Sonny Valindra
The landscape of artificial intelligence assistants is undergoing a significant shift with the launch of Underdog, a new, invite-only beta service that prioritizes user privacy by running entirely on-device. Developed by Conway Research, a startup backed by prominent venture capital firms and angel investors, Underdog represents a bold move to decouple AI functionality from extensive data collection, a practice common among many of its more established rivals. At the helm of this ambitious venture is Sigil Wen, a prodigious self-taught coder whose early immersion in the nascent AI community at a young age has shaped his vision for a more private and user-centric AI future.
Wen’s journey into the heart of artificial intelligence began at the remarkable age of 17, when he relocated to Silicon Valley and became a resident of a renowned AI hacker house. This immersive environment, famously associated with AI researcher Andrej Karpathy, served as a crucible for innovation. Within its walls, Wen not only honed his coding skills but also collaborated with a constellation of individuals who would go on to become leading figures in the AI revolution. Among his contemporaries were Aravind Srinivas, the future founder of Perplexity, and Noam Brown, a researcher at OpenAI. This unique ecosystem provided Wen with unparalleled access to the bleeding edge of AI development.
During his time in the hacker house, Wen actively engaged with early iterations of AI tools that would later achieve widespread recognition. He tested a chatbot developed by Ben Mann, a co-founder of Anthropic, which would eventually evolve into Claude. He also worked with an image generator created by David Holz, the visionary behind Midjourney, and experimented with foundational models that would become OpenAI’s GPT-3 and the groundbreaking image generator Stable Diffusion. His precocious talent did not go unnoticed by industry titans; prominent investor and entrepreneur Naval Ravikant, known for his work with AngelList, recognized Wen’s potential and hired him for Airchat, Ravikant’s social networking platform that aimed to rival Clubhouse.
Wen’s dedication to pushing the boundaries of AI was evident even in his personal projects. As a testament to his ingenuity and passion, he famously figured out how to run the GPT-2 model on an Apple Watch, a feat he described as "magical." This ability to innovate and adapt AI to unconventional platforms underscored his deep understanding of the technology and his drive to make it accessible.
From Hacker House to Thiel Fellowship: A Foundation for Innovation
Wen’s trajectory continued to ascend as he was selected as a Thiel Fellow. This prestigious program, founded by entrepreneur and investor Peter Thiel, offers young, exceptional individuals the opportunity to defer or forgo traditional higher education to pursue entrepreneurial ventures. The fellowship provides critical funding and mentorship, enabling fellows to dedicate their time and energy to developing groundbreaking projects. For Wen, this fellowship provided the ideal platform to launch Underdog, a project deeply rooted in his experiences and his commitment to user privacy.
The launch of Underdog’s invite-only beta on Monday marks a significant moment in the quest for more secure and private AI assistants. Unlike cloud-based AI services that process user data on remote servers, Underdog operates entirely on the user’s own hardware. Currently, the assistant is available for Macs and Windows PCs, with plans to extend support to Linux, iPhones, and Android devices in the near future. This on-device architecture is central to Underdog’s privacy-first philosophy, ensuring that sensitive user information remains on their personal devices.
The Engine Behind Privacy: Husky Inference
At the core of Underdog’s on-device processing capabilities lies Husky, an inference engine meticulously engineered by Wen. Husky is designed to optimize the speed and efficiency of running AI models directly on user hardware. Wen claims that Husky achieves superior performance by minimizing data transfer between a computer’s central processing unit (CPU) and its graphics processing unit (GPU), a common bottleneck in on-device AI. This optimization is crucial for delivering a responsive and seamless user experience, even when running complex AI models locally.
Beyond its core processing engine, Underdog incorporates robust security measures. The platform encrypts the access keys for any email or other accounts that users authorize the AI assistant to interact with. This multi-layered approach to security aims to build trust and provide users with a high degree of confidence in the protection of their digital identities and sensitive information.
Balancing Capability and Privacy: The Model Choice
A key consideration for any on-device AI solution is the trade-off between model size and computational requirements. State-of-the-art AI models often boast billions, or even trillions, of parameters, necessitating powerful cloud infrastructure for their operation. Underdog, however, strategically employs smaller, yet highly capable, models. The current iteration utilizes a 27-billion parameter reasoning model, fine-tuned from Qwen3.8-27B.
Wen contends that this model, despite its smaller footprint compared to large language models hosted in data centers, demonstrates performance comparable to top-tier AI systems. He points to benchmark comparisons where the Underdog model favorably measures against Claude Opus 4.6, a model that represented peak performance approximately six months prior to Underdog’s launch. This comparison suggests that Underdog is well-equipped to handle a wide array of everyday tasks that users expect from an AI assistant, such as conducting shopping research, answering complex mathematical questions, and assisting with general knowledge queries.
"You don’t need to sacrifice your privacy for the capability because they’re just as capable," Wen asserted in a statement, emphasizing his belief that the advancement of smaller, on-device models will continue to close any perceived performance gaps with their cloud-based counterparts. This perspective challenges the prevailing notion that superior AI performance is intrinsically linked to massive, centralized data processing.
A Novel Business Model: Aligning Incentives with Users
Perhaps the most disruptive aspect of Underdog lies in its innovative business model. The application will be offered free of charge initially and will eschew advertising entirely. This approach is made feasible by Underdog’s on-device architecture. Since the AI processing occurs on the user’s machine, Conway Research bypasses the substantial operational costs associated with maintaining large data centers and paying for inference services.
"I don’t have to charge you a subscription to run this because my costs are so super low," Wen explained, highlighting the economic advantage of his privacy-focused strategy. This low overhead allows Underdog to offer its core functionality without the pressure to monetize user data through advertising or other means.
Instead, Underdog is adopting a strategy inspired by the fintech sector, with significant backing from Stripe co-founder Patrick Collison, who is an angel investor in Conway Research. The AI assistant will integrate with Stripe’s secure payment infrastructure. When Underdog facilitates a payment transaction on behalf of a user, it will take a minuscule percentage of the transaction value, akin to an interchange fee. This revenue stream is designed to be transparent and directly tied to the value the AI assistant provides through financial transactions, rather than through the exploitation of user data.
This business model stands in stark contrast to the practices of many other AI assistant providers. Numerous competitors have privacy policies that permit the collection of extensive user data. This data is often subsequently sold to advertisers or third parties, or utilized for training further AI models. Such data collection can be particularly concerning for AI assistants, which, by their nature, may require access to highly personal and sensitive information, including medical history, financial details, and family information, to be truly effective.
Wen’s stance on data privacy is clearly articulated in his "AI manifesto," where he poses the provocative question: "Why should using AI require surrendering your private information?" His vision for Underdog is deeply personal, as he stated, "I honestly want to build Underdog for myself. I’m building a product that I would be proud for my future children to use." This sentiment underscores a commitment to creating technology that aligns with ethical principles and personal values.
A Strong Foundation of Support: Investors and Visionaries
The startup behind Underdog, Conway Research, has garnered significant attention and financial backing from a distinguished group of investors. Patrick Collison’s involvement is complemented by substantial investment from Andreessen Horowitz, with partner Chris Dixon leading the charge. Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund – a partnership between Menlo Ventures and Anthropic – have also contributed to Conway’s funding rounds.
Furthermore, a notable list of angel investors has thrown their support behind Wen’s vision. This includes Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown, and Deedy Das, among other prominent figures in the tech industry. This strong investor base reflects a collective belief in Wen’s capabilities and the potential of Underdog to reshape the AI assistant market by prioritizing privacy and user autonomy.
The launch of Underdog is not merely the introduction of another AI tool; it represents a philosophical statement about the future of artificial intelligence and its integration into our daily lives. By offering a powerful, yet private, AI assistant, Wen and Conway Research are challenging the status quo and paving the way for a more ethical and user-empowering AI ecosystem. The success of Underdog will likely serve as a critical barometer for consumer demand for privacy-preserving AI solutions and could inspire a broader shift in the industry’s approach to data handling and user trust.
