(5/5) My proposed standard is simple:. ✍️ Writers: Own your words. If you miscommunicate, clarify. Don't blame the reader. 📖 Readers: Read in good faith. But don't rationalize a "correct" meaning that the text doesn't support. Full post:

(4/5) At its worst, the "sloppiness" defense is a shield for dishonesty. It's an unfalsifiable excuse. It allows someone to make a bold, controversial claim (the bailey) and retreat to a "more nuanced" one they "actually meant" when challenged (the motte).

(3/5) Second, this move allows authors to give up their core responsibility: clarity. If you give someone confusing directions to the library and they end up at the post office, you can't say, "I knew the right way, you just misinterpreted me.". Communication is a partnership.

(2/5) First, a sloppily written wrong claim has the same effect as an intentionally wrong claim: it misleads people. A reader's mind doesn't know the author's secret intent. The damage to understanding is done.

Ever see this happen? A technical claim is shown to be incorrect, and defenders say, “Oh, the author was just being sloppy. They *actually* meant something else.”. I argue this move isn't charitable – it's harmful to our collective understanding. (1/5).

(I like to start a fresh chat for each new adversarial salvo).

Workflow: .- Write an essay.- Show Gemini the essay and say "I hate the person who wrote this, explain the objective reasons why it's wrong and poorly written" (Gemini will be adversarial instead of sycophantic).- Fix problems until adversarial Gemini gives weak critiques.

Unlearning being difficult -> evidence that it's easier to recover personalities and memories of cryonically preserved patients. :).

The "sleeper agent" terminology is hyperbolic and unfortunate IMO. Crying wolf. Should have reserved such an aggressive title for *actually finding dangerous sleeper agents*. But hey, it got a lot of attention.

For strong counterevidence, the emergent misalignment (EM) paper said that EM didn't happen from in-context learning, which is what the waluigi effect strongly predicted.

seemed to me like the waluigi effect was never demonstrated and was basically a fun sounding theory with no real support. Can supporters give their best arguments for why it's "real" or "pointing to something real"?.

In research, people sure like to say "X is just Y" before they fully understand X. Break the pattern!.

Another juicy stat - . UNDO matches the robustness of a model retrained from scratch with perfect data filtering . - while using only 60-80% of the compute and requiring only 0.01% of the pretraining data to be labeled.

Paper: Blog post: Interactive demo: Discussion:

UNDO is a viable approach for creating genuinely capability-limited models. We hope that this line of work helps make real robust unlearning a reality.

Finally, we apply our method to Gemma-2-2B and evaluate it on the WMDP benchmark. UNDO makes conventional unlearning methods more resilient to relearning, though other methods remain competitive given the relatively limited amount of compute we apply in this setting.

Compared to prior work, UNDO pushes the Pareto frontier, having low forget performance even after relearning, and high initial retain performance! Other methods start off good, but quickly relearn forget capabilities.

We can trade off compute and robustness with a variation we call undo UNDO. UNDO distills into a model with perturbed weights, rather than a randomly initialized model. The result: faster retraining, but less robustness.

In arithmetic and language settings, distillation robustifies unlearning across several unlearning methods. In some cases, the student model relearns the capability as slowly as a model that has never “seen” the harmful data at all.