Skip to content

Essays

  • 4 min read

Property Hunters: China Regions Mine Ledgers for Dormant Riches in 'Idle' Assets

Source: South China Morning Post

Chinese provinces are scouring their balance sheets for idle state-owned assets, turning dormant infrastructure and vacant buildings into sources of revenue as land sales — the traditional lifeblood of local government finance — remain depressed by the prolonged property downturn.

The Context: Land Revenue Collapse

For decades, Chinese local governments relied on selling land-use rights to property developers as their primary source of revenue. Land sales funded everything from infrastructure to civil servant salaries. The property crisis that began in 2021 shattered this model:

  • Land sale revenues fell by over 30% in 2022 and continued declining through 2025
  • Developer defaults left vast tracts of partially developed land in limbo
  • Local government debt ballooned, with some estimates putting total off-balance-sheet liabilities above ¥60 trillion ($8.3 trillion)

The result: local governments needed a new source of fiscal oxygen.

The Solution: Asset Revitalisation

Rather than selling future land, governments are now looking to monetise what they already own. The strategy, championed by the State Council, is called "revitalising idle state assets" — and it's spreading fast.

Jilin Province Leads the Way

Jilin, in China's northeast rust belt, has set an ambitious target: redeploy ¥100 billion ($13.8 billion) in inefficient assets. Governor Hu Yuting captured the strategy in a single line: "Transform dormant resources into development capital."

Jilin's playbook includes:

  • Underutilised public infrastructure — converting empty government office parks into commercial spaces
  • Vacant administrative buildings — repositioning former government compounds as affordable housing or tech incubators
  • Tourism assets — monetising state-owned scenic areas and heritage sites that had been operated below capacity
  • Industrial land — repurposing abandoned factory sites from the state-owned enterprise era

Sichuan and Chongqing

Other jurisdictions have followed Jilin's lead. Sichuan province and the municipality of Chongqing have launched their own asset sweeps, targeting everything from unused school buildings to underperforming toll roads.

The Mechanics: Asset-Based Financing

Asset revitalisation isn't just about selling idle buildings. The more transformative approach involves using these assets as collateral for new financing:

Mechanism Description
Asset-Backed Securities (ABS) Pool idle assets into securities, sold to investors to raise immediate cash
Infrastructure REITs List income-generating infrastructure on stock exchanges (China's REIT market has grown to ¥100+ billion)
Asset-Light PPPs Private partners operate and upgrade idle assets in exchange for revenue-sharing agreements
Direct Sales Auction underperforming assets to private buyers, often at significant discounts

This shift from land-sales dependency to asset-based finance represents a structural transformation of China's local government funding model.

Risks and Challenges

Critics point to several obstacles:

  • Valuation opacity — many state assets are carried on books at historical cost or arbitrary values, making true market pricing difficult
  • Political resistance — selling or leasing state assets can be seen as "loss of state assets," a politically charged accusation
  • Absorption capacity — the market can only absorb so many asset sales before prices collapse
  • Quality concerns — many "idle" assets are idle for good reason: they're in the wrong location, poorly constructed, or economically unviable

Broader Implications

China's asset revitalisation drive offers a window into the country's post-property-crisis economic model. The old engine — land sales funding urbanization — is sputtering. The new model relies on making existing assets work harder. It's less exciting than the construction boom, but potentially more sustainable.

For investors, the trend creates opportunities in distressed state assets being sold at discounts. For economists, it's a stress test of whether China can shift from an investment-led to an efficiency-led growth model. For local governments, it's survival.

Bottom Line

China's idle asset hunt is a fiscal necessity born of the property crisis. Whether it becomes a genuine reform — unlocking the value of trillions of yuan in underperforming state capital — or merely a stopgap that papers over deeper structural problems depends on execution. Jilin, Sichuan, and Chongqing are the early test cases. The rest of China is watching.

  • 3 min read

Deep Understanding

Source: arXiv:2603.21852

A new paper circulating on arXiv (2603.21852) undertakes a deep exploration of understanding mechanisms in artificial intelligence systems — pushing forward our grasp of how machines can develop deeper comprehension of complex problems rather than merely pattern-matching their way to plausible answers.

The Problem: Pattern Matching vs. Understanding

Modern large language models are undeniably capable. They can write essays, solve math problems, generate code, and engage in sophisticated dialogue. But a persistent question haunts the field: are these systems actually understanding anything, or are they simply extremely sophisticated pattern recognisers?

The "Deep Understanding" paper tackles this question head-on. Rather than philosophical speculation, the authors propose operational definitions and experimental protocols for distinguishing genuine understanding from superficial pattern completion.

Key Contributions

A Framework for Understanding

The paper proposes a multi-dimensional framework for evaluating understanding in AI systems:

Dimension Description
Generalisability Can the system apply a concept in contexts dramatically different from its training data?
Causal Reasoning Does the system grasp cause-and-effect relationships, or just statistical correlations?
Counterfactual Sensitivity Can the system reason about what would happen if conditions changed?
Compositional Generalisation Can it recombine known concepts in novel ways?
Explanation Quality Are the system's explanations faithful to its internal reasoning process?

Experimental Results

The paper presents experiments on state-of-the-art models (including GPT-4 class systems and open-weight alternatives) that probe each dimension. The results paint a nuanced picture: models show flashes of genuine understanding in some domains while remaining brittle pattern-matchers in others.

The Understanding Gradient

Rather than a binary "understands / doesn't understand" judgment, the authors propose a gradient view — understanding exists on a spectrum, and different models occupy different positions depending on the domain and the specific capability being tested.

Implications

If correct, the framework has significant implications:

  • AI Safety — Systems that genuinely understand concepts are safer (they can reason about edge cases) but also harder to control (they can form unexpected inferences)
  • Benchmark Design — Current benchmarks may measure pattern matching, not understanding. New benchmarks based on the paper's framework could change how we evaluate progress
  • Architecture — If understanding requires specific architectural features (e.g., causal attention, world models), this points to concrete research directions

Open Questions

The paper raises as many questions as it answers:

  • Can understanding emerge from scale alone, or does it require architectural innovations?
  • Is understanding in AI systems fundamentally different from human understanding, or are they converging?
  • How do we build systems that can reliably demonstrate understanding across all dimensions?

Bottom Line

arXiv 2603.21852 is a timely contribution to one of the most important debates in AI research. By moving beyond philosophical arguments to propose testable criteria for machine understanding, the paper provides a framework that researchers can use to evaluate — and ultimately improve — how deeply AI systems comprehend the problems they solve.

  • 4 min read

EU Wants Crisis Powers to Seize Control of Chip Supplies

Source: Financial Times

The European Commission is preparing to arm itself with sweeping crisis powers that would allow it to take direct control of semiconductor supply chains during future disruptions. The move, embedded within the European Chips Act — a €43 billion plan to bolster Europe's chip sovereignty — represents one of the most aggressive industrial policy interventions in the bloc's history.

The Problem: Overdependence on Asian Foundries

Europe currently accounts for roughly 10% of global semiconductor production, down from over 20% in the 1990s. The vast majority of advanced chips are manufactured in Taiwan (TSMC) and South Korea (Samsung). The pandemic-era chip shortage exposed the fragility of this arrangement, shuttering European auto plants and costing billions in lost output. The Chips Act, proposed in 2022 and finalised in 2023, was designed to prevent a repeat.

The Proposed Crisis Toolkit

The original draft of the Chips Act contained a controversial set of emergency measures that the Commission argued were necessary to respond to future supply disruptions:

Measure Description
Compulsory Information Requests Firms would be legally required to disclose inventories, production capacity, and demand forecasts to the Commission.
Priority Rated Orders The Commission could mandate that chipmakers prioritise crisis-relevant orders over others — effectively forcing companies to reallocate capacity.
Joint Purchasing Mechanism Member states and key buyers would pool demand to negotiate collectively, mimicking the successful EU vaccine procurement model.
Export Monitoring Enhanced surveillance of chip exports to prevent critical components from leaving the bloc during a crisis.

Industry Backlash and Concerns

The proposed powers triggered fierce pushback from the semiconductor industry. Three concerns dominated:

Intellectual Property Risks

Mandatory data sharing — particularly inventory and order-book details — could expose proprietary information to competitors or leak through Brussels' bureaucracy. Chipmakers guard their capacity utilisation data as a trade secret.

Expropriation Fears

Priority rated orders were described by one industry executive as a "backdoor nationalisation" of private manufacturing capacity. If a fab is told it must dedicate 60% of output to crisis products, who bears the financial loss from broken contracts with regular customers?

Investment Chill

Perhaps the most consequential criticism: the threat of future state intervention could deter the very mega-fab investments the Chips Act was meant to attract. Intel's planned Magdeburg facility (€17 billion) and TSMC's Dresden plant (€10 billion) — both benefiting from billions in state aid — might not proceed if investors feared unpredictable state control.

The Final Compromise

After intense negotiation between the Commission, member states, and the European Parliament, the final Chips Act emerged significantly softened:

  • Priority rated orders were scrapped. In their place is a voluntary system where the Commission can request priority treatment, subject to a qualified majority vote of member states.
  • Information gathering powers retained. The Commission kept the right to request data during declared crises — a win for central planners.
  • Coordination role preserved. The Commission will act as a crisis coordinator, matching supply and demand across member states.

The 2030 Target

The Chips Act sets a headline goal: Europe should capture 20% of global semiconductor production by 2030, up from ~10% today. Achieving this requires roughly quadrupling European chip output in less than a decade — an ambitious target that would require flawless execution of the megafab strategy and hundreds of billions in additional investment.

Bottom Line

The EU is attempting a high-wire act: it wants to attract private investment for cutting-edge chip fabs while simultaneously retaining the power to commandeer those same fabs in an emergency. The final Chips Act struck a more market-friendly balance than initially proposed, but the debate over crisis powers is far from settled. The next genuine chip shortage will test whether the compromise holds — or whether calls for harder powers return.

  • 2 min read

Free Will Is Still Undefeated

A recent Wall Street Journal opinion piece by Rob Henderson — Senior Fellow at the Manhattan Institute, PhD in Psychology from Cambridge, and author of Troubled — delivers a forceful counterargument to the growing determinist consensus in neuroscience and popular science writing.

  • 4 min read

Join the New AI Agents Vibe Coding Course from Google and Kaggle

Source: Google Blog by Anant Nawalgaria and Frank Guan

Google and Kaggle are bringing back their wildly popular AI Agents Intensive Course — now with a "Vibe Coding" focus — after the first iteration reached over 1.5 million learners worldwide. The course is free, runs for five days from June 15–19, 2026, and promises to turn participants from curious observers into builders of production-ready AI agents.

What Is Vibe Coding?

The term "Vibe Coding" captures a paradigm shift: instead of writing detailed instructions in a programming language, you describe what you want in natural language — and the AI handles the implementation. The vibe, the intent, the high-level direction becomes the primary programming interface.

This isn't about replacing coders. It's about raising the abstraction layer so high that anyone can build software, and experienced developers can build at 10× velocity.

What You'll Learn

Over five days, the course takes participants from foundational concepts to production-ready systems:

Day Topic
Day 1 Foundations of AI Agents — what they are, how they differ from chatbots, the agent loop (perceive → think → act)
Day 2 Tool Use and API Integration — giving agents the ability to call external services, query databases, and manipulate files
Day 3 Memory and Context Management — how to build agents that remember past interactions and maintain coherent state
Day 4 Multi-Agent Systems — orchestrating multiple specialised agents to solve complex problems collaboratively
Day 5 Capstone Project — build and deploy a production-ready AI agent from scratch

Each day combines conceptual deep dives with hands-on coding examples. Participants code along in Kaggle notebooks, with Google Colab integration for GPU access.

What's New in This Edition

The June 2026 edition isn't just a re-run. Google and Kaggle have refreshed the content significantly:

  • Updated curriculum reflecting the rapid advances in agent frameworks over the past year
  • New guest speakers from Google DeepMind, Anthropic, and leading open-source agent projects
  • Enhanced capstone project that participants can showcase in their portfolios
  • Expanded Vibe Coding focus — new techniques for expressing intent in natural language and letting AI handle the implementation details

Who Should Attend

The course is designed for a broad audience:

  • Software developers who want to build AI-powered features
  • Data scientists looking to automate analysis pipelines
  • Product managers who want to understand what's possible with agents
  • Students and career-changers interested in AI
  • Anyone curious about the cutting edge of human-AI interaction

No prior AI or machine learning experience is required, though basic programming familiarity helps.

Why Free?

Google's investment in free AI education serves multiple goals. It builds the ecosystem of developers who build on Google Cloud and use Google's AI tools (Gemini, Vertex AI, Kaggle Models). It democratises access to AI skills. And it creates a talent pipeline for the AI-driven economy.

The 1.5 million learners from the first cohort represent a significant expansion of the global AI-literate workforce — and Google is betting that many of them will build the next generation of AI applications on Google's platform.

How to Register

Registration is open now on Kaggle. Spots are limited and the previous edition filled quickly. The course runs live June 15–19, 2026, with recordings available afterward for registered participants.

Bottom Line

The AI Agents Intensive Course is one of the most accessible on-ramps to building with AI agents. For anyone who has been meaning to learn how to build AI agents — or who wants to see what the "Vibe Coding" future looks like — this is a low-risk, high-reward investment of five days.

  • 1 min read

Language Models Need Sleep

arXiv: 2605.26099

This paper presents an intriguing exploration into whether large language models, like humans, benefit from structured 'rest' periods.