Vorashil
vorashil.bsky.social
phd student in logic and formal verification | software engineer | mathematician | currently building https://Oyren.AI
👤 Joined on 07 September 2025
42
Followers
31
Following
57
Total Posts
📜 Kullanıcı Geçmişi
- 07 Eyl 2025, 18:45
🎉 Bluesky'e katıldı - İlk kullanıcı adı:
@vorashil.bsky.social
🚫 Blokladığı Kullanıcılar
❤️ Beğendiği Gönderiler
What is attention in a Transformer? In 42 seconds: how every token projects into Query, Key & Value vectors, scores context via QKᵀ/√dₖ, and turns softmax weights into an attention heat strip. Oyren explains #1 · Vaswani et al. (2017) #AI #LLM #MachineLearning
Show PostWhat actually happens inside one forward pass of a Transformer? In 44 seconds: tokens → embeddings + position → stacked Causal Self-Attention + Position-Wise MLP blocks on a residual stream → vocabulary logits & softmax. Oyren explains #2 #AI #LLM #MachineLearning
Show PostHow does an LLM write an answer? It predicts one token at a time in an autoregressive loop: running each new token against the KV cache, scaling logits by Temperature (T), trimming the tail with Top-p, and appending inside a fixed context window. #AI #LLM #MachineLearning
Show PostHow does an LLM read a PDF? Inside, a PDF isn't paragraphs of text. It is a stream of (x, y) drawing commands that position characters and lines on a page. Here is why basic text extraction mixes up 2-column pages, tables, and math, and why converting to Markdown first works. #AI #LLM #PDF
Show PostWhat are tokens in an LLM? LLMs don't read whole words or single characters. In 42 seconds: how Byte-Pair Encoding (BPE) merges common byte pairs into subword token IDs, why 100 English words is ~130 tokens (and 2–3x more in code or JSON), and how tokens drive cost and speed. #AI #LLM #NLP
Show Post