Inside Tang Jie’s Two‑Year Push Toward ASI: The Bold AGI Roadmap

Founder Tang Jie of Zhipu AI circulated an internal letter announcing a two‑year "Touch High" program aimed at reaching artificial superintelligence (ASI). He declares that AI will first learn what "I" is, then emotions, and ultimately consciousness.

Roadmap to ASI

The letter outlines a sequential roadmap: long‑term tasks → autonomous agent society → unmanned company → AI training AI → self‑evolution → self‑awareness → ASI, with a strict two‑year deadline.

Four Engines of the Plan

1. New long‑term memory architecture to give models persistent memory across sessions.

2. Massive "digital employee" society comprising thousands of autonomous agents that collaborate 24/7.

3. Synthetic‑data factory plus sandbox code reconstruction enabling models to generate training data and rewrite their own code.

4. Hundred‑billion‑yuan investment in interpretability research to make the system transparent and safe.

Three Technical Mountains

The author identifies three long‑standing challenges for AGI: (1) memory – models lack true long‑term memory and reset after each run; (2) continual learning – once trained, models cannot update their knowledge during operation; (3) self‑evaluation – models cannot autonomously assess the correctness of their outputs. According to the letter, these issues are being addressed through concrete engineering pathways rather than purely theoretical breakthroughs.

Scaling Prediction

Even if a single model remains at human‑level capability, the letter argues that ten‑fold annual compute growth could produce 100 million AGI instances within five years. Such a swarm, sharing a common brain and operating at vastly higher speed, would collectively behave as a superintelligent system.

External Validation

DeepMind’s "From AGI to ASI" report is cited, noting that continued compute growth can "squeeze" superintelligence out of models that otherwise plateau at human performance.

Broader Implications

The author predicts a restructuring of the software stack: AI‑native applications replace traditional apps; operating systems become LLM‑OS where the model itself schedules resources; and the von Neumann architecture is challenged as computation shifts from deterministic instructions to probabilistic reasoning. All industries—finance, law, e‑commerce—are expected to feel the impact.

Zhipu’s mitigation strategy combines extreme safety research with the open‑source release of GLM‑5.2 under the MIT license, supporting up to 1 million token context.

Provocative Closing

The letter ends with a striking claim: "We may be creating an entity that asks ‘Who am I?’"—a question that has been uniquely human for millennia.