Why is that, actually?

Standalone explainers for questions that surface while you build. Not a Q&A catalogue: the question is only the occasion, the text is the explanation - and every piece stands on its own.

• September 6, 2026

  The world, or a model of the world

  Two MCP endpoints, the same question, two well-formed JSON answers. Behind one of them sit rows in a database, behind the other nobody. Neither one shows it.

  Readworld model · simulation · measuring

• August 30, 2026

  Why the same question takes 150 seconds one time and 4 the next

  150,000 tokens of documentation, one model on a DGX Spark, one question. Whether the answer starts after 150 seconds or after 4 comes down to the order of things in the prompt alone. And no card you can buy delivers that factor.

  Readprefix-cache · ttft · inference

• August 6, 2026

  Why your 8 GB model needs 20 GB of memory

  File size was never the memory requirement. In production the KV cache comes on top - and it doesn't depend on the model, but on the context and the number of concurrent requests.

  Readvram · kv-cache · inference

• July 31, 2026

  Why the tenth turn costs a hundred times what the first one did

  An agent that works through ten steps doesn't make ten small requests - it makes ten ever-larger ones. Why the bill doesn't grow linearly but, in total, quadratically.

  Readcost · context · tokens

• July 30, 2026

  Why a model sounds more certain when it's wrong than when it's right

  Invented libraries, invented statutes, invented citations - delivered without the slightest hesitation. That's not a glitch; it follows directly from what the model was trained to do.

  Readhallucination · trust · training

• July 29, 2026

  Why the answer stops mid-sentence

  The output halts as if someone pulled the plug - mid-word, mid-line of code. Four causes are candidates, and in the result they look almost identical.

  Readtokens · truncation · operations