Underdog launches a private on-device AI assistant, backed by a16z
The privacy split in consumer AI got a new entrant tonight, GitHub turned code-review benchmarks into a vendor-bias argument, and OpenAI opened its oddest API to everyone.
Underdog launched in invite-only beta: an on-device AI assistant that keeps your data on your machine and never charges you a subscription. Self-taught coder and Thiel Fellow Sigil Wen — who moved to Silicon Valley at 17 and lived in an AI hacker house with Andrej Karpathy — built his own inference engine, Husky, to run a 27-billion-parameter model fine-tuned from Qwen3.8-27B on Macs and Windows PCs today, with Linux and phone versions promised. The backing is a who's-who: an a16z-led first round with Khosla Ventures, Hummingbird and SV Angel, plus angels including Patrick Collison, Guillermo Rauch and Noam Brown. The amount was not disclosed. The business model is the real twist — free, never ad-supported, and instead a small cut of the payments the assistant makes over Stripe's rails, an interchange fee for the AI era. That is the deliberate opposite of Instinct and Muse, whose privacy policies permit collecting user data for ads or training; we covered the sealed-VM approach in Meta ships Muse: a consumer agent inside a sealed VM. The take: if the interchange idea holds, on-device AI gets a revenue story that doesn't depend on watching you — but 27 billion parameters is a long way from a frontier model, and Wen's claim of Opus-class performance on some benchmarks is his to prove, not ours.
GitHub published ReviewBench, an open benchmark for AI code-review agents — and its own Copilot took first place at 40.1% grounded F1. The corpus is real work: 219 pull requests across 187 repositories and 19 languages, selected after analyzing 103.9 million pull requests, with a human-labeled golden set on which senior engineers agreed 96.6% of the time. The catch is who held the stopwatch — GitHub built the benchmark with Microsoft and ran every competitor's initial evaluation itself, on differing test dates the vendors neither conducted nor verified. Martian's independent Code Review Bench puts Copilot fourth online at 60.9% and fifth on its offline run. Scores aren't comparable across the two benchmarks, but the direction is: rankings swing several places depending on who designs the yardstick.
OpenAI put the Decisions API into public beta, opening a DevDay surprise to every developer. The endpoint answers bounded questions — a probability, a pick from your fixed options, a score across ordered levels — instead of free text, and OpenAI claims it is about ten times faster than the Responses API, at roughly 150 milliseconds against 1.6 seconds for a plain Luna reply. It runs on one model, gpt-6-luna, and pricing is unusually flat: $0.10 per 1 million input tokens, input-only, with no output or cache charges. The API itself is not new — it was announced in limited preview at DevDay, and independent outlets covered it then — but today's move to public beta, with pricing published, is OpenAI's own announcement to carry for now. It is a different bet from the raw text completion OpenAI built its business on: structured decisions as the product, following the route opened by OpenAI opens its Codex agent harness to every developer with the Agents API.
What to watch: whether Underdog's interchange model survives contact with payment regulators, whether independent runs reproduce ReviewBench's Copilot lead, and how long the Decisions API stays in beta before general availability.
Would you trade frontier-model quality for an assistant that never sees your data? Tell us in the comments.
