Executive Overview
In an era where artificial intelligence is predominantly defined by massive, energy-guzzling server farms, astronomical cloud computing costs, and invasive data-harvesting business models, a teenage dropout-turned-Thiel Fellow is orchestrating a quiet rebellion. Sigil Wen, a self-taught coder who was coding alongside the architects of modern AI before he could legally vote, has officially stepped into the spotlight. On Monday, Wen launched an invite-only beta for Underdog, an on-device AI assistant developed by his startup, Conway Research.
Underdog represents a fundamental architectural and economic departure from the status quo of Silicon Valley. Rather than routing sensitive user prompts, personal documents, and financial records to remote data centers—where they risk being stored, monetized, or repurposed for model training—Underdog runs entirely on the user’s local hardware. Powered by a proprietary inference engine named Husky, the assistant bypasses cloud latency entirely.
Yet, the most disruptive element of Underdog is not its technical execution, but its economic design. Eschewing the traditional software-as-a-service (SaaS) subscription model and refusing to rely on targeted advertising, Conway Research plans to offer the core assistant for free. Instead, the startup has partnered with financial heavyweights—securing backing from Stripe co-founder Patrick Collison—to monetize through transactional micro-fees. By taking a tiny percentage of payment transactions executed by the AI via secure Stripe rails, Underdog aligns its financial incentives directly with the user’s utility rather than their data footprint.
Backed by an elite cohort of venture capital giants including Andreessen Horowitz, Khosla Ventures, and the Anthropic-Menlo Ventures Anthology Fund, Wen’s project poses a direct existential question to the tech industry: Why should utilizing ambient intelligence require surrendering your private life?
Detailed Chronology: From AI Hacker Houses to the Birth of Conway Research
The Crucible of Early Silicon Valley
Long before founding Conway Research, Sigil Wen’s education was forged in the hyper-accelerated crucible of early generative AI development. At just 17 years old, the self-taught programmer packed his bags and moved to Silicon Valley, plunging headfirst into an AI hacker house that served as a breeding ground for the industry’s future titans.
Living alongside famed AI researcher Andrej Karpathy, Wen found himself rubbing shoulders with innovators who would soon define the generative AI revolution. Among his peers in those formative hacker houses were Aravind Srinivas (who would go on to found conversational search engine Perplexity) and Noam Brown (whose research at OpenAI redefined game-theoretic AI).
During this whirlwind period, Wen acted as an early-stage stress-tester for technologies that would eventually reshape global computing. He tested early beta iterations of chatbots distributed by Anthropic co-founder Ben Mann—software that would later mature into Claude. He experimented with early image-generation models developed by David Holz, which eventually evolved into Midjourney. Furthermore, he interacted with preliminary builds of OpenAI’s GPT-3 and early open-source diffusion models like Stable Diffusion.
His technical prowess was not confined to desktop environments; in what he describes as a testament to the "magical time" of early experimentation, Wen famously engineered a way to run GPT-2 locally on an Apple Watch purely for personal amusement. This exposure to cutting-edge models instilled in him a foundational intuition: the boundary between what could run locally versus what required massive cloud infrastructure was shifting rapidly.
His early exploits caught the attention of prominent entrepreneur and investor Naval Ravikant, who recruited Wen to work on Airchat, Ravikant’s short-lived audio-first social network designed to rival Clubhouse.
The Thiel Fellowship and the Conception of Underdog
Recognizing that traditional institutional education could not keep pace with the hyper-velocity of the intelligence revolution, Wen applied for and was awarded a prestigious Thiel Fellowship. The program, established by PayPal co-founder Peter Thiel, offers young innovators a stark choice: bypass or drop out of college to receive a grant and mentorship to pursue high-impact commercial or scientific projects immediately.
Freed from academic constraints, Wen turned his attention to a glaring vulnerability in the modern software landscape: the erosion of personal digital sovereignty. As everyday users began integrating AI assistants into their personal and professional workflows—entrusting them with unencrypted emails, corporate spreadsheets, private chats, and medical queries—cloud-based providers were quietly drafting privacy policies designed to harvest, analyze, and capitalize on that telemetry.
Determined to build a tool that he would willingly allow his own future children to use, Wen founded Conway Research. His mission was singular: engineer an AI assistant that delivers state-of-the-art capability without demanding a single byte of user data in return. On Monday, that vision materialized with the launch of the invite-only beta for Underdog.
Supporting Context & Metrics: The Architecture of Local Intelligence
Underdog’s Technical Infrastructure
The technical hurdles of running a capable large language model locally on consumer hardware are immense. Traditionally, local models suffer from crippling performance bottlenecks, high memory consumption, and sluggish token-generation speeds. Wen and his engineering team at Conway Research tackled these obstacles head-on by developing Husky, a proprietary inference engine optimized to squeeze maximum performance out of everyday consumer devices.
According to technical specifications released by the startup, Husky optimizes hardware communication by significantly reducing the volume of data shuttled back and forth between a computer’s main central processing unit (CPU) and its graphics processing unit (GPU). By streamlining memory bandwidth usage, Husky allows complex reasoning models to run fluidly on local hardware without sending fans screaming or batteries draining instantaneously.
The invite-only beta of Underdog currently supports macOS and Windows PCs, with expanded support for Linux, iOS, and Android slated for release in the near future.
Model Scaling and Performance Benchmarks
To fit comfortably onto consumer machines without requiring enterprise-grade graphics cards, Underdog relies on a model footprint significantly smaller than the multi-hundred-billion parameter leviathans hosted in Amazon, Microsoft, or Google data centers.
The current iteration of Underdog utilizes a 27-billion parameter reasoning model, meticulously fine-tuned from the open-weights Qwen3.8-27B architecture. While purists might argue that 27 billion parameters fall short of frontier models boasting upwards of a trillion parameters, Wen points to empirical benchmarks to prove otherwise.
According to comparative data from Artificial Analysis, Wen notes that Underdog’s fine-tuned 27B model performs favorably against models like Claude Opus 4.6 on specific reasoning benchmarks—performance levels that were considered state-of-the-art across the industry a mere six months prior. For the vast majority of everyday consumer tasks—ranging from deep shopping research and contextual document summarization to solving complex mathematics homework—the model provides more than sufficient capability.
"You don’t need to sacrifice your privacy for the capability because they’re just as capable," Wen argues, emphasizing that the performance gap between cloud-hosted giants and local models is rapidly shrinking.
Security and Data Isolation
Beyond its inference engine, Underdog incorporates rigorous cryptographic security measures designed to lock down user authorization. When users grant Underdog permission to interact with sensitive personal APIs—such as parsing personal email accounts, accessing calendar entries, or reviewing local file directories—the application encrypts all authentication keys locally. Data never traverses an external server, entirely eliminating the risk of third-party data breaches, corporate snooping, or unauthorized model-training ingestion.
Official Statements and Economic Philosophy
The Death of the SaaS Subscription
Perhaps the most radical disruption introduced by Conway Research lies not in its silicon, but in its balance sheet.
For the past three years, the generative AI industry has operated under crushing capital expenditure (CapEx) models. Startups and tech giants alike burn billions of dollars on NVIDIA GPUs and massive electricity bills to power centralized data centers. To offset these cloud inference costs, companies have invariably imposed steep monthly subscription fees (often $20 to $200 per month) or resorted to data monetization strategies.
Underdog flips this economic equation entirely on its head. Because the computational load is offloaded directly to the user’s local machine—utilizing the silicon they have already purchased and powered—Conway Research incurs virtually zero marginal inference costs per user.
“I don’t have to charge you a subscription to run this because my costs are so super low,” Wen explained to TechCrunch. Consequently, the core assistant will remain entirely free to use and will never rely on intrusive advertising.
Borrowing from Fintech: The Stripe Integration
Without subscriptions or ad-tracking revenue, how does Conway Research plan to build a sustainable, venture-backed enterprise? The answer lies in an innovative business model borrowed directly from the fintech playbook, enabled by one of Wen’s high-profile angel investors: Stripe co-founder Patrick Collison.
Underdog is designed to act not merely as a conversational chatbot, but as an active agent capable of executing digital transactions on behalf of the user—booking flights, purchasing goods, managing subscriptions, and executing micro-payments. By partnering with Stripe, Conway Research plans to take a tiny percentage of payment transactions executed by the AI assistant via secure payment rails—functioning much like a traditional credit card interchange fee.
This creates a radical alignment of interests:
- Traditional AI Models: Incentivized to collect, retain, and mine user data to sell targeted advertisements or improve external proprietary models.
- Underdog: Incentivized solely to help the user complete tasks efficiently. Underdog makes money only when the user successfully and safely transacts, aligning the startup’s revenue model with the user’s bank or credit card provider rather than data brokers.
The Investor Syndicate
The audacity of Wen’s vision has attracted an extraordinary roster of financial backers. Conway Research’s capitalization table reads like a who’s who of modern tech entrepreneurship and venture capital:
- Andreessen Horowitz (a16z): Led by general partner Chris Dixon, a vocal champion of decentralized, user-centric technologies.
- Khosla Ventures: Known for early, high-conviction bets on foundational infrastructure.
- Patrick Collison: Stripe co-founder, providing direct strategic alignment with global payment infrastructure.
- The Anthology Fund: The strategic partnership fund bridging Menlo Ventures and Anthropic.
- Hummingbird & SV Angel: Elite early-stage venture firms.
- Notable Angel Investors: Including Vercel founder Guillermo Rauch, AI researcher Noam Brown, and prominent engineer Deedy Das.
Future Outlook: The Horizon of Local AI
The launch of Underdog arrives at a critical cultural and regulatory juncture. As governments worldwide crack down on corporate data harvesting and public trust in cloud-based tech monopolies reaches historic lows, the vulnerabilities of centralized AI have never been more apparent.
As noted in recent tech industry reporting, powerful cloud assistants frequently spark fierce privacy and security alarms. When an AI assistant requires intimate access to a user’s most private details—ranging from undiagnosed medical symptoms and personal financial disclosures to private correspondence concerning children—handing that telemetry over to centralized cloud servers represents an unacceptable security trade-off for privacy-conscious consumers.
Wen’s "AI manifesto," published on the startup’s official portal, crystallizes the cultural zeitgeist driving Conway Research: “Why should using AI require surrendering your private information?”
As consumer hardware continues to advance—with Apple, Qualcomm, and AMD flooding the market with powerful neural processing units (NPUs) built directly into consumer laptops and smartphones—the technical viability of local-first AI will only accelerate. Small models will continue to expand in reasoning capability, closing the performance gap with data-center colossuses while retaining the ultimate competitive moat: absolute, cryptographic privacy.
For Sigil Wen, who spent his adolescence coding alongside the architects of the generative AI boom, Underdog is more than just a software product; it is a rectification of the industry’s original sin. By proving that high-performance AI can coexist with absolute data sovereignty, Conway Research may well be paving the path for the next generation of computing—one where the user remains the master of their own digital domain.
