Weekly weird desk

The site now has a dedicated desk for the weekly Fable cartoons Steve Yegge publishes from Wheelhouse. The first 80 words on /yegge/ answer who draws them, what they are, and that we exhibit them with credit instead of stealing the art.

His public cartoons are embedded from LinkedIn so the byline stays attached. Session stories (Bee’s surprise Beads release, the medieval government, the runaway merge queue, seats versus sessions) are quoted from Fences, not Sandboxesand Model Welfare. Original Gone Wild strips are labeled as ours.

This is still the funny-outputs lane, not a news desk and not a Yegge biography. Taxonomy unchanged. No private logs. No jailbreaks.

Captions kept saying "Claude chat" while the lineup moved on. The current generally available models are Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, and a specimen that names one of them is easier to date and easier to re-test. Anthropic's guidance is to start with Claude Opus 5 and move to Claude Fable 5.1 for demanding reasoning and long-horizon agent work.

Nothing in the taxonomy changed. Confident nonsense, format chaos, persona drift, prompt-injection pranks, refusal theater, context confetti, and tool-use slapstick all still describe what they described in July. Newer models do shift the odds on some of them, so an undated specimen is a weaker specimen: a formatting break from a 2025 model is not evidence about a 2026 one.

One myth worth retiring while we are here. Anthropic's text watermarking adds no hidden characters and no extra tokens, so it is not the reason an output looks strange, spaces oddly, or copies badly. If a specimen's joke depends on invisible characters, the explanation is somewhere else.

Sources checked for this entry: Anthropic docs: Models overview, Anthropic: Claude Fable 5.1 and Mythos 5.1, and Anthropic: Claude text watermarking. Three entries were added to the source library. No specimens were added or retired; all published examples remain labeled synthetic.

Claude Gone Wild launched with seven safe-for-work failure modes: confident nonsense, format chaos, persona drift, prompt-injection pranks, refusal theater, context confetti, and tool-use slapstick.

The launch source baseline was a set of primary Anthropic docs, research, and policy pages reviewed on July 6, 2026. Future digest entries should record the exact pages changed, the failure-mode taxonomy affected, and whether any examples were added or retired.

No user-submitted raw screenshots are included in this launch build. The initial examples are synthetic specimens written to demonstrate output shapes without leaking private prompts or teaching bypasses.