What is an NPU?
The chip your laptop and phone now advertise — what it's for, and why its headline "TOPS" number tells you less than it looks.
Author
Fiqhro Dedhen
Founder & Editor
Fiqhro Dedhen founded and edits NeuralGist. A full-stack developer with five years spent building the machinery businesses actually run on — APIs, integrations, internal tools — Dedhen writes about AI from the position of someone who has to make it work in production, where a good demo counts for nothing. The test applied to every tool covered here is not whether it is impressive, but whether it changes the job.
Writes aboutAI tooling, model releases, consumer technology, software development
24 articles
What is fine-tuning?
Teaching a model how to behave — not what is true.
What is chain-of-thought prompting?
Asking for the working, and getting a better answer as a side effect.
How to fix Ollama not using your GPU
It is slow, and it also silently gave your model a 4K context window. Both are the same bug.
Why Ollama gives you a 4K context window — and how to fix it
It is not the model. Ollama picks your context length from your VRAM at startup, and the bottom tier is very small.
What is a GGUF file?
The file format behind every local model you have downloaded — and how to read its name.
What is an AI agent?
A model that can take actions in a loop — and the word the industry has worn out.
AI Model Release Tracker
The current frontier and open-weight language models, with every spec read from a primary source.
What is an embedding?
Turning meaning into coordinates — the trick that makes semantic search work.
What is a transformer?
The 2017 architecture underneath essentially every model you have heard of.
What is an AI hallucination?
When a model states something false with total confidence — and why that is the normal case, not a glitch.
What is prompt caching?
Paying once for the part of your prompt that never changes — and the write fee nobody mentions.
What is RAG (retrieval-augmented generation)?
Giving the model the documents instead of hoping it memorised them.
What is a context window?
The model's working memory — and the reason a long chat starts forgetting things.
How to cut your LLM API bill
Four levers, ranked by how much they actually save. Switching vendor is not one of them.
What is quantization?
Shrinking a model so it fits on hardware you own — and what you give up.
Ollama vs LM Studio vs llama.cpp: which local AI runner?
They are the same engine. You are choosing an interface and a licence, not a speed.
LLM API pricing compared: OpenAI vs Anthropic vs Google
Output costs five to six times more than input — at every vendor, at every tier. That ratio, not the headline price, is what decides your bill.
What can you actually get for free?
Google gives away every Flash text model and an older Pro. It does not give away the current Pro, or a single image.
What does an AI-generated image actually cost?
Image APIs bill in tokens, not images. Only one major vendor will tell you what a picture costs.
What is a token?
The unit models read, count and bill in — and it is not a word.
What does AI voice actually cost?
Audio tokens cost eight times what text tokens cost. But the cache discount on them is close to 99%.
Which Claude model should you use?
Four current models, one 1M-token window that is not the same size on all of them, and a tokenizer change nobody announced as a price rise.
AI coding assistant pricing compared
Five tools, five different ways of quoting a price — and one of them shows different numbers to different people.