Thariq Shihipar · Anthropic · 2026-03-28
A short, unusual essay that takes a technical concept from reinforcement learning — sparse rewards — and runs it forward into something philosophical. Sparse-reward problems in RL are the ones where feedback comes infrequently and unpredictably, and the agent has to figure out which long sequence of choices led to it. Shihipar argues that the human pursuit of meaning has the same shape.
The bridge is that "enlightenment" — moments of joy that aren't caused by anything you did and aren't addictive when they arrive — is the rare-and-unpredictable reward signal of being alive. Most actions don't earn it. A few do, and you can't reliably tell which. The essay's claim is that this should change how you orient toward action: not toward optimizing the reward (impossible, the function is unknown) but toward the quality that seems to correlate with it, which he names goodness.
> The hardest problems are those with sparse rewards — where the model only receives feedback at rare, unpredictable moments.
> It makes you feel that nothing really matters, but that is also why everything matters so much.
> Truly, the only reward for goodness, is goodness itself.
Most writing from inside the labs is technical-or-strategic. Posts that take the technical vocabulary of ML and use it on the rest of life are rare and worth reading. They also do something curriculum-on-AI usually skips: they remind you that the people building these systems are thinking about meaning, not just capability. If you teach AI to anyone, this is the essay for the day they ask "but what's it all for."