Conway Research Founder Sigil Wen Launches First Product Underdog
Conway Research founder Sigil Wen has launched the first product, Underdog, positioned as a private personal AI running on users' own devices. Andreessen Horowitz confirmed it led the company's first round of funding, with amounts and terms undisclosed.
Investors include Khosla Ventures, Hummingbird, Anthology (a fund co-managed by Anthropic and Menlo Ventures), SV Angel, and Breakneck. Angel investors include Patrick Collison, Naval Ravikant, Guillermo Rauch, Noam Brown, Thomas Wolf, and Charles Songhurst. The a16z article was written by Chris Dixon, Gabriel Vasquez, and Elizabeth Harkavy, noting prior investments in Sigil Wen.
The product is currently in invite-only testing and can be downloaded on Apple silicon Macs, with plans to expand to iPhone, Windows, Linux, Android, and NVIDIA devices. Emails, schedules, meeting notes, dictation, and private messages are processed locally and can run offline; the company claims personal data is encrypted and remains on the device, with no Underdog servers storing personal data.
The public small model Woof has 4 billion parameters, with a 4-bit weight of about 2.37GB. The team developed the Husky inference engine, which is faster than MLX on M5 Max across 16 tasks, by up to approximately 4.5 times. The larger Underdog 27B is based on Qwen 3.8 27B, and the company claims it exceeds Claude Opus 4.6 from over six months ago in instruction following, document reading, mobile operations, and code fixing, but still lags in graduate-level science questions and expert problems.
Sigil Wen refers to the roadmap as Underdog’s Law: cutting-edge capabilities that require data centers today will be available on personal devices in about six months. a16z states that his motivation stems from an early experience living with Andrej Karpathy in a hacker house, where a vulnerability in an email application exposed his private emails to other users. The team's background includes Apple, Google Brain, Meta, Baidu, and the Thiel Fellowship.
The buyers are the a16z crypto and application investment teams, as well as personal capital from the stablecoin, open-source model, and payment infrastructure circles; the sellers are assistants that still place personal context in cloud APIs. The event is driven by the first round of funding. Funding is shifting from cloud inference subscriptions to edge models and local inference engines, benefiting device manufacturers and local model teams, while putting pressure on cloud personal assistants that charge for email, calendar, and credential calls.
Source: Public Information
ABAB AI Insight
Sigil Wen was born in Toronto in 2003, entered TKS in 2019 to work on early AI projects, and in 2021 generated millions in transaction volume with BitSwap, which was reported by The New Yorker. That same year, he dropped out of the University of Pennsylvania's M&T program at age 17 to move to San Francisco. His startup Monument received funding from Naval Ravikant, Elad Gil, and Justin Kan but was shut down due to regulatory issues; Serendipity achieved about $2,000 in monthly recurring revenue within two weeks of launch. In 2023, he joined Naval's Airchat as a founding engineer, integrating Whisper into the iPhone, and became the youngest member of Spearhead at 19, managing about $2 million in allocations and angel investing in over 20 companies including Etched and Friend.
This first round is not a new relationship. a16z noted that Chris Dixon had already invested in him, and Naval, Patrick Collison, and Noam Brown were all part of previous connections: he met Dixon through Naval in 2021, interacted with the Collison brothers during the Airchat period, and was in the same hacker house as Noam Brown and Marco Mascorro. The capital path initially used the visa service Extraordinary to serve Cognition, Lovable, Etched, a16z, and Founders Fund, before integrating the same group and Thiel Fellow identities into Conway. In 2026, he also created Automaton, which allowed agents to hold wallets, purchase computing power, and register domain names.
The analogy is not just another chat application, but rather how Apple integrated the neural engine into devices, and how Mistral and Hugging Face extracted small models from APIs. The current position is in a control phase: cloud labs still hold the largest models, while edge teams are competing for high-sensitivity, low-difficulty tasks like emails, calendars, and credentials. Extraordinary has transitioned to executive chairman, with the main product changing to an assistant that does not upload personal data.
Structurally, this represents a transfer of pricing power. The pricing power of cloud assistants comes from the notion that "context must be uploaded to become smarter"; Underdog keeps the context on the device, shifting inference costs from being paid per token to a one-time download plus local computing power. If the six-month lag law holds, the personal data moat of cutting-edge labs will shorten generationally, leaving advantages only in problems that can only run in data centers, rather than everyday agents.
ABAB News · Cognitive Law
- The context on whose hard drive holds the pricing power.
- The cloud sells intelligence, while the edge sells the right not to disclose data.
- The shelf life of cutting-edge capabilities is approximately one device cycle.