AI Blog
Analyses and guides from the world of AI — models, companies, market and policy.
13 posts

KV cache: why repeating a prompt is cheaper than the first one
Send the same prompt many times and part of the bill simply disappears. The KV cache is where speed and price change together — if you order your prompt correctly.

Small models move the frontier: 70B-class quality on 3 billion parameters
Multi-billion-parameter models now match last year’s ten-times-larger systems. What the quiet distillation-and-clean-data revolution means for your cost, latency and offline runs.

RLVR: when mathematical feedback replaced human taste
The new reasoning models moved training from "human opinion" to "answer right or wrong": math, code, logic. Why RLVR exploded reasoning quality — and created a new problem: tasks with no checkable "correct".

Thinking mode: on or off? A practical guide to cost-optimal decisions
Every reasoned answer means hidden tokens you pay for. A practical decision guide: which task classes justify thinking mode, where it is wasted money, and how to measure your real usage.

How MoE models work — and why their pricing is so confusing
235 billion parameters, but only 22 billion active per answer. How the mixture-of-experts architecture gives you small-model speed with giant-model knowledge — and why the "parameters" column stopped meaning much.

Agent memory: from context window to lasting memory
An agent that forgets last session is a tool, not a coworker. Agent memory has three layers — and the line between chatbot and colleague is drawn exactly there.

A2A protocol: when agents talk to each other
MCP connects agents to tools; A2A connects agents to colleagues. Two open standards side by side are building the shared language of the agent world — and "who owns the agent lingua franca?" has become the architecture question of 2026.

Context Rot: why a bigger window does not mean a bigger memory
Models now ship with million-token windows, yet research shows attention quality degrades long before the window fills. This is "context rot" — and how to tame it.

Choosing an AI coding subscription in 2026: from $6 to $200
The coding-plan market is full of pricing tiers and 5-hour windows. A guide to choosing based on the most realistic scenarios: occasional coding, full-time professional work, and building a product.

The evolution of coding agents: from autocomplete to autonomous engineer
In four years, AI coding went from single-line suggestions to agents that take whole tasks, read the repo, write code and run tests. A map of this journey and what each generation won.

Reasoning models and test-time compute: the new frontier of the model race
The race has moved from "bigger model" to "longer thinking". How giving more compute at answer time raises accuracy without retraining — and where it costs more than it is worth.

MCP: why it is called "the USB-C of AI"
The Model Context Protocol quietly became the standard that connects any AI agent to any tool. Why this open standard became last year’s most important infrastructure.

Hallucination in LLMs: why it happens and how to reduce it
The model confidently cites a source that does not exist. Hallucination is not a bug — it is a direct consequence of how these models are trained. But the mitigation toolbox improves every year.