Tokens that unlock AI access, networks that pay people to run models, GPU marketplaces paid in crypto. What each one really is, from its own documentation, which you can use today, and why a blockchain does not make a chatbot private.
Yes, an iPhone can run an AI model with nothing leaving the phone. The four apps that do it, which models fit 6, 8 and 12 GB phones, what Apple Intelligence does on the device versus in Private Cloud Compute, and the one setting that sends your request to OpenAI.
Lumo is Proton's AI assistant: zero-access encrypted history, no logs, open models on Proton's own European servers. I ran three test prompts in guest mode: what it got right, where web search fired on its own, what Plus costs in pounds, and when a local model is better still.
ChatGPT stores your conversations and, on consumer plans, uses them for training unless you switch it off. I checked my own ChatGPT, Claude and Gemini settings: what was on, the new ads controls, how Temporary Chat works, what a VPN or crypto actually changes, and how the three compare.
Six chatbots that promise privacy, checked against their own policies and, where they let me in, against the same three test prompts: what each keeps, where it runs, what it costs in the UK, how honest it is about itself, and which one to use for what.
Venice AI is a US chat service that promises no prompt logging and uncensored models. What its four privacy modes mean, what leaves your device, why UK users get no free tier and must pay by card, how the VVV token replaces a subscription, and when a local model is the better choice.
Private AI means your prompts, documents and recordings never leave hardware you control. The three ways to get it, what each costs in the UK, which models and tools to use, where it falls short of ChatGPT, and what UK GDPR requires.
MacWhisper transcribes audio on your Mac without sending it anywhere. What the free version does, what €64 Pro adds, which Mac you need, and the free alternatives: Apple's Voice Memos, whisper.cpp and Buzz.
What Claude skills are, the four built-in ones, which plans get custom skills, how to make one by describing your process in a chat, and how they differ from Claude Code skills and from Projects.
The Ollama library has 258 models. This list ranks the 30 most downloaded, gives each one's download size, context length and job, and shows which run on 16, 32, 64 and 128 GB Macs. Updated monthly.
The Ollama problems Mac users search for most, with the fix for each from the official docs: port 11434 already in use, could not connect, model running on CPU, slow first answer, context length, moving models and turning off cloud features.
What an AI receptionist costs in the UK, how IONOS, Fasthosts, Hey Jodie and Moneypenny compare on price and features, and the UK GDPR checks to make before you divert your calls.