Software fell first because it is the only knowledge-work domain that ships with its own grader. Inside the verifiable-reward loop that took Claude Code from research preview to ~$9B ARR in a year.
Coding probably moved first because it gives agents unusually fast feedback. Tests, compilers, type checkers and diffs all turn vague progress into something the loop can inspect. The interesting question is which other knowledge-work domains can build their own equivalent of that grader.
The grader framing also predicts the ordering of the next dominoes, and the axis that matters is reward latency, more than verifiability itself. Code returns a verdict in seconds. Search rankings return one in weeks, sales in months, strategy in years: the signal exists but arrives too slowly and too confounded to train against cheaply. So the domains that fall next are the ones where someone compresses the loop - simulators, proxy metrics that track the real outcome closely enough, or environments that replay historical ground truth as if live. Which suggests "who writes the next answer key" is really "who can make a slow answer key fast", and that is a different, harder engineering problem than grading code.
Since the value is in the harness and post-training, this "whoever drives the grader wins" capability diffuses relatively quickly into the AI ecosystem? Does that mean that coding becomes (or has become) a commodity, not tied to the capabilities of a frontier closed model?
Great q. The techniques eventually spread, but building robust graders, reinforcement learning pipelines, and production harnesses requires substantial engineering, data, and iteration. Coding is becoming less tied to a single frontier model, but it isn’t a commodity yet.
Thanks, Chris, I'd argue grader diffusion will likely happen faster than those tied to pre-training. Pre-training requires the $100B that only the fronter labs can muster. Plus RL techniques will like seep into pre-training also (deeper child-like reasoning becomes foundational).
Perhaps model capabilties will widen further due to the capital moat while non-model capabilities diffuse. Which means we get AGI and cost reductions simultaneously :-)
Coding probably moved first because it gives agents unusually fast feedback. Tests, compilers, type checkers and diffs all turn vague progress into something the loop can inspect. The interesting question is which other knowledge-work domains can build their own equivalent of that grader.
Absolutely correct! There will be big companies that bring this grader to other areas.
The grader framing also predicts the ordering of the next dominoes, and the axis that matters is reward latency, more than verifiability itself. Code returns a verdict in seconds. Search rankings return one in weeks, sales in months, strategy in years: the signal exists but arrives too slowly and too confounded to train against cheaply. So the domains that fall next are the ones where someone compresses the loop - simulators, proxy metrics that track the real outcome closely enough, or environments that replay historical ground truth as if live. Which suggests "who writes the next answer key" is really "who can make a slow answer key fast", and that is a different, harder engineering problem than grading code.
Since the value is in the harness and post-training, this "whoever drives the grader wins" capability diffuses relatively quickly into the AI ecosystem? Does that mean that coding becomes (or has become) a commodity, not tied to the capabilities of a frontier closed model?
Great q. The techniques eventually spread, but building robust graders, reinforcement learning pipelines, and production harnesses requires substantial engineering, data, and iteration. Coding is becoming less tied to a single frontier model, but it isn’t a commodity yet.
Thanks, Chris, I'd argue grader diffusion will likely happen faster than those tied to pre-training. Pre-training requires the $100B that only the fronter labs can muster. Plus RL techniques will like seep into pre-training also (deeper child-like reasoning becomes foundational).
Perhaps model capabilties will widen further due to the capital moat while non-model capabilities diffuse. Which means we get AGI and cost reductions simultaneously :-)
This is true, great point! I would agree for sure. Post training is much more accessible than pre training.