Personal AI heats up as a16z backs a screenless 'AI computer' from a 19-year-old founder

Ghost announced last week that it had raised an $11 million seed round led by a16z, while launching its first product: a private AI computer called "Core." The device is priced at $3,499, the first batch is already open for pre-order, shipping is expected in the last week of October, and the official website currently shows it as sold out.

The other three co-founders who started the company with Javed are 19, 24, and 23 years old respectively. What this team wants to sell is not a faster computer, but an AI assistant that is online for the long term and increasingly understands its owner: it connects apps, files, and smart devices, gathers scattered information, remembers what is happening, and then looks for opportunities to help. Users reach it through a phone or companion app, while the computer stays in place and keeps working.

It should be noted first that the above funding amount, price, shipping time, and sold-out status all come from Ghost or the reporting party's own account, and there is currently no independent third-party verification. The report itself also comes only from a single secondhand reposted source, whose original reference is TechCrunch, and it provides no first-party primary material.

Core has no screen, so users do not have to sit in front of it, nor do they have to keep giving it commands. According to Ghost's vision, users do not have to open an app and give a command every time: Core continuously receives information the user allows it to access, remembers previous instructions, follows up on tasks when needed, and then contacts its owner through the phone.

The specific scenario given in the report is a meeting: today's AI computers can basically summarize attachments according to instructions, while Ghost hopes to go further, letting Core combine relevant emails and calendars, organize meeting materials, watch for schedule changes, and update arrangements after obtaining authorization.

Javed also described a more proactive form: Core notices its owner's resting heart rate has risen, then finds from computer usage records that the owner stayed up late the night before rushing a project, and combined with information that San Francisco is in flu season, proactively reminds the owner that something may be going on with their body. He believes a computer assistant should notice key signals in advance, before its owner does.

These capabilities are currently only Ghost's vision and description; the report gives no actual performance, review, or user feedback data for Core. Whether it can truly achieve proactive reminders and cross-app task advancement has no empirical evidence yet.

Core's selling point is stated very concretely: the model and personal memory run on the device, with no subscription fee. Ghost's official website says that personal intelligence and its memory should belong to the user, and people should not first hand their lives to an AI company and then pay by usage to get help back.

This seems to partly explain the $3,499 price. Core has an NVIDIA RTX PRO 4000 Blackwell SFF graphics card built in; since the model is to run at home, Ghost must hand the corresponding computing power to users as well. Compared with using cloud AI monthly, consumers bear a larger upfront expense.

But "local" does not equal "offline." Ghost's privacy policy states clearly: conversational AI and memory are handled by Core, but online functions such as search and messaging still send necessary information to external services. What users ultimately have to trust includes both a machine that stores data and an assistant that has been granted permission to act.

The permission boundary is likewise still under construction. Javed says users can decide how much autonomy to give Core; in addition, the Ghost team is still developing a firewall that checks the assistant's outbound requests and limits how it uses accounts and passwords. This firewall is still in development and has not yet been implemented.

From these constraints it can be inferred (this is an inference): Core's "private" nature is layered — memory and conversation stay local, but once networked actions are involved, data still flows outward, and what users really need to judge is "which things can be handed over."

The more realistic question is whether agentic AI needs a new computer. The report mentions that when talking about intelligent assistants, it is hard not to think of OpenClaw, which was hot earlier this year, and the Mac Mini, which was in short supply. According to the report's description, OpenClaw already has capabilities such as memory and proactive reminders, can call both cloud models and local models, and also provides a desktop app and installation guidance.

Another route enters from the software side. The report says the representative products in this wave are Meta's Muse and Instinct, valued at over ten billion dollars; their characteristic is relying on the existing software and hardware ecosystem, entering through users' personal data and long context, and not venturing into hardware development.

Javed does not avoid the comparison. His explanation is that the Mac mini is a general-purpose computer that can be configured to run AI, while Core is designed around agents from hardware to software. In his view, what really needs to be done well is how the agent understands life, how it remembers and acts, and when it proactively contacts the user.

It should be noted that the specific capabilities and data of comparison products such as OpenClaw, Muse, and Instinct have not been confirmed by first-party sources in the report, and are presented here only as the reporting party's claims.

Some support it, while others ask in response: what exactly is the difference between this and building your own home server. The report does not give a conclusion, only presenting both sides of the argument.

Judging from its product positioning, Core is more likely to first attract people who use AI frequently and are willing to pay for privacy and convenience. For this type of consumer, what really determines whether they keep Core may be some very small things: whether it can remember the user's preferences, whether it can stop reminding after something is done, and whether it can ask first when uncertain.

Putting the model locally helps reduce dependence on remote AI services, but it cannot replace these judgments. Whether users are willing to bear the $3,499 upfront expense for privacy and convenience is currently uncertain.

The report also does not disclose Core's sales, production capacity, or supply chain details; how the boundary between long-term memory and autonomous action permissions is defined is only said in the report to be decided by the user, with details unclear.

When talking about competition from big companies such as Apple, Javed appears calm: giants have all kinds of advantages for building products, but that does not mean the product already exists. As for whether Ghost succeeds or fails, his statement is also very direct — if it cannot make a product people love, failure should not be blamed on competition from giants. These entrepreneurs are all under 25, but what they want to redefine is the computer and the way people and computers interact.