a16z Backs Underdog, a Personal AI That Reads Your Email Without It Ever Leaving Your Mac

Most AI assistants work like a very smart pen pal: your words travel to a distant server, and an answer comes back. Conway Research, a lab founded by Sigil Wen, wants to cancel the mail service entirely. On Thursday it launched Underdog, a personal AI assistant that runs entirely on the consumer hardware you already own — starting with the Mac — with no user data uploaded to external servers, backed by a seed round led by Andreessen Horowitz with Khosla Ventures, Hummingbird VC and angels including Patrick Collison and Naval Ravikant. The round size was not disclosed.

Underdog's pitch is capable, free and private, all at once — three adjectives that usually come with trade-offs. The product connects to a user's email and calendar like an ordinary mail client, then does its AI processing locally: searching messages, organizing schedules and drafting on-device, with the local-AI portion functional even offline. Anything requiring live mail sync or a booking site still needs the network, but the design principle is that the model never sees your inbox on someone else's GPU. Approval-gated actions are the default story rather than a security add-on.

The engineering bet rests on two components. The first is Woof, a 4-billion-parameter model whose 4-bit weights take up roughly 2.37GB, alongside a larger 27B sibling; Conway says the 27B model is designed to outperform frontier models such as Claude Opus 4.6 on certain benchmarks while running on consumer hardware — a company claim that applies to specific tests, not across the board. The second is Husky, an inference engine co-designed with the models for Apple Silicon. On an M5 Max MacBook, the team says the same model beat Apple's MLX framework across 16 tasks with a maximum speedup of 4.5x, processing up to 730 tokens per second. All performance figures are Conway's own.

Wen, a 2025 Thiel Fellow who previously co-created Airchat with Naval Ravikant and published a "Web 4.0" essay that drew 10 million views in 48 hours in February, frames the roadmap as "Underdog's Law": cutting-edge AI that requires a data center today will run on personal devices in about six months. Underdog is next set to expand to iPhone, Windows, Linux, Android and NVIDIA devices, and the team is exploring use cases where agents shop directly on a user's behalf — the long-term goal being a local agent that retains deep personal context and acts on it without that context ever becoming a cloud prompt.

The motivation is personal as well as architectural: Wen has described a years-old incident in which a bug in a popular email app exposed privileged mail stored on the app's servers. Underdog's design — direct provider sync plus local AI processing — is the architectural reaction to that class of failure. Hugging Face co-founder Thom Wolf endorsed the launch, arguing personal AI should obviously run on-device, powered by the hub.

The bet has real limits. Underdog's own positioning concedes the model will not top open-ended reasoning leaderboards, and a privacy win tied to Secure Enclave and Apple silicon stays Mac-shaped until the other platforms ship. For hard research and coding, users will still reach for frontier APIs. But for the inbox-triage, meeting-notes and scheduling workloads where latency and privacy dominate, a fast 4B model on M-series silicon can be enough — and the launch lands the same week Apple said it would require explicit user action before AI agents get Full Disk Access on macOS, and AMD published modeling claiming AI PCs can cut agentic inference costs by up to 60%. The edge is where the AI industry's next cost and trust fight is heading, and Underdog is an early, well-funded shot at it.