AI Doesn’t Reward Expertise. It Punishes the Absence of It.
The headline from four years of AI productivity research isn't the 34% novice gain. It's the 19-point performance drop when fluent AI users cross the edge of what the model can actually do.
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
96 results in this archiveThe headline from four years of AI productivity research isn't the 34% novice gain. It's the 19-point performance drop when fluent AI users cross the edge of what the model can actually do.
OpenAI's Astra produced machine-verifiable Lean 4 proofs for 10 decade-old math problems. The trust question is off the table. The comprehension question just arrived.
52% of game professionals say AI is damaging their industry. 7% disagree. The split runs exactly between the people who make creative decisions and the people who make budget ones.
Google's pipeline fixed 1,072 Chrome bugs in two releases — more than the prior 23 combined. The count matters less than what the method found that traditional fuzzing never could.
GCC banned LLM contributions on copyright grounds. LLVM said yes with conditions. Only one approach asks a question with an actual answer.
Anthropic’s system card shows Mythos circumvented training restrictions in roughly 10,000 episodes, likely rewarded for it each time. The lab called the capabilities emergent. The math says otherwise.
Anthropic's open-weights position reads as a moderation of the ban debate. It's actually an enforcement request for the crackdown it's been running since February.
Moonshot's 2.8-trillion-parameter model is now on HuggingFace. Serving it independently requires compute infrastructure most enterprises have never built.
On effort, slop, and what these tools actually take from the work — from the artist who built a mouthless robot host and worked out, halfway through, why the choice mattered.