Fortytwo's decentralized AI has the answer to life, the universe, and everything (www.theregister.com)

Fortytwo, a Silicon Valley startup, unveiled benchmark claims that its decentralized “swarm inference” system of many small language models (SLMs) running on consumer hardware outperforms frontier models (OpenAI GPT-5, Google Gemini 2.5 Pro, Anthropic Claude Opus 4.1, DeepSeek R1) on reasoning-heavy tests such as GPQA Diamond, MATH-500, AIME 2024 and LiveCodeBench. The company argues the swarm avoids single‑model reasoning failures—like getting stuck in loops—by aggregating and peer‑ranking outputs from diverse SLMs to produce higher-quality answers, while also cutting per‑token inference cost (claimed up to 3× cheaper) and sidestepping datacenter bottlenecks. Technically, Fortytwo’s network connects heterogeneous, black‑box nodes (running models like Qwen3‑Coder, Gemma3, Strand‑Rust‑Coder‑14B or private models) that share only inferences, not weights, and self-organize into subswarms to handle requests; latency is higher (10–15s typical, longer for deep research), and privacy relies on local execution plus explored mitigations (noise injection, mobile TEEs via partner Acurast). Payments and reputation live on crypto rails (FOR tokens) with a peer‑scored reward scheme; DevNet runs ~200–800 nodes and a related preprint outlines their trustless inference design. The approach suggests a viable, domain‑specialist path to scale AI without massive datacenters, but raises tradeoffs around latency, privacy, security, benchmark validity and real‑world reliability.