might play a big role like there's a lot of solution. You saw the robot here which is able [20:01] to climb the steps. That's the problem the pots to solve. Yeah. So, I think this is going to be [20:06] a huge opportunity. All right. Well, at Axios, we like to end our events with one fun thing, [20:12] and I didn't prep for this, so we're going to put you on the spot, which is, do you have the [20:16] Lyft app on your phone? I do. They do. And what do you think? I think it's I think they do a [20:22] good job. I like the pink color. Excellent. All right. so much that thank you so much for being