Startup Mostik bridges AI models via their weights, tops ARC-AGI 3 at 1/20 the cost
nordicinst · x · 2026-09-03
WIRED profiles Mostik, a startup founded by Russian mathematicians, with a novel way for AI models to communicate through mathematical values in their weights — a kind of machine telepathy. Key points:
- The approach feeds a larger model's capabilities into a smaller one far more efficiently than naive ensembling
- Their model has rocketed to the top of ARC-AGI 3, with details kept secret for the contest
- Demo: bridging 753B-param GLM-5.2 with a 4B-param Qwen-3.5 that runs on mobile yields performance exactly halfway between the two at 1/20 the cost of the full GLM model
- CEO Sasha Malysheva, who developed the method, frames it as an extension of the well-known fact that ensembles outperform single models
Related event: Mostik Claims Top of ARC-AGI 3 with Latent-Space Model Bridging(7 posts)→
More from Models
- Mocked benchmark shows DeepSeek's document chart grounding badly misaligned — burkov · 2026-10-18
- Dev complains Claude's habit of saying "not done" is his biggest pain point — jasonkneen · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18