A field-first notebook on AI systems https://sushant-97.github.io/ en Wed, 29 Jul 2026 22:52:41 +0000 https://sushant-97.github.io/hidden-language-tax/ https://sushant-97.github.io/hidden-language-tax/ Sun, 26 Jul 2026 00:00:00 +0000 The same 8,192-token window holds roughly 7,600 Hindi words under one tokenizer and about 1,600 under another. The whole difference comes from one small file, written by decisions taken before training began. https://sushant-97.github.io/an-llm-has-never-seen-a-letter/ https://sushant-97.github.io/an-llm-has-never-seen-a-letter/ Tue, 21 Jul 2026 00:00:00 +0000 A visual tour from raw bytes to BPE: how one invisible translation layer sets model cost, context capacity, and many of the bugs that look like model failures.