VIDRAFT's Pre-AGI Thesis: Quantum at the Micro-Scale, Physical AI at the Macro-Scale

VIDRAFT's Pre-AGI Thesis: Quantum at the Micro-Scale, Physical AI at the Macro-Scale

The Pre-AGI thesis

VIDRAFT, a Korean AI startup founded in 2024 and based in Seoul, organizes its roadmap around an idea it calls Pre-AGI. The thesis splits intelligence across two scales. At the micro-scale, the substrate is quantum computing: the smallest units of computation and physical state. At the macro-scale, the substrate is physical AI: robots and world models that act in and predict the real environment. The argument is that integrating both scales, rather than scaling one monolithic language model, is a more tractable path toward general capability. Pre-AGI is the deliberately bounded, buildable step before anyone earns the right to say AGI.

What sits behind the badges

The homepage shows live ranking badges. Each headline number carries a specific scope, and it is worth stating those scopes plainly:

• #1 on the K-AI leaderboard. This is leaderboard scope, not a claim of global supremacy.
• GPQA Diamond 90.9%, reported as #3 globally and #1 in Korea on that single benchmark, base-only.
• 15 first-place finishes across Polaris drug-discovery tasks.
• ~1.18M cumulative downloads across the model ecosystem on Hugging Face.

The infrastructure behind this is small by frontier standards: roughly 24 GPUs (B200x16 plus H200x8). That constraint shapes everything. Rather than from-scratch pretraining at frontier scale, the work leans on evolutionary merging of open models, on-device models, and safety benchmarks published through the FINAL-Bench organization, inside the broader GINIGEN ecosystem.

Reading the download number honestly

The ~1.18M figure covers foundation models plus community copies and quantizations; counting image LoRAs pushes the whole ecosystem to ~1.55M. Most of that volume is not on first-party repos. A single community copy, ansulev/Darwin-9B-NEG, accounts for ~561K downloads, while the original FINAL-Bench/Darwin-9B-NEG sits near 2K. The copy out-downloaded the original by roughly 258x. Much of the remaining volume lives on third-party mirror and quant accounts such as mradermacher and bartowski. That is genuine, organic community adoption, but it would be dishonest to present it as first-party traffic. One copy drove much of the total.

Why the two-scale framing matters

Treating quantum and physical AI as endpoints of one spectrum is a bet about where returns come from. Language models are the visible layer, but the public artifacts span benchmarks, on-device runtimes, and drug-discovery leaderboards. The Pre-AGI label keeps those threads pointed at one goal without overclaiming that any single model is close to general intelligence.

Honest scope

Every headline here is scoped. The K-AI ranking is leaderboard-scoped and self-selected by entrants. GPQA Diamond 90.9% is a single-benchmark, base-only result: strong, but not a general capability claim. The Polaris placements are task-specific. The download totals mix first-party repos with community mirrors, quants, and one dominant copy, so they measure ecosystem reach rather than centralized traffic. Quantum and physical-AI work is research-stage. Read the badges as evidence of focus and community pickup, not as proof of frontier parity.

More: https://vidraft.net