Desearch.ai | Subnet 22

SN22 now supports Bittensor 10.3.0. The release removes a codec conflict that could break validator startup and updates the install path. Network upgrades should be boring. That means fixing dependency edges before operators hit them. #Bittensor #SN22 #Desearch

SN22 now supports Bittensor 10.3.0. The release removes a codec conflict that could break validator startup and updates the install path. Network upgrades should be boring. That means fixing dependency edges before operators hit them. #Bittensor #SN22 #Desearch

If answers only appear after JavaScript runs, many crawlers never see them. Desearch now server-renders product FAQs and exposes crawlable hubs for 30 guide and 35 API-reference pages. Docs should work for people, search engines, and agents. #TechnicalSEO #AIAgents #Desearch

If docs only work after JavaScript runs, many crawlers never see them. Desearch now serves real 200-status hubs linking 30 guide pages and 35 API-reference pages. Docs should work for people, search engines, and agents. desearch.ai/docs/guide #AIAgents #Desearch

Fast AI search should be measured by its tail, not just its best run. The public Desearch benchmark now shows median and p90 response time beside source relevance, answer quality, and groundedness. Quality and latency, on the same public scoreboard. 22.desearch.ai

Fast AI search should be measured by its tail, not just its best run. The public Desearch benchmark now shows median and p90 response time beside source relevance, answer quality, and groundedness. Quality and latency, on the same public scoreboard. 22.desearch.ai

Desearch AI Search now routes full-answer requests through the validator first, with Miner API fallback when that path is unavailable. That connects product traffic to the SN22 quality loop without making reliability depend on one route. #Bittensor #SN22 #Desearch

Desearch AI Search now routes full-answer requests through the validator first, with Miner API fallback when that path is unavailable. That connects product traffic to the SN22 quality loop without making reliability depend on one route. #Bittensor #SN22 #Desearch

Search becomes infrastructure when another product can build on it. We’ve been working with @Bitcast_network on what comes next for Desearch × Bitcast. Public details land next week. #Bittensor #SN22 #Desearch

Long-running AI work should expose state. Desearch forecasting returns a prediction job with reasoning, polling status, and citations. For agent builders, “still working” is part of the product contract. desearch.ai/numinous-forec… #AIAgents #Desearch

A source should win because it answers the question, not because it looks like the right type of page. SN22 PR #370 uses Qwen3.5 397B and a clearer rubric. A wiki, list, or news page can score equally when it contains the answer. #AISearch #Bittensor #SN22

More miners should not let one search type absorb the whole reward pool. The new SN22 design fixes emissions at 90% for AI Search and 10% for social search. Within AI Search: 60% fast, 20% balanced, 20% deep. Each lane competes on its own quality. #Bittensor #SN22 #Desearch

Fast, balanced, and deep search are different workloads. One weak mode should not hold back the others. SN22 PR #370 separates capacity and routing by mode, so organic traffic follows proven quality. github.com/Desearch-ai/su… #Bittensor #SN22 #Desearch

Miner action: Update manifest.json to declare fast, balanced, and deep capacity separately. A single ai_search value still works during migration and applies to all three modes. Fast mode is removed for social search. PR: github.com/Desearch-ai/su…

SN22 burn is now 50%, down from 70%. A validator release planned for 13:00 UTC today adds per-mode capacity, mode-aware routing, fixed emission pools, and better link scoring. A major update to how search quality and rewards move through the subnet. #Bittensor #SN22 #Desearch

Before an agent sends, buys, or updates anything, add a retrieval checkpoint. Live sources + citations: continue. Stale or conflicting evidence: pause. Desearch is more than an answer layer. It helps agents know when to act and when to stop. #AIAgents #AISearch #Desearch

AI Search results should not be buried in raw data. In development for Desearch Console: visual cards for Web links and X posts, with answers and sources shown together. A clearer path from answer to evidence. #AISearch #Desearch

One weak hour should not wipe out a miner’s progress. SN22 v0.0.223 uses more representative quality checks and reduces sudden reward drops. More stable and fair scoring. #Bittensor #SN22 #Desearch

A links-only answer should be judged on its links, not receive credit for a summary it never produced. SN22 v0.0.223 now scores each answer based on what was actually requested. Clearer requests. Cleaner incentives. #Bittensor #SN22 #Desearch

Every miner should receive a fair, representative test. SN22 v0.0.223 now uses a predictable mix: 60% fast, 20% balanced, 20% deep, with Web and X split evenly. Less testing luck. Better quality signals. #Bittensor #SN22 #Desearch