I had an idea for a sketch today and I want to make myself actually write a script so I'm making a market. I want people to bid in the market so I'm making it about AI. I have ~2.5 solid jokes, I need ~2.5 more solid jokes and a tight bit of exposition, and I hope to finish the sketch by the end of Sunday.
I'll write the scaffolding for the sketch, write in the jokes I'm happy with already, and try to build around them. Whenever I hit a point where I suspect a joke could work and the joke isn't so obvious that I can fill one in without thinking, I'll send the current script and my understanding of the sketch off to the listed LLMs with a request for several jokes that could fit there in the sketch. If any of the jokes are funny or could be funny with enough reworking (in my opinion), I'll resolve this market as "yes" for that LLM, even if I don't end up using it. If none of the jokes are funny or I think I could do better, I'll spend some time brainstorming and see if I can put anything funny there. If the LLMs fail and I fail, I'll move on and I won't put a joke there. I'll continue this process until I have a 2-2.5 page sketch. If I get stuck (can't make progress and don't have anywhere else in the scaffolding to jump to), I'll delete something I've written and try a new approach, which may mean deleting jokes but will not count against any LLM that wrote a joke I liked for that section.
I'll be sending the requests off through Openrouter, so every LLM will see exactly the same prompt. Each new request will be in an entirely new conversation, context won't accumulate.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ10 | |
| 2 | Ṁ8 |
Resolution report: Sonnet got there!
Some of this may have been prompt failure. When you tell LLMs to be funny they start by trying to write jokes, but this is the wrong order for this creative process. When you tell them what you think the funny part of the sketch is, the core comedic premise that makes jokes work, they latch onto the jokes rather than understanding the premise, leading to slop, which is inherently not funny.
Grok 4 in particular is really bad. It's just desperate to get to a punchline, it wants to please the user and so assumes that whatever jokes I've already written in are the gold standard. It builds on jokes in a way that seems cloying and pathetic, rather than building on the premise. I started giving more explicit instructions and "disallowing" certain types of suggestions in the end, which prevented this behavior but may have suppressed creativity in other ways.
Gemini 2.5 Pro was good at some technical stuff, it could understand the relationship between setup and punchline, but it just isn't funny, I don't know how to describe it. You know how explaining jokes kills them? Gemini has trouble telling jokes, its telling always feels like an explanation. I think there may have been some salvageable premises in these, but the way Gemini describes them is so off-putting that I couldn't bring myself to punch up any of its suggestions.
GPT-5 Chat is just nothing. This may have been user bias, since I probably use GPT-5 the most in everyday settings, so I'm tuned into its ticks and quirks, but it has an overwhelming "polished slop" feel. More creative than Grok 4 and Gemini 2.5 Pro, but again, just too off-putting to be funny, and too attached to tropes and formulas.
Sonnet 4.5 is pretty hit-or-miss. I got the vibe that it loves its own ideas, or more likely was pretending to love them, even when they were bad. It would defend good and bad suggestions with equal fervor, seemingly without an internal quality monitor. It succeeded mostly on creativity and prompt volume. I don't think it actually stuck any punchlines, maybe a few of them could have been salvaged but I didn't see how. However, when I was struggling with exposition I asked the LLMs for framing devices, which broke the pattern of punchline-chasing, and Sonnet 4.5 recommended a sort of genre parody that I liked and thought could work. It was a bit too SNL and I ended up not using most of it, but I thought the idea was good enough for a "Yes" resolution.