News for hackers. Stuff that matters.

armanified

I'm a professional web scraper and have pretty much exhausted most open-source stealth browsers. The paid ones aren't a very economical solution when you want to scrape at scale and have to deal with heavy bot mitigation.

The issue I repeatedly faced with most tools is that almost all of them run on Linux and try to pretend they aren't, which might work for a while, but given enough samples, it's pretty obvious to large bot-mitigation infrastructures, and they start blocking or at least push a bit harder.

So I tried to build a browser myself, which does a few things:

* Randomize profiling based on the given seed * Match the timezone based on the proxy exit IP * Can outsource Canvas to a different machine ( from the target OS )

This canvas outsourcing is what really makes it different, because pretending isn't enough. Some of the toughest bot-mitigation infrastructures probe deep into the canvas, a depth that a simple patch can't bypass on a different OS. But this is optional and should only be done if everything else fails.

Right now, this is an early release. Any feedback is highly appreciated

True, but most would ignore LM if it weren't LLM.
OMG! Why didn't I thought fo this first :P
OMG! You just gave me the next idea..
Pretty neat! I'll definitely take a deeper look into this.
Uppercase letters were intentionally ignored.
My initial idea was to train a navigation decision model with 25M parameters for a Raspberry Pi, which, in testing, was getting about 60% of tool calls correct. IMO, it seems like around 20M parameters would be a good size for following some narrow & basic language instructions.
It mostly doesn't, at 9M it has very limited capacity. The whole idea of this project is to demonstrate how Language Models work.
I haven't compared it with anything yet. Thanks for the suggestion; I'll look into these.
I intentionally removed all optimizations to keep it vanilla.

Show HN: I built a tiny LLM to demystify how language models work

https://github.com/arman-bd/guppylm

Built a ~9M param LLM from scratch to understand how they actually work. Vanilla transformer, 60K synthetic conversations, ~130 lines of PyTorch. Trains in 5 min on a free Colab T4. The fish thinks the meaning of life is food.Fork it and swap the personality for your own character.

( Read More... | 137 comments | 167 points | on HN )
Good to see someone contributing to the Bangla space.
A lot was generated with LLM, but not vibe-coded. I have a clear expectation and project plan.
JS/TS could actually be my next venture! Would love to hear about your use case to help prioritize features if I build it out.
Fair concern. Most of these projects are abandoned because keeping up with major browsers' changes is exhausting. I genuinely want to maintain this and plan to update it with major Chrome releases as part of my work depends on it. Things move fast in this space, but I am committed to keeping it functional. I'll do my best to stay on top of it.