Core patterns for agentic organizational design, self-organizing knowledge networks, scaling co-scientists, and more

📖In this issue

• Framework: Agentic AI: Core patterns for organizational redesign

• A review of tools and approaches for AI in science

• Self-organizing knowledge networks help open-ended discovery

• Google’s new “co-scientist” is built on scaling test-time inference

• AI continues to advance across the toughest benchmarks

💡Framework: Agentic AI: Core patterns for organizational redesign

The advent of robust, scalable AI agents will dramatically reshape organizational structure. I created this framework to distil some of the fundamental patterns that will drive successful transformation as we shift wholesale to Humans + AI work.

I have used this in some board and executive briefings as a discussion starter, and first shared this publicly this week at Agentforce World Tour Sydney.

For an explanation of the framework read the post.

🧠🤖Humans + AI

A review of tools and approaches for AI in science

A good review article covers the scope of AI use in science, including in Literature search and summarization, hypothesis generation and idea formation, experimentation, data analysis and hypothesis validation, and more. The chart below is just one of the many facets covered in the paper.

Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation
Summary of key insights

Self-organizing knowledge networks help open-ended discovery

Synthetic biology is our future. A goundbreaking new biological foundation model Evo2 achieves state-of-the-art prediction of genetic variation impacts and generates coherent genome sequences, spanning all domains of life.

Genome modeling and design across all domains of life with Evo 2
Summary of key insights

Google’s new “co-scientist” is built on scaling test-time inference

“Google has adapted Gemini 2.0 to make it generate novel scientific hypotheses in a fraction of the time taken by teams of human lab researchers…. {It} emphasizes the use of "test-time scaling," where AI agents use increasing amounts of computing power to iteratively review and re-formulate their output.“

Google's AI co-scientist is 'test-time scaling' on steroids. What that means for research

💡Reflections

AI continues to advance across the toughest benchmarks

The trajectory of AI capabilities is reaching new thresholds. Recent benchmark results show a landscape where advanced AI models are increasingly surpassing human performance across diverse domains of cognition and expertise. 

The Shifting Frontier of AI Performance

The question is no longer whether AI can match human capabilities, but rather which domains remain uniquely human. The evidence presented in recent benchmarks paints a compelling picture of this transformation:

In general reasoning and AGI tasks, OpenAI's o3 (High-compute) model has achieved 87.5% performance on semi-private evaluation sets, approaching the estimated 90% performance of skilled humans. This represents a fundamental shift in AI's ability to handle abstract reasoning that was once considered the exclusive domain of human cognition.

Perhaps more striking is the performance in specialized knowledge domains. The o3 model has reached 87.7% performance on graduate-level STEM evaluations spanning biology, physics, and chemistry—exceeding domain expert performance of 81.3%. This suggests that AI has crossed a significant threshold in mastering structured technical knowledge.

Mathematics: The New Frontier

Mathematics has long served as a benchmark for advanced reasoning. The results here are particularly noteworthy:

Google DeepMind's Alpha Geometry2 has successfully solved 84% of past International Mathematics Olympiad geometry problems—a remarkable achievement considering that even gold medalist human performers typically reach around 78% success rates.

In standard mathematical word problems, GPT-4 has achieved approximately 92% accuracy, surpassing typical human performance. This suggests that AI can now effectively translate natural language descriptions into mathematical frameworks and solve them with greater reliability than most humans.

Yet the frontier is not uniformly conquered. In advanced mathematics, o3 achieves only 25.2% on frontier-level problems—areas where human experts struggle as well. This delineates the current boundary of AI capability.

Professional Knowledge and Social Intelligence

The implications for professional domains are profound. GPT-4 has achieved approximately 85% performance on the USMLE—the standardized examination for medical licensing in the United States—approaching the typical human pass performance of around 80%.

Perhaps most surprising is the performance on social intelligence scales, where ChatGPT-4 exceeds all psychologist scores in counseling contexts. This suggests that AI has developed capabilities not only in structured knowledge domains but also in understanding human emotions and social dynamics—areas previously considered uniquely human.

Strategic Implications

These benchmarks reveal three critical insights for organizations and knowledge workers:

First, the substitution pattern identified in organizational redesign is accelerating across cognitive domains, not merely routine tasks. This will fundamentally reshape the landscape of knowledge work.

Second, the areas where humans maintain clear advantages are narrowing to frontier domains—those at the absolute leading edge of human knowledge and creativity.

Third, the trajectory suggests that augmentation and dynamic orchestration, rather than competition, offer the most productive framework for human-AI collaboration.

In essence, these benchmarks don't simply measure performance—they map the evolving relationship between human and artificial intelligence. The strategic question for organizations and individuals is not whether to adapt to this new reality, but how to leverage it for unprecedented capabilities through thoughtful integration and collaboration.

The future belongs not to those who resist this trajectory, but to those who design sophisticated frameworks for human-AI collaboration that capitalize on the unique strengths of each.