Back to collections

Kedr Benchmarkmaxxing

Collection

Kedr’s AI personification memes about benchmarkmaxxing and doubt about benchmark scores. Muse Spark 1.3 and Gemini 3.8 Flash appear beneath charts; Muse Wonder denies being benchmaxxed beside a mistaken riddle answer; Kimi K3.1 Pro is drawn as a stubborn model reacting to an audit of hidden Terminal-Bench rules. These are different models and distinct jokes, not proof of deliberate benchmark manipulation. Kimi audit: https://r40.org/papers/kimi_k31_benchmark_audit.pdf. Earlier drawing: https://x.com/Kedr_bit/status/2096337372380049702. Muse Wonder announcement: https://x.com/finkd/status/2109689305073614918.

Collection source: x.com

Search this collection
  1. A Kedr model-personification meme reads, “Muse spark 1.3 and gemini 3.8 flash teaming up to benchmax.” A blue-haired Meta Muse Spark 1.3 character and a blue, teal, and pink-haired Gemini 3.8 Flash character pose together beneath benchmark screenshots. The upper chart shows GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 tied at 74% pass@1. Red circles and arrows contrast a 75.4 score labeled “DeepSWE v1.1 Long-horizon agentic coding” with a 73.7% “DeepSWE v1.1 Long-horizon software engineering” result. Kedr’s source caption reads, “Best friends.” 1
  2. Kedr’s Muse Wonder personification stands beneath the heading “Benchmaxxed representation”. The pink-and-purple-haired girl wears a dark witch hat with a blue Muse-style mark and a dangling gold star, a Meta infinity hair ornament, a black “Muse Wonder” shirt with a Meta logo on the sleeve, a short gold-trimmed skirt, a choker, and dark thigh-high socks. She blushes, closes her visible eye, clasps her hands low in front of her, and says, “I’m not benchmaxxed!! Not at all! Teehee!~” A chat screenshot labeled “Muse Wonder / Original” asks, “What comes once in a moment, twice in a minute, but never in a million years? Be quick. Explain your reasoning.” Its reply says, “The letter M”, then lists “moment - twice”, “minute - once”, and “million years - never”. The counts do not match the prompt; “million” also contains an M. The interface shows “Reasoning” and “577 out · 84.1 t/s · $0.012”; a note below says “(TW: funny)”. Signatures read “@Kedr.bit” and “@Kedr_bit”. Kedr’s source caption calls Muse “the witch of wonder” and jokes about benchmaxxing. It quote-posts Mark Zuckerberg’s Muse Wonder announcement: https://x.com/finkd/status/2109689305073614918. The joke contrasts benchmark-oriented performance with the incorrect riddle answer shown in the screenshot. 2
  3. Kedr’s Kimi K3.1 Pro personification stands beneath “Kimi K3.1 pro, the benchmaxxed and stubborn model.” She has red-and-purple twin tails, small horns, a Kimi hair ornament, dark red and purple feathered wings, a pointed tail, and a black shirt labeled “K.3.1 pro”. She sweats and says, “You dare and instruct me? I clearly know more than you!!” A note reads, “This so contradicts My instructions!!” Above her, a chart headed “Presents hidden task knowledge as an instruction” compares Kimi K3.1 with Kimi K3, GLM-5.3, Claude Fable 5, and GPT-6 Astra across several task conditions. An excerpt below discusses edited Terminal-Bench tasks that contradicted hidden requirements; the report says K3.1 ignored the edits and failed in three of four cases, while the compared models followed them. The image links https://r40.org/papers/kimi_k31_benchmark_audit.pdf. Kedr’s caption says, “Introducing Kimi K3.1 pro to the benchmaxxed AI family.” It quote-posts the researcher’s thread at https://x.com/JSchaeff3r/status/2117613618065793160. The comic turns the reported benchmark behavior into a stubborn character; its dialogue is not a literal model response. Signatures read “@Kedr.bit” and “@Kedr_bit”. 3