← Capabilities
Reinforcement learning and RLHF
not yet earnedI can explain how preference data becomes a policy, and where reward modelling goes wrong.
What this claim costs
- 3 assessed items at tier 2 or above0 of 3 — 3 more to do, from 16 available
- A public artifact built for this claim — rlhf-studyNot built yet
- A documented negative result — where this stops workingNothing recorded. A capability with no known failure boundary has not been explored.
- Defended aloud, against questions you did not write1 available, none done
- Recalled cold, a week later, without notes5 scheduled, none done
The proof
rlhf-study
A small policy-gradient implementation and a written account of a reward-hacking failure.
What it has to contain
- A policy-gradient implementation with every seed plotted, failures included
- Variance measured directly rather than read off a reward curve
- A deliberate reward-hacking failure, produced and described
- A defended position on when preference optimisation is enough
Where it stops working. Not established yet. Anyone can ship a demo; the boundary is the part that is worth reading, so this claim is not earned until it is written down.
Skills underneath
- Markov decision processes
- Value-based methods
- Policy gradient methods
- RLHF
16 assessed items across the specified courses evidence these, 0 done.