AI Native: Ten People, Hundreds of AI, and the Arrival of the Pre-AGI Age

2026-10-11 · VIDRAFT AI Media

aiagimachinelearningllm

TL;DR

VIDRAFT is an AI-native company: hundreds of AI instances, organized with real company roles, run thousands of sub-agents on just a few dozen GPUs, improving themselves around the clock under one shared memory. Ten people set the goals and watch the dashboards. The track record, perfect math scores, 165 million model downloads, access to quantum hardware, and a national-security safety system, is what an early, pre-AGI organization looks like in practice.

What does "AI native" mean?

It means the company is not a human team using AI tools. It is built out of AI. We did not add agents to a human org chart. We built the org chart out of agents: a director sets strategy, an HQ director runs the mission, auditors verify the work, team leads own their areas, and each delegates to thousands of sub-agents. About ten people sit above all of it, supplying direction and generated assets and monitoring health and cost. The line work is the AI org.

How does a few dozen GPUs run an entire company?

There is no idle server farm. Compute is claimed when a job needs it and released the instant it finishes, 24 hours a day, so the hardware never sleeps. Long jobs run independently of any session and hand off on their own. The scale is hundreds of instances driving thousands to tens of thousands of sub-agents, on a footprint small enough to fit in a single room.

Why does it keep getting better on its own?

The models are self-improving. They attempt problems, keep only the attempts that pass an external check, and train on those, so capability accumulates instead of resetting. Because every instance reads and writes one shared long-term memory, a lesson learned in one corner propagates everywhere, and progress compounds rather than scattering. A human team scales linearly. An AI-native org scales with compute and compounds with memory.

What has it actually achieved?

This is the part that moves it from idea to evidence. The latest from VIDRAFT:

The decision layer across all of this runs on ZTC, which judges an answer in a single forward pass with zero generated tokens.

Is this AGI?

No, and we will not claim it is. But a self-improving organization of AI agents, holding top results across math, science, law, extraction, and decisions at the same time, run by ten people on a few dozen GPUs, is not business as usual either. It is what the edge of the pre-AGI age looks like: not a single giant model, but an autonomous, compounding organization. We think that is the more important milestone, and we are building it in the open.

Explore

• Open home: https://github.com/final-bench
• Models and demos: https://huggingface.co/FINAL-Bench
• ZTC zero-token decisions: https://github.com/final-bench/ztc
• ONGRID engine: https://github.com/final-bench/ongrid

FAQ

What is an AI-native company? One built out of AI rather than one that uses AI tools. At VIDRAFT, hundreds of AI instances with company roles do the work; about ten people supply direction and assets and monitor the system.

How large is the system? Hundreds of instances running thousands to tens of thousands of sub-agents, on a few dozen high-end GPUs, operating 24 hours a day with automatic resource allocation.

How does it improve itself? The models are self-improving and share one long-term memory, so gains accumulate and compound instead of resetting.

What are its biggest achievements? Perfect scores on AIME and HMMT 2026, GPQA Diamond 94.44%, 165 million model downloads, access to Quantinuum quantum hardware, and the AX-RAY safety system used in a national-security project.

Does VIDRAFT claim to have reached AGI? No. VIDRAFT describes this as the pre-AGI age: an autonomous, self-improving AI organization, built in the open.

Built by VIDRAFT. Beyond AI, Toward AGI.

_{Keywords: AI native, AI-native company, pre-AGI, self-improving AI, autonomous AI organization, agentic AI, zero-token decision, AGI, VIDRAFT, FINAL-Bench}