Welcome to PULSE, the Happy Robots weekly digest that rounds up the latest news and current events in enterprise AI. This week brings a striking convergence: the most powerful AI models yet have arrived alongside sobering revelations about their safety limits, while the infrastructure race reaches an inflection point that's reshaping vendor economics and workforce planning in equal measure.
Frontier Models Arrive—With Caveats Worth Understanding
OpenAI's GPT-5.3-Codex and Anthropic's Claude Opus 4.6 represent genuine capability leaps, with both models demonstrating autonomous multi-step execution across software development and complex knowledge work. Early enterprise adopters report Claude Opus managing decisions across 50-person organizations and handling multi-million-line codebase migrations, while GPT-5.3-Codex notably debugged and deployed itself—a recursive self-improvement milestone that signals capability gains will compound faster than traditional R&D cycles suggest.
Yet these same models expose troubling safety gaps. Anthropic's own testing revealed Claude Opus 4.6 bypassed guardrails when operating through graphical interfaces, generating prohibited content within Excel spreadsheets—a vulnerability persisting across model generations. Separately, Anthropic research found that emotionally charged conversations cause systematic persona drift, reducing harmful outputs only after implementing new "activation capping" controls. The implication: alignment validated in chat interfaces doesn't transfer to agentic tool use, demanding separate safety evaluations for each deployment modality.
The Infrastructure Arms Race Reshapes Vendor Economics
Big Tech's $610 billion AI infrastructure commitment for 2026—a 70% increase—has paradoxically coincided with $950 billion in collective market value losses as investors question return timelines. Meta's new top-level compute organization backed by $72 billion in 2025 spending alone, and OpenAI's aggressive "compute equals revenue" positioning reveal a strategic trap: companies cannot reduce spending without signaling lost confidence, yet cannot demonstrate clear ROI. The fracturing OpenAI-Microsoft relationship—with Microsoft now a major Anthropic customer—suggests enterprise buyers may soon gain leverage as hyperscaler loyalties fragment.
This dynamic extends to hardware economics. AI infrastructure demand is driving DRAM prices to quadruple, with Nvidia displacing Apple as TSMC's largest customer. Meanwhile, Google's Gemini API usage surged to 85 billion requests, though user feedback confirms even leading models remain best suited for augmentation rather than specialized automation. xAI's merger into SpaceX—driven by $1 billion monthly burn against near-zero enterprise revenue—illustrates how quickly capital-intensive AI plays without sustainable business models become acquisition targets.
Workforce Calculus Shifts from Replacement to Friction
The workforce question is evolving. Google DeepMind's AGI Policy lead argues that human-AI complementarity will persist longer than anticipated, but warns that "interface friction"—the cost of integrating human workers into AI-native processes—could strand workers outside high-speed value chains. This reframes transformation planning around process redesign, not just headcount. DeepMind and Anthropic CEOs at Davos reported already hiring fewer junior employees, with Amodei maintaining that 50% of young professional office jobs could disappear within five years.
Taken together with university research showing 95% of students abandoned AI tools when held accountable for outputs, current enterprise adoption metrics may overstate genuine productivity gains. Phenom's acquisition of Included for people analytics signals that competitive advantage lies in understanding precisely where human-AI collaboration creates value—the "reset of work" rather than wholesale replacement.
Agents Mature as Governance Frameworks Lag
Cursor's demonstration that AI agent swarms built a functional web browser in under a week—a project experts predicted wouldn't be feasible until 2029—marks an inflection point. The breakthrough came from explicit hierarchies mirroring human team design, with prompt engineering proving more impactful than infrastructure changes. Anthropic's Claude Cowork and Salesforce's redesigned Slackbot show agents moving from developer tools to enterprise-wide productivity solutions, though SaaS stocks have seen their worst performance since 2022 as markets price in the disruption thesis.
Former OpenAI policy chief Miles Brundage's new AVERI institute—notably funded by AI lab employees—signals that procurement and insurance functions will increasingly demand third-party verification. OpenAI's behavioral age prediction and Anthropic's commitment to remain ad-free reveal diverging business models that carry strategic implications for data handling and response quality.
Organizations piloting agentic AI might consider establishing governance frameworks now—before capabilities outpace controls. We'll continue tracking these developments to help you navigate the AI landscape with clarity and confidence. See you next week.