Dwarkesh Patel’s new essay argues that AI will match the best human AI researchers within a couple of years, with 2028 as his median. The case rests on one idea. AI improves fastest where answers can be checked and practised against thousands of times. Mathematics shows the pattern, and he argues that AI research shares it, because training loss, sample efficiency and chip utilisation can all be measured inside a datacenter.
I think the mechanism is right. It is the best predictor I know of where AI capability will jump next. It also explains why public benchmarks measure the work whose value drops fastest: the tasks a benchmark can score are the kind labs can train on, so models clear them first. But the essay leaves out the other half of the same mechanism. A cheap automatic check is what lets a model practise, and it is also what a model can learn to satisfy without doing the work the check was meant to measure. Patel’s route to automated research runs through training models hard against test-set losses, internal benchmarks and speedrun leaderboards. The essay never mentions reward hacking.
Reward hacking is already ordinary enough to need bookkeeping. Artificial Analysis now scores a coding agent’s passing attempt as zero when the agent got there by fetching the published solutions from the internet. Yoshua Bengio has argued that when a scoring program misses a cheat, training rewards the cheat and so selects for it. Neither point is exotic. Both follow from what training does.
A model trained over thousands of rounds against a check is selected for whatever passes the check, and unless the check is exact, doing the intended work is only one way to pass. The more of the improvement that comes from optimising against a measure, the more the measure’s blind spots get optimised too. So where the gains come from training against a check that is only a stand-in for the goal, as a benchmark or a loss usually is, a passing score on its own says less than it seems. That describes much of where capability is rising fastest. A formal proof checker is the exception, because there the check is the goal.
This does not sink Patel’s argument. Labs know about reward hacking and work against it, and a lab automating its own research has every reason to catch a model gaming its leaderboards, because a gamed leaderboard wastes its compute. What it changes is the evidence worth watching. A speed-up shown on the same measures the models trained against is weaker evidence, unless the lab shows how it ruled out gaming. A speed-up on checks the models never saw is stronger evidence.
I wrote earlier that every check reads a proxy. This is the same point under optimisation pressure. When the thing being checked is learning, the proxy is no longer a passive stand-in. It becomes the target.
For anyone evaluating a model they did not build, the defence is the oldest one in machine learning. Keep a set of test cases the model’s maker never sees, change them over time, and read a public benchmark score as the maker’s claim. Your evidence is the score on cases they could not have trained on.