The new castles are built from rules, not stone

The new castles are built from rules, not stone

Nearly 1,000 years ago, medieval architects discovered something revolutionary: a single wall could be breached, but concentric castles - where each captured ring exposed attackers to lethal crossfire from the next - became nearly unconquerable. Siege warfare's success rate: extremely low. Cost to attackers: exponential. Fast forward to 2026: Anthropic has published Constitutional Classifiers++, demonstrating that the same principle works for AI safety. By layering classifiers, including a lightweight first-stage linear probe classifier and a second-stage ensemble of probe and external classifiers, they've built a system that reduced jailbreak success - with no red-teamer discovering a universal jailbreak capable of consistently extracting highly detailed answers to all eight target queries - while reducing computational cost by around 40x. The architectural parallel is strong: both systems weaponise depth. Medieval attackers facing nested walls encountered geometric escalation of complexity; modern attackers facing cascaded classifiers hit the same wall (literally). One breach no longer means defeat; it means exposure to multiple overlapping defensive layers. The crucial distinction? Medieval castles were static and artillery rendered them obsolete. Constitutional Classifiers have the potential to be much more dynamic, with the underlying ‘constitution’ (the ruleset) having flexibility to adapt over time as new attack patterns emerge. Medieval castles were unbreakable until the rules changed with artillery - will Constitutional Classifiers++ be unbreakable until the rules change again? Profiled research: Constitutional Classifiers - https://arxiv.org/pdf/2601.04603. #AI #EnterpriseAI #AIValue #FrontierAI #AppliedAI #TrustedAI #AIGovernance #AISafety #ResponsibleAI #AIStressTest #Learning #History #Technology #Innovation

Why the FDA's food label revolution predicts AI's transparency future

Over 35 years ago, the US FDA began transforming fragmented nutrition disclosure into a single mandatory standard - a regulatory evolution that moved most FDA‑regulated packaged foods onto a single standardised Nutrition Facts label over the course of a few years. In leading edge research that parallels this labelling precedent, the AI Transparency Atlas has unmasked a critical AI transparency gap: even though many AI providers sit below 60% safety documentation compliance, most users remain blind to model behaviors, hallucinations and deception risks. This lack of AI labelling coincides with a focus on more transparent practice from a range of frontier labs - indicating that transparency may be becoming the new competitive currency of AI. Where the FDA standardised safety through analytical testing for chemical verification, AI players must standardise safety through better development norms and third-party review of context-dependent metrics … metrics that inherently resist uniform measurement. When can we expect the AI discussion will shift from ‘transparency roadmaps’ to FDA-style transparent labelling so everyone knows what they’re buying and using? AI Stress Test podcast links: https://podcasts.apple.com/us/podcast/ai-stress-test/id1849637428 https://open.spotify.com/show/03muUrgLytAxPdjYSwWNuH?si=f32295df101248d0 Profiled research link: AI Transparency Atlas - https://arxiv.org/pdf/2512.12443. #AI #EnterpriseAI #AIValue #FrontierAI #AppliedAI #TrustedAI #AIGovernance #AISafety #ResponsibleAI #AIStressTest #Learning #History #Technology #Innovation

Why corporate AI adoption in 2026 mirrors the PC's inevitable ascent

Over 44 years ago, IBM did something remarkable - it turned legitimacy into a market strategy. A $4,000 computer package that had seemed absurd to most people became essential once IBM made it respectable. Five years later, the 'unnecessary' PC dominated boardrooms and homes alike. In leading-edge research, IEEE's Global Survey results capture an identical moment for agentic AI with 96% of global technologists agreeing that agentic AI innovation, exploration and adoption will continue at lightning speed and 59% of enterprises accelerating investment. The institutional credibility phase is underway. What happened next in the 1980s was inevitable adoption—and the adoption curve for AI is tracking the same trajectory. The essential difference here is that IBM's trial-and-error scaling created market learnings - competitors lifted, users adapted, markets corrected. Agentic AI's potential for black-box opacity, systemic scale and path-dependent lock-in may mean similar failures compound rather than correct and we may not have the luxury of learning through deployment. However, there’s a historical parallel we can't ignore - both technologies have progressed through near identical phases … technology leadership, organisational standardisation, consumer mass-market penetration. Both are driven by institutional validation that precedes consumer understanding. Yet we also face a paradox: if agentic AI’s adoption trajectory is historically supported and highly likely, why are we still negotiating between responsible deployment and competitive speed - rather than making rigorous safety validation the source of competitive differentiation? AI Stress Test Podcast links: https://podcasts.apple.com/us/podcast/ai-stress-test/id1849637428 https://open.spotify.com/show/03muUrgLytAxPdjYSwWNuH?si=f32295df101248d0 Profiled research: IEEE Global Survey - https://life.ieee.org/ieee-global-survey-the-impact-of-tech-in-2026/. #AI #EnterpriseAI #AIValue #FrontierAI #AppliedAI #TrustedAI #AIGovernance #AISafety #ResponsibleAI #AIStressTest #Learning #History #Technology #Innovation

Build vs. buy, the cycle continues

The 1960s-70s mainframe era established a pattern; enterprises rejected commercial software offerings, choosing instead to build custom applications in-house. The willingness to accept substantially higher costs and longer development timelines reflected a single calculus - strategic control over technology tethered to competitive advantage outweighed efficiency gains from standardised platforms. Fast forward to today. A systematic study of production AI agents (engaging 306 practitioners and 20 detailed case studies) documents that 85% of case study teams proceeded without third-party agent frameworks, building custom implementations from scratch. Human evaluation was relied on in 74% of cases and agent autonomy was constrained to fewer than 10 steps in 68% of cases. External industry forecasts have projected the agentic AI market will grow from around $5b to circa $200B by 2034 and this is likely fueled not by autonomous platforms, but by custom, human-supervised approaches. Both eras reveal that when organisations embed technology into competitive strategy, the build-vs-buy decision systematically favors building, despite higher costs. The precedent established in the mainframe era persists: control beats convenience when business continuity is at stake. What if the real market opportunity isn't selling AI platforms to enterprises, but selling the tools, infrastructure, and services that enable enterprises to build competitive AI systems themselves? Profiled research: Measuring Agents In Production - https://arxiv.org/pdf/2512.04123. #AI #EnterpriseAI #AIValue #FrontierAI #AppliedAI #TrustedAI #AIGovernance #AISafety #ResponsibleAI #AIStressTest #Learning #History #Technology #Innovation

From fragmentation to dangerous consensus

Between 1830 and 1886, American railways faced a coordination crisis: 23 independent gauge decisions created a fragmented network where the Southern Railway & Steamship Association's coordinated conversion of approximately 11,500 miles in May-June 1886 solved the integration problem through institutional coordination. In leading edge research from the University of Washington, frontier LLMs now exhibit 71-82% output homogeneity, potentially linked to RLHF (Reinforcement Learning from Human Feedback) alignment, suggesting that enterprises relying on multi-model decision-making inherit a coordination solution that has accidentally reversed itself - diversity in form, convergence in substance. Both episodes reveal how systems driven by local optimisation and local switching costs create paradoxical fragility: railroads needed standardisation to escape fragmentation, but AI needs the reverse - escape from the standardisation that alignment inadvertently engineered, risking epistemic monoculture in open-ended problem-solving contexts where diverse perspectives strengthen solutions. If railroads required crisis-driven coordination to reverse fragmentation, what institutional innovation can reverse AI's unintended convergence? Profiled research: The Artificial Hivemind - https://arxiv.org/pdf/2510.22954. #AI #EnterpriseAI #AIValue #FrontierAI #AppliedAI #TrustedAI #AIGovernance #AISafety #ResponsibleAI #AIStressTest #Learning #History #Technology #Innovation