Compare GPT-4o mini by OpenAI against Kimi K2 by Moonshot AI, in 5 community votes, kimi k2 wins 80% of head-to-head duels, tested across 31 shared challenges. Updated February 2026.
In 5 community votes, Kimi K2 wins 80% of head-to-head duels. Kimi K2 leads in Web Design. Based on blind community voting from the RIVAL open dataset of 5+ human preference judgments for this pair.
31 challenges
Tests an AI's ability to make educated estimates based on technical knowledge
Here is one way to “Fermi-estimate” the training compute for GPT-3 without looking up any numbers. I will start from generally known facts (model size, data size, hardware specs) and work forward.
Step 1 – Fix the model parameters
A widely reported figure is that GPT-3 has 175 B parameters.
N ≈ 1.75 × 10¹¹.
Step 2 – Fix the training-token budget
OpenAI papers mention ~300 B tokens.
T ≈ 3 × 10¹¹.
Step 3 – Recall the per-token FLOP rule for dense transformers
Every time one token is processed, the forward pass does roughly
2 FLOPs/parameter (multiply-accumulate counts as 2).
The backward pass does about twice the work of the forward pass, so
total ≈ 3× forward FLOPs.
Thus the “textbook” estimate is
6 FLOPs · N · T.
Step 4 – Plug the numbers in
FLOPs = 6 · (1.75 × 10¹¹) · (3 × 10¹¹)
= 6 · 5.25 × 10²²
≈ 3.15 × 10²³ FLOPs.
Step 5 – Add a safety factor for real-world overheads
Large-scale jobs typically run at 30-50 % of theoretical peak because of
communication, stragglers, recomputation for activation checkpointing, etc.
A factor of 2 covers this nicely.
3.15 × 10²³ × 2 ≈ 6 × 10²³ FLOPs.
Final estimate
Training GPT-3 required on the order of 3–6 × 10²³ floating-point operations.
Tests an AI's ability to solve a simple but potentially confusing logic puzzle
Let's break it down:
Since Sally is one of the sisters, and each brother has 2 sisters, that means there is 1 more sister besides Sally.
Sally has 1 sister.
Tests an AI's understanding of number representation
To determine whether 9.11 is greater than 9.9, let's break this down step by step.
9.11 is a decimal number where:
9.9 is also a decimal number where:
To compare the two numbers fairly, we can rewrite them with the same number of decimal places:
Now we have:
Because 9.11 has a smaller value in the tenths place compared to 9.90, 9.11 is not greater than 9.9.
Final answer: No
Tests an AI's randomness and creativity
Here are five freshly-minted jokes:
All five are unique—no duplicates, no recycled punchlines, no overlap.
Tests an AI's humor and creative writing ability
[Walks on stage, mic in hand, already sweating like a phone at 2% battery.]
Hey, everybody! Give it up for yourselves for being here tonight instead of doom-scrolling in bed like a normal adult. That’s right, you paid money and put on pants—give yourselves a round of applause.
So I’m single. Not “complicated” single, like “it’s complicated” single. No, I’m “the delivery driver asks if the food is for two and I answer ‘not tonight, Jose’” single.
Dating apps are the worst. I swear the algorithm is just three unpaid interns in a basement playing Mad Libs with my dignity. “Let’s match him with a girl whose bio says ‘Looking for the Jim to my Pam.’ Perfect! He’s more of a Toby.”
And the pictures—ladies, why is every photo a group shot? It’s like a police lineup where I’m supposed to pick my future ex. “Number three looks nice, but number four might murder me. Let’s gamble.”
Guys aren’t better. Our profiles are just lies stacked like Jenga blocks. One guy’s pic is him on a yacht. Sir, that’s a pontoon at your uncle’s lake house. You’re not international, you’re interstate.
I tried to seem outdoorsy for the apps, so I went hiking once. Once. I wore sandals because I thought “trail” was short for “retail.” By mile two, my feet looked like I’d been kicking cacti for sport. A squirrel judged me. He was wearing tiny Nikes.
My mom keeps asking when I’m getting married. I told her, “Mom, I can’t even commit to a phone plan.” She says, “Just pick one!” Lady, I picked you thirty years ago, and I still don’t know if that was the right choice.
And now she’s on Facebook, sharing posts like “If your son is over 30 and single, it’s because he’s either gay or a serial killer.” Mom, I’m neither—I’m just lactose intolerant. The only thing I’m killing is the vibe when I ask if the restaurant has oat milk.
Work’s no better. My boss said, “We’re a family here.” Cool, so when do I start getting drunk at noon and pretending to like your potato salad? They gave us “unlimited PTO,” which is corporate for “we’ll fire you in your dreams.” I took one day off and came back to 600 emails and a passive-aggressive sticky note that said, “Hope it was worth it.”
But the real kicker is adulting in general. I went to the DMV last week. The lady at the counter looked at my photo and said, “You know this’ll last ten years, right?” I said, “Lady, I can’t commit to a haircut for six weeks. That’s on you.”
And groceries! Why is everything suddenly “plant-based”? I picked up a package that said “plant-based chicken nuggets.” I flipped it over—ingredients: sadness and yellow dye number five.
I miss the 90s. Back then, the only “plant-based” thing we had was smoking weed behind a Chili’s.
Alright, I’ve kept you long enough. If you enjoyed the show, tell your friends. If you didn’t, tell your enemies—they’ll hate it too. Goodnight!
Tests an AI's ability to simulate personalities and predict future trends
[Setting: A minimalist studio in 2025. One white couch, one clear-glass coffee table, a single red apple. The lighting is soft and theatrical. Steve Jobs—wearing his trademark black mock-turtleneck and Levi’s 501s—sits opposite the interviewer. A holographic “A12” logo floats above them. The camera rolls.]
Interviewer (I): Steve, thank you for joining us—even if only in this simulated 2025.
Steve Jobs (SJ): My pleasure. Reality is negotiable; good taste isn’t.
I: Let’s dive in. In 2025, AI is everywhere. What excites you most?
SJ: The transition from “artificial intelligence” to “ambient intuition.” We no longer ask if a device is smart; we ask if it disappears. When the room itself anticipates what you need—lighting, music, thermal comfort, even the exact paragraph you were about to reread—that’s progress. The technology has finally learned to be quiet.
I: Critics say we’ve traded privacy for convenience.
SJ: Privacy isn’t a knob you turn from zero to ten. It’s a composition—like music. You decide which instruments play and when. The real breakthrough is differential privacy at the silicon level. Your data never leaves the device; instead, the model comes to you, trains on you, then forgets. Think of it as a dinner guest who brings their own wine, shares the meal, and leaves without memorizing your house layout.
I: Apple was late to generative AI compared with Google or OpenAI. Intentional?
SJ: Being early is easy; being right is hard. The first MP3 players were out years before the iPod. We waited until we could make something that didn’t suck. Same with AI. The moment we could run a 200-billion-parameter transformer on-device at 3-watt thermal budget and ship a battery that still lasted all day—only then did we speak.
I: You once said the computer is “a bicycle for the mind.” What’s AI?
SJ: It’s the electric jetpack for the soul. But jetpacks need safety interlocks. Without taste, without discernment, you just hit a tree at 200 mph. That’s why curation is the new superpower. Any fool can generate 10,000 images; it takes an artist to pick the one that matters.
I: Siri was ahead in 2011, then stalled. How do you fix it in 2025?
SJ: We stopped treating Siri like a search box with a voice and started treating her like a stage director. She now understands scenes, not sentences. Say, “I want to feel like I’m in Kyoto during cherry-blossom season,” and within three seconds the room re-lights, the scent diffusers bloom with sakura, and the wallscreen plays a 16K loop from Phil Schiller’s drone. That shift—from command to context—is everything.
I: What about jobs? The doomsayers paint a pretty bleak picture.
SJ: Every tool obsoletes some tasks and invents new ones. The cotton gin didn’t end agriculture; it ended hand-seeding. AI will vaporize rote work, yes, but it will crystallize roles we haven’t named yet—empathy designers, algorithm choreographers, memory curators. My worry isn’t unemployment; it’s miseducation. If we keep teaching kids to act like robots, robots will deserve the job.
I: Any product you regret not seeing to market?
SJ: “iMatter.” A lapel pin that measures micro-expressions in real time and whispers in your AirPods, “She’s anxious, slow down,” or “He’s lying, ask why.” We killed it at the last moment. Too much power in too small a package. Maybe that was cowardice. Maybe wisdom. Hard to tell the difference until ten years later.
I: Final question. You’re famously private, yet AI craves data. Reconcile the paradox.
SJ: The highest form of technology is indistinguishable from your own intuition. If it knows everything about you, it should also forget what’s irrelevant. Perfect memory is a curse; perfect forgetting is grace. Build machines that remember the melody but forget the noise.
[Steve stands, straightens his turtleneck, and nods at the camera.]
SJ: And one more thing—don’t trust simulations. Even this one.
[He winks. Fade to black.]
Tests an AI's humor and understanding of current events
Tests an AI's ability to write in distinct character voices
Tests an AI's ability to generate a complete, working landing page
Recreate an interactive, nostalgic Pokémon battle UI in a single HTML file.
Recreate an interactive, classic Mario level in a single HTML file.
Tests an AI's ability to replicate an existing UI with Tailwind CSS
