Language models Published 24 August 2026 · Updated 19 September 2026 · 11 min
Archive
Articles
Technical articles, essays and historical notes, selected for English readers.
All topics and categories
Archive layout
Page 2 / 4 · 46 articles
Language models 
Language models Published 12 August 2026 · Updated 19 September 2026 · 12 min
Choosing the generator, embedding model and reranker for enterprise RAG
A practical three-part model stack for document assistants: small-model candidates, retrieval quality, vector and KV budgets, document preparation and a diagnostic table for finding the failing component. Read article →
Language models Published 13 July 2026 · Updated 19 September 2026 · 11 min
What four-bit quantization saves—and what it changes
Real Qwen3-8B file sizes, a worked KV-memory budget and the differences between quantization methods, precision, quality and serving speed. Read article →
Language models Published 2 July 2026 · Updated 19 September 2026 · 12 min
LLMs on RTX 4090: what changes between 24 and 48 GB?
Exact weight files, a worked KV-cache budget, replicas versus model splitting, and the point at which professional or datacenter GPUs become useful for inference and adaptation. Read article →
Language models Published 18 June 2026 · Updated 19 September 2026 · 13 min
How large a language model does your task actually need?
Choose between specialized models, small generators and larger LLMs by separating task quality, language evidence, retrieval, memory and serving capacity—with published comparisons and a worked memory example. Read article →
Security Published 18 September 2026 · 24 min
ZTAI for autonomous agents: bounded authority throughout execution
Removing direct human access to confidential data must not hand unlimited authority to an agent. How can task permissions, RAG, memory, MCP, tool execution, shutdown and independent evaluation keep that authority bounded and testable? Read article →
Language models Published 8 September 2026 · Updated 17 September 2026 · 14 min
From one user to enterprise LLM serving: when to change the model, replicas or GPU
Diagnose queues, memory and availability separately. Worked replica and cloud-capacity examples show how to grow a service while preserving quality, latency and failure tolerance. Read article →
Language models Published 29 July 2026 · Updated 17 September 2026 · 12 min
AirLLM and layer-wise inference: large models, little VRAM, and the cost of waiting
Where streamed weights live, how transfer time limits generation, why KV still matters, and which research, batch and adapter workloads can benefit from layer-wise execution. Read article →
Security Published 15 September 2026 · 14 min read
Preserving data confidentiality when using public APIs
Is removing a customer's name, or a provider's promise not to use data for training, enough to preserve confidentiality? The problem becomes harder when a useful model response depends on the very meaning and relationships an organization cannot disclose. This article compares approaches from minimization and pseudonyms to semantic transformation, trusted execution environments and homomorphic encryption. What does each method hide, how well does it work with an ordinary API, and which risks remain? Read article →
AI infrastructure Published 12 September 2026 · 14 min
Why the fastest GPU does not necessarily deliver the fastest response
More compute and a higher token rate do not always mean a faster response. This article examines time to first token versus completion time, memory and concurrency, the division of work between GPUs and LPUs, and communication costs—so infrastructure choices reflect the capacity to serve requests at the required quality and latency. Read article →
Security Published 9 September 2026 · 14 min read
When people cannot see the data, has their access really been removed?
Removing direct access to confidential data also requires control over code changes, output release and infrastructure administration. ZTAI must address those indirect paths. Read article →
AI infrastructure Published 8 September 2026 · 10 min