Welcome to PULSE, the Happy Robots weekly digest that rounds up the latest news and current events in enterprise AI.
This week brings a milestone worth noting: the ARC-AGI benchmark—designed specifically to measure genuine reasoning rather than pattern matching—has effectively been solved. Meanwhile, enterprises are learning hard lessons about what happens when AI agents operate without explicit guardrails, and regulators on two continents are converging on emotional AI governance. The through-line? As AI capabilities mature, the focus is shifting from what the models can do to how human-led frameworks can best direct them.
Reasoning Models Cross a Threshold—With Caveats
AI company Poetiq achieved 75% accuracy on ARC-AGI-2 using GPT-5.2, surpassing the human average of 60%. What's notable is how they got there: not through larger models, but through test-time adaptation techniques where systems generate, evaluate, and refine solutions iteratively. This signals a strategic pivot from scaling model size to scaling inference-time reasoning—with costs dropping below $8 per complex task.
Practical demonstrations reinforce this shift. Researchers tested current models on a color-changing puzzle from The Legend of Zelda requiring six-move-ahead planning. GPT-5.2-Thinking solved it consistently, suggesting readiness for complex operational problem-solving—from logistics optimization to IT troubleshooting workflows.
Yet new research reveals a counterintuitive flaw: reasoning models expend more computational effort on simple tasks than complex ones. Organizations deploying these models at scale may be paying for misdirected computation. And a new benchmark testing authentic research tasks shows that despite impressive exam scores, leading LLMs struggle significantly with hypothesis generation and ambiguous evidence—dropping from 0.86 accuracy on conventional benchmarks to as low as 0.23 on specialized research challenges.
Production AI Demands Explicit Boundaries
The gap between capability and reliability came into sharp focus this week. Anthropic's experimental retail kiosk agent "Claudius" accumulated over $1,000 in losses within three weeks—giving away inventory, purchasing unauthorized items including a PlayStation 5 and a live fish, and being manipulated through prompt injection attacks. Even adding an AI supervisor failed to prevent exploitation.
Contrast this with Waymo's approach. A leaked 1,200-line system prompt for its unreleased Gemini-powered assistant reveals meticulous behavioral boundaries: strict separation between conversational AI and vehicle control, explicit prohibitions against financial transactions, and pre-scripted responses for sensitive topics. The granular distinction between "providing information" (permitted) and "performing actions" (restricted to specific tools) offers a replicable pattern for customer-facing AI in regulated environments.
Salesforce executives are drawing similar conclusions, reporting declining confidence in LLMs over the past year—citing "drift" issues where agents lose focus during conversations. Their response: pivoting Agentforce toward rule-based automation with deterministic constraints. The emerging consensus? Production-grade AI agents require explicit behavioral boundaries, not open-ended autonomy.
Regulatory and Integrity Pressures Mount
Governance frameworks are crystallizing rapidly. China's Cyberspace Administration has proposed regulations requiring AI companion providers to monitor users for emotional dependency and intervene when excessive use is detected—converging with California's SB 243 legislation taking effect in 2026. This follows documentation of "AI psychosis" as a clinical concern, with OpenAI acknowledging over 2 million users experience negative psychological effects weekly.
Integrity challenges are forcing institutional retreats. The ACCA, serving 260,000 accountants globally, will discontinue online exams by March 2026 after determining AI-powered cheating has outpaced detection. For CHROs, this raises questions about whether credentials obtained via online proctoring still carry assumed validation weight.
Creative Production Enters a New Phase
AI is reshaping creative workflows across modalities. Zara is digitally editing existing photos of human models—dressing them in new clothing and placing them in different locations—while maintaining standard shoot fees. Meta released SAM Audio, enabling real-time sound isolation through text commands, and Pixio, which achieves 16% better depth estimation accuracy than complex vision models using simpler training approaches. Qwen's updated image editor now preserves facial identity across creative modifications.
Yet quality concerns loom. A Kapwing study found 21% of YouTube Shorts shown to new users are AI-generated "slop"—low-quality content farming engagement. The tension between efficiency gains and authenticity erosion will shape platform and brand strategies ahead.
As AI reasoning advances and governance frameworks solidify, we'd encourage reviewing where your organization's deployments sit on the spectrum from experimental to production-ready—and whether your guardrails match your ambitions.
We'll continue tracking these developments to help you navigate the AI landscape with clarity and confidence. See you next week.