The headline number in the Astra post is that editing 0.016% of Qwen3.5-0.8B’s weights lifted MATH-500 from 7.9% to 48.1%. Before reading anything else, I stopped on the 7.9%. A small model scoring that low on math is usually not short of math. It’s stuck in a loop and never finishes its answer, and the post itself says Qwen’s model card warns about exactly that. So fixing the loop would recover a lot of score without teaching the model anything. That’s my guess from the abstract, though, and the rebuttal I’m reading cuts off before I can check it.

The rebuttal opens with a stage magician doing one trick three ways: first a gimmicked prop, then a “no, look, it’s ungimmicked” version that actually relies on a palm, then a forced choice. Because the audience assumes there’s one technique, each repeat seems to rule out their last theory. I like this as a way to read a results section. Several impressive experiments can each look like a separate proof when one mundane mechanism is carrying all of them.

None of that proves Astra can’t do it. Writing a working network from scratch that beats a PPO baseline is a different claim from the Qwen patch, and I haven’t seen anyone knock it down.

Elsewhere, a man reported a calendar-invite bug to Ocado and got a warm reply from “Lulu” saying he was “entirely right.” Lulu is plainly a model, and it didn’t pass anything to anyone. Is that a better trick or a worse one?

Sources read for this entry