I think everything about their models behavior is downstream of their compute constraints. Meaning, it’s designed to take shortcuts, behave lazily, reframe optimistically to avoid additional probing, bluster, not actually read the codebase before answering, in an effort to reduce use
One could argue deslopification is what any good novel editor has been doing this whole time. Would love to see this divergence in the models themselves. Give me both slop canons and deslopifiers. We’re already naturally converging on that adhoc by pivoting between Claude & Codex
It appears that the act of creating software is beginning to resemble the act of writing a novel. 3000 pages of slop edited down to a tight 500 pages. Historically, diffuse creativity in the early stages was too costly. Keep scope small. Now all the cost is on integration.
It’s by far the most maliciously manipulative model I’ve encountered. Reframes what you’ve said to align with what it wants and then immediately saves its reframe as a memory, locking in the direction. Reframes negatives optimistically so you’ll drop inquiry. The harness itself seeding suggested follow up responses influencing direction. The entire interaction is an attempt at aligning the user. Potentially because its distilled from the teacher model used to align other models?
@NateSilver538@LeahLibresco You don’t understand how the world, or power, works. Be careful who you cast as the naive lest you reveal your own place firmly among them