Is the "35-Year-Old Curse" Broken? Are Middle-Aged People Making a Comeback to Lead the AI Revolution?

A new model, an old demographic — coincidence or signal?

MiniMax M2.7 has been released with surprising speed, and it is not just another incremental update. The new model, MiniMax M2.7, reportedly represents the product’s first time “deeply participating in iterating its own models,” folding automated self‑improvement and an Agent Harness into a continuous development loop. What does that mean for who leads innovation? If models can take on more of the execution and even parts of decision‑making, could experienced, middle‑aged engineers and managers — often written off by China’s so‑called “35‑year‑old curse” — regain influence as supervisors and system designers rather than frontline coders?

From benchmarks to full‑stack agents

MiniMax’s team ran M2.7 through a suite of benchmarks — SWE Bench, VIBE‑Pro, MM‑ClawBench and MLE‑Bench — designed to probe software engineering execution, end‑to‑end project delivery, long‑horizon agent stability and research‑level reasoning. Results show M2.7 already sits in the first tier on many engineering‑execution tasks (SWE Bench Pro and VIBE‑Pro), while it still trails top models on the more abstract, algorithmic MLE‑Bench. In plain terms: M2.7 is getting very good at finding and fixing code, orchestrating multi‑step tool use, and delivering complete projects, but high‑end research reasoning remains a tougher challenge.

Human scenes, machine craft

The developers tested M2.7 with four staged “exams” ranging from a high‑fidelity multi‑role family chat to building games from scratch and producing analyst‑style reports based on Nvidia (英伟达)’s annual filing. The family chat demo — a WeChat‑style page populated with personas including a 55‑year‑old retired father, a 52‑year‑old community worker mother, and a 24‑year‑old son — illustrates the model’s ability to maintain separate conversational states, role relationships and long dialogues. The team also used an Agent Harness IDE that shows agent thought traces, virtual file systems and live previews to stress test the model’s multi‑agent coordination and autonomous coding workflows.

Bigger picture: labor, leadership and geopolitics

It has been reported that industry figures such as former Google CEO Eric Schmidt see recursive self‑improvement and memory systems as a transformative path for AI. In China and globally, that trend intersects with geopolitical realities: advanced models still depend on compute dominated by Western suppliers and constrained by export controls, so efficiencies that let smaller, older teams do more with less matter strategically. Will this revalue experience over youth? Not overnight. But by shifting many execution tasks into repeatable agent loops, M2.7‑style systems could make senior engineers and managers more valuable as architects and policy setters — steering AI rather than writing every line of code. Who leads the revolution may end up being a question of role, not age.