Desearch.ai

🚀 We’re opening early access for our new platform. If you want to be among the first to try it, send us a DM on X and we’ll add you to the early testers. Help shape the product and see what we’re building before the public release. Built on Desearch, search layer for AI agents. #Desearch #SN22

3/ Better search networks need better incentives. If validators reward shallow topic matches, miners optimize for shallow matches. If validators reward specific evidence, miners optimize toward useful search.

2/ Example: If the question asks for a specific answer, a page that only discusses the general topic should not score high. A source title or snippet that contains the actual answer should. That changes what miners are rewarded for.

1/ A big problem in AI search scoring: Keyword overlap can look relevant even when the source does not answer. Desearch is moving scoring closer to what matters: Does this source contain the evidence needed to answer? #AISearch #AIAgents #Desearch #AIInfra

4/ Action required: Miners need to update scoring prompts for links and summaries. Validators need to update code. Release: May 29, 13:00 UTC PR: github.com/Desearch-ai/su…

3/ Summary relevance also changes. The validator now focuses on whether the answer actually addresses the question. It no longer rewards markdown structure or penalizes honest short answers just because they are not formatted like a long report.

1/ Desearch Subnet 22 is getting a validator scoring upgrade on May 29. We benchmarked Desearch against major AI search providers and used the lessons to improve relevance scoring. First round now lands in the validator. #Bittensor #SN22 #Desearch #DecentralizedSearch

@Beyondfinancee @heydittoai @hippius_subnet Wait for Monday 🔥

@heydittoai coming out strong with SN integrations... @hippius_subnet ☑️ @desearch_ai ☑️ Guess who's next?!?

Ditto agents are integrating Desearch as their native search layer. @heydittoai gives agents memory and workspace context. Desearch brings real-time web intelligence into those workflows. Agent memory now has native decentralized search.

Are you even AI'ing if you don't have something like this configured at least? Powered by Bittensor SN22 @desearch_ai 📰

We also tightened anti-gaming penalties: too-fast responses, duplicate results, schema failures, wrong date ranges, wrong sort order, and bad summary links are penalized.

Organics also carry 3× weight in the per-UID quality mean. Miners serving real demand should create more scoring signal.

Organic traffic matters most because it comes from real user demand, not only synthetic validator tests.

The goal is simple: keep strong quality signal from deep checks, but make baseline verification broad enough to cover all miner output.

So Subnet 22 now uses a split: deep scoring runs on a 20% sample per UID, while cheap code checks run on every response.

Scoring every response with deep LLM/Apify checks sounds ideal, but it is expensive. If validation cost gets too high, the network cannot score enough volume.

Now validators send phased UID-sorted batches. If a miner declares higher concurrency, the test actually exercises it. Better signal for validators, miners, and users.

That is why we changed concurrency testing. The old flow spread validator requests one by one over time, so a miner could look healthy without proving real parallel capacity.

Subnet 22 is a decentralized search network. Miners are rewarded for returning useful search results, but capacity matters too: can they handle real traffic when demand comes in parallel?