AI · 2h ago
Interns beat frontier AI labs on Terminal-Bench, now founders owe a private jet
Backboard interns secretly built a codebase called "PJ Branch" and achieved 84.3% on Terminal-Bench 2.1, the highest score ever, beating models from top labs. They used an older model and worked nights and weekends without the founders' knowledge. Now co-founders Jon and Rob face a moral and legal obligation to charter a private jet as promised.
Meridian48 take
This is a feel-good startup story, but the real tech takeaway is that a small team with limited resources outperformed trillion-dollar labs, raising questions about the efficiency of massive compute investments.
ai-benchmarksstartup-culture