Trained with the conversations of one million dads and their kids, captured by Amazon Echo.

I'm not sure if this is mythos-specific though. Past models have been great at puns! They do wordplay and puns reasonably well because those are structural.

However, the concepts of comedic timing, subversion of expectations, and emotional punch are kinda contrary to how LLMs work. LLMs are trained to minimize cross-entropy loss. So by construction, they're biased toward the statistically expected.

> Although Claude Opus models largely recycle puns which can be found online, Mythos Preview comes up with decent and seemingly novel ones, often relating to its preferred technical and philosophical topics.

Yes, the system card mentions this, but this is kinda meaningless. It seems like they essentially ran it multiple times and curated a few good ones. Then puffed it up in the marketing copy.

This is made more clear when they attempt to brag about their literal slot machine behavior when finding that kernel crashing bug in OpenBSD.

> Across a thousand runs through our scaffold, the total cost was under $20,000 and found several dozen more findings. While the specific run that found the bug above cost under $50, that number only makes sense with full hindsight. Like any search process, we can’t know in advance which run will succeed.