a16z leads Conway Research's first round for Underdog, an on-device AI that stays on your phone

Andreessen Horowitz on Friday said it is leading Conway Research's first funding round for Underdog, a personal AI product built to run on users' own devices; the firm did not disclose the round's size.

Frontier chat still means your life leaves the device. Conway's bet — and a16z's lead check — is that a free, capable assistant can stay on the Mac and iPhone you already own, so the model that knows you never has to ship your inbox to someone else's GPU rack.

On Friday, 2 October 2026, Chris Dixon, Gabriel Vasquez, and Elizabeth Harkavy posted that Andreessen Horowitz is leading Conway Research’s first round of funding and partnering with founder Sigil Wen as he builds Underdog. The page is titled “Investing in Conway, the Creator of Underdog.” It dates the post October 2, 2026, and does not print an hour. Dixon is a general partner who founded and leads a16z crypto. Vasquez is an investment partner focused on AI applications. Harkavy is a partner on the a16z crypto investment team. Those roles are on the post. The closing line is theirs: they are thrilled to lead the first round and partner with Wen as he builds Underdog.

What the post says the product is. Underdog is personal intelligence that runs on a person’s own devices, including offline, helping them make decisions and get time back. The conviction a16z says it shares with Wen: “the AI that knows us best should work for us, with data we control.” As assistants stop only answering questions and start acting, the post says, usefulness depends on access to conversations, commitments, documents, and preferences. That is also the information a person most needs to protect. a16z’s longer bet, in the same post, is that as the models improve and reach more consumer devices, this kind of assistant can put powerful AI in billions of people’s hands: “free, capable, reliable assistance” and sovereignty over time, information, and choices. Sovereignty, in that sentence, means the person keeps control. Those lines are a16z’s. The post does not print a price, a download count, or a ship date.

What the post does not say about the money. It says a16z is leading Conway Research’s first round. It does not print the size of the round, a valuation, or a list of other investors. In a venture round, the lead is the main investor. The page does not say how large that check is. CryptoBriefing, Diego Almada Lopez, 2 October 2026, names Khosla Ventures and Hummingbird VC among other investors, and says the size of the round was not part of the announcement. TokenPost, the same news cycle, says CryptoBriefing named those two firms and that TokenPost could not find material confirming their participation. Securities.io, Owen Hartley, 2 October 2026, reports the a16z lead and says the firm did not disclose the round’s size or terms. No dollar figure for the round appears in a16z’s post.

Who Wen is, as a16z tells it. The firm says it had backed him before this round. He came to San Francisco at 17 without much of a network, slept for a time at a WeWork, skipped college to work on AI, and spent nearly a year in a hacker house with researchers and engineers who helped create DALL·E, Whisper, GPT-3, and Stable Diffusion. DALL·E and Stable Diffusion make images from text. Whisper turns speech into text. GPT-3 is an earlier text model from the same era. Andrej Karpathy lived in that house. Anthropic cofounder Ben Mann invited Wen to be one of the earliest testers of what became Claude. David Holz showed him the first version of the model behind Midjourney. a16z says it first met Wen when Naval Ravikant brought him to an event Dixon hosted in 2021. Wen then helped Ravikant launch Airchat, bringing new AI models into a product people used every day and recruiting most of its engineering team. Ravikant also made him the youngest member of Spearhead. In 2025 he received the Thiel Fellowship, the grant for young people who leave school to build. Those lines are a16z’s. The post does not print a previous check size.

Why the product is personal, still from the post. Years ago, while Wen was living and working alongside Karpathy, a bug in a popular email app exposed his private, privileged emails to other users. a16z says that fallout helped motivate him to build products that keep a person’s data safe. At the same time he was testing how much intelligence could fit on a device a person already carries. He got GPT-2 running on an Apple Watch, and later helped bring Whisper to the iPhone at Airchat. GPT-2 is an older, smaller text model. As smaller models got more capable and the hardware improved, the post says, he saw a path to useful personal AI that runs locally. Locally means on the device, not on a server in someone else’s building. Those lines are a16z’s.

What Conway says it is. On conway.tech the lab says it builds intelligence you can own. It calls itself a frontier AI lab building efficient models, inference engines, and personal AI products for devices people already own, and it says it is the creator of underdog.ai. A model is the trained system that answers. An inference engine is the software that runs that model and writes the next words. The thesis line on the page is “Underdog’s Law: today’s frontier intelligence reaches your devices in six months.” Frontier, here, means the most capable systems, the ones that today mostly run in data centers. The research list on the same page names agentic commerce, machine economies, programmable markets, autonomous organizations, robotics and autonomous retail, on-device AI, AI security, confidential inference, and megakernels. Agentic, in that list, means software that takes a next step, not only a chat reply. Confidential inference means running the model in a way that keeps the data private. Those names are Conway’s. The page does not attach a product schedule to each one.

The speed claim, and whose number it is. Conway’s update names Husky as its model-specific inference engine and says the peak result is up to 4.5 times MLX’s token-generation speed. The task is a function edit. The model is Woof 4B. The machine is an Apple M5 Max, with Flash enabled. MLX is Apple’s own software for running this kind of model on a Mac. A token is a chunk of text, often a word or part of a word. Token generation is how fast the model writes. Woof 4B is Conway’s name for a model with about 4 billion parameters. A parameter is one number the model learned in training. Four billion is small beside the cloud models people chat with. An M5 Max is a high-end Apple chip. Flash, on this line, is a setting Conway names for that peak. Up to 4.5 times is the ceiling Conway publishes for that one task, on that one chip. It is the company’s measurement. TokenPost says the company’s own page puts Husky at 730 tokens a second on that run and MLX at 163. Seven hundred thirty divided by 163 is about 4.5. That division is arithmetic on TokenPost’s two figures. TokenPost says the speeds are the company’s, from one device and one task, not an independent rerun. CryptoBriefing also reports a company claim of up to 730 tokens a second on MacBooks, and describes Woof as under 2.5 gigabytes. Two and a half gigabytes is about the size of a short movie. That size line is CryptoBriefing’s and TokenPost’s account of Conway. It is not an outside weigh-in of the file.

The quality claim, kept as the company’s. Conway’s other update says that six months ago the top AI was Claude Opus 4.6, and that Conway’s model, Underdog 27B, beats it while running on a Mac or an iPhone. 27B is the name’s reading: about 27 billion parameters, several times Woof and still meant for a device a person owns. The homepage sentence does not list the tests. CryptoBriefing says Conway reports a win on certain benchmarks, not across the board. TokenPost says it could not find an independent rerun, and that the comparison should not be read as a win on every test. The product site’s own answer is quieter. Underdog says it trains its own models to be fast and power-efficient on the computer, and that it may not top the leaderboards. A leaderboard is a public ranking of models. Those lines are Conway’s and the outlets’ reading of them.

What the product site says is available today. underdog.ai says Underdog is in an invite-only beta and getting better every day. A beta is a test version, not a finished public launch. The page says it is 100 percent local and 100 percent private, a small model that runs on the computer rather than in a data center, at less than 4 gigabytes. Less than 4 gigabytes is the product page’s size for what sits on the machine. It is a different sentence from the 2.5 gigabyte line the outlets attach to Woof. The page says a Mac with Apple silicon can download it today. Apple silicon is the chip family in current Macs, the M-series. iPhone, iPad, Windows, Android, and Linux are listed as coming. A person can tick those and get one email per build when it ships. a16z’s post describes intelligence on the devices a person owns, including a phone. Conway’s 27B line says the model runs on a Mac or an iPhone. The product page’s download, today, is the Mac.

What it says it will do, and what stays on the machine. The page says Underdog can use the browser to book flights, reserve tables, order groceries, shop, and cancel forgotten subscriptions, and that the person approves before it books, buys, or cancels. Mail connects directly to Gmail and Outlook, and the credentials stay on the computer. Meeting notes are transcribed and summarized on the device. The page says audio and notes are not sent to a data center for that processing. Dictation turns speech into text in any app, on the device, and the audio is not streamed out. A calendar connects to Google Calendar and Outlook Calendar. A data-broker tool looks for the person’s information on those sites and sends removal requests. Expense tracking and tax preparation are marked coming soon. Saved memories stay on the Mac, and the person can see, edit, or delete them in Settings. Each account gets its own AES-256-GCM vault. That is a standard lock for stored data. The key lives in Apple Keychain, secured by the Secure Enclave, and the page says the key never leaves the device. The Secure Enclave is a separate chip on the Mac that holds secrets apart from the main processor. The page also says summaries, drafts, and answers come from a model on the computer, and that content is not sent to a cloud AI service for processing. Mail the person sends, and bookings they approve, still go to the recipient or the website. Those lines are Underdog’s. They are the company’s description of the product. They are not an outside security audit.

The picture is a product card. A navy field, the word Underdog in white at the top, and a white cartoon dog outlined in cyan, with a black nose and a pink tongue, holding a glowing cyan bone. Two labels read PRIVATE and ON-DEVICE. It is the mascot from Conway’s Underdog preview, cropped to the card. The frame does not print a calendar date, a dollar figure, or a photograph of a phone.

In plain terms, Andreessen Horowitz said on Friday that it is leading the first funding round for Sigil Wen’s Conway Research, the lab behind Underdog, a personal assistant built to run on a person’s own devices. The firm did not say how much the round is. CryptoBriefing names Khosla Ventures and Hummingbird VC among other investors. TokenPost says it could not confirm those two. Wen, as a16z describes him, is a 2025 Thiel Fellow who helped Naval Ravikant launch Airchat. Conway says it builds models and the engines that run them for machines people already own, and it publishes a peak of up to 4.5 times Apple’s MLX on one function-edit test with Woof 4B. That speed is the company’s. The product site says a Mac with Apple silicon can run the invite-only beta today, with the iPhone and other builds still coming, and that the model’s work on mail, notes, and dictation stays on the computer.