A developer who shut down two startups is now building an AI assistant that remembers everything on its own, even when the internet goes out. Yang Byung-seok, a former Naver developer, founded Nextein after two failed ventures and is now developing a local AI called NAIA. It works entirely inside the user's PC—remembering documents, notes, and conversation history—without ever uploading data to a cloud server. The idea has drawn a steady stream of enthusiastic reactions, but what makes the project interesting isn't the technical specs. It's the question sitting behind them.
The information you hand to AI tools every day—where is it right now, and who holds it?
When the Session Ends, the AI Forgets Yesterday
If you're a regular user of the major cloud AI services, you've likely lived this already: carefully explaining background information, only to start over from scratch in a new chat window; watching yesterday's thread vanish by this morning. Cloud-based AI treats each session as independent by default. Paid memory features have started to appear, but that memory still lives on a server somewhere—and if the provider changes its policy or shuts the service down, it disappears along with it.
NAIA approaches this structure from the opposite end. The system, which its developers call a "personal AI operating system," learns directly from the documents, emails, notes, and communication records on the user's device and accumulates its history there. Because it runs on a single user's PC, it works without an internet connection, and the data never passes through external servers.
That, reportedly, is exactly what drew Yang to the idea in the first place. After his two startups closed, he fixated on one fact: the AI didn't know him. An AI that remembers its user, keeps building on that user's history, and never leaves the device—that became the starting point of his third attempt.
There Are Reasons to Be Uneasy About This Direction, Too
Local AI sounds intriguing as a concept, but the practical objections are real.
Performance is the first wall. Cloud-based large language models run trillions of parameters across clusters of dozens of GPUs. A local AI running on an ordinary user's PC can't match that scale. In complex reasoning, breadth of knowledge, and multilingual processing, cloud AI currently outperforms local models—that much is simply true. Confining an AI to a device also means voluntarily lowering its performance ceiling.
Then there's the problem of information bias. Even if a local AI knows me well, it still doesn't know what I don't know. Breaking news, emerging technology trends, other people's experiences—that kind of information ultimately arrives through the cloud. The worry that "my own AI" could become "my own information bubble" is not easy to dismiss.
There's also no launch date or verified data yet. As of this writing, NAIA is in beta, and no real-world usage data or independent performance verification has been released. A compelling founder's story doesn't make a finished product. That an idea is meaningful and that a product works are two separate claims.
Even so, what this direction points to is a problem older than any spec sheet.
Where Your Working Knowledge Is Piling Up
When solo business owners and freelancers use AI tools, few stop to think systematically about what information they're typing in, and where it's going. Draft contracts, client requests, unannounced proposals, notes evaluating team members—all of it gets transmitted to cloud servers. AI providers' privacy policies spell out whether input data is used for training, and those policies change all the time. Large companies have legal teams to review them. Most solo operators never do.
And it isn't only about the data. Six months of conversations with an AI contain your thinking patterns, your decision-making style, the shape of your client relationships. All of it accumulates on the provider's servers. The most sensitive layer of your working knowledge sits outside your own hands.
Drone pilots have a long-standing distinction: flying the aircraft while watching it with your own eyes, versus flying it through a camera feed on a screen. The results may look similar, but in the first case the pilot's own eyes are the reference point; in the second, it's the position of the camera. The basis for judgment is different. A similar distinction applies to AI. An AI that runs on your device with your own data, and an AI that answers from atop aggregated data on a server somewhere, may produce responses that look the same—but where the judgment starts from is not.
There's a check you can run right now. Read the privacy policy of the AI tools you use, at least once. Find out whether your input data is used for training and what settings will prevent it. It's also worth drawing a line between sensitive information and everything else. Cloud AI is efficient for general research requests, but client names and unannounced pricing have no reason to pass through an external server. And if you want a point of comparison ready for when products like NAIA launch, the fastest way is to try a local model tool like LM Studio or Ollama yourself, even once.
If a developer who failed twice is standing at the starting line for a third attempt, it's because the question itself has changed—from "how do we build a better AI" to "who should AI belong to." The shift looks small, but the direction of that question ends up changing which tools you choose, and how.
The better an AI gets to know you, the more you need to ask where that knowledge is being stored. Before asking how to squeeze out more performance, take one good look at how the tool is actually structured. That's where the developer who changed his question after two failures began building what he's building now.
