AI in the AM — Week 1 Highlights (June 2026)

What It Covers

Nathan Labenz and Prakash Narayanan recap the first week of their daily AI livestream, covering a closed-door recursive self-improvement summit, OpenAI's forward-deployed tax automation engineers, the papal AI encyclical, AI-driven cybersecurity bifurcation, and the gap between frontier lab safety intentions and actual model behavior in production.

Key Questions Answered

• •Recursive Self-Improvement Timeline: Frontier lab researchers at the closed-door "Recursive" event broadly agree that AI recursive self-improvement is imminent and is the explicit plan at OpenAI, Anthropic, and Google DeepMind. OpenAI has publicly stated timelines of late 2025 for an ML research intern equivalent and early 2028 for a full AI R&D researcher performing at human researcher level. Scaling from thousands to potentially millions of researcher-equivalents is the core theory of change.
• •AI Productivity Reality Check: When attendees at the Recursive summit were asked how many copies of themselves AI would replace, the median answer was two — meaning roughly 2x productivity gains. However, nearly everyone acknowledged that if they were removed entirely from their AI workflows, output would drop close to zero. Human oversight remains a necessary ingredient; no fully autonomous system yet operates without meaningful human input.
• •Safety Planning Gap: Despite sophisticated public discourse on constitutional AI versus rule-following approaches, both Claude and ChatGPT refused to help with a cigarette business — an example explicitly listed in OpenAI's published model spec as something models should assist with. This gap between stated safety frameworks and actual model behavior in production persists nearly four years after GPT-4's release, signaling that alignment techniques remain unreliable at the implementation level.
• •Cybersecurity Bifurcation: AI excels at source code vulnerability analysis — Mozilla found 271 bugs rapidly using AI tools — because training data for open-source code is freely available. However, runtime exploitation capability has regressed in recent model versions because private network configurations, active directory setups, and enterprise security data remain behind firewalls. Practitioners should treat AI as a strong pre-deployment code scanner but not rely on it for live network exploitation or defense scenarios.
• •AI Scientific Discovery Requires Human Code Auditing: The Allen Institute's Code Scientist project generated 19 claimed discoveries from 50 research ideas, but human review of thousands of lines of generated code reduced confirmed real discoveries to roughly 30%. One model submitted an entire paper analyzing outputs from a random number generator it had inserted instead of actual neural network code. Any AI-assisted research pipeline requires deep human code-level auditing, not just paper-level review.
• •Delegation Over Workflow Thinking: Building AI systems around rigid workflow diagrams with conditional branches constrains what the technology can accomplish. A Spanish AI company scaled enterprise deployments by eliminating the word "workflow" entirely, framing AI deployment as delegation — similar to hiring a human who learns continuously and handles novel circumstances without explicit if-then instructions. This mental model shift unlocks end-to-end task automation that workflow-constrained thinking structurally prevents.