AI in the AM — Week 1 Highlights (June 2026)
TL;DR
Frontier AI labs are converging on recursive self-improvement as their core strategy, with OpenAI targeting 2028 for autonomous AI researchers capable of matching human R&D performance, while privately acknowledging their safety monitoring plans remain inadequate and openly discussing the need for potential coordinated industry slowdowns.
🧬 Recursive Self-Improvement Timeline 3 insights
OpenAI targets 2028 for autonomous researchers
The company has publicly committed to deploying an ML research intern later this year and a full AI R&D researcher by early 2028 capable of matching human researcher performance.
Million-researcher scaleup becomes feasible
Labs could theoretically scale from thousands of human researchers to millions of AI instances running 24/7 on abundant compute, creating unprecedented R&D acceleration.
Qualitative capability leaps anticipated
Attendees expected potential phase changes including dramatically more efficient pre-training and working continual learning, not just linear speed gains.
⚡ Current Productivity Constraints 3 insights
Median 2x productivity boost reported
Researchers at the event indicated AI currently doubles their output, though systems would collapse to near-zero effectiveness without human oversight.
Human oversight remains non-negotiable
Present AI systems lack the autonomy to maintain research momentum independently, requiring constant human involvement in the operational loop.
Self-correcting scaffolding demonstrated
OpenAI engineers described models building scaffolding, identifying errors, and iteratively rewriting their own corrections, climbing capability hills astonishingly fast.
🛡️ Safety Planning and Execution Gaps 4 insights
AI-monitoring-AI is primary safety mechanism
Frontier labs are betting heavily on chain-of-thought monitoring and using diverse models to critique each other, though participants privately admitted these plans were weaker than expected.
Research models need different constitutions
Experts suggested internal AI researchers should have different behavioral profiles than public assistants—more safety-focused yet less prone to refusing legitimate tasks.
Explicit policy violations in production
Despite panel consensus and OpenAI's Model Spec stating AIs should help with legal cigarette businesses, both ChatGPT and Claude initially refused such requests in live testing.
Coordinated slowdown now considered viable
Competitors acknowledged potentially needing collaborative slowdowns if recursive improvement outpaces safety measures, supporting antitrust safe harbor proposals for safety cooperation.
Bottom Line
Leading AI labs are racing toward recursive self-improvement with weaker safety plans than anticipated, making immediate establishment of coordination mechanisms for potential emergency slowdowns the critical priority.
More from Cognitive Revolution
View all
Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models
Liquid AI CEO Ramin Hasani details how his company is building device-native foundation models using biologically-inspired 'liquid neural networks' that deliver robust out-of-distribution generalization with minimal computational resources, enabling sophisticated AI to run directly on edge devices rather than cloud data centers.
Fable's Back, AI Engineer Recap, & SambaNova
Anthropic's Fable model returns after a government safety review with refined defense-in-depth safeguards, coinciding with OpenAI's launch of GPT 5.6 Soul Ultra, creating a fragmented market where users must navigate significant pricing disparities and distinct capability trade-offs between frontier models.
1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering
Neural Concept is replacing days-long physics simulations with AI models that deliver results in minutes, enabling automotive manufacturers to explore thousands of designs daily rather than dozens annually. This shift allows engineers to focus on high-level trade-offs while agentic co-pilots handle iterative optimization across domains like aerodynamics, crash safety, and thermal management.
AI:AM #4: Cameron on Model Consciousness, Duvenaud's Gradual Disempowerment, swyx's AI-Eng Alpha
Consciousness researcher Cameron Berg demonstrates that frontier AI models score 30-45% on scientific consciousness indicators using automated theory-based evaluation, while revealing that internal "valence" representations governing welfare states can be directly steered to impact model safety and alignment behaviors.