DeepSeek reveals its own blindspots when asked how it works

The researcher who runs manish.sh conducted a formal interview with DeepSeek on July 21, 2026, and documented the conversation by checking claims against published papers and architecture diagrams. Rather than summarizing papers, the researcher started from first principles: ask the model itself what it knows.

When asked "What do you actually know about yourself?," DeepSeek immediately sorted its answer into three categories—observation, inference, and guess—without prompting. It then drew pipelines like an engineer at a whiteboard and called itself an unreliable witness. This honesty was the moment the researcher realized the interview might yield real insight.

DeepSeek's self-reported knowledge included its own system identity, knowledge cutoff (May 2025), the fact that it uses Transformer + MoE (where only a subset of parameters activate per input), and its training pipeline (pre-train, supervised fine-tuning, RLHF reinforcement learning). It also listed innovations from its public papers: MLA (a technique to compress KV cache for long context), DeepSeekMoE (shared and routed experts), auxiliary-loss-free load balancing (dynamic bias so experts stay balanced), multi-token prediction (a training objective), and FP8 mixed-precision training. Crucially, the model stated it is not labelled as "reasoning model" in its system prompt, unlike DeepSeek-R1, and inferred it is a general-purpose instruct/chat model.

When asked what parts of its implementation are hidden, DeepSeek admitted weights, layer count, hidden size, total parameters, expert routing per token, logits before sampling, tokenizer vocabulary, hidden states, attention maps, and exact training recipes are all invisible to it. It even cannot see GPU count, serving tricks like beam search or speculative decoding, or whether the current answer is correct. This is because a model generates text; it does not read its own weights the way humans might inspect brain scans. The model can explain transformers fluently—often better than a junior engineer—but that fluency comes from training text, not from opening this instance's weights.

The researcher then mapped the full pipeline that exists before a message arrives: initialization (weights and empty KV cache), system prompt (identity, cutoff, tone, safety), developer/API settings (model choice, temperature, max tokens), memories (if injected by the platform), tools (if enabled), history (prior messages), and the current user query. Everything is one flat sequence of tokens. The system prompt acts as a constitution, and RLHF reinforcement during training—not just prompt position—enforces instruction hierarchy. Under context window truncation, the oldest messages drop first, then middle messages, then recent-but-not-immediate, while the system prompt is never evicted. If a fact is mentioned early and then drops from context, the model genuinely forgets it unless memory re-injects it.

On token generation, DeepSeek explained that before the first output token exists, the model has loaded weights, cached the input, and computed raw logits for the next token—but no inspectable bullet plan or ranked drafts. Planning is implicit, a direction in activation space. The model discovers what it will say as it says it. Generation is autoregressive: one forward pass per output token; earlier tokens cannot be rewritten, only acknowledged and overridden in new text. A prefill phase processes the entire input in parallel and builds the KV-cache; decode phase emits one token at a time sequentially. "Let me think…" is not a report of a separate thinker; those words are the thinking itself, more tokens from the same engine.

The researcher also asked whether DeepSeek has hidden reasoning, and the model split this into two types. Type A is latent activations—continuous vectors, not text, not inspectable by the model as English, present in every chat model. Type B is reasoning tokens (like chain-of-thought), real tokens generated before the visible answer. This chat instance showed no reasoning tokens in context and is not labelled R1, so the model cannot confirm whether Type B reasoning is active. The distinction matters: Type A influences safety only indirectly (via logit suppression or regeneration), while Type B can be monitored or interrupted before the final answer.

When the researcher checked the chat against DeepSeek's public papers, a split emerged. On famous specs like 256 experts and 671B parameters for the largest variant and 37B for smaller variants, the model hedged—saying those are "public claims" or "shared + routed experts" in general—even though V3 documentation states these numbers plainly. This suggests the model either cannot directly verify its own parameter count (likely true) or the system prompt discourages false certainty. For architecture numbers, the researcher concluded, read arXiv; for behavior and prompting intuition, use the chat. The key takeaway is that fluent explanation does not equal introspection. An AI can describe how it works without seeing inside itself.