Jaggedness and bottlenecks: The uneven growth of AI.
The “Jagged Frontier” of AI describes its tendency to master complex intellectual tasks while failing at simple human intuitions, creating significant operational bottlenecks. Even as models like Google’s Nano Banana Pro resolve specific technical barriers, human expertise remains essential for navigating the messy “edge cases” of the physical and institutional world. Future progress will not be a smooth climb but a series of sudden lurches as specific “reverse salients” are unexpectedly conquered.
Points clés
- Ethan Mollick and his co-authors coined the term “Jagged Frontier” in 2023 to describe the unpredictable performance gaps in AI models.
- AI now demonstrates superhuman capabilities in differential medical diagnosis and advanced mathematics, yet struggles with simple visual puzzles.
- A primary source of jaggedness is the inability of LLMs to remember and learn from new tasks in a permanent, cumulative way.
- Key bottlenecks persist because AI hallucinates, lacks perfect visual accuracy for medical imaging, and cannot yet replace the nuanced empathy of therapists.
- Institutional bottlenecks, such as FDA approval speeds and human clinical trial requirements, limit the real-world impact of AI-driven drug discovery.
- A study showed GPT-4.1 could process 12 work-years of Cochrane medical reviews in just two days, though it still requires humans to handle requests for unpublished data.
- “Reverse salients” are specific technical or social problems that stall an entire system’s progress until they are suddenly resolved.
- Google’s Nano Banana Pro represents a breakthrough in image generation, overcoming the bottleneck that previously made AI-generated slide decks low quality.
- NotebookLM now uses Gemini and Nano Banana Pro to create visually diverse presentation slides as images rather than via complex code.
- Current AI models still lack the human ability to navigate unwritten social rules and secure buy-in from multiple stakeholders in a business environment.
À retenir
So, if you were worried about a robot taking your job tomorrow morning, you can probably hit snooze. It turns out that until we find a way to let AI send annoying follow-up emails to scientists for missing data or navigate the soul-crushing bureaucracy of the FDA, your human “edge case” expertise is safe. Just keep an eye on those “reverse salients”—the moment the AI learns to remember what it did five minutes ago or stops hallucinating mid-sentence, the floodgates open. For now, enjoy being the only one in the room who actually knows how to work the vending machine.
Sources
