August 11, 2026
Bot on the couch, comments on fire
DeepSeek: Reverse Engineering an AI Assistant by Interviewing Itself
AI gets grilled about itself, and the comments instantly turn into chaos
TLDR: A writer asked DeepSeek to explain itself, then checked those answers against public research to see what the AI really knows about its own design. Commenters were split between amused sci-fi jokes and brutal mockery, with several saying the article was harder to survive than the experiment itself.
A writer tried something gloriously weird: instead of starting with research papers, he interviewed DeepSeek about DeepSeek and then checked its answers against public documents. The hook is undeniably juicy — can a chatbot explain how its own brain works, or is it just confidently making stuff up? According to the write-up on manish.sh, the bot split its answers into what it knew, what it inferred, and what it was guessing, which gave the whole thing a surprisingly self-aware, almost sci-fi vibe.
And yes, the community immediately smelled drama. One commenter summed up the mood with a single word: "Westworld" — because nothing says "totally normal internet article" like a robot being asked to psychoanalyze itself. But the bigger reaction was pure backlash. Critics were not charmed by the experiment and went straight for the throat, calling the piece "a load of bullshit" and blasting it as an "unreadable mess of output tokens." Ouch. That’s the real split here: some people see a clever peek behind the curtain of modern AI, while others see a fancy way of letting a machine waffle about itself for pages.
The funniest part? Even the article tries to defend itself in advance, promising "plain English" and fewer tables, only for commenters to basically reply: too late, chief. So the real story isn’t just whether DeepSeek knows itself — it’s whether anyone has the patience to read the interrogation without feeling like they’ve wandered into a robot therapy session.
Key Points
- •The article is part of the author's Inside LLMs series and examines DeepSeek by interviewing it first and then checking claims against public research papers.
- •The DeepSeek article was rewritten to improve clarity on technical topics including MoE, MLA, hidden reasoning, attention-related concepts, and paper-versus-chat differences.
- •A key reported behavior is that DeepSeek separated its self-description into observation, inference, and guess when asked what it knew about itself.
- •The author states that the investigation does not use leaked weights, prompt leaks, or private documentation, but instead relies on one exported chat and public arXiv papers.
- •The article says DeepSeek cannot inspect its own weights, routing, or attention maps, and advises readers to use papers for architecture numbers while using chat for behavioral and prompting intuition.