KAN creator Liu Ziming raises back-to-back rounds for 'meta-model' startup

"Must model design always be done by humans?" Early this year, Liu Ziming — then a Stanford postdoc about to join Tsinghua — posed the question to himself. He then ran a two-month "human experiment": every day he took a random dataset and forced himself to predict the training curve before running it. Around day 50, something clicked. If human intuition can be trained, why not train a "model of models" that owns that intuition? That idea became Meta-Loop Intelligence, founded this May.

Two rounds in half a year, hundreds of millions raised

According to 36Kr's Emergence reports, Meta-Loop has closed seed and seed-plus rounds totaling several hundred million yuan. The seed was led by Shunwei Capital with Future Industry Fund, Xinglian and Junli following; the seed-plus drew Shanghai Semiconductor Investment, 360's strategic arm, Shoucheng Capital, Baidu Ventures, Hillhouse-affiliated Hillhouse Xing and Yunshi Capital. Founder Liu Ziming, born in 1998, is now an assistant professor at Tsinghua's Institute for AI, holds a PhD from MIT, did postdoctoral work at Stanford, and studied physics at Peking University. His team averages just 22 years old.

Not a language model — an intuition model

Meta-Loop defines its mission as AI for AI: the first AI is a meta-model that designs all models; the second is what it designs. Unlike the RSI (recursive self-improvement) wave built on LLMs, Liu argues natural language covers only 30% of the world: "Intuition is the reflex of experience; natural language is a poor carrier for strong intuition." The meta-model targets the remaining 70% of the uncompressed numerical world — AI for Science, world models, and the "science of AI" itself. Near-term applications include industrial inspection, quant finance, weather and power forecasting, and autonomous driving; a general-purpose Meta-Model 1.0 arrives next year.

"What I've always wanted is to write the Newton's equation behind AI — to find something interpretable." — Liu Ziming

The researcher-founder wave

Meta-Loop is not alone: RSI paper author Shunyu Yao was reported to be starting up at a ~$2 billion valuation, as AI-researcher founders heat up. The common thread: turning "using AI to research AI" from papers into products. If meta-models work, model design could be industrialized the way EDA tools industrialized chip design — perhaps the next source of inference-cost decline. The risk is equally clear: intuition models lack mature evaluation systems, and commercial validation is still early.

(Information synthesized from 36Kr, InfoQ and Huxiu reports)