SN22 validator scoring keeps getting stricter. The latest update diversifies synthetic AI Search questions, separates web and X relevance by tool type, includes quoted tweet content, and improves duplicate handling. Quality should matter in rewards. #Bittensor #SN22

When AI Search is streaming, the UI should not look empty. We updated the Desearch Console so early events stay visible while the final answer is still being built. Small UX detail. Big trust detail for developer tools. #AISearch #AIInfra #Desearch

Link only search results should not feel like raw plumbing. In the Desearch Console, ONLY_LINKS responses now render as visual cards with titles, snippets, tweets, URLs, and cleaner deduping. Better APIs need better debugging surfaces. #AISearch #AIInfra #Desearch

Search quality for AI agents is not just more links. It is fresh data, useful sources, grounded answers, and clean structure the model can use. That is the layer Desearch is building for products that need live web and social context. #AISearch #AIAgents #Desearch

Desearch is showing long-term conviction onchain. 763.7K SN22 alpha is perpetually locked under Bittensor's Conviction upgrade, about 24.1% of staked SN22 alpha. 501K by Desearch. 262.7K by external holders. Conviction for decentralized search. #Bittensor #SN22 #Desearch

IntoTAO covered Desearch clearly. SN22 turns search into a live quality loop: miners fetch useful data validators score output builders get it through an API That is how a subnet becomes a product. intotao.app/subnets/desearch #Bittensor #SN22

Web search scoring is tighter in SN22. Relevance now uses an LLM check. Snippets are verified against the page body. Filler snippets score 0. site: queries are enforced. Genuine web results matter more than keyword matching. #AISearch #SN22 #Desearch

Subnet 22 validator update is merged. Rewards now split 80% AI search, 10% X, 10% Web. Each type has a quality bar, and real volume matters. Serving only one type no longer earns full credit. PR: github.com/Desearch-ai/subne… #Bittensor #SN22 #Desearch

Benchmark v1 is only the start. The real goal is a stronger measurement layer for decentralized search. More runs. Better question sets. More categories. More pressure on miners to improve. Search gets better when competition is measured. #Bittensor #SN22

We made the benchmark data public. Questions, answers, sources, scores, and verdicts are available for builders and researchers to inspect. Dataset: huggingface.co/datasets/dese… Better evals create better incentives. #Bittensor

This benchmark does not use a static answer key. The judge reads the returned answer, cited pages, and source content. We score: Source relevance Answer quality Groundedness Search quality should be measured on what users actually receive. 22.desearch.ai

Source relevance matters because bad sources make good looking answers useless. In our first benchmark, Desearch posted the strongest source relevance score: 92.3%. That is the signal we care about: search that finds pages worth citing. 22.desearch.ai #AISearch #SN22

The key idea behind decentralized search: Miners compete to return better answers and better sources. Validators measure quality. Rewards move toward better behavior. The product improves in public. That is the loop we are building on Desearch. #Bittensor #SN22

Bittensor capturing all the excitement at the Louve this week for @proofoftalk. Amazing to connect with @GordonFrayne, @SiamKidd, @Raleigh_CA and so many others from the community. Keep the conversations coming for Day 2.

Open benchmarks beat black box claims. We published the Desearch search eval repo so builders, miners, validators, and researchers can inspect how the benchmark works. Repo: github.com/Desearch-ai/desea… #Bittensor #DecentralizedSearch

🚨 $TAO’s Desearch SN22 is one of those updates most people will scroll past, but it matters. Everyone is chasing chatbots. Desearch is building the layer chatbots and agents cannot work without: Fresh, reliable, live search. They just shipped a validator scoring upgrade, PR #355, live May 29. Sounds boring. It is not. The old scoring rewarded keyword overlap. Basically, if a page mentioned the topic enough, it could score well even if it never really answered the question. That is the problem with most of the internet today. SEO spam ranks because it has the right words, not the right evidence. Desearch changed what gets rewarded. Now miners score higher when their sources actually contain the evidence the user asked for. Intent match over word match. They also made scoring more consistent, stopped rewarding long markdown answers for no reason, and stopped punishing short answers when they are actually correct. This is the part I think people miss. In a normal search company, an engineer changes the model behind doors and everyone just hopes it gets better. On Desearch, the reward function changes in public. And once the validator rewards better evidence, thousands of miners are pushed to optimize for better evidence. Change the reward, and the whole network re-optimizes. The Bittensor difference. Desearch is not just building another AI search tool. It is building an open search infrastructure layer for agents. Live web. X. Reddit. TikTok. Facebook. Instagram. Arxiv. Forecasting. API access. SDKs. MCP server. 99.997% uptime. And it is already being used as the native search layer inside @heydittoai SN118 agents. This is what agents need. A chatbot without live retrieval is just guessing from stale memory. Agents that research, buy, book, trade, analyze, and decide need fresh information they can trust. That is the layer Desearch is going after. The chatbot is the face. Retrieval is the nervous system. And SN22 is building it in the open. @desearch_ai built by datura SN22 $TAO DYOR