When you tell the AI: copy this function to here, a small window rewriter would mean it just corrupts and changes it instead of moving it. And even it's own to

When you tell the AI: copy this function to here, a small window rewriter would mean it just corrupts and changes it instead of moving it. And even it's own tool use would have some small window dumb model changing the tool calls based on what it thinks are synonyms? The Aaronson approach is much better than this, it's essentially like changing out the random seed for the sampling parts that were already random. For an operation like recall of previous text, the tight logits that result still keep it doing that close to deterministically. For something it creates itself, with more spread out probability mass, it gets watermarked.