Andreessen Horowitz announced today that it is leading Conway Research's first funding round, backing founder Sigil Wen as he builds Underdog — a personal AI assistant that runs entirely on users' own devices. The announcement, authored by general partner Chris Dixon, investment partner Gabriel Vasquez, and partner Elizabeth Harkavy, did not disclose the round's size or valuation.

The investment thesis is privacy-first personal computing. A16z says it shares Wen's conviction that "the AI that knows us best should work for us, with data we control." As assistants graduate from answering questions to acting on our behalf, the firm argues, their usefulness depends on access to our conversations, commitments, documents, and preferences — exactly the information we most need to protect. Underdog's answer, in a16z's words: "personal intelligence that runs on your own devices, helping you make decisions and reclaim your time, even offline."

A founder shaped by a privacy failure#

The product's origin is personal. Years ago, while Wen was living and hacking alongside AI researcher Andrej Karpathy, a bug in a popular email app exposed his private, privileged emails to other users. The fallout, a16z writes, "helped motivate him to build products that empower people while keeping their data safe."

Wen's route into AI runs through the frontier's inner circle. He came to San Francisco at 17, slept at a WeWork for a time, and skipped college. He spent nearly a year in a hacker house alongside researchers and engineers who helped create DALL-E, Whisper, GPT-3, and Stable Diffusion — Karpathy lived there, Anthropic co-founder Ben Mann invited Wen to become one of the earliest testers of what became Claude, and Midjourney's David Holz showed him the first version of the model behind Midjourney. He went on to help Naval Ravikant launch Airchat, recruiting most of its engineering team, became the youngest member of Ravikant's Spearhead, and received the Thiel Fellowship in 2025.

A hacker's desk at night with a glowing laptop and hand-drawn neural network sketches
A builder's origin: Wen ran GPT-2 on his Apple Watch and helped bring Whisper to the iPhone. AI-generated illustration — AI Frontier Post

The on-device stack#

Conway's technical bet is that frontier-capable intelligence can be compressed onto consumer hardware. Wen has been exploring how much intelligence fits on the devices we carry for years: he got GPT-2 running on his Apple Watch and helped bring Whisper to the iPhone during his time at Airchat, according to a16z.

The product lineup reported by Dealroom includes Woof, a 4-billion-parameter model that takes up less than 2.5GB, and a larger Underdog 27B, which Conway says is designed to outperform frontier models such as Claude Opus 4.6 — reportedly beating it on certain benchmarks while running on consumer hardware. Conway built custom inference engines tuned for Apple Silicon, which the company says run up to 4.5 times faster than Apple's own MLX framework, processing up to 730 tokens per second on MacBooks. TokenPost's coverage notes the peak figure comes from Conway's own benchmark on an Apple M5 Max for a function-edit task, and that results varied across the 16 tested tasks. Conway describes Underdog as running on Mac and iPhone, including without Wi-Fi, with agentic capabilities — it can take actions on a user's behalf.

A computer chip with glowing data streams looping within it, visualizing local AI inference
Local inference: Conway's custom engines claim up to 4.5× the speed of Apple's MLX on its own benchmarks. AI-generated illustration — AI Frontier Post

Why local-first is the bet#

Cloud assistants have a structural privacy problem: to answer you, they must see what you ask. Keeping that material on the device sidesteps the issue rather than managing it — there is no server breach to fear if the server never sees the data. Running models locally also shifts compute costs onto hardware users already own, which is how Conway can pitch an assistant as capable, private, and free at the same time. Conway's website describes a frontier AI lab working on efficient models, inference engines, and personal AI products, built on the thesis that frontier intelligence can reach users' devices within six months, reports RuntimeWire.

A16z's closing line frames the ambition in civilizational terms: "AI should be a liberating equalizer: free, capable, reliable assistance that gives everyone with a phone or computer sovereignty over their time, information, and choices."

What the announcement does not include is any user, revenue, pricing, or retention data — and the performance claims are Conway's own, not independently reproduced. The round establishes a16z's backing for Wen's thesis: that the most valuable AI of the next decade won't live in a data center, but in your pocket.