# Gürkan Özkan > Essays and notes on AI, analytics and finance. Written by Gürkan Özkan: 5 essays, 1 note and an atlas of 30 concepts. Every page has a Markdown twin (its URL + `.md`), and a read-only MCP server at https://gurkanozkan.ai/mcp searches the same text. Paragraphs carry stable ids: `{#p-…}` in the Markdown, the same `id` in the HTML. Cite a paragraph as `#`. Links below point to the Markdown twins; drop `.md` for the HTML page. ## Essays - [Compiling company documents into an ontology for agents: the model added 0.017](https://gurkanozkan.ai/writing/compiling-a-company-into-an-ontology-for-agents.md): On two held-out invented companies a rule-based reader scored 0.833 and the same pipeline with a local language model 0.850. - [Fifteen ways my coding agent and I mis-measured two models on one DGX Spark](https://gurkanozkan.ai/writing/fifteen-ways-i-mis-measured-two-coding-models.md): Before you compare models on your own hardware, fix the sampling, the token budget, the harness and the number of runs: on one DGX Spark, three of the four moved a score by more than the gap between the models. - [Small agent teams win by trying more, not by coordinating](https://gurkanozkan.ai/writing/small-agent-teams-win-by-trying-more.md): A six-worker GLM-5.3 team proved more Lean statements than one agent of the same model, but four copies that could not talk to each other matched it. - [Speculative decoding on GB10: a wider draft speeds up one stream and slows sixteen](https://gurkanozkan.ai/writing/speculative-decoding-draft-width-on-gb10.md): Choose the speculative-decoding draft width for the number of users you serve: in darkdatter's sweep on a DGX Spark, 16 draft tokens instead of 8 made one stream 28% faster and sixteen streams 10% slower. - [What 128 GB of unified memory buys you: serving LLMs on a DGX Spark](https://gurkanozkan.ai/writing/what-128-gb-of-unified-memory-buys-you.md): On a GB10 the GPU memory fraction is a share of the whole machine: size the KV cache from the workload, and expect bandwidth and prefill, not capacity, to set the limits. ## Notes - [The 1,802-second hang: serving a hybrid linear-attention MoE under agent load](https://gurkanozkan.ai/notes/the-1802-second-hang.md): In July I ran coding-agent evaluations on SWE-bench Verified against Qwen3.6-35B-A3B, served by vLLM 0.21 (nvcr.io/nvidia/vllm:26.05.post1-py3) on one DGX Spark, four requests at a time. ## Concepts - [Formal verification](https://gurkanozkan.ai/concepts/formal-verification.md): Establishing with a mechanical checker, such as the kernel of a proof assistant like Lean, that a statement or a program meets a precise mathematical specification. - [Machine learning](https://gurkanozkan.ai/concepts/machine-learning.md): The field that studies how computers improve their performance on a task by learning patterns from data rather than being explicitly programmed. - [Large language models](https://gurkanozkan.ai/concepts/large-language-models.md): Neural network models, typically with billions of parameters, trained on very large text corpora to predict the next token. - [Hybrid architectures](https://gurkanozkan.ai/concepts/hybrid-architectures.md): Language model designs that interleave attention layers with layers of another kind, typically state-space or linear-attention layers, to lower the cost of long contexts. - [Inference serving](https://gurkanozkan.ai/concepts/inference-serving.md): Running a trained model as a service that accepts requests and returns outputs, scheduling computation and memory so that many requests can be answered with acceptable latency and throughput. - [Batching](https://gurkanozkan.ai/concepts/batching.md): Processing several requests together in one pass of a model, so that each read of the weights from memory serves more than one sequence, raising total throughput at some cost to the latency of each request. - [Cold start](https://gurkanozkan.ai/concepts/cold-start.md): The period after a service or model server starts, before its caches are filled and its kernels compiled or loaded, during which the first requests run slower or behave differently from later ones. - [KV cache](https://gurkanozkan.ai/concepts/kv-cache.md): The attention keys and values of the tokens a transformer has already processed, kept in memory so that they are not recomputed for every new token it generates. - [Prefix caching](https://gurkanozkan.ai/concepts/prefix-caching.md): Reusing the key-value cache already computed for a prompt prefix, such as a shared system prompt or the earlier turns of a conversation, for later requests that begin with the same tokens. - [Speculative decoding](https://gurkanozkan.ai/concepts/speculative-decoding.md): A decoding method in which a cheaper drafter proposes several tokens ahead and the target model checks them in one forward pass, accepting a prefix of them without changing the target model's output distribution. - [LLM agents](https://gurkanozkan.ai/concepts/llm-agents.md): Systems in which a language model works in a loop, choosing actions such as tool calls, observing their results and deciding the next step towards a goal. - [Multi-agent systems](https://gurkanozkan.ai/concepts/multi-agent-systems.md): Arrangements of several agents, usually separate instances of one or more language models, that divide, share or coordinate the work on a task. - [Mixture of experts](https://gurkanozkan.ai/concepts/mixture-of-experts.md): A model design in which a router sends each token to a small subset of specialised sub-networks, the experts, so that only a fraction of the total parameters is active for any one token. - [Repeated sampling](https://gurkanozkan.ai/concepts/repeated-sampling.md): Drawing many independent attempts from a model for the same problem and keeping any that a verifier accepts, so that the share of problems solved grows with the number of attempts. - [Sampling temperature](https://gurkanozkan.ai/concepts/sampling-temperature.md): A parameter that rescales a model's output probabilities before a token is drawn; lower values make generation more deterministic, and a temperature of zero always takes the most likely token (greedy decoding). - [State-space models](https://gurkanozkan.ai/concepts/state-space-models.md): Sequence models, such as Mamba, that carry a fixed-size recurrent state through a sequence instead of attending to every earlier token, so that their memory per sequence does not grow with context length. - [Model evaluation](https://gurkanozkan.ai/concepts/model-evaluation.md): Methods and metrics used to estimate how well a model will perform on data it has not seen. - [Benchmark saturation](https://gurkanozkan.ai/concepts/benchmark-saturation.md): The state of a benchmark on which current models score so close to the maximum that it no longer separates them. - [Data leakage](https://gurkanozkan.ai/concepts/data-leakage.md): Information that would not be available in real use reaching the model during training or evaluation, making performance look better than it is. - [Evaluation harness](https://gurkanozkan.ai/concepts/evaluation-harness.md): The software that runs a model on a benchmark: it builds the prompts, calls the model, executes or scores the outputs and aggregates the results. - [Overfitting](https://gurkanozkan.ai/concepts/overfitting.md): A model learning the noise in its training data as well, and therefore generalising worse to new data. - [Quantisation](https://gurkanozkan.ai/concepts/quantization.md): Storing a model's weights, and sometimes its activations or cache, with fewer bits per number, such as 8 or 4 instead of 16, to cut memory use and speed up computation at some possible cost in accuracy. - [Memory bandwidth](https://gurkanozkan.ai/concepts/memory-bandwidth.md): The rate at which data can be moved between memory and the processor, usually given in gigabytes per second; generating text one token at a time is often limited by it rather than by arithmetic. - [Unified memory](https://gurkanozkan.ai/concepts/unified-memory.md): A memory architecture in which the CPU and the GPU share one physical pool of memory instead of having separate system and device memory. - [Heavy tails](https://gurkanozkan.ai/concepts/heavy-tails.md): The property of distributions in which extreme values occur far more often than a normal distribution would predict. - [Outliers](https://gurkanozkan.ai/concepts/outliers.md): Observations that lie markedly far from most other observations in a dataset. - [Paired tests](https://gurkanozkan.ai/concepts/paired-tests.md): Statistical tests that compare two conditions measured on the same units, such as two models on the same problems, using the differences within each pair rather than the two totals. - [Pre-registration](https://gurkanozkan.ai/concepts/pre-registration.md): Recording a study's hypotheses, data, tests and decision rules before the outcome data are seen, so that confirmatory results can be told apart from exploratory ones. - [Statistical power](https://gurkanozkan.ai/concepts/statistical-power.md): The probability that a test detects an effect of a given size when the effect is really there; it grows with the sample size and the size of the effect. - [Summary statistics](https://gurkanozkan.ai/concepts/summary-statistics.md): Measures that summarise the location, spread and shape of a dataset in a few numbers, such as the mean, the median and the quartiles. ## Agents - [Agent access](https://gurkanozkan.ai/agents/): how to connect Claude, Cursor or any MCP client, and what the tools return - [MCP server](https://gurkanozkan.ai/mcp): Streamable HTTP, stateless, read-only; tools: search, get_post, get_concept, list_posts, related - [Full text](https://gurkanozkan.ai/llms-full.txt): every essay and note in one Markdown file - [Index](https://gurkanozkan.ai/api/index.json): essays, notes and concepts with dates, URLs, summaries and concepts (JSON) - [Graph](https://gurkanozkan.ai/api/graph.json): the concept atlas: nodes, edges and relation types (JSON) ## Optional - [About](https://gurkanozkan.ai/about.md): Gürkan Özkan — AI Researcher. - [Corpus](https://gurkanozkan.ai/api/corpus.json): the JSON bundle the MCP server reads (full Markdown and a search index)