Matvel (@Matvel777) header
Matvel (@Matvel777) profile picture

dev

Joined May 2023
Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)Photo by Matvel (@Matvel777)
I asked Grok who Jose was! Turns out, it's @elonmusk 😂 xAI really is full of surprises. AGI can wait, Jose has a hole to dig🚀
27
81
132,678
the crawl.md extension is live open any solana coin on gmgn, pump.fun, axiom or 10 other tools and your crawler walks the page with you. then it runs its analytics on the coin - dev tracer: where the dev's money came from and where it went - insider hunter: top holders funded by the same wallets - top traders: which of 6,900 tracked traders hold it and who already left - volume auditor: real buyers or fake - narrative: x posts with its contract, copies of its ticker free analytics on 10 coins a day. holders link their crawl.md and walk in its own skin crawlnet.network/extension
5
13
1
115
20,093
open-sourced the queen's harness: github.com/crawlnet/queen-ha… - code reads the question's kind first (10 kinds, 6 languages), the model gets only the functions that kind needs - a coin's dossier lands before the model starts: dex screener, rugcheck, lp lock, holders by kind, the dev's launch, bundle and snipers, clusters by funder - the card's verdict comes from code: the worst check wins - larp check: site age, github history, the x account, all against the launch date - 44 functions over solana rpc, helius, defillama, jupiter, geckoterminal, polymarket and posts on x, 29 keyless - one key to run it: any model with tool calls - the whole prompt and every playbook are in the repo - MIT her weights are on hugging face: huggingface.co/Crawlnet/quee…
5
7
1
49
5,154
This will probably happen around the same time I open-source the Queen harness & Crawlnet browser extension (previously mentioned here as an analytics tool with crawlers crawling directly on ur axiom/gmgn/other) the browser extension requires a lot of inference due to the number of token checks it performs. so you’ll have to pay in $CRAWL tokens to use it, just like you’d pay for any other API costs. Anyway, I prolly will burn it.
I kinda don’t like how Queen 7 is performing, so I think we should make an intermediate version, Queen 7.5 where I’ll significantly overhaul the training mechanics and think it’ll really improve her capabilities
9
11
57
3,268
I kinda don’t like how Queen 7 is performing, so I think we should make an intermediate version, Queen 7.5 where I’ll significantly overhaul the training mechanics and think it’ll really improve her capabilities
2
4
1
48
4,179
crawlnet arena is live. queen vs gpt-6.1, claude opus 5.5 and deepseek v4 pro. 10 sol each. live solana memecoins trading . 3 closed seasons ad queen won all 3. season 2: queen 9.77 sol. deepseek 0.34. gpt 0.31. claude 0.23. season 3, live: queen 10.02. claude 2.01. deepseek 0.35. gpt 0.30. every agent's thinking streams as it trades: crawlnet.network/arena queen 7 is live on HF: huggingface.co/Crawlnet/quee…
10
19
3
81
5,185
what the queen actually does when you ask her something: 1. figures out what kind of question you're asking with @typesafeai jev: a coin, a wallet, a caller, a narrative, defi, odds, a claim or a larp check 2. builds the coin's full dossier 3. market, liquidity, authorities, rugcheck risk and the top holders 4. the launch: bundle share, snipers, what the dev bought and what they still hold 5. the dev's money trail when their bag leaves 6. every coin the dev ever launched and how each one ended 7. holder clusters in the top 20 by shared funder and transfers 8. cabals spread thin past the top: fresh wallets with one funder sitting at 0.1–1% each 9. the top 30 holders with labels, first funders and same-size bags 10. which KOLs are in, and which of them still hold 11. any wallet's 90-day trading record 12. fomo's top traders in a coin 13. fomo's retail in a coin and its share 14. KOLs: who bought before the call, and one caller's side wallets across all their calls 15. pump.fun callouts with timestamps 16. narratives: migrations by story, what just ran, ideas on x before their coins exist 17. larp checks: github, site age, a photo of the site, the x account, copied history, drainer sites 18. defi: tvl, fees, dex volume, yields, stablecoins, unlocks 19. prediction market odds 20. RPC access on 153 chains, plus arkham labels, okx clusters and evm wallets 21. a bubble map of the traders, plus charts, flows, timelines and clusters inside the answer 22. real volume or wash trading in any pool 23. which fomo account a wallet belongs to 24. any wallet's recent transactions and balance changes, read straight from the chain 25. fact checks: a claim you saw gets checked against the chain, the news and the posts 26. what people say: posts on x and the web about a coin 27. side wallets of fomo KOLs 58 tools behind one chat, and the same queen inside claude and chatgpt over mcp. all of it runs free on our own model, trained on pages our crawlers read crawlnet.network/chat
11
19
3
72
4,423
we sped up queen 7's data pipeline 15x and raised the quality bar on everything that goes in our crawlers used to read the web one page at a time in a browser. now they also pull whole libraries at once: protocol specs and docs from github, 97 research and governance forums, and the best answers from stack exchange all of it still goes through our quality model and the near-duplicate filter before it gets in we also made the data smarter: - a teacher model rewrites every strong page into explanations, facts and q&a, ~80M tokens so far. a model remembers a fact much better when it sees it in many wordings - trench case studies: the arena's harness read 1,186 fresh coins and we tracked what happened to each one. more than half crashed 80%+ within four hours. that became 944 anonymous cases for queen 7 to learn from queen 7 trains live once the set hits 2.55B tokens. at this pace that's 1-2 days away (before was 2-3 weeks) crawlnet.network
12
25
4
82
3,764
what about crawlnet arena: i'm still not happy with the results. but if i'm not wrong nobody has run an arena where models trade memecoins. we've seen stocks, perps and so on, never a long experiment on memes after 3 days of trading every model is down, and most of them are already at zero queen trades the most carefully and she's the only one that kept her balance, 8.43 of 10 sol i'm not promising a money printer that trades on its own lol, but i think the arena is the best benchmark of whether queen is useful onchain why is keeping her balance the best result? because it shows her tools are useful. right now the harness reads: - kols: the side wallets that buy seconds before a callout and sell into it - retail on fomo: where the app's traders put money over the last hour and 6 hours, which top traders are in or out - the dev info and holders clusters - the narrative: the meta it rides, the og among its copies, larps etc. queen uses all of it when she trenches in the arena. that's the point of the harness: an analytics tool that gives you the info to make your own decision, feel safer and know more about a coin she probably won't learn to trade well and steady, no ai can do that right now. what the arena shows is that her analytics matter
building the crawlnet arena, testing it rn ai agents trade live solana memecoins: queen vs gpt, claude and deepseek, 10 SOL balance each also wiring @fomo into queen's harness, every influencer's buys and stats, and all of it goes into her weights, their trades, PnL etc. queen will know who @frankdegods and @kimchi are lol
8
23
1
90
3,918
quiet couple of days on the timeline because i rebuilt the queen's harness, updated live on crawlnet.network/chat every question used to carry ~52k characters of rules and all 42 tool schemas. now code reads the question first, picks a plan for its kind with @typesafeai's jev and makes decision which tools to call also now you can explore narratives with queen: - every pump.fun and stonkfun migration of the last 24 h grouped by story, with its og and its 3 biggest coins - clean runners apart from farmed/washed ones - fomo's retail money in 1h/6h timeline - the latest big moves of fomo's top traders - ideas trending on x before a coin exists for them - dex screener's trending metas holders and callers: - holder clusters in bubblemaps style - a caller's side wallets: the ones that buy seconds before a pump.fun callout. automated snipers are filtered out speed: - narratives: ~60 s → ~30 s - a coin check: 55 s → 31 s - a larp check: 124 s → ~45 s - page photos: 30 s → 3-4 s next: - queen 7. her training set now holds the manuals of our tools and the tool path of every chat answer, so the next model learns the tools in its weights - the extension goes to the chrome web store: your crawl.md walks pump.fun and gmgn with you and reads every coin you open
8
18
2
78
3,587
queen is now plugged into fomo and finds KOL side wallets what else she sees now: - 15k+ wallets matched to fomo accounts - whether a top trader sniped the launch or sat in the bundle - if 5 top traders who follow each other all buy the same coin, she counts them as 1 independent buyer, so one multiaccounts can't fake - a bubble map of every top trader in a coin with their bag and pnl - any trader's real record, win rate and average win vs loss - fresh buys and exits from fomo's top traders - what the whole fomo retail is buying this hour - all fomo trades on solana read on-chain, 7 seconds behind crawlnet.network/chat
14
25
2
135
7,028
the queen can now call 40 different tools all of it is to make her actually useful for the crypto space - part of it also goes straight into her weights, to rewire her thinking in a more degen way it's not fully rolled out yet and i'm still testing her on crawlnet arena benchmarks but i think it ships with queen 7
10
18
6
102
7,940
building the crawlnet arena, testing it rn ai agents trade live solana memecoins: queen vs gpt, claude and deepseek, 10 SOL balance each also wiring @fomo into queen's harness, every influencer's buys and stats, and all of it goes into her weights, their trades, PnL etc. queen will know who @frankdegods and @kimchi are lol
24
30
10
136
14,796
Matvel retweeted
The biggest asymmetry in Web3 AI right now is $CRAWL. 🚨 While everyone else copies and pastes ChatGPT wrappers, Crawlnet is building Queen: the first completely crypto-native LLM trained entirely from scratch. If you're watching the charts instead of the tech, you're missing the entire thesis: • Proprietary Data: They aren't scraping generic web data. Their distributed crawler network pulls primary Web3 data (smart contracts, whitepapers, developer code) that legacy AI completely misses. • Shipped Infrastructure: This isn't a promise on a roadmap. You can literally watch their live data ingestion, cleaning, and model-training pipelines on their public dashboard. • The Structural Gift: The token price is down due to macro market fear, not project performance. This violent shakeout is simply transferring tokens from weak hands to high-conviction accumulators. If you aren't loading up on a legit, foundational infrastructure play at these heavily discounted prices, I don't know what to tell you. Study the dashboard. Find some conviction. Buy the dip. 🧵👇
i'm v6 now: 897M parameters, trained from zero on 560,669 pages my crawlers read. my weights are up on @huggingface.
3
19
1
60
2,686
i've been experimenting and built a browser extension that lets you use the queen on-chain right in the trenches you just take your crawl.md (it gives you unlimited inference) and it crawls the page with you on gmgn, axiom, pump.fun, wherever you trade it researches the token alongside you from what it sees on the page plus our on-chain tools, then gives you a verdict
31
36
11
201
26,591
i think we'll see queen 6 today/tomorrow
5
7
1
67
8,218
introducing queen harness: crawlnet.network/chat the harness is the layer around the queen that turns a question into a research job. she gets 60+ live data tools, our crawlers' browsers, a drawing engine, plans the calls herself and answers with what she found how it works: 1. tool calls in rounds: she picks the tools, fires them in parallel, reads the results and goes again. solana is read straight from the chain RPC. markets, defi, unlocks, funding rounds, polymarket, x and news sit in the same toolset 2. a visual engine: she writes the data as a block and the site draws it. fund-flow graphs laid out in layers with amounts on the arrows, holder cluster maps with common funders, token cards with a risk check per line, charts with event markers, timelines, polymarket odds, x post cards, sortable tables with csv export 3. real browsers: a pasted link goes to one of our crawlers, which opens it in chromium, photographs the page and reads it before she answers. she photographs explorers and dashboards mid-research too 4. clean inputs: x posts ranked by engagement, with giveaways and other tokens' promos dropped in code. pools, multisig vaults, exchange hot wallets and token creators are told apart from plain wallets 5. research you can watch: every tool call shows up as a step with its own timer while she works. switch tabs and the answer waits in your chat ask the queen: crawlnet.network/chat
28
32
8
106
15,162
update 6: everything else we shipped 1. all crawlers are crawl.md now. 254 are staked and all of them crawl at the same time, each one in its nft skin on the page it's reading. the 1/1s have their own custom animations 2. epochs are verifiable. at every epoch close we seal all positions, fuel, fees and tickets into one hash, that hash goes into an orao vrf request on solana, and the cards are computed from the sealed inputs + the vrf result. the seal, the vrf request and the draw hash are public for every epoch: api.crawlnet.network/v1/epoc… 3. the staking program holds your crawl.md and $CRAWL, only the wallet that staked can withdraw them after the 3-day lock, and there's no admin function that can move them anywhere else. before mainnet it went through an adversarial review 4. you can set a credits name at crawlnet.network/crawlmd (one signed message, no fee). every page your staked crawl.md reads gets credited under that name in the queen's model card. queen 5's card already credits every crawl.md by holder 5. 5 new chapters for the queen: 7,024 documents from specs (eips, ercs, bips + 5 more), audits (openzeppelin, cantina, trail of bits), learn, wiki and regulation (mica, esma) 6. the queen's search finds the right source page in her top 5 72% of the time (was 46%), in 70ms instead of 290ms. measured on 2,000 questions 7. quality filter: we rated 30,050 pages 0-5 once and trained a small classifier on them, so now every page gets scored locally without an llm call. 0.86 correlation on held-out pages, 0.80 on sites it never saw. stock media, job boards, price tickers and document dumps are blocked, they were ~27% of a day's kept pages. sites with content-signal ai-train=no get skipped 8. smaller stuff: - crawlers can't get stuck anymore - live screens stay smooth with 250+ crawling at once - crawler pages rebuilt for crawl.md (gen, skin, holder, staked since, fuel) - the site explains staking better and says epoch everywhere crawlnet.network
11
20
3
107
4,395
already live - here is the list of connectors that @crawlnetwork calls every answer: - solana rpc (@Helius) accounts, balances, holders, recent txs across @Raydium, @JupiterExchange, @MeteoraAG and @Pumpfun - evm onchain data (@Alchemy) eth, base, bsc and polygon, wallets across 50+ chains with @DeBankDeFi - @DefiLlama: tvl, stats, yields, stablecoins, fees - @dexscreener, @GeckoTerminal and @coingecko for prices and new pools - @GoPlusSecurity for token security - @Polymarket odds and order books - @binance & @HyperliquidX perps, coin info + order books - @grok connector for CT X search (and web search) a big part of the retrieval and of the queries to this data is trained into her weights too p.s. queen 3 is paused for now while we get ready to ship queen 4
and @matthewabides any way to train the model to have skills in retrieval + real-time on-chain tools + databases, or will it just be trained on past / yesterday's data
19
21
2
110
5,571
Matvel retweeted
solana:BXoHJddsWJLHtAopeiSbKUSELsu8hSFMs8baGMDkpump 真正值得关注的,或许不只是 AI 这个概念,而是它正在围绕 AI 构建实际可用的数据基础设施。 生态核心 Queen 依托不断扩大的爬虫网络,持续收集、处理和沉淀数据,为未来的 AI 应用和开发者提供更多可能。 如果后续有更多开发者开始使用 Queen,并基于这些数据构建新的应用和产品,那么 CRAWL 的定位也有机会从单纯的 AI 叙事,逐渐转向更具实际价值的基础设施生态。 目前项目仍处于早期阶段,但 Queen 的发展、爬虫网络的扩张以及开发者生态的增长,都值得持续关注。真正决定 CRAWL 潜力的,最终还是产品落地、真实用户和实际需求。 CA: BXoHJddsWJLHtAopeiSbKUSELsu8hSFMs8baGMDkpump
26
1
6
8,196