--- title: "Technology and operations" source: https://bitvaria.com/en/blog/topics/technology-and-operations --- Articles # Technology and operations How models work and what turns a model call into a reliable application. ## Good starting point - [![](https://bitvaria.com/_astro/keine-datenbank.NEfw5zCf.jpg)](https://bitvaria.com/en/keine-datenbank/) 27\. Aug. 2026 6 min read ## [Why a Language Model Is Not a Database](https://bitvaria.com/en/keine-datenbank/) An extraction reliably pulls values out of contracts, until one field is missing. The same call still returns a number, cleanly formatted, just invented. Why a language model is not a database, and what that means for checking its output. Daniel Schleipfer - [![](https://bitvaria.com/_astro/token.CoJHAWcZ.jpeg)](https://bitvaria.com/en/was-ist-ein-token/) 15\. Juni 2026 6 min read ## [What Is a Token? Why a Language Model Sees No Text](https://bitvaria.com/en/was-ist-ein-token/) A token is the smallest unit a language model splits text into. Why the model sees no letters, and what that one fact explains about cost and errors. Daniel Schleipfer ## All posts on this topic - [![](https://bitvaria.com/_astro/prompt-caching.Cl9GOLol.jpg)](https://bitvaria.com/en/prompt-caching/) 6\. Okt. 2026 8 min read ## [Why the Same Long Instruction Sometimes Costs Almost Nothing and Sometimes Full Price](https://bitvaria.com/en/prompt-caching/) A server can reuse the opening of a request instead of recomputing it. That only works if the opening is identical, token for token. From the first differing token on, the saving is gone. Daniel Schleipfer - [![](https://bitvaria.com/_astro/ist-das-modell-aus-china-ein-problem.Dpppcpzz.jpg)](https://bitvaria.com/en/ist-das-modell-aus-china-ein-problem/) 1\. Okt. 2026 10 min read ## [Is an AI Model from China a Problem?](https://bitvaria.com/en/ist-das-modell-aus-china-ein-problem/) The question hides three different questions: where the data goes, how the model behaves, and what the company allows. Only the first one is settled by where the model runs. - [![](https://bitvaria.com/_astro/speculative-decoding.9Qt7bc5r.jpg)](https://bitvaria.com/en/speculative-decoding/) 28\. Sept. 2026 8 min read ## [Why an Answer Arrives Slower Than the GPU Allows](https://bitvaria.com/en/speculative-decoding/) For every single token, the GPU re-reads the entire set of model weights just to compute that one token. Speculative decoding lets a small model guess ahead and saves exactly that wait. Daniel Schleipfer - [![](https://bitvaria.com/_astro/quantization.D5prDg1p.jpg)](https://bitvaria.com/en/quantization/) 18\. Sept. 2026 6 min read ## [Why Fewer Bits Can Change the Answers](https://bitvaria.com/en/quantization/) Fewer bits save space for model weights. Comparing the intended tasks shows whether answers stay good enough and whether the model actually runs faster. Daniel Schleipfer - [![](https://bitvaria.com/_astro/paged-attention.BDNW7Nm9.jpg)](https://bitvaria.com/en/paged-attention/) 8\. Sept. 2026 6 min read ## [Why GPU Memory Is Already Full Before It Has To Be](https://bitvaria.com/en/paged-attention/) A short answer can tie up a lot of GPU memory. Paged attention splits the KV cache into blocks that a request receives when it needs them. Daniel Schleipfer - [![](https://bitvaria.com/_astro/continuous-batching.BYEIgzcb.jpg)](https://bitvaria.com/en/continuous-batching/) 31\. Aug. 2026 6 min read ## [Why a Finished Answer Still Has to Wait](https://bitvaria.com/en/continuous-batching/) Eight requests start at the same time on the same server. The shortest one is done after three tokens. Its answer still does not arrive until the longest of the eight has written its last word. Daniel Schleipfer - [![](https://bitvaria.com/_astro/qualitaetsluecke-offener-modelle.DHkMttWi.jpg)](https://bitvaria.com/en/qualitaetsluecke-offener-modelle/) 14\. Aug. 2026 8 min read ## [The Quality Gap in Open AI Models, Honestly Quantified](https://bitvaria.com/en/qualitaetsluecke-offener-modelle/) In June an open model cracked 50 points on the leading independent AI benchmark for the first time; six weeks later a second one came within three points of the frontier. Why the average number is still almost irrelevant for your own decision. - [![](https://bitvaria.com/_astro/sampling.7v5WKIte.jpg)](https://bitvaria.com/en/sampling/) 12\. Aug. 2026 8 min read ## [Why the Same Model Almost Never Gives the Same Answer Twice](https://bitvaria.com/en/sampling/) Same prompt, same model, twice, two different answers. Not a bug: the model draws from a probability distribution at every word. What temperature, top-p, and top-k actually set. Daniel Schleipfer - [![](https://bitvaria.com/_astro/prefill-decode.DNxM7bFf.jpg)](https://bitvaria.com/en/prefill-decode/) 22\. Juli 2026 6 min read ## [Why the Same AI Request Has Two Different Speeds](https://bitvaria.com/en/prefill-decode/) One request answers instantly, then drips out word by word. Another takes a while before the first word appears, then runs through fast. The reason: two phases with different bottlenecks. Daniel Schleipfer - [![](https://bitvaria.com/_astro/wie-sieht-ki-ein-bild.CwlFrMzB.jpg)](https://bitvaria.com/en/wie-sieht-ki-ein-bild/) 13\. Juli 2026 6 min read ## [How Does an AI See an Image? Why It Reads the Screenshot but Misses the Tiny Number](https://bitvaria.com/en/wie-sieht-ki-ein-bild/) An AI model never sees an image as pixels. It cuts it into a grid of tiles that become numbers, and reads them like tokens. Why the fine print disappears first. Daniel Schleipfer - [![](https://bitvaria.com/_astro/ki-modelle-hosten-oder-mieten.DCbdN9Pz.jpg)](https://bitvaria.com/en/ki-modelle-hosten-oder-mieten/) 30\. Juni 2026 12 min read ## [Hosting Your Own AI Models or Renting? The Honest Math](https://bitvaria.com/en/ki-modelle-hosten-oder-mieten/) Is your own server for open AI models worth it, or is a subscription cheaper? The honest cost and hardware math for the Mittelstand, as of 2026. Daniel Schleipfer - [![](https://bitvaria.com/_astro/kv-cache.DoXy8WTH.jpg)](https://bitvaria.com/en/kv-cache/) 20\. Juni 2026 5 min read ## [Why a Model Gets More Expensive the Longer It Talks](https://bitvaria.com/en/kv-cache/) Same model, same question. Every answer costs more the longer the conversation runs. The reason has a name: KV cache. Daniel Schleipfer [All topics](https://bitvaria.com/en/blog/topics/) [View All Posts](https://bitvaria.com/en/blog/)