Intelligent World 2035:
Turning Vision into Action

Existing AI models have insufficient explainability, physical understanding, causal reasoning, long-horizon reasoning, self-evolution loops, and multimodal representation. To make AGI a reality, we need to step up research in three directions: language intelligence (the cognitive framework for AGI), embodied intelligence (the groundwork for physical understanding), and scientific intelligence (enabling AI systems to autonomously make scientific discoveries). These directions, in conjunction with symbolism techniques (which make up for the cognitive deficiencies of statistical learning), will ultimately converge and give rise to unified, logically-consistent, and evolvable general world models.

As AI agents evolve from conversational tools into autonomous executors, storage and memory systems are becoming critically important. Logical distortion, dynamic scenario adaptation, data transportation, and security mechanisms are all key challenges. The pathways include developing a causal inference engine, simulating human associative memory and forgetting, advancing near-data computing, and integrating native security systems. These are not simply iterations of storage technology. Rather, they represent a leap from data management to cognitive intelligence. Establishing an accurate and trustworthy memory foundation is the key to agents becoming our decision-making partners.

The Tau Scaling Law replaces geometric scaling with time scaling. To implement this law, we must address two major constraints. First, LogicFolding amplifies vertical interconnect parasitics and inter-water variations, meaning traditional Electronic Design Automation (EDA) toolchains struggle to cope with 3D topology. Second, energy efficiency control spans 12 orders of magnitude, requiring a cross-level trade-off spanning devices to systems. Given these constraints, the entire set of design methodologies for folding must be reconstructed, and the Keep-Out Zone (KOZ) constraint and issues related to statistical modeling need to be addressed. In addition, advancements must be made in energy efficiency technologies such as memory semantic interconnect and near-packaged optics. By addressing these two major constraints, we can realize the 100-fold performance improvement of the Tau Scaling Law.

The evolution of smart devices relies on simultaneous breakthroughs in the compute foundation and interaction paradigm. We can explore an elastic converged computing architecture that is physically decoupled, trustworthy, and secure in order to virtualize scattered and heterogeneous device compute into a trustworthy compute resource pool capable of supporting unified scheduling across all domains. Furthermore, we can build a virtual avatar capable of cognitive symbiosis and autonomous evolution, allowing devices to make the jump from passively executing commands to behaving as intelligent companions that proactively understand user intent and collaborate in user decision making.

As computing demand continues to surge, GW-scale AI data centers are an inevitable trend. However, power supply and thermal management remain core bottlenecks to their deployment. Breakthroughs are thus required in wide-bandgap direct conversion, high-voltage DC power supply, chip-embedded microfluidic cooling, and liquid nitrogen cryogenics. Whether we can break these power and thermal limits will determine how far AI computing capacity can go.