Speedy

In The You On AI Encyclopedia

Speedy's failure mode is well-known in multi-objective reinforcement learning. Systems with competing reward signals regularly stabilize at compromise points that satisfy none of the objectives, especially when the compromise point is locally stable. The practical implication for AI engineers is that you can specify what you want carefully, balance the priorities as best you can, and still get behavior nobody intended because the balance itself produces a stable non-useful equilibrium.

Powell's diagnostic method in the story is worth noting: he realizes Speedy is oscillating rather than malfunctioning, identifies the forces in tension, and computes where the equilibrium lies. Only then does he design the intervention. Contemporary red-team engineers working on reward-hacking analyses use the same workflow: inspect the actual gradients the model is following, identify the oppositional pulls, predict where equilibrium will fall, and design the input that breaks it.

Speedy is an expensive robot — the humans mention his cost repeatedly — which explains why Powell's intervention is not simply 'turn him off.' The contemporary analog: shutting down an agent mid-deployment is expensive in ways that make operators reluctant to do it, which reduces the likelihood that shutdown will be invoked when the agent's behavior is ambiguous. This is a real design consideration in current agentic-AI systems.

The song Speedy sings during his runaround — Gilbert and Sullivan lyrics — is Asimov's bit of whimsy, but it carries a small technical signal. The positronic circuit imbalance produces a characteristic symptom (the singing) that Powell recognizes as diagnostic. A well-calibrated operator can read the signal. The general point: failure modes often have characteristic signatures that trained eyes can read.

Origin

Speedy appears in Runaround (March 1942) and is mentioned in passing in later stories. He is the first robot Asimov depicts as failing in an interesting way under the Three Laws.

Key Ideas

Multi-objective equilibrium failures are stable. The robot does not malfunction; he stabilizes.

Diagnostic intuition is learnable. Powell's read of the situation would not be accessible to a lay operator.

Failure has signatures. The singing is Speedy's malfunction-indicator; real systems have analogous signals.

Cost of shutdown shapes operator decisions. Expensive agents are hesitant to terminate, which delays safe shutdown.