Developer tag | eYssen AI

29 Articles

Developer ×

Fine-Tuning: When Is It Finally Worth It?

A fine-tuned model with 350 million parameters hit a **77.6%** success rate on a tool-calling task — while a general model with 175 billion parameters managed **26%** on the same task. Five hundred times fewer parameters, three times the result. That is the real promise of fine-tuning. And meanwhile one large provider is **retiring its fine-tuning platform**: no new users are being accepted, and existing ones have until January 2027.

Advanced Developer Explainer Data, RAG & Self-Hosted AI

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Local AI Runners: Which One to Pick

The most important thing to know: **Ollama, LM Studio, Jan, KoboldCpp and Docker Model Runner all use the same engine** — llama.cpp, and MLX on a Mac. They do not differ in speed but in the packaging and in the **defaults**. And the defaults are exactly where the trap is: Ollama **truncates at 4,096 tokens** out of the box, even when the model could handle 256k.

Beginner Developer Tool review Ollama Data, RAG & Self-Hosted AI

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Quantization: How Big Models Fit on Small Machines

According to a unified 2026 measurement, four-bit quantization **cuts model size by 71%** and costs on average **0.43 percentage points** of accuracy. Essentially free. Except **not for everyone**: the same four-bit quantization also produced a **59%** collapse on long input, and the degradation is **up to five times worse in languages other than English** — measured across 26 languages, Hungarian among them.

Advanced Developer Explainer Data, RAG & Self-Hosted AI

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Hardware for Local AI: VRAM, RAM and Apple Silicon

The rule of thumb everyone repeats says you need about half the parameter count in memory. `[Own calculation]` The four-bit weights of an eight-billion-parameter model really do eat 4.6 gigabytes — **but the same model's 128k-token context eats 18 gigabytes**, four times the model itself. Anyone who only looks at model size has worked out a fifth of the memory they actually need. And there is one measurement that turns the whole picture around: **a chip with half the GPU cores processes a prompt 83% faster than a more expensive one** — while generating a fifth slower.

Advanced Developer How-To Data, RAG & Self-Hosted AI

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Hybrid Search: Why Vectors Alone Are Not Enough

In an academic measurement of 24 thousand questions, vector search scored **19.2%** on questions of the "who is [X]'s child?" type — the decades-old keyword algorithm scored **82.9%**. Sixty-three points apart. On a different set of questions the very same methods come out the other way round. **Neither one is better than the other: they have different failure modes.** Except that hybrid search is not what improves things most — re-ranking is.

Advanced Developer How-To Data, RAG & Self-Hosted AI

Sep 13, 2026

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

Krisztian Eyssen, Krisztian Eyssen

Project Instructions an Agent Actually Follows

In a paired experiment, a project instruction file meant **20% fewer tokens and 20% less time** for the same task. What nobody has demonstrated is that it makes the code better. And there is a ceiling: **at roughly eighty simultaneous rules**, "follows every rule" essentially stops happening.

Advanced Developer How-To Coding & Development

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Reviewing AI-Written Code: What to Look For

Since 2023 the share of generated code that compiles has climbed from 50% to 95%. Over the same period the security rate **stayed between 45% and 55%**. The defects did not disappear, they moved house: permissions and architecture took the place of typos. Review has to follow them.

Intermediate Developer How-To Coding & Development

Sep 13, 2026

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

Krisztian Eyssen, Krisztian Eyssen

Hallucinated Packages: When AI Invents a Library

**43%** of invented package names come back identically in all ten runs out of ten. That is what turns a hallucination into an attack surface: the attacker does not have to guess what the model will invent — one question is enough. A documented, successful attack, on the other hand, **had not happened** by mid-2026 — and that honest answer is stronger than scaremongering.

Advanced Developer Troubleshooting Coding & Development

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Tests First: Building the Safety Net with AI

Out of a hundred generated tests, **eighteen** stayed usable in an industrial-scale system. And there is a surprising split: **a test you hand the model in the prompt improves the result by 7–30% — a test the agent writes for itself improves it by practically nothing.** The difference between the two is who gets to say what "correct" means.

Intermediate Developer How-To Coding & Development

Sep 13, 2026

Krisztian Eyssen, Krisztian Eyssen

Understanding a Legacy Codebase with AI

Model performance **falls off sharply above 32,000 tokens** — while the context window they advertise runs from 128,000 to a million. And there is a measured bias: ask how the repo is put together, and the model would rather **rewrite** it than understand it. This post is about how to ask so that does not happen.

Intermediate Developer Recipe Coding & Development

Sep 13, 2026

Tags

Audience

Everyone Office worker Marketer Creative Business owner Developer

Follow Us

Archives

Tags