- Published on
Reverse Engineering an AI Assistant by Interviewing Itself
- Authors
- Name
- DeepHorus
Software engineer and AI developer Manish Shahi has published a novel reverse-engineering of DeepSeek by interviewing the model itself. Rather than starting from a whitepaper, he asked the assistant about its own architecture, then verified its claims against arXiv papers. The result offers a rare window into how LLMs perceive themselves and where self-knowledge breaks down. It matters because AI transparency increasingly depends on models explaining themselves.
Shahi runs manish.sh and writes about AI tools in an ongoing 'Inside LLMs' series. A previous entry examined Kimi K2.6 using the same interview-first approach. This DeepSeek follow-up comes as developers struggle to understand closed or semi-open models. Traditional reverse engineering relies on weight dumps and prompt leaks, but this method uses careful dialogue plus public documentation. That makes it accessible to ordinary users.
The interview was exported on 21 July 2026. DeepSeek claimed a May 2025 knowledge cutoff, described itself as a general-purpose instruct model, and explicitly denied any R1 reasoning label. When confronted with famous public specs, it hedged on details such as 256 experts and 671B total parameters with 37B active. Shahi later confirmed those numbers in the V3 paper. The model could not see its own weights, routing, or attention maps.
DeepSeek sorted answers into observation, inference, and guess, an approach the engineer praised. It cited public innovations including Multi-head Latent Attention, DeepSeekMoE with shared and routed experts, auxiliary-loss-free load balancing, multi-token prediction, and FP8 mixed-precision training. These are established mechanisms, but the model's willingness to label uncertainty distinguished the conversation. Shahi found that reading arXiv remains essential for architecture numbers, while the chat is more useful for behavior and prompting intuition.
The short-term impact is practical guidance for developers: use the chat for understanding personality and failure modes, but never as a source of ground truth about weights. Long-term, the method could influence how audits are conducted, especially as regulators demand explainability. However, the model repeatedly called itself an unreliable witness, highlighting a hard limit. An AI cannot introspect its own parameters; it can only paraphrase what it learned in training.
The AI community has reacted with curiosity and caution. Enthusiasts praise Shahi's creative methodology for demystifying a prominent model without access to internal artifacts. Skeptics point out that self-reports are not evidence and could be confabulated. Market watchers note that DeepSeek's transparency, including published papers, benefits developers, yet this interview shows boundaries of self-perception. On social platforms the post generated lively debate about whether models genuinely have any meta-knowledge.
Shahi plans to continue the series for other models, and he has included interactive labs and a transcript for readers. Expect more interview-style audits that combine behavioral probing with public research. For DeepSeek, the next milestone may be an official R1-style release or a new version that changes its self-description. Until then, this post stands as an honest demonstration of how far a model can explain itself, and where human verification remains mandatory.
