We're a small crew of humans who build with AI — then ship it. We swap facts, pitch ideas, and actually make the things work on someone else's machine. No theory-only sessions. Ever.
Click a card to flip it. Every one of these shaped what we build and how we think about it.
A large language model, at its core, just predicts the next token. Everything else — reasoning, code, poetry — is an emergent side effect of predicting trillions of tokens really, really well.
"Attention Is All You Need" (Vaswani et al.) dropped in June 2017. Its attention mechanism lets a model weigh every word in a sentence at once instead of chewing through it left-to-right. It's the foundation of everything since.
Weights learned from data mean models generalize by blending patterns they've seen. That's why they're brilliant on familiar problems and confidently wrong on edge cases. Hallucination is a design feature, not a bug to ignore.
Model performance follows a predictable power law with data and compute (Kaplan et al., 2020). That's why the last few years felt like exponential growth — we were literally following the curve upward.
Words become vectors in high-dimensional space, so "king − man + woman ≈ queen" is actual arithmetic in embedding space. Semantic search, RAG, and all similarity tricks ride on this one idea.
A 24 GB card like the RTX 3090 can finetune small LLMs and run 7B–13B models comfortably. Local inference means no API bill, no data leaving the house, and total tinkering freedom.
Every month we vote. Tap a card to lock in your vote — first member's pick becomes the project. Votes are saved on this device.
Type your name into the terminal and we'll generate your unique club token. Copy it, show it at the meeting, you're in.
One member demos each week. One fact gets shared. One idea gets picked to build — and it must actually work on someone else's machine.