PromptZone - Leading AI Community for Prompt Engineering and AI Enthusiasts

Noor Krishnan
Noor Krishnan

Posted on

How DeepSeek Reverse Engineered Itself

A post on Hacker News last week described a technique for reverse engineering DeepSeek by prompting the model to interview itself about its own architecture and training.

The approach uses the model as both subject and interrogator. One instance generates questions about weights, tokenization, and safety layers while another responds, producing structured output without external tools.

Model: DeepSeek | Method: Self-interview | Engagement: 16 points, 2 comments on HN
Source: manish.sh article

How the Self-Interview Process Works

The technique starts with a system prompt that instructs one DeepSeek instance to act as an interviewer focused on technical internals. The second instance answers under constraints that force concrete details rather than generic refusals.

Questions target training data mixtures, context window handling, and alignment mechanisms. Responses are logged and cross-checked across multiple runs to identify consistent claims.

What the Hacker News Thread Shows

The discussion received 16 points and 2 comments. Participants noted the method's low cost compared with API scraping or weight inspection.

One comment questioned output reliability, while the other suggested combining self-interviews with activation patching for verification.

Comparison with Traditional Reverse Engineering

Standard approaches require weight access or heavy API querying. Self-interviewing needs only chat access and runs locally or via cheap endpoints.

Method Access Needed Cost Level Output Structure Verification Ease
Weight inspection Full weights High Precise High
API scraping Rate limits Medium Noisy Low
Self-interview Chat only Low Structured Medium

Pros and Cons of the Approach

  • Requires no special infrastructure beyond standard inference
  • Produces readable transcripts that can be parsed automatically
  • Risks hallucinated internals that must be validated elsewhere
  • Limited to what the model is willing or able to disclose

Who Should Try Self-Interviewing

Researchers studying closed models with chat access will find it useful for initial mapping. Teams with weight access should skip it in favor of direct inspection.

Developers building evaluation harnesses can use the transcripts as seed data for more rigorous tests.

Practical Next Steps

Clone the prompting pattern from the linked post and run paired instances on a local DeepSeek deployment. Log outputs in JSON for later analysis.

Cross-reference claims against public benchmarks or smaller open models where internals are known.

Bottom line: Self-interviewing offers a low-resource entry point for mapping model behavior when weights remain unavailable.

The method is likely to spread to other chat-only models as teams seek cheaper ways to document internals before investing in heavier analysis.

Top comments (0)