Explainer tag | eYssen AI

60 Articles

Explainer ×

Why Run AI on Your Own Machine?

Krisztian Eyssen, Krisztian Eyssen

Why Run AI on Your Own Machine?

Of the four usual reasons, two hold up and two do not. `[Own calculation]` Against the cheapest API tier, a $3,500 machine breaks even at **2.6 billion tokens a year** — 133 days of continuous generation. And here is the twist that holds the whole article together: **the volume at which local running pays off is exactly the volume at which the local model is no longer good enough.** There is a fifth reason, though, that nobody states out loud, and measurably it is the strongest of the lot.

Beginner Business owner Explainer Data, RAG & Self-Hosted AI

Sep 13, 2026

Vector Databases Without the Jargon

Krisztian Eyssen, Krisztian Eyssen

Vector Databases Without the Jargon

A widespread myth says the vector index eats a lot of memory. According to the vendor's own capacity-planning formula, **the graph is 5% of total memory** — 0.29 GB of index alongside 5.72 GB of vectors. It is the vectors that eat memory. And there is an even more important number: **ten thousand vectors are 75 megabytes.** That fits on a phone. So the question is not whether you need a separate vector database, but **when** — and the answer is usually "later than you think".

Intermediate Developer Explainer Data, RAG & Self-Hosted AI

Sep 13, 2026

Teaching AI Your Own Documents: RAG Explained

Krisztian Eyssen, Krisztian Eyssen

Teaching AI Your Own Documents: RAG Explained

Start with a 910-question benchmark: the same model scored **48.1%** out of the box, **50.4%** after fine-tuning, and **87.5%** with RAG. On another model fine-tuning did not help but **hurt** — from 35.3% down to 21.9%. So RAG is not one option among several: for new knowledge it is an order of magnitude better. It just happens **not to be teaching** — and most of the disappointment comes from believing that it is.

Intermediate Business owner Explainer Data, RAG & Self-Hosted AI

Sep 13, 2026

The Security Bill of AI-Generated Code

Krisztian Eyssen, Krisztian Eyssen

The Security Bill of AI-Generated Code

From 11 September 2026, anyone placing software on the market in the European Union **has to report an actively exploited vulnerability within 24 hours**. Meanwhile roughly **45%** of generation tasks produce code with a known vulnerability when there is no explicit security instruction. Those two numbers together are the bill.

Advanced Developer Explainer Coding & Development

Sep 13, 2026

Vibe Coding: What It Is Good For and Where It Stops

Krisztian Eyssen, Krisztian Eyssen

Vibe Coding: What It Is Good For and Where It Stops

By the original definition, the point of vibe coding is that you "**forget that the code even exists**". So the question is not how good the model is — it is whether **you can afford to forget that this particular code exists**. In nine documented cases the answer was no.

Beginner Business owner Explainer Coding & Development

Sep 13, 2026

From Prototype to Production: The 70% Wall

Krisztian Eyssen, Krisztian Eyssen

From Prototype to Production: The 70% Wall

There is **no measurement** behind the famous "70%" — the author himself writes that he took it from a tweet. What has been measured: **37%** of companies can show any bottom-line effect at all from AI, and that share has **not moved** in a year — while 80% feel personally more productive. That is the wall.

Intermediate Business owner Explainer Coding & Development

Sep 13, 2026

Where AI Saves Real Time in an Office Day

Krisztian Eyssen, Krisztian Eyssen

Where AI Saves Real Time in an Office Day

Seven task types, with measurements. In some the saving is large and real: in a 444-person experiment a writing task fell from 27 minutes to 17. In some it is exactly zero: in a 6,000-person experiment measured by telemetry, time spent in meetings did not move. And in some the machine slows you down. The difference is not luck, it is the type of task.

Beginner Office worker Explainer Work & Productivity

Sep 13, 2026

Autocomplete, Chat, Agent: Three Modes of AI Coding

Krisztian Eyssen, Krisztian Eyssen

Autocomplete, Chat, Agent: Three Modes of AI Coding

The difference between the three modes is not the model — the same model runs in all three. The difference is **whose hands the filesystem is in**. In the first it is yours, in the second it is on the clipboard, in the third it is the tool's — and from there the only question is when it asks.

Beginner Developer Explainer Coding & Development

Sep 13, 2026

Self-Hosted Personal Agents and Their Risks

Krisztian Eyssen, Krisztian Eyssen

Self-Hosted Personal Agents and Their Risks

The security policy of the most widely used open personal agent lists what does **not** count as a vulnerability. This is on that list: **"prompt injection, if it does not bypass policy, authentication, approval, the sandbox or a tool boundary."** And this: **"anyone who can operate an agent can make it do anything that agent is able to do."** This is not a flaw in the system — this is the system. This article is about what you need to know about that before you start it up.

Intermediate Everyone Explainer Agents & Automation

Sep 13, 2026

Multi-Agent Systems: When More Agents Help

Krisztian Eyssen, Krisztian Eyssen

Multi-Agent Systems: When More Agents Help

Five real multi-agent systems were measured in the same study. The best one scored **84.8%**, the worst **13.3%** — and in the middle of the field sit two systems that got **three quarters of the tasks wrong**. The difference did not come from the model. It came from the pattern. This article is about **which four patterns pay back the cost of coordination, and which two do not.**

Advanced Developer Explainer Agents & Automation

Sep 13, 2026

Agent Memory: What to Store and What to Forget

Krisztian Eyssen, Krisztian Eyssen

Agent Memory: What to Store and What to Forget

Memory is the weakest agent capability there is. On the best independent measurement, **the average of 109 models fails to reach 30% on any memory strategy** — the very same models score 86–91% on simple tool calls. That measurement also shows that **there is no universally good memory architecture**. And there is a second reason for caution: memory is an **attack surface** — in one study an attacker planted a persistent false memory with a **98.2% success rate**, using nothing but questions.

Advanced Developer Explainer Agents & Automation

Sep 13, 2026

Agent Reliability Math: Why Twenty Steps Fail

Krisztian Eyssen, Krisztian Eyssen

Agent Reliability Math: Why Twenty Steps Fail

0.98 to the twentieth power = **66.8%**. That one multiplication explains the feeling that an agent is "almost right, but it always breaks something somewhere". And the reverse is harsher still: **at twenty steps, a 90% end-to-end success rate demands 99.475% per-step accuracy.** Not 90%. This whole topic lives in the gap between those two numbers.

Intermediate Business owner Explainer Agents & Automation

Sep 13, 2026

1
2
3
4
5

Follow Us

Archives

Tags

Audience

Everyone Office worker Marketer Creative Business owner Developer

Category

AI Basics Safety & Risk Regulation & Policy Infrastructure & Chips Business & Deals Prompting Models & Market Writing & Content Images & Design Video & Audio Marketing & SEO Coding & Development Work & Productivity Agents & Automation Data, RAG & Self-Hosted AI Business, Law & Security

Format

How-To Explainer Comparison Tool review Recipe Template Troubleshooting Case study Glossary News

Level

Beginner Intermediate Advanced

Tool

OpenRouter Perplexity HeyGen ElevenLabs Suno Obsidian MCP n8n Ollama EYAS

Others

Odoo