CSU Chico Builds Edge AI Lab to Prevent Dot-Com-Style Compute Bust

Panel 7 - 4: Design for an Interactive Software Artificial Intelligence Lab for Test Driven Design — Ai4

Ai4

Surging capital investment in centralized AI factories risks repeating the 2000s telecom crash unless engineers deploy physical, interactive edge applications that generate real compute demand.

AI-assisted summary of Panel 7 - 4: Design for an Interactive Software Artificial Intelligence Lab for Test Driven Design — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 14, 2026, this presentation features Professor Sam Siewert from California State University, Chico introducing the interactive Software Artificial Intelligence Lab (i-SAIL). The talk outlines how the lab's facilities and curriculum support test-driven design, sensor fusion, and hands-on edge AI research across newly established graduate and undergraduate degree tracks.

## TL;DR Professor Sam Siewert details the architecture and motivation behind CSU Chico's i-SAIL facility, which bridges high-end cloud model training with physical edge deployment. Drawing parallels to the telecom crash of the early 2000s, Siewert argues that sustaining today's AI compute investments requires training a workforce focused on practical, physical applications rather than speculative infrastructure alone. He showcases an array of student and collaborative projects—including SLAM mobile robotics, acoustic sensor networks to prevent elephant poaching, automated roadway maintenance, and baseball vision analytics—while advocating for test-driven physical design and cost-effective industry research partnerships.

## Atmosphere & style The presentation has a pragmatic, engaging, and seasoned academic tone, blending dry humor with deep domain expertise in aerospace and embedded systems. Siewert's narrative transitions smoothly between cautionary historical reflections on tech bubbles and enthusiastic celebrations of student ingenuity in the lab. The vibe is down-to-earth and encouraging, highlighting the gritty, hands-on reality of physical hardware development over polished marketing hype.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) CSU Chico established the i-SAIL laboratory to support dual graduate tracks in AI Science and AI Engineering, bridging physical hardware and software development. - [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) Capital redirection into large-scale AI factory infrastructure has contributed to a short-term contraction in entry-level computer science job postings. - [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) Current investments in AI infrastructure mirror the late-1990s telecom buildout, necessitating real-world interactive applications to avoid a similar collapse in compute demand. - [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) Interdisciplinary collaboration with civil engineering applies edge computer vision to scan, map, and dispatch automated pothole repair vehicles. - [08:10](https://www.youtube.com/watch?v=-ylzCcnatI4&t=490s) Student projects successfully bridge the gap between high-end cluster training and physical edge deployment on platforms like the NVIDIA Jetson Nano. - [08:10](https://www.youtube.com/watch?v=-ylzCcnatI4&t=490s) Over a three-year span leading up to the lab's creation, student and faculty efforts produced 16 research publications and extensive open-source codebases. - [11:42](https://www.youtube.com/watch?v=-ylzCcnatI4&t=702s) Participating in competitive challenges yielded notable results, including a student-developed model outperforming OpenAI's CLIP baseline in a Qualcomm challenge. - [11:42](https://www.youtube.com/watch?v=-ylzCcnatI4&t=702s) Early recruitment of sophomore undergraduates enables students to construct applied AI and robotics systems before completing advanced theoretical prerequisites. - [13:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=784s) A six-year collaboration with Cornell University applies edge acoustic sensor networks and AI classification to reduce illegal elephant poaching in Africa. - [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) Physical edge AI requires a dedicated test-driven design methodology and isolated infrastructure, such as private 5G networks, to bridge the sim-to-real gap. - [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) Traditional desktop workstations have become obsolete, replaced by a dual paradigm of centralized cloud AI factories and distributed edge devices. - [17:00](https://www.youtube.com/watch?v=-ylzCcnatI4&t=1020s) Low-overhead industry sponsorships and campus-based remote internships provide companies with cost-effective applied AI research while mentoring graduate and undergraduate students.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) Launching i-SAIL and Bridging the AI Systems Gap

Sam Siewert introduces himself—noting that while some students call him "Dr. Sam," the audience is welcome to just call him Sam—and explains that he is presenting the culmination of a year-long initiative to establish a dedicated, interactive laboratory space. Supported by modest seed funding, the project was created to address a perceived gap in AI education and provide hands-on infrastructure for California State University, Chico's newly launched graduate programs: a Master’s in AI Science and a Master’s in AI Engineering. Modeled directly after the classic distinction between Computer Science and Computer Engineering, the curriculum spans both software and hardware. Rather than treating AI purely as an abstract mathematical discipline, the lab centers heavily on physical systems and robotics, including real-time simultaneous localization and mapping (SLAM) navigation carried out by mobile robots mapping the university's engineering center.

```mermaid flowchart TD accTitle: Edge AI Pipeline from Training to Physical Deployment subgraph Cloud["Cloud Infrastructure & Training"] A["AI Factories / HPC"] --> B["High-End Acceleration (DGX / Thor / Spark)"] end

subgraph Lab["i-SAIL Development & Validation"] B --> C["Model Optimization & Pruning"] C --> D["Sensor Fusion & Simulation"] end

subgraph Edge["Physical Edge Applications"] D --> E["Autonomous Systems (Jetson Nano / Mobile Robots)"] D --> F["Infrastructure Maintenance (Roadway Pothole Repair)"] D --> G["Interactive Sports & Recreation (Visual Odometry & Coaching)"] end ```

Addressing the current economic landscape, Siewert highlights a paradox troubling today's students: despite widespread enthusiasm for AI, employment databases show that job postings across STEM and computer science specifically have trended downward over recent years. While the market is beginning to show early signs of recovery, the dip has generated real anxiety among students. Siewert observes that corporate capital is currently being redirected into massive "AI factory" hardware infrastructure, leading to short-term hiring cutbacks. To ensure graduates remain competitive and resilient, universities must modernize their degree offerings and establish physical development spaces where students, faculty, and industry partners can collaborate on practical AI implementations. Siewert draws on his past experience doing exactly this 26 years prior at the University of Colorado Boulder, where he created a Master’s in Embedded Systems Engineering within the Department of Electrical and Computer Engineering despite initial skepticism, turning it into one of the department's most enduring programs.

- CSU Chico offers dual graduate tracks: a Master’s in AI Science and a Master’s in AI Engineering. - SLAM navigation research in the lab allows autonomous mobile robots to construct internal maps of the engineering building in real time. - Heavy capital investment in AI factory hardware has contributed to a temporary contraction in entry-level computer science hiring.

---

[03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) Lessons from the Dot-Com Era and the AI Winters

Reflecting on the history of technology hype cycles, Siewert recounts his own journey as a graduate student during the dot-com boom and the onset of the second AI winter around the year 2000. While funded by a NASA Space Grant fellowship at CU Boulder and collaborating with the Jet Propulsion Laboratory (JPL) in Los Angeles, he frequently encountered deep skepticism from mission engineers whose prevailing sentiment was "don't mess up my spacecraft." He humorously recalls presenting in packed lecture rooms where audience members would vigorously shake their heads in disapproval throughout his talk. When the subsequent AI winter set in, the stigma was severe enough that recent graduates had to scrub the term "AI" from their resumes and thesis titles to find employment—though Siewert kept the word "agent" in his dissertation title, a choice that has aged remarkably well.

Siewert draws a direct historical parallel between today's AI buildout and the massive overbuilding of telecom and datacom fiber-optic infrastructure in the late 1990s. At the time, vast bandwidth was deployed without existing high-volume consumer use cases; streaming platforms like Netflix did not yet exist, ultimately triggering the telecom crash and forcing Siewert to transition from networking into High-Performance Computing (HPC). To avoid a similar crash in modern AI factories, the industry urgently needs a specialized workforce capable of inventing and deploying practical, interactive applications that generate genuine compute demand. While mainstream attention remains fixated on large language models (LLMs) and token capacity, Siewert emphasizes physical AI, sensor fusion, computer vision, and human-machine teaming. A prime example is CSU Chico’s collaborative work with their civil engineering department, which utilizes computer vision for automated roadway maintenance—envisioning a system where an initial vehicle scans and catalogs potholes so an automated trailing vehicle can patch them directly.

- Siewert conducted his doctoral research with CU Boulder's AI lab and JPL under a NASA Space Grant fellowship. - The early-2000s telecom crash was driven by laying high-capacity fiber networks prior to the invention of bandwidth-heavy applications like video streaming. - CSU Chico's civil engineering projects use edge computer vision to scan, map, and dispatch automated roadway pothole repair vehicles.

---

[08:10](https://www.youtube.com/watch?v=-ylzCcnatI4&t=490s) Student Innovations: From Climbing Gyms to Edge Deployment

Over an intensive summer teaching schedule encompassing computer vision, real-time embedded systems, and an online sensor fusion course, Siewert saw students channel theoretical concepts into tangible edge applications. Student creativity produced a surge of unexpected projects, including several computer vision trackers for climbing gyms and outdoor recreation tools such as a visual odometry system designed for smartphones to track hiking routes without relying on satellite navigation. A standout commercial success originated from a year-long sponsored research program that developed an AI-driven baseball coaching assistant by fusing radar telemetry with iPhone visual data. The sponsoring company subsequently hired the lead student developer to manage the resulting commercial product.

To bridge the gap between cloud-scale training and real-world deployment, the lab provides an end-to-end hardware pipeline. Students train heavy models on high-performance compute clusters and enterprise accelerators before distilling, optimizing, and deploying them onto compact edge hardware such as the NVIDIA Jetson Nano for integration into advanced driver-assistance systems (ADAS) and autonomous vehicles. Siewert emphasizes that establishing a physical "sense of place" fosters immense student dedication, noting that in past embedded systems labs, students frequently lost track of time and skipped meals. In one memorable instance, a student fainted from exhaustion; Siewert caught her before she hit the floor, prompting her grateful family to gift him a hand-carved wooden elephant. Over the past three years leading up to and establishing the lab, the program has yielded 16 research publications, numerous supported grants and contracts, and extensive GitHub codebases that serve as public portfolios for students entering the tech workforce.

- Student projects include smartphone visual odometry for hiking and combined radar/iPhone computer vision for baseball coaching analytics. - Deployment workflows target edge devices like the NVIDIA Jetson Nano alongside higher-tier training platforms (DGX, Thor, Spark). - Over a three-year period, student and faculty efforts have generated 16 academic publications and multiple grant-funded software deliverables.

[11:42](https://www.youtube.com/watch?v=-ylzCcnatI4&t=702s) Student Competitions and Early Undergraduate AI Engagement

The interactive lab model emphasizes direct, hands-on competitive challenges and early student exposure. During the spring semester, students were given the option to participate in an AI competition in lieu of sitting for a traditional midterm exam. While only five out of thirty students chose the contest route due to the immense workload involved, the results were remarkable: one student outperformed OpenAI's baseline model for CLIP in a Qualcomm-sponsored contest, alongside several other students who placed favorably.

Beyond upper-division coursework, the lab intentionally recruits younger students early in their academic journey. Each summer, a cohort of roughly ten sophomore undergraduates joins the lab to construct applied AI projects. Even if these students lack advanced foundational prerequisites like advanced circuit design or higher-level programming, they aggressively leverage modern AI developer tools to build real-world systems. Introducing them to practical computer vision, robotics, and edge computing early builds sustained momentum for junior, senior, and graduate-level research.

- **Contest vs. Exam**: 5 of 30 students opted to replace their midterm with the Qualcomm-sponsored AI challenge. - **Milestone Performance**: A student model successfully surpassed the OpenAI CLIP baseline. - **Early Talent Pipeline**: Around 10 sophomore undergraduates are onboarded each summer to build applied AI systems before entering advanced coursework.

---

[13:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=784s) AI for Science: Acoustic Sensor Networks to Prevent Elephant Poaching

A premier showcase of the lab’s applied AI capability is a six-year research initiative aimed at preventing illegal elephant poaching in Africa. Conducted as a direct data collaboration with Cornell University—originating after an open competition failed to produce actionable solutions—the project applies edge-based sensor networks and machine learning models to acoustic monitoring in Central Africa's Dzanga National Park. The speaker highlights master's student Riley Den in the audience, noting that this caliber of applied research motivates students to pursue doctoral programs.

```mermaid flowchart TD accTitle: Elephant Poaching Prevention AI Architecture subgraph DataCollection["Field Acoustic Sensors"] A["Acoustic Sensor Network (Dzanga National Park)"] --> B["Edge Sensor Processing"] end subgraph ModelProcessing["i-SAIL AI Pipeline"] B --> C["Cornell Data Collaboration"] C --> D["Machine Learning Models & Acoustic Classification"] end subgraph Validation["Deployment & Simulation"] D --> E["Simulation at Scale (Quantified Poaching Reduction)"] E --> F["Field Edge Deployment"] end ```

The anti-poaching system originated unexpectedly from prior research conducted for federal agencies. After seeing the lab's acoustic and tracking solutions for the Federal Aviation Administration (FAA) and the U.S. Coast Guard, collaborators asked if the same algorithmic tracking principles could be adapted to protect wildlife. Extensive simulation models now demonstrate that deploying these acoustic AI sensor networks at scale produces measurable reductions in poaching activities.

---

[14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) Test-Driven Design and Physical Edge Lab Infrastructure

Translating algorithmic models into reliable physical systems requires an interactive "test-driven design" methodology tailored specifically for AI and machine learning. Just as rigorous unit testing transformed software engineering, physical AI requires iterative, real-world hardware verification to bridge the sim-to-real gap, validate digital twins, and test reinforcement learning models. While conventional enterprise IT often relies solely on virtual cloud infrastructure, edge AI demands physical facilities equipped with dedicated local connectivity, such as an independent private 5G network isolated from standard university IT.

```mermaid flowchart LR accTitle: Evolution of AI Computing Paradigm subgraph Legacy["Legacy Computing"] A["Desktop Towers & Workstations (Dinosaurs)"] end subgraph Modern["Modern Edge-to-Cloud Paradigm"] B["Cloud Data Centers / HPC"] <--> C["Mini AI Factories (NVIDIA DGX Spark)"] C <--> D["Physical Edge Devices (Robotics, Sensors, 5G)"] end ```

Addressing the rapidly shifting technological landscape, the speaker shares a photo of his high school freshman son standing beside a Tesla humanoid robot to illustrate the difficulty of predicting the future workforce. Preparing students for this shifting paradigm requires hands-on familiarity with compact supercomputing hardware and mini "AI factories"—such as the NVIDIA DGX Spark—alongside embedded robotics. Traditional desktop towers have become obsolete "dinosaurs"; modern development instead demands fluid, bidirectional interaction between centralized cloud models and distributed edge hardware.

---

[17:00](https://www.youtube.com/watch?v=-ylzCcnatI4&t=1020s) Industry Partnerships, Remote Internships, and Academic Roadmaps

To sustain and scale the i-SAIL initiative, industry partnerships are structured around low-overhead research sponsorships and cost-effective applied internships. Rather than sending students to high-cost metropolitan centers like Silicon Valley or Chicago, companies can sponsor bounded, IP-safe summer projects executed directly on campus. Under the mentorship of university faculty, cohorts of ten undergraduates and six graduate students deliver production-grade applied AI solutions remotely while keeping corporate costs low.

- **Lab Sponsorship**: Industry donors can sponsor a dedicated workbench in the i-SAIL facility for $2,000. - **Summer Research Model**: Mentorship programs support 10 undergraduate and 6 graduate researchers per summer on scoped industry projects. - **Space Exploration Project**: Developing an initiative with NASA and an R1 research university utilizing AI computer vision to identify prospective lunar landing sites. - **Curricular Integration**: Continual rollout of updated coursework in computer vision, agentic systems, and edge ML to lay the theoretical foundation for applied lab projects.

## Notable quotes > [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) don't mess up my spacecraft — Sam Siewert

> Sense of place for research with robotics, computer vision, sensor networks and sensor fusion, simulation, and interactive testing of AI applications is critical for success even as more services are pushed up into the cloud and work becomes more virtual due to interest in physical AI. — Sam Siewert

> Enabling test driven AI edge design has proven to be transformative for new graduate and undergraduate AI curriculum and new partnerships with industry, government, and non-governmental organizations benefiting students and faculty be allowing quick “what if” testing of ideas that the cloud alone cannot fulfill. — Sam Siewert

> While LLMs (Large Language Models) have reignited intense interest in AI and inference, the advent of agentic AI and the pursuit of more general AI, or AGI (Artificial General Intelligence) is a long-term vision driving human-machine teaming for emergent AI applications. — Sam Siewert

> [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) Dr. Sam — Sam Siewert

> [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) agent — Sam Siewert

> [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) dinosaurs — Sam Siewert

> [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) test-driven design — Sam Siewert

## Mentioned Sam Siewert, California State University, Chico, University of Colorado Boulder, NASA Space Grant, Jet Propulsion Laboratory (JPL), Netflix, NVIDIA Jetson Nano, NVIDIA DGX Spark, NVIDIA Thor, OpenAI CLIP, Qualcomm, Cornell University, Dzanga National Park, Federal Aviation Administration (FAA), U.S. Coast Guard, Tesla, Riley Den, i-SAIL

## Video description Sam Siewert, Professor, California State University Chico

Motivation for student, industry and faculty collaboration for software AI (Artificial Intelligence) innovation is enhanced by building facilities and concentrating resources required to enable test driven design and development. Sense of place for research with robotics, computer vision, sensor networks and sensor fusion, simulation, and interactive testing of AI applications is critical for success even as more services are pushed up into the cloud and work becomes more virtual due to interest in physical AI. While LLMs (Large Language Models) have reignited intense interest in AI and inference, the advent of agentic AI and the pursuit of more general AI, or AGI (Artificial General Intelligence) is a long-term vision driving human-machine teaming for emergent AI applications. For example, the i-SAIL (interactive Software AI Lab) we have created is attractive for research projects such as AI coaching assistants using sensor fusion, sensor networks to prevent illegal animal poaching, use of AI computer vision methods for Lunar and Mars prospecting as well as computer vision for roadway maintenance automation and 3D printing monitors and scanners. This paper details the design for the i-SAIL at California State University Chico and early projects that are helping to define resources required for AI at the edge. Following an organic approach, we have used modest seed funding to start the i-SAIL and intend to grow it through industry sponsors, grants, and services for courses that require research project space and tools. Enabling test driven AI edge design has proven to be transformative for new graduate and undergraduate AI curriculum and new partnerships with industry, government, and non-governmental organizations benefiting students and faculty be allowing quick “what if” testing of ideas that the cloud alone cannot fulfill.

[Source](https://www.youtube.com/watch?v=-ylzCcnatI4)

中文简报

CSU Chico打造边缘AI实验室以防互联网式算力崩盘对集中式AI工厂激增的资本投资恐重蹈2000年代电信业崩盘的覆辙，除非工程师部署能产生真实算力需求的实体交互式边缘应用。 Sam Siewert教授详细介绍了CSU Chico旗下i-SAIL设施的架构与建设初衷，该设施将高端云端模型训练与实体边缘部署相连接。Siewert将其与2000年代初期的电信业崩盘相提并论，指出要维持当前的AI算力投资，需要培养专注于实用实体应用而非仅着眼于投机性基础设施的专业人才。他展示了一系列学生及合作项目——包括SLAM移动机器人、防大象盗猎的声音传感器网络、自动化道路养护以及棒球视觉分析，同时倡导测试驱动的物理设计与高性价比的产业研究合作。加州州立大学奇科分校（CSU Chico）建立了交互式软件人工智能实验室（i-SAIL），旨在支持“AI 科学”与“AI 工程”双硕士培养方向，搭建连接物理硬件与边缘软件开发的桥梁。 (0:04) 当前资本大规模涌入“AI 工厂”基础设施导致了初级计算机招聘的短期收缩；这种繁荣酷似 20 世纪 90 年代末的电信建设泡沫，亟需发明可落地的物理与交互式应用来支撑真实的算力需求。 (3:36) 实验室打通了从高性能集群训练到 NVIDIA Jetson Nano 等微型边缘设备部署的落地流程，相关师生在过去三年内已产出 16 篇学术论文及多项开源代码库。 (8:10) 在实践与竞赛驱动机制下，该校学生在高通 AI 挑战赛中构建的模型超越了 OpenAI 的 CLIP 基线水平，实验室同时通过早期招募大二学生构建实战型人才培养梯队。 (11:42)

Reading summary

Sam Siewert argues that to prevent a dot-com-style infrastructure collapse and counter contractions in entry-level tech hiring, universities must shift AI education toward physical, edge-based systems validated through interactive, test-driven laboratories. Navigating the AI Infrastructure and Hiring Paradox (3:36) Siewert highlights an economic paradox: while corporate capital heavily funds massive 'AI factory' hardware, entry-level computer science and STEM job postings have trended downward. Drawing direct parallels to the late-1990s telecom crash—where extensive fiber-optic infrastructure was laid without high-bandwidth consumer applications like video streaming—he warns that current overinvestment in large language models and raw compute capacity risks a similar downturn if practical demand lags. Recalling his experience as a graduate student during the onset of the second AI winter in 2000, Siewert notes that graduates back then were forced to scrub 'AI' from their resumes. To avoid another cycle of disillusionment, he argues that the industry urgently requires engineers skilled in physical AI, computer vision, and sensor fusion who can create practical edge applications that generate authentic compute demand. Closing the Sim-to-Real Gap with Edge Pipelines (14:56) Translating algorithmic research into physical hardware requires an interactive test-driven design methodology to bridge the sim-to-real gap, validate digital twins, and test reinforcement learning models. Siewert asserts that traditional desktop workstations are obsolete dinosaurs. Modern AI workflows instead require dynamic coordination between centralized cloud accelerators and compact edge devices, supported by physical infrastructure such as dedicated private 5G networks. To support this workflow at CSU Chico's interactive Software Artificial Intelligence Lab (i-SAIL), students train complex models on high-performance compute clusters and mini AI factories (such as the NVIDIA DGX Spark) before pruning and optimizing them for deployment on embedded hardware like the NVIDIA Jetson Nano. Over three years, this lab framework has produced 16 academic publications, open-source code repositories, and applied software suites across dual graduate tracks in AI Science and AI Engineering. Field Applications from Infrastructure to Conservation (13:04) The lab's research prioritizes applied, cross-disciplinary physical systems over purely theoretical software. In municipal infrastructure, a partnership with civil engineering uses edge computer vision to scan and catalog roadway potholes so automated trailing vehicles can patch them. In consumer sports, students built an AI baseball coaching assistant fusing radar telemetry with iPhone camera feeds, leading the corporate sponsor to hire the lead student developer to manage the commercial product. In conservation technology, the lab maintains a six-year collaboration with Cornell University deploying acoustic sensor networks in Central Africa's Dzanga National Park to combat elephant poaching. Adapted from tracking solutions originally built for the Federal Aviation Administration and the U.S. Coast Guard, extensive simulation models demonstrate that scaling these edge-deployed acoustic AI networks produces measurable reductions in illegal poaching activity. Early Talent Pipelines and Remote Industry Partnerships (17:00) To build an early pipeline, the lab introduces cohorts of roughly ten sophomore undergraduates each summer to hands-on computer vision and robotics, allowing them to leverage modern AI tools before taking advanced prerequisites. Coursework also incorporates competitive benchmarks; five students opted to replace a standard midterm with a Qualcomm-sponsored contest, where one student successfully outperformed OpenAI's baseline model for CLIP. To finance and scale this training sustainably, Siewert advocates low-overhead industry collaborations where companies sponsor dedicated lab workbenches for $2,000 or fund bounded summer research cohorts (ten undergraduates and six graduate researchers) directly on campus. This remote model delivers production-grade solutions—such as an initiative with NASA to identify lunar landing sites—without requiring students to relocate to expensive metropolitan hubs. Siewert concludes that the long-term viability of AI hardware investments depends on training engineers who can ground algorithmic models in the physical world. However, scaling these programs requires institutions to secure external lab sponsorships, manage unpredictable shifts in robotic platforms, and overcome legacy departmental silos.

精炼阅读摘要

面对企业资本大规模涌入人工智能算力基建而初级计算机岗位招聘短期收缩的现状，加州州立大学奇科分校教授Sam Siewert指出，高校必须突破纯数学算法教学的局限，建立连接云端超算与物理边缘部署的交互式软件AI实验室（i-SAIL）。Siewert主张以测试驱动设计理念贯穿物理AI研发，通过培养兼顾软硬件集成的工程人才并催生真实算力需求，以防行业重蹈2000年代电信光纤过剩与AI寒冬的历史覆辙。泡沫周期警示与软硬件双轨学位改革 (3:36) Siewert结合自身在2000年互联网泡沫破裂及第二轮AI寒冬期间攻读博士并与NASA喷气推进实验室（JPL）合作的经历指出，当前AI资本高度集中于“AI工厂”硬件建设，导致计算机科学就业岗位出现阶段性缩减。他将现状与1990年代末期电信行业盲目铺设光纤网络进行了类比：当时由于流媒体等高带宽消费级应用尚未出现，超前建设最终引爆电信危机；如今的AI行业若要避免类似硬着陆，迫切需要能开发实用交互应用、创造真实计算需求的复合型人才。为应对这一挑战，加州州立大学奇科分校参照计算机科学（CS）与计算机工程（CE）的经典划分，同步推出了AI科学硕士与AI工程硕士双轨学位。该架构旨在打破将AI仅视为抽象理论的教学传统，依托i-SAIL实验室构建软硬件一体化的实践环境，引导学生与产业伙伴共同开展基于物理系统与人机协同的工程开发。端云协同与测试驱动的物理AI实验室 (14:56) 实验室将传统软件工程中的测试驱动开发（TDD）方法延伸至物理AI领域。Siewert指出，物理AI系统必须通过高频度的真实硬件迭代测试，才能有效弥合仿真到现实的差距（sim-to-real gap）、验证数字孪生模型及强化学习算法。为了摆脱传统企业IT虚拟化环境的约束，实验室配备了独立于校内常规IT的私有5G网络，为移动机器人和传感器网络提供物理连接保障。在硬件计算架构上，实验室淘汰了过时的传统台式机工作站，建立了中央云端集群与微型AI工厂（如NVIDIA DGX Spark）、边缘计算设备（如NVIDIA Jetson Nano）双向互通的流水线。学生在高性能算力平台上完成大规模模型训练后，通过模型剪枝与蒸馏优化，将轻量化算法部署至无人车高级辅助驾驶系统（ADAS）及各类边缘设备中进行端到端闭环验证。聚焦边缘部署的学生创新与竞赛成果 (8:10) 在密集型课程与实验室孵化机制下，学生将计算机视觉与传感器融合技术应用于多样化的边缘场景。典型成果包括用于攀岩馆的视觉追踪器、无需卫星导航即可在智能手机端运行的徒步视觉里程计，以及与土木工程系合作的道路坑洼自动化巡检与修补系统——该系统通过先导车辆识别标记路面破损，进而调度后导无人车实施自动化修补。此外，一项融合雷达遥测与iPhone视觉数据的棒球AI辅助执教系统成功完成商业化，主导该项目的学生已被资助企业直接录用为主管。实验室亦鼓励学生通过竞赛替代传统考核并吸收早期本科人才。在春季学期的一项高通赞助竞赛中，5名自愿放弃传统期中考试的学生取得了优异成绩，其中一名学生构建的模型性能超过了OpenAI的CLIP基线。实验室每年夏季招募约10名大学二年级本科生，引导其在尚未修习高级电路或深奥编程课程前借助现代AI开发工具上手实际项目。过去三年内，实验室累计产出16篇学术论文、多项开源软件套件及广泛的GitHub代码库。野生动物声学反盗猎与低成本校企合作 (13:04) 在重大科研应用层面，实验室与康奈尔大学展开了为期六年的数据合作，将此前为美国联邦航空管理局（FAA）和美国海岸警卫队研发的声学追踪算法，迁移至中非共和国赞加国家公园（Dzanga National Park）的非法大象盗猎监测。模拟研究表明，在野外部署这套基于边缘计算与机器学习分类的声学传感网络后，能够量化降低盗猎行为的发生率。目前，实验室还与NASA及一所R1顶尖研究型大学推进合作，利用AI计算机视觉筛选月球潜在着陆点。为确保实验室长期运营的可持续性，i-SAIL推行了一种低开销的校企合作模式。企业无需将学生输送至生活成本高昂的硅谷或芝加哥，只需出资2,000美元即可在实验室赞助专属工作台，或以低预算定向资助夏季研发课题。在高校教师指导下，由10名本科生与6名研究生组成的科研团队在校内远程交付具备生产级标准、知识产权边界清晰的AI工程解决方案。 Siewert强调，应对生成式AI基础设施建设过热与初级就业市场波动的关键，在于建立将边缘硬件、物理验证与产业实际需求紧密结合的应用型工程教育体系。然而，当前AI工厂巨额资本投入与终端实际应用价值产出之间的脱节现象能否在短期内消除，仍取决于行业能否加速孵化出足够多具备商业盈利能力的物理AI落地场景。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2026 年 8 月 14 日发布在 YouTube 平台上，主讲人为加州州立大学奇科分校的 Sam Siewert 教授。视频内容主要介绍了该校创立的交互式软件人工智能实验室（i-SAIL），重点阐述如何通过测试驱动设计（TDD）、物理 AI（Physical AI）与多传感器融合技术，推动高校师生与产业界的边缘 AI 产学研协同创新。 ## TL;DR 加州州立大学奇科分校的 Sam Siewert 教授介绍了专注于物理 AI 与边缘计算的 i-SAIL 实验室的构建背景与落地实践。针对当前计算机科学就业紧缩与产业界巨额 AI 算力投资之间的脱节，实验室绕开单纯的文本大语言模型，聚焦移动机器人、传感器融合、市政道路自动维护以及非洲大象防偷猎等真实物理交互场景。通过推行面向 AI 的测试驱动设计（TDD）、独立 5G 实验网络和早期本科生科研机制，该实验室打通了从云端训练到边缘端落地的全链路，并探索出低成本、高回报的校企协同孵化新模式。 ## Atmosphere & style 演讲整体基调务实、幽默且极具历史纵深感。主讲人以亲历互联网泡沫破裂与第二次 AI 寒冬的技术老兵视角，既风趣地分享了与大模型“斗智斗勇”的日常，又真诚地剖析了当下学生的就业困境与技术狂热。整场报告兼具前沿探索的极客激情与对实体实验室“场所感”的人文温情，在娓娓道来中展现出脚踏实地将 AI 落地于物理世界的教育情怀。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) 加州州立大学奇科分校创立 i-SAIL（交互式软件人工智能实验室）并设立 AI 科学与工程双硕士项目，致力于解决工程教育与产业应用脱节的痛点，重点聚焦软硬件结合与移动机器人技术。 - [01:13](https://www.youtube.com/watch?v=-ylzCcnatI4&t=73s) 面对产业界巨额投资“AI 工厂”算力基建而初级就业岗位收紧的矛盾，高等教育需建立人文与技术交融的实体研发空间，强化学生面对面的系统工程开发能力以对抗就业周期。 - [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) 结合亲历 2000 年互联网泡沫与第二次 AI 寒冬的历史教训，技术演进呈现螺旋上升规律，大学的核心使命在于为学生构建跨越技术寒冬的长期抗风险能力。 - [05:47](https://www.youtube.com/watch?v=-ylzCcnatI4&t=347s) 避开单纯的 LLM 文本生成热潮，实验室全面转向物理 AI（Physical AI）、多传感器融合与人机协同，探索如市政道路自动化巡检修补等能产生实体经济闭环的落地应用。 - [08:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=484s) 打通“云端训练到边缘落地”的工程链路，将大模型压缩优化部署至低功耗嵌入式设备，并通过实体实验室提供的“场所感”凝聚学生的沉浸式科研热情。 - [11:42](https://www.youtube.com/watch?v=-ylzCcnatI4&t=702s) 采取积极的早期科研培养策略，从大二阶段吸纳本科生并引导其直接运用前沿 AI 工具研发原型，甚至以高水平 AI 算法竞赛替代传统期中考试以激发实战能力。 - [13:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=784s) 践行“科学人工智能”（AI for Science），将原用于 FAA 与海岸警卫队的目标追踪技术跨界迁移至非洲国家公园的声学防偷猎系统，实现显著的生态保护实效。 - [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) 推行面向 AI 的测试驱动设计（TDD），搭建独立 5G 专网与“微型 AI 工厂”，为强化学习、数字孪生及虚实迁移（Sim-to-Real）提供快速的闭环验证环境。 - [17:00](https://www.youtube.com/watch?v=-ylzCcnatI4&t=1020s) 创新推出校内异地实习与低门槛工位赞助模式，降低企业差旅成本与学生异地生活负担，同时拓展与 NASA 及顶尖高校在月球智能着陆选址等尖端课题上的合作。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) 创立 i-SAIL 实验室与加州州立大学奇科分校的 AI 项目主讲人 Sam Siewert（学生常称他为 Dr. Sam）开场介绍了他在过去一年中推动的新项目。依托少量的启动资金，他在加州州立大学奇科分校（California State University, Chico）建立了一个专注于交互式软件人工智能的实验室（i-SAIL），旨在填补当前工程教育与产业应用之间的脱节，并为学校新设立的人工智能硕士项目提供实践支持。奇科分校参照计算机科学（CS）与计算机工程（CE）的经典划分模式，同步开设了人工智能理学硕士（Master's in AI Science）与人工智能工程硕士（Master's in AI Engineering）两个专业。该实验室定位于软硬件结合与系统级开发，尤其聚焦于移动机器人技术。 Sam 展示了学生在工程中心利用移动机器人进行 SLAM（同步定位与建图）导航的实景图。面对即将到来的会场招待酒会（Happy Hour），他幽默地表示不会讲太多枯燥的技术细节，并打趣道：“SLAM 简单来说就是边搞清楚自己在哪里、边把周围的地图建出来——正如各位参会者今天在展馆里做的一样。”这套实验室平台的核心，正是让学生在真实的物理与系统环境中直观理解并验证这些算法。 --- ### [01:13](https://www.youtube.com/watch?v=-ylzCcnatI4&t=73s) 计算机科学就业低谷与“AI 工厂”硬件投资的脱节 Sam 剖析了促使该实验室成立的核心痛点：当前学生群体中普遍蔓延的就业焦虑。尽管各方对宏观经济数据的解读不一，但招聘平台的数据清晰地显示，计算机科学及整个 STEM 领域的初级岗位招聘需求出现了明显下滑。虽然招聘量近期已有轻微回升，Sam 个人也坚信未来会强劲反弹，但这种就业紧缩感在学生中间是切实存在的。与此同时，产业界的资本正以前所未有的规模涌入以算力硬件为核心的“AI 工厂”（AI Factory）基础设施建设，许多企业为了采购高昂的硬件甚至不惜缩减人力开支。这种“重基建投资、轻应届招聘”的现状，倒逼高等教育必须求变，提供更加多元化的课程与培养路径。奇科分校正是通过落地 AI 硕士学位与强化课程体系来应对这一挑战。Sam 强调，建设一个人文与技术交融的实体研发空间至关重要——在这里，学生、产业界伙伴与教师能够面对面共同开发 AI 应用。这种产学研协作模式并非纸上谈兵，早在 26 年前，Sam 就在科罗拉多大学博尔德分校（CU Boulder）创立了嵌入式系统工程硕士项目；尽管当时许多人断言该方向没有市场需求，如今它已成为该校电气与计算机工程系最成功的学位项目之一。 ```mermaid flowchart TD accTitle: 从云端算力基础设施到边缘端物理AI应用的研发生态 subgraph Infrastructure["基础设施层（AI 工厂 / 云端）"] A["高密度算力中心 / HPC"] B["大模型与通用算法训练"] end subgraph iSAIL["i-SAIL 交互式实验室（桥梁与转化）"] C["产学研协作空间（学生 / 教师 / 产业导师）"] D["测试驱动边缘设计与模型压缩 (DGX / Thor)"] end subgraph PhysicalEdge["物理边缘端落地（应用与需求牵引）"] E["移动机器人 / 自动驾驶 (SLAM / ADAS)"] F["多传感器融合应用 (雷达 + 视觉)"] G["市政与道路自动化巡检系统"] end A --> C B --> C C --> D D --> E D --> F D --> G ``` --- ### [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) 历史的回声：经历互联网泡沫、AI 寒冬与 NASA JPL 的往事面对当下的 AI 狂热，Sam 分享了他与 ChatGPT “斗智斗勇”的日常乐趣：他喜欢在对话中先持有一种立场诱导模型附和，随后突然反转立场看模型是否再次顺从，最后再戏谑地告诉模型“哈哈，我骗你的，我其实坚持最初的观点”。这种对技术热潮的冷静审视，源于 Sam 亲身经历过 2000 年前后的互联网泡沫破裂（Dot-com Bust）与第二次“AI 寒冬”。当年，他通过 NASA 太空资助奖学金项目攻读博士学位，往返于洛杉矶的喷气推进实验室（NASA JPL）与科罗拉多大学博尔德分校的 AI 实验室之间。在那个时期，JPL 的航天工程师对 AI 抱有极大的排斥与戒备心理，普遍态度是“别用你的 AI 把我的航天器搞砸了”（"don't mess up my spacecraft"）。Sam 回忆道，自己当年在一个座无虚席的大会议室作报告时，台下就有资深工程师一边听一边全程毫不客气地使劲摇头，这给初出茅庐的博士生带来了巨大的心理压力。随着第二次 AI 寒冬接踵而至，AI 概念迅速沦为行业毒药，Sam 在毕业求职时甚至不得不把所有与 AI 相关的字眼从简历中抹去，险些连博士论文标题都做了删改——唯一幸存的是标题中的“Agent”（智能体）一词，如今这一概念重新火遍全球，让他倍感欣慰。Sam 指出，即便行业未来可能迎来第三次 AI 寒冬，技术演进的总趋势依然是螺旋上升的，大学的核心使命是为未来培养具备抗风险能力的人才。 --- ### [05:47](https://www.youtube.com/watch?v=-ylzCcnatI4&t=347s) 摆脱单纯的 LLM：物理 AI、传感器融合与基础设施应用需求 Sam 从 2000 年光纤网络过度建设中汲取了深刻教训：当年全美过度铺设了海量的电信和数据通信基础设施，却因为缺乏杀手级应用而陷入泡沫崩盘——因为在 2000 年，像 Netflix 这样能吃满光纤带宽的视频流媒体服务根本尚未诞生。泡沫破裂后，Sam 原本规划的网络通信职业生涯被迫转向了高性能计算（HPC）。今天的 AI 算力中心同样面临类似的挑战：巨额投资建造的硬件必须找到真正的终端应用场景，才能形成经济闭环，而这需要产业界提供就业岗位来吸纳开发这些应用的工程师。与绝大多数聚焦于大语言模型（LLM）与 Token 生成的讨论不同，Sam 明确表示 i-SAIL 实验室将绕开单纯的文本大模型，全面转向“物理 AI”（Physical AI）、传感器融合、计算机视觉以及“人机协同”（Human-Machine Teaming）。实验室追求的是能够与人类产生真实物理交互的落地工具。 - **土木工程跨学科合作**：奇科分校土木工程系成为了 i-SAIL 最活跃的合作伙伴之一。双方联合争取科研资助，研发道路维护自动化系统：设想由前方巡检车利用计算机视觉自动识别路面坑洼，后方相距 10 英里的工程车则根据数据自动填补沥青修复路面，解决奇科本地严重的道路破损问题。 - **产学研生态闭环**：通过实验室将真实痛点转化为研发课题，由教师担任导师、产业界提供支持，让学生在解决具体工程问题中锻炼系统集成能力。 --- ### [08:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=484s) 边缘端部署、学生创新应用与实验室的“场所感” 为了快速激发学生的创造力，Sam 在夏季高强度讲授了三门课程（两门计算机视觉与一门实时嵌入式系统），并在春季上线了传感器融合课程。高密度的理论输入迅速转化为一系列新颖的学生项目：从攀岩馆视觉分析辅助工具，到户外徒步中的视觉里程计（Visual Odometry）轨迹追踪系统。其中一项历时一年的资助研究项目，将毫米波雷达与 iPhone 视觉数据深度融合打造出“AI 棒球教练助手”，该方案不仅成功转化为商业产品走向市场，主导该项目的学生毕业后直接被资助企业聘为产品经理。这些项目的共同特征是打通了“云端训练”与“边缘端落地”的链路：学生利用高性能计算中心训练模型，在高端边缘开发平台（如 NVIDIA DGX、Thor 架构平台）上验证性能，最终将其精简优化并部署到 Jetson Nano 等嵌入式边缘计算模组中，服务于高级辅助驾驶（ADAS）与自主机器人。 Sam 特别强调了为沉浸式开发建立“场所感”（Sense of Place）的价值。由于软件与 AI 极易沦为纯虚拟的屏幕作业，一个设备完备的实体实验室能够给学生提供无可替代的归属感与创造激情。学生经常废寝忘食地沉浸在调试中，甚至重现了当年嵌入式实验室里学生因熬夜低血糖晕倒的插曲——当时 Sam 在学生倒地前一把将她接住，事后学生的家人特意赠送了一尊手工雕刻的木象以表达感激。在过去三年的筹备与初期运行中，实验室已积累了丰硕的成果： - **学术产出与开源积累**：累计发表学术论文 16 篇，沉淀了大量面向工业界的开源 GitHub 代码仓库，成为学生向潜在雇主展示工程实操能力的核心资产。 - **科研经费与横向课题**：依托加州州立大学奇科分校，成功承接并交付了多项来自企业和政府机构的研发合同与课题资助。 ### [11:42](https://www.youtube.com/watch?v=-ylzCcnatI4&t=702s) 竞赛突破与早期本科生科研培养加州州立大学奇科分校（California State University Chico）的 i-SAIL（交互式软件人工智能实验室）在实际竞赛和科研教学中取得了显著成效。在春季的一场由高通（Qualcomm）赞助的 AI 竞赛中，实验室的一名学生成功击败了 OpenAI 的 CLIP 基线模型，获得了非常亮眼的成绩，其他学生也斩获了诸多名次。演讲者幽默地提到，他当时允许全班 30 名学生自主选择：是参加这场 AI 竞赛还是参加期中考试。最终有 5 名学生选择了竞赛挑战，虽然工作量远超普通期中考，但学生们全力以赴并交出了惊艳的答卷。实验室的核心理念之一在于“尽早吸纳学生参与 AI 研究”。每年夏季，实验室都会招募大约 10 名大学二年级左右的本科生。虽然低年级学生在基础知识储备上尚显不足，但实验室采取了非常积极（Aggressive）的策略——直接引导学生使用前沿 AI 工具来构建 AI 解决方案。相较于电路或入门级编程等容错率较低的基础课程，在高年级应用与项目研发中大胆使用 AI 工具能让学生更早体验端到端开发，激发深厚的科研兴趣。 - 竞赛成果：学生在高通赞助的比赛中超越 OpenAI CLIP 基线。 - 教学创新：提供竞赛替代期中考试的机制（30 人中有 5 人选拔参赛）。 - 人才梯队：每年夏季吸纳约 10 名大二本科生进组，利用 AI 工具直接开展原型设计。 --- ### [13:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=784s) AI 保护野生动物：非洲防偷猎项目与多方合作演讲者特别介绍了现场参会的硕士研究生 Riley Den，并指出实验室已经培养出不少有志于继续攻读博士学位的优秀科研苗子。随后，他详细介绍了实验室持续进行了六年的旗舰项目——利用“科学人工智能”（AI for Science）保护非洲大象、打击非法盗猎。该项目源于与康奈尔大学（Cornell University）的数据合作：康奈尔大学最初发起过一项防偷猎挑战赛但进展有限，随后双方转为深度数据共享与合作。 ```mermaid flowchart TD accTitle: 非洲国家公园 AI 防偷猎项目流程图 accDescription: 展示从以往政府项目迁移到非洲大象保护项目的数据流与端到端部署流程。 subgraph Past["技术来源与经验迁移"] FAA["FAA / 美国海岸警卫队项目"] --> SensorTech["声学传感器网络与目标跟踪"] end subgraph Collab["数据与算法研发"] Cornell["康奈尔大学数据协作"] --> ModelDev["i-SAIL 算法建模与优化"] SensorTech --> ModelDev ModelDev --> Sim["大规模仿真验证 (有效降低偷猎)"] end subgraph Deploy["实地边缘部署"] Sim --> Field["中非共和国 Dzanga 国家公园"] Field --> EdgeAI["边缘检测设备实地布设"] end ``` 这项研究的灵感源于实验室此前为美国联邦航空管理局（FAA）和美国海岸警卫队（Coast Guard）开发的技术，当时有人提出“能否将为空管开发的技术用于保护大象”，看似不可思议的想法最终取得了重大突破。目前该算法已经在中非共和国的赞加国家公园（Dzanga National Park）进行部署准备。大规模仿真结果提供了强有力的证据，证明该系统在实地应用中能显著降低盗猎行为。 --- ### [14:56](https://www.youtube.com/watch?v=-ylzCcnatI4&t=896s) i-SAIL 实验室设计理念与物理基础设施 i-SAIL 的核心理念是**以“交互性”（Interactive）为本**，全面推行面向 AI 和机器学习的“测试驱动设计”（Test-Driven Design, TDD），其逻辑与传统软件工程的测试驱动开发一脉相承。通过将强化学习、数字孪生（Digital Twin）、虚实迁移（Sim-to-Real Gap）等技术与物理机器人和传感器网络结合，学生可以在真实的闭环环境中快速进行“What-if”假设验证。为了摆脱传统大学网络设施的束缚并支持边缘 AI 实验，演讲者积极游说学院院长和后勤部门，成功建立了完全运行在独立专网上的 5G 实验室环境。演讲者展示了自己刚上高一的儿子与 Tesla 人形机器人的合影，坦言面对未来瞬息万变的技术形态，即便作为业内人士也难以准确预测未来的就业格局。传统的台式机和塔式工作站正在成为“恐龙”，取而代之的是由 DGX 等构成的“微型 AI 工厂/盒中超算”与端侧边缘设备的直连架构，而承载这种架构的物理实验室对于培养实战人才不可或缺。 --- ### [17:00](https://www.youtube.com/watch?v=-ylzCcnatI4&t=1020s) 产学研赞助模式与前沿航天项目在报告尾声，演讲者向与会企业发出了务实的合作邀请：企业只需赞助 2,000 美元，即可在实验室的工作台上获得冠名权。针对学生跨州前往硅谷或芝加哥实习所产生的高昂通勤与生活成本，实验室提出了一种新型的“校内异地实习”孵化模式： | 维度 | 传统异地实习模式 | i-SAIL 校内合作实习模式 | |:--- |:--- |:--- | | **企业成本** | 需承担高额差旅、住宿及异地实习薪资补贴 | 极具性价比，依托校内师生团队完成 | | **学生负担** | 面临生活成本压力与异地安顿挑战 | 在校内利用专业实验设备即可开展研发 | | **项目形态** | 受限于企业内部 IP 隔离与合规流程 | 企业拆分独立模块，由教师带队攻坚 | | **指导力量** | 完全依赖企业内部导师日常带教 | 大学教授与研究生联合提供日常指导 | | **协同方式** | 线下集中办公 | 远程视频会议（Zoom）+ 敏捷汇报 | 此外，演讲者还透露实验室正在与一所 R1 顶尖研究型大学及美国国家航空航天局（NASA）合作开展全新的前沿研究，利用计算机视觉和边缘 AI 技术辅助月球着陆点的智能选址。演讲者强调，通过持续完善 AI 新课程体系、结合论文发表与真实科研项目，学生才能真正具备应对未来技术变革的核心竞争力。最后，演讲者向全场致谢，并邀请对大象保护算法感兴趣的参会者与现场的 Riley 交流，共同探讨 i-SAIL 赞助与产学研合作机会。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=4s) SLAM 简单来说就是边搞清楚自己在哪里、边把周围的地图建出来——正如各位参会者今天在展馆里做的一样。 — Sam Siewert > [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) 别用你的 AI 把我的航天器搞砸了 — NASA JPL 航天工程师 > [03:36](https://www.youtube.com/watch?v=-ylzCcnatI4&t=216s) 哈哈，我骗你的，我其实坚持最初的观点 — Sam Siewert > [13:04](https://www.youtube.com/watch?v=-ylzCcnatI4&t=784s) 能否将为空管开发的技术用于保护大象 ## Mentioned Sam Siewert, 加州州立大学奇科分校 (California State University, Chico), i-SAIL (interactive Software AI Lab), 科罗拉多大学博尔德分校 (CU Boulder), NASA, NASA JPL (喷气推进实验室), ChatGPT, Netflix, NVIDIA DGX, NVIDIA Thor, Jetson Nano, Qualcomm (高通), OpenAI CLIP, Riley Den, 康奈尔大学 (Cornell University), 美国联邦航空管理局 (FAA), 美国海岸警卫队 (Coast Guard), 赞加国家公园 (Dzanga National Park), Tesla ## Video description Sam Siewert, Professor, California State University Chico Motivation for student, industry and faculty collaboration for software AI (Artificial Intelligence) innovation is enhanced by building facilities and concentrating resources required to enable test driven design and development. Sense of place for research with robotics, computer vision, sensor networks and sensor fusion, simulation, and interactive testing of AI applications is critical for success even as more services are pushed up into the cloud and work becomes more virtual due to interest in physical AI. While LLMs (Large Language Models) have reignited intense interest in AI and inference, the advent of agentic AI and the pursuit of more general AI, or AGI (Artificial General Intelligence) is a long-term vision driving human-machine teaming for emergent AI applications. For example, the i-SAIL (interactive Software AI Lab) we have created is attractive for research projects such as AI coaching assistants using sensor fusion, sensor networks to prevent illegal animal poaching, use of AI computer vision methods for Lunar and Mars prospecting as well as computer vision for roadway maintenance automation and 3D printing monitors and scanners. This paper details the design for the i-SAIL at California State University Chico and early projects that are helping to define resources required for AI at the edge. Following an organic approach, we have used modest seed funding to start the i-SAIL and intend to grow it through industry sponsors, grants, and services for courses that require research project space and tools. Enabling test driven AI edge design has proven to be transformative for new graduate and undergraduate AI curriculum and new partnerships with industry, government, and non-governmental organizations benefiting students and faculty be allowing quick “what if” testing of ideas that the cloud alone cannot fulfill. [Source](https://www.youtube.com/watch?v=-ylzCcnatI4)

Source quotes

[0:04] Sam Siewert: SLAM 简单来说就是边搞清楚自己在哪里、边把周围的地图建出来——正如各位参会者今天在展馆里做的一样。 [3:36] NASA JPL 航天工程师: 别用你的 AI 把我的航天器搞砸了 [3:36] Sam Siewert: 哈哈，我骗你的，我其实坚持最初的观点 [13:04] 能否将为空管开发的技术用于保护大象

Source quotes

[3:36] Sam Siewert: don't mess up my spacecraft Sam Siewert: Sense of place for research with robotics, computer vision, sensor networks and sensor fusion, simulation, and interactive testing of AI applications is critical for success even as more services are pushed up into the cloud and work becomes more virtual due to interest in physical AI. Sam Siewert: Enabling test driven AI edge design has proven to be transformative for new graduate and undergraduate AI curriculum and new partnerships with industry, government, and non-governmental organizations benefiting students and faculty be allowing quick “what if” testing of ideas that the cloud alone cannot fulfill. Sam Siewert: While LLMs (Large Language Models) have reignited intense interest in AI and inference, the advent of agentic AI and the pursuit of more general AI, or AGI (Artificial General Intelligence) is a long-term vision driving human-machine teaming for emergent AI applications. [0:04] Sam Siewert: Dr. Sam [3:36] Sam Siewert: agent [14:56] Sam Siewert: dinosaurs [14:56] Sam Siewert: test-driven design

Original transcript

[0:04] So, I I am Sam. Uh some of my students call me Dr. Sam. You can call me just Sam. [0:10] Um and I'm here today to talk about a project I've worked on for a year now. And uh I have [0:15] some uh small funding to get a lab going to fill in a gap, a gap that I've perceived um and to also [0:25] give a little boost to our new master's in AI science program. So Chico actually has a [0:31] AI science program, masters as well as a master's AI engineering program kind of following the [0:37] computer science and computer engineering pattern if if that makes sense. And so um this there'll [0:42] be hardware and software here. This is more about systems. Uh we do robotics in this lab. [0:48] The picture on the uh right as you face it is slam navigation of our engineering center using mobile [0:54] robots and things like that. And I'm not going to get super technical today, especially right [0:58] before happy hour. So, uh, slam just means simultaneous localization mapping, figuring [1:04] out where you are and building a map while you're figuring out where you are. You've all been doing [1:08] that all day here at the conference, so you know what I mean. Uh, so we'll get into it now. And [1:13] so I I'm going to explain what the gap is. Uh, and this is something that students perceive. [1:18] I don't know exactly how real this is. I'm not an economist, so you know, I'm not going to get real [1:23] deep in this, but I just picked these because they were readily available. I think I might have even [1:27] used chat DPT to help me find these, but one of the uh graphs comes from an employment database. [1:34] And they're just pointing out that in computer science, it's maybe kind of STEM in general, but [1:39] in particular computer science, uh you know, job postings are down, right? And even if that isn't [1:48] statistically true, our students feel that way, right? So it's you know, it's it's a and uh you [1:54] know, so and that's by year. It's starting to come back up a little bit. And I personally I I don't [2:01] usually act as a pundit, but I personally think it's going to come back up pretty strongly here in [2:05] the future. But I think we also need to realize, for example, in computer science that we need uh [2:11] more diversity of programs that we offer. So, we just launched our AI master's degree, for example, [2:18] and we're strengthening our our coursework and curriculum in AI, and we're uh our goal is to [2:23] help build the workforce of the future in AI. And uh what I'm really here to talk about is a [2:31] development space where students, industry, and faculty meet up to do AI. [2:37] And that sounds kind of simplistic, but uh I've used this basic idea before at another university [2:44] and 26 years ago I started a master's in embedded systems engineering which [2:48] everyone said wasn't needed. It's one of the most successful programs at the University of Colorado [2:54] in the in the department of electrical computer engineering. Okay. So I'm going to go to the next [2:59] slide. I won't go through the rest of that. So the chart on the right is the AI factory investment [3:04] and the uh implication there is that companies are investing in hardware right now in a big way [3:11] and so they have to make some cuts to do that and so you know there's all kinds of perception [3:17] on why that is maybe jobs are being taken etc. I'm not going to get into that but there is a [3:23] there's a measured uptick in investment in AI factories and there is a measured downward trend [3:29] in hiring computer student science students and that's you know obviously uh scary to our students [3:36] okay so uh you know the classic thing is after our other speakers this is [3:41] a little bit dangerous but I had to do it let's ask AI what it thinks about AI [3:46] that's always fun I I enjoy you know debating with uh chat GBT and I use a lot of the suggestions [3:53] made by the prior speakers like you know I'll try like I like one of the games I like to play is see [3:58] if I can take a position and get it to agree with me and then I'll change my position on [4:01] it and see if I can get it to agree with that too and then I say haha I tricked you I actually had [4:06] the original position that's always kind of fun I don't know if you guys do that maybe it's just me [4:10] so uh what does AI have to say about AI well so I lived through the uh year 2000B bomb as it was [4:18] called Right. I just got my PhD and when I went back to graduate school to get my PhD, uh, I went [4:26] through the NASA space grant fellowship program and I worked with the Jet Propulsion Lab in LA, [4:32] but I did my work at University of Colorado Boulder and I worked with the AI lab [4:38] and it was really interesting because a lot of people at the Jet Propulsion Lab [4:42] didn't really like the idea of AI like their attitude was like don't mess up my spacecraft. [4:47] I remember literally giving presentations to a room this size and it was packed and seeing [4:51] someone in the audience just shaking their head no while I was talking like [laughter] you know so [4:56] it was a little intimidating as a new graduate student. Uh then it then the second AI winter set [5:02] in and I literally had to hide that I did AI on my resume as a new graduate. I I just and I almost [5:09] took it completely out of my title my thesis. It actually does have the word agent in it and I'm [5:14] kind of glad I preserved that because that's that's pretty popular now. So, um, okay. So, [5:20] you know, how important is this for preparing the future workforce? Uh, definitely important. So, [5:27] you know, uh, we don't know what's going to happen with AI, but even if there's a third winter, [5:31] even if it's down, it's overall an upward trend, right? So it it we need to prepare students for [5:38] this and we need to uh meet uh in what industry needs in that regard as well. And so uh the key [5:47] is one of the keys I see the one of the ways to avoid a futurebomb kind of incident is [5:53] having lived through that we overbuilt telecom dataccom infrastructure right and we didn't create [6:01] a use case for that Netflix didn't exist in the year 2000 right most of the things that [6:07] use all the bandwidth we have uh on fiber optic networks were non-existent I remember at the time [6:13] working at companies thinking that's going to be a problem, right? And of course it was, right? So, [6:20] um, that caused me I thought after I quit my uh gig with NASA, I would go work on uh networking, [6:28] but instead I wound up working in HPC because of the dot bump. So, [6:32] you need to be resilient and adaptive. And more so, I think we need a workforce that will build [6:37] the apps that will create the demand for this new AI factory infrastructure. and someone's going [6:43] to have to hire them to put it bluntly, right? And my suggestion to everyone from industry is [6:50] we need to do that so that we can use this new AI infrastructure. And I I know there's things like, [6:56] you know, do we have enough tokens and all that kind of thing. Uh most people are here today to [7:00] talk about LLMs. I'm going to kind of ignore them. I'm here to talk about sensors and physical AI [7:07] and computer vision and so forth and and apps that humans interact with. So I'm very big on [7:14] human machine teameing and interactive AI use and my belief that this will create a lot of demand [7:21] for our new AI infrastructure. Uh the images on the bottom are one of our most popular customers [7:28] which is our civil engineering department. uh they just can't get enough of us in this new lab [7:33] in that they you know they have grants to you know they imagine like one truck drives along [7:38] the road and finds all the potholes and then like 10 miles behind is another truck where [7:42] it's automatically fixing the potholes. I can't wait for that. I don't know about you guys [7:48] in Chico we have a lot of potholes. Okay. So you know we need these apps. We can build them [7:54] but we need a lab to do it. That's the bottom line message. And we need to get the students in there. [7:59] We need faculty in there as mentors and we need your help out in industry and we need to, [8:04] you know, work together to make this happen. So, here's some exciting things my students [8:10] did just this summer. For some crazy reason, I decided to teach three classes this summer. [8:15] Two on computer vision and and uh one on real-time embedded systems and then I taught sensor fusion [8:22] for the first time in the spring online. So, I'm a glutton for punishment, but uh [8:27] you know, just presenting students to this theory, these algorithms in classes [8:34] gets them building apps that to me are pretty amazing, right? And they're things that [8:38] don't exist yet. They might not even make sense. They might not ever get anywhere, [8:42] but they're interesting. I I've had an outbreak of climbing gym AI computer vision apps. [8:50] Um, and while this might seem like, do you really need this? We had a year-long uh [8:56] research sponsored research program to create a a AI uh baseball coaching assistant combining radar [9:05] with iPhone visual data and that's being sold as a product. They hired our student. [9:11] He's managing the product. Um, you know, so that a success story. So, you know, [9:16] it's kind of a little bit of a playground, uh, to be honest, but it's a great playground where [9:21] students can can put these all this theory into use. Uh, they're actually working between the [9:28] cloud where the AI factories are and the edge and they're trying to get things out of the cloud into [9:35] the edge. So they're very focused on deploying models and integrating them with systems uh like [9:44] autonomous vehicles ADAS. Some of you may do that here. Uh the the app on the top right is [9:51] uh it's one of my favorites this summer. It's for uh when you go hiking visual odometry and keeps [9:58] track of where you hiked, right? Uh so just neat apps for a phone. Uh and then of course we have [10:04] the climbing uh gym apps. Okay. So uh the key is uh I think AI is kind of virtual. So I'm trying to [10:15] create a sense of place, a place to go to do this, right? So um the point is the students love it. [10:24] I've had this problem before with when I created the embedded systems engineering lab. I literally [10:28] had to kick students out. I had one student faint and I caught her before she hit the floor. [10:34] Her family gave me a handcarved wooden elephant as a thank you for catching their daughter. Um, [10:40] you know, it's because they weren't eating, they weren't sleeping. They get so into this when they [10:44] have the lab with the right stuff so they can go train their models on a high performance computer. [10:50] they could test it on a something like a DGX Spark or a Thor or you know whatever the latest greatest [10:56] thing is and then they could deploy it to Jetson Nano or whatever kind of smaller device they have [11:03] to put into uh a system. So we've had a lot of different sponsors [11:09] so far in our first year and we produced a lot of papers. So this is kind of the numbers slide. [11:15] I won't spend a lot of time on it. It's over actually over the past three years. So we started [11:19] doing this before we had the lab. So you know 16 publications, a lot of software dev projects, [11:26] more GitHubs than I could list there, you know. So GitHubs are great because then [11:31] the students get to showcase what they did and potential employers get to see their code and [11:35] so forth. Um and you know various contracts and grants that we've supported at Cal State Chico. [11:42] And uh this one I'm pretty proud of. So this spring we started entering some AI contests [11:49] and we actually had one student place ahead of open AAI baseline for uh clip. So beating clip [11:55] in a Qualcomm sponsored uh competition. We had other students place and so on and so forth. [12:01] Uh but I thought that was pretty cool. I gave them the opportunity to do the contest instead [12:06] of taking a midterm. I thought they might all do it. Five out of the out of the 30 did it. But [12:11] uh you know they worked really hard on it. is probably more work than the midterm. Um, [12:16] also the idea of this lab is to engage students early. So, every summer I have [12:21] about 10 students that are undergraduate, more like sophomore year, and get them into AI early. [12:28] Um, they're they're not as prepared, but these are all images of things that they built. So, [12:35] um, you know, and we are pretty aggressive on using AI tools to build AI solutions. [12:41] um you know and that's probably less risky than the more foundational classes like circuits or you [12:47] know the first programming class you have. So this is a little further on. These are all like junior, [12:52] senior year or graduate level. Uh so they're getting a lot of lot of things done and starting [12:58] them early gets them interested in building these kinds of applications uh early as well. [13:04] Uh this is one that uh Riley Den is back there and I'll just point out Riley, she's waving. [13:11] Uh one of our master's students. Uh so I'll just use this slide to point out that we have uh [13:17] not a large number of students but a a reasonably good number of students who want to go on and get [13:21] their PhD after working on AI. It's a good area and uh Riley is one of those students [13:28] who's thinking about it, I think. Hopefully. Yeah. Okay. She's nodding. Yes. And uh you know so this [13:34] is a project we've done for six years to use AI science AI for science to try to help uh prevent [13:42] poaching of elephants in Africa. This is a joint project data collaboration project we have with [13:48] Cornell University. They originally uh put this out as a contest to see if they could get progress [13:55] and they didn't get a lot of progress. So they agreed to just do a data collaboration with us and [14:01] we think we've kind of solved it. We're still writing papers and trying to you know get more uh [14:07] constructive feedback and so forth. What the next thing to do would be to try to deploy this. Um so [14:15] this is where it would be deployed in Africa uh in the Zanga National Park. Those are the performance [14:22] of some of the models. Uh we have a poster, we have a couple posters here, but one of our posters [14:27] is on this project and we have a simulation uh that gives really strong evidence that it actually [14:33] reduces poaching at scale. So, and we've done some projects similar to this in the past with FAA and [14:40] Coast Guard and things like that. And and that's how we actually kind of got into this. Someone [14:44] came to us and said, "What you're doing for the FAA, could you do that for elephants?" Which [14:48] I know sounds a little crazy, but it it has worked out. So, uh, the themes of this lab is interactive [14:56] is key. So, we've got to have people building apps, building things, testing them, using [15:02] testdriven design for AI and machine learning very much like you might do that for software, [15:08] right? So, you know, I kind of view this as we've been down this road before, but we have new tools [15:13] that, you know, can allow us to do things maybe faster. uh we do have to be careful with them but [15:21] uh you know this is really enabling uh projects like this in robotics and sensor networks and so [15:26] on so forth and um you know some of it requires building a new lab so I've had to convince [15:34] facilities the dean a lot of people that you know let me do this I'm going to do it all with 5G [15:40] networking on our own network rather than university infrastructure things like that [15:45] if anyone's interested it does require I I believe building new new lab facilities. [15:51] So this is just a that's actually my son next to the Tesla robot. So my point here is you know the [15:57] future workforce is going to be different. I don't even know what to advise my son who's you know a [16:01] freshman in high school what it's going to be like. I I really honestly don't know and I'm [16:06] in tech like you know a few years ago I would have said yes. I know what to advise him to do. [16:11] Um, so you know, and you've got completely new uh computer infrastructure like the DGX Spark, [16:18] which is basically kind of like a supercomputer in a box or a little mini AI factory, right? [16:23] And uh, you know, we definitely need to expose students to this. So not only cloud-based systems [16:29] but more edgebased systems as well uh for things like reinforcement learning and uh [16:36] simtoreal gaps and digital twin all those kind of things that probably many of you are working on. [16:41] So in the you know this is just kind of I guess the obvious slide. These uh towers and desktops [16:47] are kind of the dinosaurs and the new thing is going straight from the cloud to the edge, [16:51] right? But you do still need a lab to facilitate that is the argument here. And uh this is the ask. [17:00] If anyone wants to write out a check for $2,000, I will put your company name on a bench in our lab. [17:05] I know it's just I thought I'd ask. [laughter] And then uh the other thing is we have this idea at [17:13] Chico that many of our students wind up going to Chicago or Silicon Valley at great expense to the [17:18] company and great expense to the student for an internship. If you could carve off something that [17:24] you know gets around whatever IP restrictions you have and so forth and have a student work [17:28] on it over the summer. I do that every summer with 10 undergraduate students. I do that every [17:34] summer with a half dozen graduate students and you know uh the mentoring is pretty affordable [17:41] with our faculty and that kind of thing. So you get applied AI done you know right at Chico [17:47] uh probably saves your company a lot of money. I guess the only downside might be less contact [17:52] with the student but certainly in our world we can zoom and things like that. So and we that's been [17:58] successful with a number of the companies we're working with. So, we're looking to broaden that. [18:04] One of, you know, we have some exciting new ideas for a joint project with a R1 research university [18:10] and NASA that we're working on right now uh for uh lunar landing uh find spots and so on and so [18:18] forth. I won't go into it too much because it's kind of new. Uh the courses are important to lay [18:23] the groundwork. So, the students need the courses. So creating the new AI curriculum [18:29] that uh enables them to have the knowledge to go ahead and do these kinds of things. [18:34] And I think I'm kind of running out of time here. Here's just more curriculum stuff and papers and [18:38] I will be done. So I don't know if I'm am I eating in my question time. [18:44] Oh, there's no questions. Okay. So if you have questions, uh you know, find Riley or myself [18:51] and uh Riley would be glad to answer questions. is the elephant expert in the room. And uh you know, [18:59] I'm glad to talk to you about the Isel. If you want to get your name on one of those benches, [19:03] you know where to find me. All right. Thank you so much. I appreciate it.