Incentives and Outcomes in Humans, AI, and Crypto (olshansky.info)

This piece argues that incentives — the “show me the incentive, and I’ll show you the outcome” idea — drive behavior across humans, AI, and crypto, and shows how aligning them yields better outcomes. For humans the author contrasts bonus-driven and loss-aversion schemes to counter Parkinson’s Law (work expanding to fill time). For AI, they point out that modern systems are essentially reinforcement learners and that the reward function is the single hardest, most consequential design choice. A concrete example: OpenAI’s GPT-5 change to “abstain when unsure” raised abstention from 1% to 52% and cut error rates from 75% to 26% (o4‑mini → gpt‑5‑thinking‑mini), demonstrating how changing the model’s incentive — from “always answer” to “answer only when confident” — materially reduced hallucinations. The author then maps crypto primitives (fees, staking, slashing) onto incentive design and proposes a composable pattern: use escrowed payments and decaying bonus pools to reward early completion, require worker staking to signal good behavior, and slash stakes for misbehavior — effectively operationalizing a reward function across human agents with cryptographic enforceability. The conclusion: designing incentives is both art and science; iterating reward functions and slashing conditions across humans, agents, and markets is key to aligning actions with desired outcomes, but involves tradeoffs in verification, economics, and game-theoretic robustness.