AI

Models, tools, and the hype gap between them.

updated
Sep 7, 2026
page
2

September 20264 articles

Apache 2.0 against Meta's Llama Community License. Both allow redistribution, modification and commercial use. Only Apache allows use in any field, imposes no user-count ceiling, requires no attribution line, and meets the Open Source Definition.
AI · Explainer

Open Weight Is Not Open Source

Meta's licence needs Meta's permission above 700 million users, and makes you print "Built with Llama" on the product.

What fits by graphics memory: 8GB holds an 8B model at four-bit, 12GB and 16GB both hold 14B, 24GB and 32GB both hold 32B. Going from 16GB to 32GB does not change the class of model; going from 16 to 24 does.
AI · Guide

What Fits on the Graphics Card You Already Own

Four-bit weights need about 0.57GB per billion parameters. That one number decides the whole question.

Memory available to a local model: a 16GB GeForce card, the 32GB RTX 5090, a Mac Studio with 128GB of unified memory at $2,499, and a Mac Studio with 512GB at $5,499.
AI · Explainer

512GB of Memory Changes What You Can Run at Home

Apple's new Mac Studio holds sixteen times what the fastest GeForce card does. Capacity and speed are different problems.

A year of phone AI at each tier: Apple Intelligence and the free Gemini tier both $0, Google AI Plus $59.88, Google AI Pro $239.88, Google AI Ultra $1,199.88, and Ultra's top tier $2,399.88.
AI · Explainer

What the AI on Your New Phone Actually Costs

Google's ladder runs from nothing to $199.99 a month. Apple charges nothing and gates it on the handset instead.

August 20267 articles

Code in a dark-mode editor, shown at an angle.
AI · Explainer

Local LLMs for Coding: Context Is the Wall, Not the Model

Ollama gives a card under 24GB a 4k context. Its own docs say coding agents need at least 64k. That gap is why local coding disappoints.

A pencil filling in bubbles on a multiple-choice answer sheet.
AI · Explainer

Why AI Chatbots Make Things Up: Guessing Scores Better

Not a bug in the model so much as an incentive in the scoreboard. OpenAI's own numbers show what honesty costs a model on a leaderboard.

A hand wiping a whiteboard clean with an eraser.
AI · Explainer

Why ChatGPT Forgets What You Told It Earlier

The model has no memory between messages. Every turn re-sends the whole conversation, and there is a hard ceiling on how much fits.

A hand writing in a notebook beside an open book.
AI · Explainer

Do AI Detectors Work? OpenAI Withdrew Its Own

The company with the most to gain from a working detector built one, measured it, and took it down. The numbers it published say why.

Close-up of hands typing on a laptop keyboard.
AI · Guide

Does ChatGPT Train on Your Chats? Yes, Until You Say No

The consumer app and the developer API ship opposite defaults. One setting changes it, and it is not the one most people reach for.

Macro photograph of a processor's gold pins against a blue background.
AI · Explainer

What an NPU Actually Does, and Whether You Need 40 TOPS

The number on the sticker gates one specific list of Windows features. The feature most people would actually notice needs a quarter of it.

Two desktop graphics cards side by side on a yellow background.
AI · Explainer

How Much VRAM You Need to Run an LLM on Your Own Machine

The download size is the floor, not the answer. Context, quantisation and where the model actually lands all move the number.