Author(s): Mandar Karhade, MD. PhD.

Originally published on Towards AI.

When long-horizon tasks, exploratory freedom, some time, and the ability to self-improve are given to a good model, Great things happen.

Qwen 3.7 max, it’s a big boy model. It is supposed to be fairly capable of running long-horizon tasks. In a recent experiment by Alibaba Group, this model was given a task of improving the performance of a software system on unknown hardware. The model was given an unlimited budget of time to improve the kernel. It made more than a thousand tool calls, ran hundreds of evaluations, and found incredible opportunities to speed up the kernel one step at a time.

The article explains how long-horizon, recursive improvement works: instead of answering once, an AI is given a goal and a looping process (hypothesize, test, inspect, modify, repeat) so it can iteratively refine performance even in unfamiliar settings. It highlights that such success depends on tolerance for failure, information and behavioral persistence, and rapid iteration—capabilities that matter especially when documentation is sparse and hardware is unknown. The author contrasts this with human limitations shaped by specialization and narrow “tunnel vision,” arguing that AI can explore neglected gaps and even enable discoveries across domains like drugs, materials, protein engineering, and chip design. It also notes that this approach isn’t a silver bullet: uncontrolled long runs can waste energy, optimize wrong metrics, accumulate hidden errors, and be hard to audit, so safeguards like sandboxing, cost limits, permissions, checkpoints, logs, and intermittent human review are essential. Overall, the piece argues that Qwen 3.7 Max demonstrates the value of letting well-instrumented long-running experiments proceed.

Read the full blog for free on Medium.

Published via Towards AI