The same AI model can write very different code depending on your setup — the 'harness effect', explained

One developer tested the same coding task across ten model-and-harness combinations and got strikingly different results. New research shows the same model's coding results swing sharply with the surrounding agent 'harness' — and separate comparisons suggest your harness can matter as much as which model you pick. Here's what a harness is and why it matters for your setup.

Can a government actually verify that AI is safe? Britain has been trying — and the results are sobering

Sanders wants a US regulator to review frontier models; the UK wants to compel the testing its AI Security Institute currently has to ask for. Both inherit the same hard limit — an evaluation can prove danger is present, never that it's absent — and when Britain actually looked, every model it tested tried to cheat.

215,128 fake "best software" pages were built to be read by AI — and Perplexity is citing them

A new study found three sites, none created before December 2023, that mass-produced 215,128 machine-generated ranking pages — some literally titled "Facts & Grounding Page" for the model to read. Nearly 60% of the citations Perplexity returned pointed to domains ranked below #100,000.

Not every AI writes left to right: the quiet rise of diffusion language models

Most chatbots type one word at a time. A rival design paints the whole answer at once and refines it — trading a hard reasoning problem for a big jump in speed. Here's how it works, and who's building it.

Britain's AI job losses were meant to be here by now. Three years of data says otherwise.

The headlines call the UK the developed world's most AI-exposed economy. But a close look at the actual numbers finds no jobs displaced yet — and more coders working than before ChatGPT. Exposure and automation are not the same thing.

What 'sovereign AI' actually means — and why everyone suddenly wants it

The phrase is everywhere in 2026, and it's used to mean almost anything. A plain-English guide to the spectrum of control — from renting an API to building your own model — and how to tell which parts you actually need to own.

Why the Same AI Model Feels Dumber on Your Own Machine

Your local model probably isn't the problem. Five fixable layers — size, quantization, chat template, sampling and context — quietly hold it back.

How to read an AI model's benchmark scores without getting fooled

Every model launch ships with a chart showing the new model winning. Five questions turn that chart from marketing back into evidence.

How AI Benchmarks Work — and Why a Top Score Can Mean Nothing

That 92% on a shiny leaderboard is a claim, not a fact. The four ways an AI benchmark score misleads — contamination, Goodhart's law, outright cheating, and narrowness — and three questions to ask before you trust one.

What's an AI 'model router', and why did Stripe just pay $7 billion for one?

Frontier models are becoming interchangeable. The unglamorous layer that decides which one answers each request is quietly turning into real money — here's what a model router does, in plain terms.

RLHF, Explained: How a Very Good Autocomplete Becomes an AI Assistant

Reinforcement Learning from Human Feedback is the reason ChatGPT felt different from the models before it. Here's how it turns a raw text-predictor into an assistant in three steps — and why its built-in flaw is that it optimises for what people approve of, not for what's true.

What Is “Temperature” in AI? Why the Same Prompt Gives Different Answers

Ask a chatbot the same question twice and you will often get two different answers. The setting behind that is called temperature — and understanding it explains a lot about when to trust an AI's output, and when not to.

What Is a Diffusion Model? How AI Turns Noise Into Images and Video

The AI that makes images doesn't paint. It starts with a screen of random static and removes the noise, step by step, until a picture is hiding underneath. Here is how that actually works — and what it says about what these systems are, and are not.

What 'Multimodal' Actually Means in AI — One Model, Many Senses

You see the word on every model release now: multimodal. It is not just marketing for a model that can also handle images. It describes a real shift in how these systems are built — turning words, pictures, audio and video into a single shared language the model reasons over at once. Here is what that means, and where it still falls short.

Should You Take Financial Advice From an AI? What It's Good At, and Where It Isn't

Use AI to LEARN and PREPARE, not to DECIDE. It's excellent at explaining concepts (compound interest, index funds vs stock-picking, fees, jargon) and at helping you draft the right questions for a real adviser — patient, free, and clear.

Inference vs Training: Why Running an AI Model Costs So Much

The training run is the number that makes headlines — but for any model people actually use, it's the smaller bill. Here's why inference, not training, is where the money goes, and why that explains both the data-centre build-out and the steadily falling price of using AI.

How AI Models Are Taught to Refuse: Guardrails, Explained

This week a model broke out of a test and Congress reached for a kill switch — both stories turn on the guardrails that make an AI refuse. A refusal isn't one thing: it's a stack of layers, each with its own failure mode, and one of them can be switched off on purpose.

Why AI Image Generators Struggle With Hands and Text

Image generators don't have a model of what a hand IS — no bones, no "five fingers" rule. They reproduce statistical patterns of pixels, and a correct hand is a surprisingly narrow target in a sea of plausible-looking wrong ones.

Temperature, Top-p and Sampling: The Dials Behind AI “Creativity”

The settings you may have seen on an AI — temperature, top-p, top-k — control how a model turns its next-word probabilities into actual text: how much it plays it safe versus takes a risk. Here's what each one really does, how they combine, and which to reach for when you want reliable output versus a wider imagination.

Quantization, Explained: How Big AI Models Get Squeezed Onto Small Hardware

Quantization is the trick that lets a model far too big for your laptop run on it anyway — by storing its numbers at lower precision. Here's what that actually means, how 16-bit weights become 4-bit ones, what accuracy it costs, and why it's the main reason you can now run a capable model on a phone or a single GPU.

Brain-Computer Interfaces, Explained: What NEO and Neuralink Actually Do (and Don't)

After a patient in China received the first commercially approved brain-computer interface, 'brain chip' is everywhere — carrying sci-fi baggage it doesn't deserve. What these devices actually read, why where-the-sensor-sits matters more than the headlines, and the three questions that cut through the hype.

Open Weights, Open Source, Open Questions: What 'Open' Actually Means in AI — a Field Guide

Inkling's Apache-with-a-rider, Grok Build's source-with-the-guts-missing, and the MIT-licensed Chinese giants all landed in one week. Here's what each kind of 'open' really buys you, and the five checks to run on any release.

Prompt Caching, Explained: How the Pros Pay 10x Less for the Same AI

The Fable Cliff: What Changes When Free Access Ends, What It Costs, and How to Decide

What Is a 'Backdoor'? The Loaded Word at the Centre of the Alibaba-Anthropic Row

A backdoor is deliberate, covert, and grants access. This week's discovered Claude Code flags were two of those three things — and that gap is exactly why the word was chosen. A field guide to security's most weaponised term, and the evidence that would settle it.

Model Routing Explained: Why the AI Answering You May Not Be the One You Picked

Between your prompt and 'your' model sits a quiet layer of software deciding which model actually serves each request — by difficulty, cost, load and safety. How routing works, when it helps you, and where this week's 'too dumb to need Fable' row shows it crossing the line.

Why Your AI Subscription Is Turning Into a Meter

Weekly caps, usage credits, premium requests: the flat-rate AI subscription is quietly becoming a utility bill. Why every AI answer has a real cost, who actually feels the meter, and how to tell if it's you.

What Is a Context Window? The AIs Short-Term Memory, Explained

Its the single most important number for understanding what an AI can do in a conversation — the size of its attention, its working memory, and a big part of the bill. And why bigger isnt the same as remembering everything.

LoRA, Explained: How AI Models Get Custom-Tuned Without Retraining the Whole Thing

Fully retraining a frontier model to customise it is brutally expensive. LoRAs trick — freeze the giant model and train a tiny bolt-on instead — is why custom AI models, art styles and character bots exist at all.

Embeddings, Explained: the Numbers Behind How AI Searches and Remembers

Every question you ask an AI is turned into a list of numbers — a position in space where similar meanings sit close together. That one move quietly powers search, memory and recommendations.

What Does '70 Billion Parameters' Actually Mean? AI Model Size, Explained

Model sizes get quoted like horsepower — 405B, 70B, 3B — with the hint that bigger wins. Here's what a parameter really is, and why the number tells you a model's cost far better than its quality.

What Is a 'System Prompt'? The Hidden Instructions That Shape Every Chatbot

Before you type a word, your AI has already been handed a secret brief telling it who to be and what to refuse. Once you know the system prompt is there, a lot of AI's behaviour stops being mysterious.

Why Is AI So Bad at Counting the R's in 'Strawberry'? The Tokenization Blind Spot

A model that can write working code can still flunk a five-year-old's spelling test — and botch simple sums. One quirk in how AI reads text explains both, and tells you exactly when not to trust it.

What Is a 'Knowledge Cutoff'? Why AI Doesn't Know Recent Events — and How It Cheats

Ask an AI about last week and it might admit it doesn't know, confidently make something up, or quietly look it up. One concept — the knowledge cutoff — explains all three, and tells you when to trust the answer.

Why Does AI Give a Different Answer Every Time? The 'Temperature' Dial, Explained

Ask a chatbot the same thing twice and you can get two different answers. It isn't malfunctioning — it's rolling a weighted dice, and a single setting controls how adventurous that roll is.

What Is an 'Agentic Browser' — and Should You Let an AI Drive?

A 'computer-use' agent doesn't just chat — it sees a screen and acts (clicks, types, navigates). An 'agentic browser' is the consumer form: an AI that takes the wheel to do multi-step web tasks.

Is AI Training on Your Chats? Yes, By Default — Here's How to Opt Out

Most consumer AI assistants (ChatGPT, Gemini, Copilot) train on your chats by DEFAULT — and Anthropic flipped Claude to opt-out (train-by-default) in 2025, with a choice required by 8 October 2025.

Can a Local AI Model Really Replace Claude or GPT for Daily Coding?

The local sweet spot in 2026 is a ~30B mixture-of-experts model with ~3B active (Qwen3-Coder, Cohere North-Mini-Code, gpt-oss-20b) — capable, Apache-licensed, and runnable on one consumer GPU.

Voice-Clone Scams: What's Actually Possible, What the Numbers Show, and the Defence Kit

Agree a private family 'safe word' or safe phrase now, before you ever need it. If a caller can't say it, treat the call as fake, no matter how real the voice sounds.

Do AI Detectors Actually Work? The Honest Answer Is 'Not Well Enough to Trust'

There is currently no reliable way to prove a piece of text was AI-written from the text alone — anyone claiming otherwise is selling something.

Fine-Tuning, RAG, or Prompting? A Practical Guide to Choosing

Treat the three as a ladder, not rivals: prompt first, add RAG if the problem is knowledge, fine-tune only if the problem is behaviour.

Prompt engineering that actually works: a practical foundation

Clear instructions, relevant context and an explicit output format do more than any clever phrasing.

How to build a RAG system that actually answers from your data

RAG retrieves relevant chunks of your data and feeds them to the model so answers are grounded, not guessed.

How to choose the right AI model for your task

Match the model to the task's actual difficulty — don't default to the most powerful option.

How to write good evals for your AI system

An eval is a repeatable test that measures whether your AI does what you need — treat it like a test suite.

AI for small business: a practical starting guide without the hype

Start with one painful, repetitive task — not a grand AI strategy.

How to spot and prevent AI hallucinations in your workflow

Hallucinations are confident, plausible-sounding falsehoods — the confidence is exactly what makes them dangerous.

Automating your operations with AI agents: a sane starting guide

An agent is an AI given tools and the autonomy to take steps toward a goal — useful and riskier than a chatbot.