这段故事始于近12年前，当时塔约·阿德桑耶开始从事微芯片和人工智能处理器相关工作。他主要帮助大型制造商决定其硬件应采用哪些芯片。他告诉TechCrunch，这些年的工作让他较早洞察到人工智能计算市场的需求将走向何方。他说：“创办Lola Vision Systems，是对世界未来走向以及我所看到趋势的一次押注。”
The story begins almost 12 years ago, when Tayo Adesanya started a career working with microchips and AI processors. He mainly helped large manufacturers decide which chips to use in their hardware. Those years, he told TechCrunch, gave him early insight into where demand in the AI computing market was headed. “Starting Lola Vision Systems was a bet on where the world was headed and what I was seeing,” he said.
2024年，他创办了人工智能基础设施公司Lola Vision Systems，该公司开发用于在设备上运行人工智能模型的软件和芯片。其核心产品是一种软件，可将人工智能模型转换为特定芯片能够运行的指令。阿德桑耶称这套软件为“编译器工具链”，并表示它构成了一个巨大的瓶颈：在新硬件上手动设置人工智能模型，仅仅为了开始测试就可能需要“大约200小时”。
In 2024, he launched Lola Vision Systems, an AI infrastructure company that builds software and chips for running AI models on devices. Its core product is software that translates AI models into instructions a specific chip can run. Adesanya calls this software a “compiler toolchain,” and he says it is a massive bottleneck: manually setting up an AI model on new hardware can take “roughly 200 hours” just to begin testing.
Lola Vision表示，公司已重新构建了这一软件层，同时还在开发自己的半导体芯片，目标是进一步实现更多流程的自动化。客户提供其代码和希望使用的人工智能模型，无论是定制开发还是开源模型，软件都会将二者转换为客户芯片能够执行的指令。
Lola Vision says it has rebuilt that software layer and is also developing its own semiconductor chips, with the goal of automating more of the process. A client provides its code and the AI model it wants to use, whether custom-built or open source, and the software translates both into instructions the client’s chip can execute.
“速度只是其中的一部分，”阿德萨尼亚说道。他解释说，更快的启动速度能让航空航天以及“其他任务关键型公司”有时间“在自身数据上运行更精确的模型，同时降低功耗”。
“Speed is only part of it,” Adesanya said. He explained that faster setup gives aerospace and “other mission-critical companies” time to “run more accurate models on their own data, at a lower power.”
“对于这些客户而言，”他说，“准确性和可靠性并非可有可无的附加功能，而是决定产品能否通过监管审查，以及能否在实际应用中稳定运行的关键因素。”
“For these customers,” he said, “accuracy and reliability aren’t nice to have. They determine whether a product passes regulatory review and whether it works reliably in the field.”
总部位于华盛顿特区的Lola Vision是多家试图为NVIDIA的设备端AI运行技术提供替代方案的初创公司之一。阿德萨尼亚表示，目前许多公司最初会选择NVIDIA的Jetson系列——这是一套用于设备端运行AI的紧凑型计算模块——或者开源AI模型。他声称，这些方案“往往刚部署就出现故障或运行效果不佳，团队需要花费数天甚至数周时间才能让它们正常运行，之后还要再花更长时间进行调试，直到模型可用为止。”
Lola Vision, based in Washington, D.C., is one of several startups trying to offer an alternative to NVIDIA’s technology for running AI on devices. Right now, Adesanya said, many companies start with Nvidia’s Jetson, a line of compact computing modules for running AI on devices, or with open-source AI models. Adesanya claimed these “often break or run poorly out of the box, so teams spend days or weeks getting them to run at all, then even more weeks debugging until the models are usable.”
“即便如此，”他继续说道，“功耗常常超出边缘计算的预算，或者主板无法为产品成功运行所必需的中大型模型提供足够的计算能力。这会导致识别模型的表现落后于目标，或对物体产生误判。”（边缘计算指的是直接在设备（如摄像头或无人机）上运行AI，而不是在远程数据中心运行。识别模型则是用于识别物体的AI系统。）
“Even then,” he continued, “power consumption often blows edge computing budgets, or the board can’t deliver enough compute for the medium to large models the product actually needs to run successfully. This leads to the recognition models lagging behind targets or misreading objects.” (Edge computing means running AI directly on a device, such as a camera or drone, rather than in a remote data center. Recognition models are AI systems that identify objects.)
据该公司称，已有十几家企业客户签署了意向书，表示希望在Lola Vision的芯片上市后购买，而且目前已有至少一家客户正式签约。此外，公司还与一家名为SCALE的微电子人才发展计划合作，以接入更多半导体实验室。“为了更早实现收入，我们现在将把我们的软件授权给现有硬件使用，”阿德萨尼亚说。（换句话说，公司不会等待自己的芯片问世，而是允许客户付费使用其软件，运行现有的芯片。）他还补充说，截至目前，公司总共已筹集到略高于100万美元的资金。
According to the company, a dozen corporate customers have signed letters expressing interest in buying Lola Vision’s chips once they are available, and it already has one signed customer. It has also partnered with SCALE, a microelectronics workforce development program, to work with more semiconductor labs. “To get revenue sooner, we will now license our software on existing hardware,” Adesanya said. (In other words, rather than waiting for its own chips, the company will let customers pay to use its software on chips that already exist.) He added that the company has raised just over $1 million in total funding to date.
Lola Vision 入选了今年的 TechCrunch Battlefield 200，这是该计划选出的 200 家初创公司之一。“我还在普渡大学读书时，TechCrunch 就一直是我最喜欢的媒体之一，”他说。在花费大约一年时间开发产品并赢得首个客户后，他觉得是时候申请 Battlefield 计划，让公司被更广泛的受众所了解。
Lola Vision was selected for this year’s TechCrunch Battlefield 200, a group of 200 startups chosen for the program. “TechCrunch was a favorite when I was a student at Purdue,” he said. After about a year of building the product andg its first customer, he said, he felt it was time to apply to Battlefield and get the company in front of a wider audience.
至于他对这次活动最期待的是什么，他表示：“希望能建立有意义的联系，并尽可能多地了解我们所在领域以及周边正在发生的事情。”“而且，坦率地说，我也期待投资者能签下支票。”想要了解更多关于 Lola Vision Systems 以及下周将加入 TechCrunch Disrupt 的其他数十家经过严格筛选的初创公司（还有前来考察的风险投资家）的信息，欢迎于下周 10 月 13 日至 15 日前往旧金山与我们一同参与。
As for what he’s most excited about when it comes to the event, it’s “making meaningful connections and learning as much as I can about what’s happening in and around our space,” he said. “And, to be direct, I’m looking forward to investors writing checks.”To learn more about Lola Vision Systems and dozens of other highly vetted startups that are joining us at TechCrunch DIsrupt next week (along with the VCs coming to check them out), join us in San Francisco next week, October 13-15.