Here's my impressions of your algorithm:<p>1. read each site<p>2. rent a 4090 with <a href="https://vast.ai" rel="nofollow">https://vast.ai</a> to run vllm<p>3. let llm model invent its own category and tag names freely<p>4. save 1KB of metadata each<p><pre><code> a. a small local language model that reads each one and writes a name, two or three sentences, a category, and a handful of tags.
</code></pre>
5. `code is going up as open source` soon (TM)
Code appears to already be up: <a href="https://github.com/alexmorleyfinch/marlin" rel="nofollow">https://github.com/alexmorleyfinch/marlin</a>
The technical details are on another page: <a href="https://alexmorleyfinch.github.io/marlin/history/v1/article/technical.html" rel="nofollow">https://alexmorleyfinch.github.io/marlin/history/v1/article/...</a><p>Your impressions seem about right, but there are a few control steps it seems.
They really needn't have specified "in a weekend" cause yeah we can tell.<p>Since when has low effort become a selling point anyhow?
TS;DR: Too Sloppy; Didn't Read.
Interesting project. Website discovery is indeed in a pretty dire spot, definitely a space that needs innovation. An auto-labeled website directory isn't <i>that</i> silly of an idea.<p>I have a 400 GB sqlite database with samples of rendered root document DOMs I use for ad detection in Marginalia Search I've been meaning to explore similar ideas using.
I have been wanting do do this. The biggest source of domains is certificate transparency logs. Also ICANN zone files. According to some scientific papers these cover 88% of all registered domains. You could crawl dns for CNAME records with all ipv4 IPs by distributing requests across dozens of DNS servers, the internet archive or the common crawl but doing it for the internet archive is a dick move without giving them money<p>There's about 200 million active domains currently. That's about 66% of all businesses worldwide of which there are around 300 million. Around 100 to 150 million have active webpages
It feels like we've hit a point where search engines can become what "todo list apps" were for devs 10 years ago.<p>What a homebrewed solution lacks in coverage it excels in indexing and serving a small slice of the internet really really well.
To be fair they are a supremely interesting problem to hack away at, and one that will meet you where you are.<p>Almost anyone can put together a basic search engine in a few thousand lines of code, it's just not very hard to make a program that will index a few million documents better than Confluence.<p>Then, between that first ansatz and a working scalable internet search engine, you have a pile of interesting problems touching every aspect of computer science and computer hardware and networking, enough so that hundreds of people will have gotten PhDs in narrow sub-problems of those problems you'll be facing.<p>It's great because you can just tackle the stuff you feel comfortable approaching and leave the rest for later.
Sorry I have a lot of trouble understanding what this is useful for. Like, I am never going to replace it with Google, DuckDuckGo, ChatGPT or even Bing.
How do you build a list of domains you want to index ? I see there is a fetcher and a spider in the code but so for I haven't found how to build that list.
This is actually where I see software going in the short term -- cloud moving to local.<p>A few years ago, if you wanted translation, you'd use Google Translate. If you wanted to search the web, you'd use Google search.<p>But for a few gigabytes, you can now install nllb-200-distilled-600M, and get translations for almost any language locally. You can have your computer crawl the web, create abstracts and categorizations for websites, and build search exactly as you want it.<p>The main limiter now is hard drive space (and to an extent, local compute) -- but right now it feels like the 70s again where the terminal into a remote server turned into building applications locally.
From the screenshot, it's very funny that one of the indexed sites is www.llresearch.org, which looks like it's run by a crackpot.
Sometimes I think people forget how capable computers are. 500k is not much. You can just slap that in a Lucene instance. This is a solved problem.
Domains are way more than just 40M though.
Like a personal Google? How do you bypass all the captcha, ip bans, cloudflare turnstile antibot stuff etc?
Thats the fun part, the user just went with happy path. Javascript, captchas, cloudflare protected content did not made to the catalogue. This sort of use case exists in LLM training data a lot which makes it easier. The data gathered by the user is not really practically useful cause there are way too many gotchas when it comes to web scraping and building a catalogue (source: I have done scraping for a particular domain data and had to do at least 10+ iterations to get it >90 right)
They don't: "skips the model entirely if the page is empty, parked, or a bot-challenge wall"
I think Kagi Small Web filter would give you very similar results.
Check out my latest project! You can fork it, tweak the policy manually or with AI, run the system and watch the data come in! It's engineered to keep a low data footprint, so 500k domains fits into 1GB on disk. If you have local models it's free! You just might not get the best throughput depending on your GPU. My production data is not exposed anywhere yet, and I may never expose it. The point is for you to fork and make your own policy, and thus your own personal search engine! The article covers basic analysis on my data, so it's worth a read if you're interested! A deeper analysis may arrive with V2 if I ever do it
[dead]
[dead]
[dead]
[dead]