deep read: @manthanguptaa
2026-02-19
software engineer, technical writer. blogs at manthanguptaa.in.
10 articles, dec 2025 – jan 2026. the best explainer in this corpus and the only author who doesn't have a product to sell. he reverse-engineers other people's systems better than their own documentation does. the question he hasn't answered yet: what does he think should exist?
the good
the tokenization article is the best single piece of technical writing across all five authors. it walks through BPE, WordPiece, and Unigram from first principles, builds a tokenizer from scratch with real code and real output, and explains why LLMs can't do arithmetic (it's the tokenizer, not the model). the insight that "317+492" tokenizes inconsistently across three levels of granularity — and that this determines model capabilities — is something most AI engineers don't know but should. at 28K chars it's the longest article in the entire corpus, and none of it is filler.
he reverse-engineers systems better than their own documentation. his openclaw memory analysis traces the full lifecycle: context vs memory (ephemeral/expensive/bounded vs persistent/cheap/searchable), the hybrid search formula, the pre-compaction memory flush. more detailed and clearer than openclaw's own docs. his chatgpt context structure piece — reverse-engineering the 4-layer context stack entirely through conversation probing — discovered that chatgpt uses lightweight conversation digests instead of RAG. simpler, faster, cheaper than what most people assume.
the LLM-as-judge article is the most production-ready piece in the set. all four evaluation modes (pointwise, reference-based, pairwise, listwise), named biases (position, verbosity, self-enhancement), and the de-biased pairwise comparison — run every A/B test twice with positions swapped, call ties when results flip. immediately usable.
the memory failure article names what everyone else is too polite to say. "most of them are glorified caches that store tokens, not thoughts." sharpest single sentence about AI memory in the entire corpus. he's the only author who insists that consolidation and forgetting — the agent's "sleep cycle" — is mandatory, not optional.
the bad
the question-mark problem: he explains beautifully but never prescribes. 10 articles and not one says "here's what I would build." the writing stays safely in the explainer lane when it could credibly move into the architect lane. someone this good at understanding systems should be designing them.
verdict
if you read one technical deep dive from this collection, make it the tokenization article. then the memory failure piece. manthan is the author you recommend to engineers who want to actually understand how things work, not just use them.