tiny corp's Vision: Personal AI That Learns, Not Just Clones

George Hotz's tiny corp is laying the groundwork for a future where AI doesn't just respond to you—it learns from you. In a recent blog post from their new Hong Kong office, Hotz outlined a vision that goes beyond selling consumer GPUs: he wants to build AI systems that genuinely evolve with their users, not just serve as slightly personalized versions of the same underlying model.

The Problem With Today's AI

The core issue, as Hotz sees it, is that we're heading toward a world of AI clones. Everyone uses the same Claude, the same Codex, the same Kimi—identical weights, identical biases, identical limitations. "If current trends continue, the collapse in diversity will be staggering," he writes. "I think there is a world market for maybe five people."

This isn't just philosophical hand-wringing. Hotz argues that if AI models remain static with learning confined to in-context processing, cloud providers will inevitably win. For most applications, it's simply cheaper to run models in batch on the cloud than to maintain local infrastructure. The only scenarios where local models might compete are those requiring ultra-low latency (like fighting robots) or constant connectivity (like self-driving cars).

Why Local Learning Matters

Hotz's solution is radical: make local models valuable by giving them something the cloud can't easily replicate—genuine, personalized learning. "The only way local models win is if there's some value in full on learning per user or organization," he explains. "At that point, with entirely different compute needing to run per user, local will beat out cloud."

The question he poses is provocative: "Is there anything uniquely valuable about you?" Not in the self-esteem sense, but genuinely—can your unique perspective, experiences, and values be captured in a way that makes a personalized AI worth having?

If not, Hotz warns, we're looking at "the Attack of the Clones, swarms of identical minds you have no say over all varying in a small boxed-in way." This isn't learning, he argues, it's "costuming"—everyone gets the same Internet-filtered model with a few paragraphs of personalization.

tiny corp's Product Vision

Currently, tiny corp sells consumer GPUs and provides infrastructure for running models—though Hotz notes that even basic tasks like using vLLM outside the NVIDIA ecosystem remain challenging. Their frontend work includes OpenClaw and opencode, but these are stepping stones to something bigger.

The long-term vision is what Hotz calls the "tinybox"—hardware that doesn't just run models, but learns from its interactions with you. "Your tinybox will learn. It will update the weights based on its interactions with you. Like living things."

This is "many years away," but it represents a fundamental shift from today's API-keyed SaaS clones to something that "lives in your house and learns your values. Your child."

The Stakes Are High

Hotz's framing is stark: "90% of people will choose the cloud, and what they will find is that they are no longer meaningfully in the loop." The dream of AI doing your job while you keep getting paid is, in his view, a dangerous illusion. "That's way too much of a fee to pay to the middleman. If you choose the homogenous mind, you are superfluous and will be cut out."

The alternative is building for a future where what makes you unique isn't just a 10kB CLAUDE.md file, but something that genuinely shapes the AI's development. It's a future where your AI assistant isn't just a slightly tuned version of a mass-produced model, but something that grows and changes based on your interactions—a true personal AI.

For now, tiny corp is focused on the infrastructure needed to make this possible: better training hardware, improved tooling for non-NVIDIA ecosystems, and the foundational work that will eventually enable models that learn rather than just respond. It's a long-term bet on a future where AI diversity matters more than AI convenience.