This paper provides a complete, ground-up explanation of how Large Language Models (LLMs) work. A central section walks through the entire forward pass with concrete numbers: embedding lookup, Query/Key/Value matrices, the dot-product score matrix, softmax weights, weighted Value sums, layer-by-layer refinement, and the final word prediction. It explains how backpropagation adjusts every matrix in response to prediction errors. It also covers fine-tuning and RLHF in detail. The paper concludes by applying this understanding to explain precisely why LLMs fail at constraint satisfaction.

https://vzocca.substack.com/p/how-large-language-models-work

Musulmonlar qiladigan eng bema’ni ishlardan biri musulmon davlatlari hukumatlarini "qo‘g‘irchoq" deb atashdir. Musulmon dunyosida qo‘g‘irchoq rejimlar yo‘q. Bil’aks, shunday rejimlar borki, ular oʻta poraxo‘r va xudbin boʻlganidan, AQSh, Rossiya yoki Xitoy ularga pora berib, o‘z ta’sir doirasiga kiritib olgan. Biroq bu rejimlar o‘z xalqini majbur bo‘lganidan emas, balki o‘zlari shuni istagani uchun sotadi.

Aslida esa, bu rejimlar AQShdan ancha mustaqil, shuningdek, o‘z manfaatlariga ham ega va ularni AQShdan mustaqil ravishda ko‘zlaydi. Biz koʻrib turgan butun reallik yovuz AQSh butun dunyo ishlarini yashirincha boshqarib turgan ulkan fitnadan iborat emas. Aksariyat hollarda, eng yomon qassoblar va o‘g‘rilar o‘zimizdan chiqqan bo‘ladi va AQSh ulardan o‘z manfaati yo‘lida foydalangani kabi, ular ham AQShdan o‘z manfaatlari yo‘lida foydalanadi.

(C)

@Tafakkur_chizgilari

"if quantum computers start breaking cryptography a few years from now, don’t you dare come to this blog and tell me that I failed to warn you. This post is your warning. Please start switching to quantum-resistant encryption, and urge your company or organization or blockchain or standards body to do the same."

Scott Aaronson
(also one of the 3 scientists who said no to Epstein)


https://scottaaronson.blog/?p=9718

TGPages Atlas (https://tgpages.com/atlas/map/)

An interactive map of Telegram channels that helps you quickly find communities by topic and assess their interests. This is a convenient visual catalog for a basic analysis of the Telegram ecosystem. It's a useful resource for quickly finding similar channels, tracking thematic connections, and understanding audience distribution around a specific topic.

Link to the service (https://tgpages.com/atlas/map/)

#OSINT

Bugun bir qiziq jumlaga duch keldim:

O‘zingizni ko‘p yil yashagan deb o‘ylaysiz,
Lekin aslida bir xil yilni qayta-qayta takrorlab turibsiz.

Yangilanish, o‘zgarish ko‘proq harakat qilish orqali paydo bo‘lmaydi.
balki o‘zimizning eski fikrlarimiz va odatlarimizdan voz kechish bilan keladi.

@botirziyatov