Bittensor guard model gains 8 F1 points in 4 weeks to near-SOTA via miner attacks
bittingthembits · x · 2026-10-06
Trishool's specialized guard model on Bittensor subnet SN23 climbed from 73.9% to 78.53% to 81.87% F1 in four weeks — less than a point from SOTA, with the gap to perfect shrinking from 26.1 to 18.1 (30% less error).
The mechanism is unusual:
- Adversary-driven iteration: miners across the network attack the guard nonstop; every successful attack becomes training data, so it improves weekly, not yearly
- A general model does everything; a specialist does one job and gets attacked on it relentlessly
- Trishool says its input guard is already top of its class; the output guard is near SOTA
- Astroware, the team running SN23, was accepted into OpenAI's Trusted Access for Cyber program
The author argues most people haven't priced in that a paid-to-find-weaknesses network can improve specialist models faster than the general-model race.
More from Models
- llama.cpp beats Ollama by ~30% prompt eval speed on RTX 5060 Ti with Clef Flash 9B — ngxson · 2026-10-06
- Rumored Anthropic November IPO raises questions about Fable 5.5 October timing — haider1 · 2026-10-06
- Dev measures app security in 'minutes until GLM escapes the sandbox' — current record: 2 — lucasmeijer · 2026-10-06
- Dev says OpenAI models ignore existing design conventions, Opus still preferred for product engineering — almmaasoglu · 2026-10-06
- Economist: since GPT-5, labeling disagreements are almost always the LLM being right — soumitrashukla9 · 2026-10-06
- Dev argues Persimmon is the most underrated Western lab release: realistic user simulators may be quietly powering RL environments — willcb · 2026-10-06