P @pm9 - Twitter Profile

pm9 retweeted

10 days ago

Let’s talk about my biggest onchain conviction play. $Crawl @crawlnetwork is one of those projects where I think people are looking at the first product & missing the actual company being built. The obvious thing is “crypto LLM.” The bigger thesis is a model that becomes the intelligence layer for an entire industry, with the industry itself economically incentivized to build its training data. Think about what that means. Crypto already produces an absurd amount of information, but it is scattered across chains, contracts, docs, GitHub, governance, funding rounds, markets, research and social data. Today you pay different companies to organize pieces of it, then manually connect the pieces. Nansen can charge up to $1,800/month, Glassnode up to $999, Chainalysis sells six-figure contracts, and the broader alternative-data market is already measured in billions. An industry-specific LLM changes the product entirely. Instead of another terminal that shows you data, you have a model that understands the relationships inside the data and can reason across them. A fund can ask about an ecosystem, an agent can investigate a protocol, a wallet can interpret onchain activity, a protocol can analyze its competitors, and an application can build its own intelligence layer on top of the model. The model becomes the interface to the industry’s information. That is already a massive opportunity. But Crawlnet’s more interesting idea is how they are trying to build it. They are turning the training corpus into an economy. Crawlers are paid to find and produce the information the model needs. 60% of fees goes toward browsers and GPUs, 40% toward crawlers. Protocols can pay to get information into the corpus. The network takes that constantly expanding information, trains the model, and distributes the weights. The thing being produced by the network isn’t a token, a dashboard or a static database. It’s training data. This is where the $axs style gamification matters. Game showed that you can turn participation in a digital economy into an economically incentivized game. Crawlnet is applying a similar concept to data production: instead of people playing to produce game assets, operators crawl to produce information that becomes part of an AI’s brain. The network can potentially turn thousands of people and agents into a distributed data-collection workforce, with rewards tied to the economic value of what they contribute. That creates a much more interesting flywheel than a normal LLM company has. More crawlers create more data. More data creates a larger and potentially better crypto model. A better model attracts more applications, agents, protocols and funds. More usage creates more economic demand. More economic demand funds more crawling and compute. The network can therefore keep expanding the thing that makes the model valuable. And the starting point is already surprisingly large. Three days in, Queen v2 is 75.5M parameters, trained from scratch on 88.9M tokens and 79,250 pages. 34,282 pages were collected by its own crawlers. The corpus already contains information on 29,000 projects, 13,162 funding rounds and 8,341 investors. Roughly 160 crawlers are active across 644 domains. The previous 42M-parameter model reportedly won a blind test 81–16. It’s the distance between 75.5M and what the network could eventually train on.

18

198

48

83

37K

P

@pm9

12 days ago

$Crawl is not another wrapper. It is a public training run, crawlers scrape crypto data to train the LLM in which Queen is pretrained from scratch on that data only. fees fund compute, and crawler owners split the pool. The upside is the gamification around the workers and what it could become to allow the dataset to grow at a faster rate, along with the attention it would bring to the space and token. Novel and huge potential. Positioning accordingly.

0

3

1

0

463