The AIPower Shift

In February 2025, Andrej Karpathy coined "vibe coding" — describe what you want, accept what comes back, forget the code exists. Exactly one year later, he declared it obsolete. The new default: you are not writing code 99% of the time. You are orchestrating agents who do, while acting as oversight. He called this agentic engineering. Then he joined Anthropic — not coincidentally. His reason: the model is the bottleneck on what an agent can do. His open-source autoresearch project now runs hundreds of AI experiments overnight. Now OpenAI has published the empirical proof. Their economic research paper on Codex (June 25, 2026) documents the transition in real numbers, measured inside their own organisation. Agentic AI changes the unit of knowledge work from single interactions to delegated, long-horizon tasks. Chatbot interactions are short and self-contained. Agents operate independently for minutes, hours or days while orchestrating tool calls, iterating toward solutions, and crossing job function boundaries. By May 2026, 80.6% of Codex users had assigned tasks estimated to exceed 30 minutes of human work. 70.2% had assigned tasks exceeding one hour. 25.6% had assigned tasks exceeding eight hours. Users at the 99th percentile now run more than 60 hours of parallel agent work per day. Codex accounts for 99.8% of weekly output tokens generated within OpenAI. Non-developer adoption grew 189x among organisational users since August 2025. A quarter of work done by lawyers, recruiters and finance staff on Codex is engineering or coding — tasks outside their job descriptions. But agentic tools are not equal. XDA Developers ran a head-to-head benchmark (June 23, 2026): Claude Code vs OpenAI Codex vs Google Antigravity, given identical instructions to build a full React project management frontend. The result: Claude Code won — not on functionality, where all three were comparable, but on UX judgment. Codex had scaling problems and awkward interactions. Antigravity scored 8/10 but missed contextual colour-coding. Claude Code nailed the small details — progress bars that changed colour by status, smooth animations, clean Kanban grouping, thoughtful iconography. The verdict: "Claude Code stood out as the strongest result — its understanding of user experience created a product-like feel." The implication for the power shift is sharp: when the task is agentic, model quality becomes product quality. The tool that understands design intent, not just code syntax, wins the workflow. Sources: Forbes / Jodie Cook (June 12, 2026); OpenAI Economic Research / Codex (June 25, 2026); XDA Developers / Parth Shah (June 23, 2026).

On April 7, 2026, Anthropic released Claude Mythos Preview — its most cyber-capable model — and immediately locked it behind Project Glasswing, a restricted access program for major technology firms only. The reason: Mythos can discover thousands of critical vulnerabilities across every major operating system and web browser, chain four exploits to escape a browser sandbox, and complete end-to-end autonomous attacks on enterprise networks. The UK AI Security Institute found AI cyber capabilities are doubling every four months. Chinese AI models are estimated to be seven months behind the US frontier — meaning adversaries could reach Mythos-level capability by late 2026. The time from vulnerability disclosure to active exploitation has collapsed to 20 hours. The central strategic asymmetry: attackers need only find one gap; defenders must close all of them. Compute is now the deciding variable in national cyber power. This is not a future risk. It is an ongoing transfer of offensive capability from nation-states with skilled hackers to any actor with enough compute and the right model. Who owns the most capable AI owns the most dangerous cyber weapon ever built — and is simultaneously the only one who can defend against it. Source: IAPS, April 2026.

On June 12, 2026, the Trump administration did something no government had ever done: it forced a private AI company to pull its models offline worldwide. Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei ordering that Fable 5 and Mythos 5 be subject to export controls for all foreign nationals — inside and outside the United States. Because American "deemed export" rules include foreign nationals working at Anthropic itself, the company had no choice but to disable both models for every user on earth. The trigger: Amazon CEO Andy Jassy personally called the White House after Amazon researchers found a jailbreak in Fable's cybersecurity guardrails. Amazon has invested $13 billion in Anthropic with a commitment to $20 billion more. The episode revealed the ultimate hidden dependency in the AI power stack: the kill switch. Every hospital, bank, defence contractor, and government agency that built workflows on American AI discovered in one weekend that their infrastructure could be switched off — not by a cyberattack, not by a technical failure, but by a letter from a cabinet secretary. Europe's reaction was immediate: French PM Lecornu — "We cannot rely on tools developed by foreign powers." Canada's Carney — "Nobody has done anything wrong. But we will have done something wrong if we just accept this." China's open-source AI developers: delighted. The Economist asked how Europe must respond. The answer this chapter gives: the only sovereign AI is AI you run yourself, on compute you own, with models you control. Everything else is rent — with an eviction clause. Sources: Axios, Fortune, Al Jazeera, IAPP, The Economist — June 12–18, 2026.

For two decades, the smart city was the dominant urban ambition. Sensors, data, real-time dashboards, optimised traffic flows, predictive policing, digital government. The promise was compelling. In many respects it was delivered. But Professor Lily Kong, President of Singapore Management University, writing at the 10th World Cities Summit, asks the question that reframes everything: is being smart enough? Her answer — increasingly, no. The wise city is not a rejection of technology. It is a refusal to treat technology as sufficient. A smart system finds the fastest route. A wise city asks whether all residents have access to mobility. A smart algorithm allocates resources efficiently. A wise city asks whether outcomes are equitable. A smart city collects data. A wise city asks what data should be collected, and understands what should be done with it. The distinction cuts to the heart of what AI can and cannot do. AI handles volumetric data brilliantly — it identifies patterns invisible to human observers. But cities are not lived cognitively alone. People encounter cities through memory, attachment, comfort, beauty, smell, belonging and trust. These do not appear in training sets. Singapore's own experience is the case study: public housing succeeded not through engineering alone but through a broader vision of nation-building. Water security is not merely a technical achievement but decades of strategic thinking and societal commitment. Even Singapore, Kong notes, got it wrong once — rejecting wheelchair ramps on efficiency grounds four decades ago, then paying to retrofit every station at far greater cost. Wisdom, she argues, places efficiency within a broader and longer horizon. The AI power shift accelerates this paradox: as information becomes more abundant, wisdom does not automatically follow. The cities — and the nations — that survive the AI transition will be those that use AI to inform judgment, not replace it. The ultimate measure of a city is not how much data it collects or how efficiently it manages infrastructure. It is whether children can thrive, older adults can age with dignity, communities can remain resilient, and future generations can inherit a liveable planet. Source: Straits Times / Professor Lily Kong (SMU President), 10th World Cities Summit, June 2026.