What happens when you lock 12 PhDs in a room for four months? #1 on the ARC-AGI leaderboard This project is by Mostik AI (http://mostik.ai, 'mostik' means 'little bridge' in Russian). The core idea: enabling different models to communicate directly in **latent space**, rather than converting internal computations into text before passing them to another model. Here's how it works: The large model (e.g., the 753B GLM-5.2) handles 'reading the question and reasoning,' passing its hidden states directly to the small model. The small model (e.g., the 4B Qwen that can run on edge devices) handles 'writing the answer and generating text' (decoding). No text is involved in the middle, and neither model needs fine-tuning. They only train a small 'bridge.' In terms of results, they claim the small model achieves about 80% of the large model's accuracy, runs roughly 20x faster, and costs about 1/20th of the large model. Using a similar approach, they’ve climbed to #1 on the ARC-AGI leaderboard. The team consists of about 15 people (12 PhDs + a Fields Medalist), founded roughly four months ago, with backing from General Catalyst. Their goal is to prevent everyone from being locked into a few proprietary, cutting-edge large models, enabling open-source/small models to leverage the reasoning capabilities of large models. Note: General Catalyst (GC) is a top-tier U.S. venture capital firm founded in 2000 by Joel Cutler and David Fialkow in Cambridge, Massachusetts. They now describe themselves as more of a 'transformation company' rather than a traditional VC. GC invests across seed to growth stages, managing assets in the $30–43 billion range, and has backed over 800 companies. Notable investments include Airbnb, Stripe, Snap, HubSpot, Canva, Samsara, Ramp, as well as AI/defense-focused companies like Anthropic, Mistral, Anduril, Helsing, and Applied Intuition.