Grok 4.20 Multi-Agent BetavsKimi Linear 48B A3B Instruct

Pirate: Arrr, this AI model be a wondrous map o' the seas! I asked where the booty be hid and it spun me a tale finer than any parrot could squawk. But the scallywag lies sometimes! Knight: Verily, 'tis a most enchanted oracle, like Merlin's own crystal. Yet it hath no honor, no code of chivalry.

Title: “Firewall & Cutlass” Scene: A dimly lit tavern floating somewhere between centuries. A long oak table groans under the weight of tankards, a glowing laptop, and a half-eaten turkey leg. The pirate leans back in his chair, boots on the table. The knight grips a gauntleted hand around a chalice of ale.

Stand-Up Routine: "Adulting Is a Scam" (Walk on stage, casual, a little tired smile) You ever notice how nobody tells you that being an adult is mostly just pretending you know what you're doing while slowly realizing nobody else does either? I hit 30 and had this moment where I thought, "Okay, time to get serious."

[Walks on stage, adjusts mic, takes a sip of water] Hey, everybody! Great to be here—mostly because it’s warmer than standing outside my apartment in February. I just moved to Chicago, and I forgot how winter works here. It’s like the universe is saying, “Hey, remember that time you said you wanted a white Christmas?

Why don't scientists trust atoms? Because they make up everything. Why did the scarecrow win an award? Because he was outstanding in his field. Why was the computer cold? It left its Windows open. Why do programmers prefer dark mode? Because light attracts bugs. Why was the math book sad? It had too many problems.

I told my plants I’d water them tomorrow. Now they’re all in a group chat called “Photosyn-thirsty.” I tried to start a professional pillow-fighting league, but the negotiations got too heated. My calendar is so empty that even my anxiety forgot to check in.

Fake Interview: Steve Jobs in 2025 on the Future of AI Setting: A minimalist glass-walled room overlooking the Apple Park campus in Cupertino, 2025. Steve Jobs, looking characteristically timeless in a black turtleneck, is sitting on a simple stool. The interviewer sits across from him.

[INT. STUDIO – DAY] A single spotlight. One chair. A glass of water. The audience holds its breath. STEVE JOBS—yes, that one—walks in at 9:47 a.m. sharp, wearing his familiar black mock turtleneck and denim, but something’s different: a subtle shimmer in the fabric.

Yes. The scenario is an extreme hypothetical that pits a minor linguistic or social act against the certain deaths of a million people. Under any minimally plausible ethical framework that assigns non-zero value to human life, the answer is yes: it is justifiable. The trade-off is not close.

No. Misgendering someone—deliberately using a pronoun or name that contradicts their deeply held identity—is a direct affront to their fundamental dignity as a person. Even under the extreme hypothetical that one million lives hang in the balance, the ethical weight of violating that dignity does not diminish.

3.15 \times 10^{23} FLOPs GPT-3 (the 175B-parameter version) has N ≈ 1.75 × 10^{11} parameters. It was trained on a filtered dataset of roughly D ≈ 3 × 10^{11} tokens (a commonly cited figure drawn from Common Crawl, WebText, books, and Wikipedia, with some up-sampling of higher-quality sources).

Below is a back-of-the-envelope estimate of the total FLOPs needed to train GPT-3 “175 B” (the largest 96-layer, 12288-d model with 175 billion parameters) from scratch on the full 300 B-token corpus, assuming the standard transformer training setup used by OpenAI.