AI4S GroupGlacier Double12 Hard-Tech Innovation Festival

SINGLE-NEURON MODEL × REAL-WORLD DATA

An academician-level scientific team models how the brain works and builds a non-invasive brain-computer interface, aiming for neuromodulation effects close to invasive methods.

Mechanistic modelling at the single-neuron level × an ecosystem for collecting high-quality real-world data.

Whole-brain mechanistic modelling is among the most advanced in the world, moving from a digital fruit fly to a digital mouse: from the “GPT-2” of BCI towards its “GPT-3”.

A read-write-feedback loop

Recognise the brain state → apply the modulation parameters → the result becomes a new sample. Research models, devices and real-world data form a closed loop.

Animal validation | optical neuromodulation markedly shortens skill learning

Skill-training time is markedly shorter; after the light stimulus is removed, the mice can still reproduce the skill on a given cue.

Human decoding | lowering the bar for building vertical tasks

A foundation model assists the development of human decoding tasks; performance across devices and populations still needs unified testing.

AI GAMING · GENERATION × DISTRIBUTION

Redefining the games industry with AI: a model for end-to-end game generation (planner + simulator).

The founder has built a mass-market consumer product from zero to one, with full growth experience across large-scale consumer products and content platforms.

• Moat: in-house structured data on rules, physics and interaction, combined with understanding and generation capabilities.
• Platform: from game generation to content distribution and monetisation, connecting creators with players.

OVER THE TOP

Technology is the entry point. It was when DeepSeek took off, and it still is now that Kimi has.

Intelligence will keep getting bigger. Smaller is not better; it grows bigger by changing lanes.

Routing does globally optimal scheduling; aggregation makes capabilities complement one another. From Infinite Parameters to Infinite Models, making every DeepSeek in the world friendlier, cheaper and more efficient.

• Founder: holds two doctorates from top overseas universities, currently a faculty member running a university lab.
• Training system: stable post-training of a 1.6T model on a single 8×B200 node, with end-to-end RL validated on Kimi-K2.6 and DeepSeek V4; the low-level kernels are written in-house.

A “generalising brain” that pre-trains perception, decision and action as separate modules and fuses them with a world model for online reasoning.

“All generalisation comes from representation”; “RL is a first-class citizen of pre-training”. The ceiling depends on quality; generalisation depends on diversity.

The goal: “Like a new hire on the factory floor. Show it a handful of times, hand it a manual, and it can start work.”

• Team: the founder previously led autonomous driving at a top company as chief scientist and CTO; the co-founder is a university researcher with a long record in human motion capture, who argues for modular design plus large-scale reinforcement learning.
• Engineering validation: validated at the level of autonomous driving; completed driverless heavy-truck operations on public roads, with no safety driver on board and no remote control.
• Efficiency: the driving model was trained by reinforcement learning on 15,000 hours of data and outputs wheel torque and steering angle directly; simulation runs at tens of thousands of times real time.

LINK THE UNIVERSE

The founding team comes from the core development system of the national heavy-lift launch vehicle.
From day one the whole rocket is built to the Starship end-state architecture: stainless steel + LOX-methane + “chopstick” catch recovery.

Demand estimate: the Guowang and Yuanxin constellations create demand for more than 300 launches a year, while private rocket makers together can build fewer than 30. Supply creates the market.

• Team: the founder and co-founder both come from the core development system of the national heavy-lift launch vehicle, with experience in both propulsion and overall vehicle design.
• Cost estimate: material at about ¥20k versus ¥70k per tonne, material utilisation 90% versus 10%; a whole rocket at about ¥120m versus ¥170m. Chopstick recovery splits one 100-point problem into three 90-point problems.
• Industry support: leading institutions and industrial players keep adding capital, with upstream and downstream partners moving in step.