#VideoReason We are open-sourcing the entire VBVR stack to speed-up the arrival of video reasoning as the next fundamental paradigm of intelligence - 150+ synthetic generators - 1 million training clips - Cloud-scale data factory - Unified EvalKit - 100 rule-based evaluators - Strong baseline model Checkout at video-reason.com

‼️Video models start to reason, let's build-in-public scaled eval framework together 🚀 Early results: grow-ai-like-a-child.com/vid… Codebase: github.com/hokindeng/VMEvalK… (Apache 2.0) 1⃣One-click inference across ALL available models 2⃣Unified API & datasets & auto resume + error handling + eval 3⃣Plug new models and tasks in <5 lines of code Currently, VMEvalkit hosts 39 models across 11 families, and 5 reasoning tasks-Chess, Maze, Sudoku, Mental Rotation, and Raven Matrices-which maximally create 20K+ unique question pairs. With in-house wares, we conducted experiments on the 5 reasoning tasks, and in a glance 1⃣ all video models show non-trivial reasoning ability 2⃣ clearly performance hierarchy across models 3⃣ tasks differ in difficulty, and models display idiosyncratic strengths. For a long time, people wonder if there is such thing as "language-free" reasoning. Aphasia patients could play chess and solve math puzzles, and fMRI🧠studies show distinct brain networks of reasoning and language (Fedorenko, 2022). But counterargument exists as most patients actually lose rather than never acquiring language, and one never acquire is rare (Penfield and Roberts, 1959). The Fodorians believe we reason about the world through some sorts of mental language, the "Mentalese"🧩(Pinker, 1994), while the opposite camp believes the only way to reason meaningfully about the world is through world representations (Hassabis & Maguire 2009; Lecun, 2022), and otherwise it's all tautology (Wittgenstein, 1921). 👇 Let's dig into the results: a thread (1/n)