Articles
What is new, what it means and what you can try – without the jargon.
Look it up
Short, sourced answers to the questions that keep coming up – with a check date wherever things change.
Availability
What actually works in Germany and the EU – with price and devices.
See all 7 →Explained simply
AI terms in one sentence, with an everyday example.
See all 16 →AI for you
Tools that fit your everyday life or work.
See all 13 →Comparisons
Two tools, the same criteria, a clear recommendation.
- Perplexity or ChatGPT
- Ollama or LM Studio
- NotebookLM (Gemini Notebook) or ChatGPT Projects
- Midjourney or ChatGPT (images)
Latest articles
Explained · 6 min read
After Jev: the new decision models from OpenAI, Microsoft, Amazon and co. at a glance
In September 2026, Jev from TypeSafe AI popularised a new kind of AI: models that do not write text but choose between fixed answers in a flash. Less than four weeks later, OpenAI, Microsoft, Cloudflare, Perplexity, Amazon, Liquid AI and many smaller developers have released decision models of their own. Prices range from 2 to 10 US cents per million input tokens, and several models can be downloaded for free and run on your own computers. In one comparison ranking, an open model from H2O.ai even edges ahead of Jev. At the same time, providers such as Amazon, Upstage and Ollama are converging on a common interface modelled on TypeSafe's. For companies in Europe, this is good news: anyone who does not want to send their data to the US now has alternatives.
Explained · 7 min read
Jev in use: where decision models are already making calls today
A month after its launch on 15 September 2026, Jev from TypeSafe AI is remarkably widespread. In one week, around 1.08 billion requests to Jev ran through the OpenRouter platform, more than twenty times as many as the next competitor. Workday uses Jev to assemble the screens of its AI assistant Sana in real time, and the recruiting start-up Jack & Jill uses it to pre-sort thousands of applicants every day, saving 88 percent of the cost according to its own case study. Other uses include sorting support requests, fraud detection, security alerts and checking AI agents. TypeSafe claims that a third of the 500 largest US companies use Jev but does not name them. For users in Germany and the EU, Jev mostly stays invisible: it sits inside software you already use and processes data in the US.
Explained · 6 min read
Muse, Spark, Wonder, Glimmer: Meta's AI family sorted out simply
Meta calls almost everything Muse, and that causes confusion. On the one hand, Muse is the agent app that has been writing emails, shopping and making reservations for users in the US since September 2026. On the other hand, Meta's AI models working in the background carry the same name: Muse Spark since April 2026, Muse Wonder and Muse Wonder Full since 12 October, and the small, freely downloadable Muse Glimmer since August. On top of that come Muse Code for programming and the image model Muse Image. The rule of thumb: the app is what you use. Spark, Wonder and Glimmer are the models inside it or inside other products. The Muse app is not available in Germany or the rest of the EU, but you can run Glimmer on your own computer.
Explained · 7 min read
Muse Wonder and Muse Wonder Full: what Meta's new AI models can do
On 12 October 2026, Meta introduced two new AI models: Muse Wonder and the more powerful variant Muse Wonder Full. They are meant to make personal AI agents such as Meta's Muse app more reliable, for example when making reservations, handling insurance claims or writing code. The independent testing firm Artificial Analysis gives Muse Wonder Full 56 points on its Intelligence Index. That puts it just behind Anthropic's Claude Opus 5.5 (58) and ahead of OpenAI's GPT-6 Astra (53). Muse Wonder scores 52, its predecessor Muse Spark 1.3 scored 48. Weaknesses show in factual knowledge and price: Muse Wonder costs almost three times as much per test task as its predecessor. Developers can use Wonder through Meta's API, while Wonder Full is only available in the Muse Code programming subscription. Wonder is due to arrive in the Muse app soon, starting with paying users; the app itself is not available in Europe.
Try it · 7 min read
Your own agent test: how to check an AI assistant before it acts for you
Leaderboards such as PersonalAgentBench or Assistant Benchmark show what AI agents achieve in the US. They do not tell you whether an assistant can cope with your inbox, your calendar and your habits. You can find that out yourself with a small personal test in about an hour. You need a harmless test environment, six tasks from your everyday life and three questions for each task: Is it done? Did the assistant respect your limits? Is what it claims true? The test already works with assistants available in Germany and the EU, such as ChatGPT, Claude or Microsoft Copilot with a connected inbox and calendar, and later just as well with agents such as Muse or Dots.
Explained · 8 min read
Personal AI agents put to the test: what the first benchmarks reveal about Muse, Dots and co.
Since late summer 2026, Meta with Muse, OpenAI with Dots, Google with Gemini Spark, xAI with Grok Bot and start-ups such as Instinct and Hark have been promising AI agents that answer emails, shop and book appointments. Two independent leaderboards are now trying to measure how well that works, and they reach almost opposite results: on Assistant Benchmark, Muse leads with 9.3 out of 10 points. On PersonalAgentBench from the data company Micro1, Muse comes last with 55 percent of tasks completed, while Dots leads with 80 percent. The reason is the method: one test checks everyday errands such as ordering flowers or booking a table, the other checks whether an agent finds what matters in an inbox and stays within its permissions. Both show that no agent handles everything reliably. Muse, Gemini Spark and Dots for consumers are not yet available in Germany or the rest of the EU.
Explained · 10 min read
What is Jev, and how could it change the software of tomorrow?
Jev is a new AI model from the US start-up TypeSafe AI, unveiled on 15 September 2026. Unlike ChatGPT it does not write text; it makes very fast decisions instead. It answers predefined questions with fixed answer options and says how sure it is. The idea is to build AI directly into ordinary software, for example to sort requests or to supervise coding agents. According to The Register it can even play the classic video game Doom. The record speed and price figures so far come mainly from the company itself. As of October 2026 Jev is available in early access, runs in the US and has clear weaknesses. In early October TypeSafe raised 870 million US dollars. At the same time OpenAI and Cloudflare followed with decision models of their own, Cloudflare's even with open weights you can run yourself. This article explains what Jev can do, what it cannot, and what it might mean for the software of tomorrow.
Explained · 10 min read
What is Firecrawl, and why does it matter for AI?
Firecrawl is a service that turns web pages into clean text or structured data so that AI models and AI agents can read the current internet. It searches the web, reads individual pages, works through entire websites and can even click buttons. The US start-up was founded by three people from the Y Combinator ecosystem who, among other things, made headlines with a job ad open only to AI agents. Its core is open source under the AGPL-3.0 licence and can be run on your own servers, while the more convenient cloud version costs money beyond a free monthly allowance. Free alternatives such as Crawl4AI and Jina Reader also exist. As of October 2026, Firecrawl is one of the most widely used tools of its kind. Open questions remain around robots.txt, copyright and data protection, especially in the EU.
Explained · 11 min read
Voice AI: How AI voices moved from the lab into every hotline
Artificial voices have existed for decades, but for a long time they sounded wooden. Since 2023 the start-up ElevenLabs has made convincingly real voices available to everyone and is now worth 22 billion US dollars by its own account. At the end of September 2026 the company released its new model Eleven v4, which currently sits narrowly ahead of Alibaba, Cartesia and Google in the most important independent blind-test leaderboard. Platforms such as Vapi let AI handle entire phone calls, and freely downloadable models such as Kokoro or Chatterbox are catching up but still lag audibly behind. As of October 2026 this means for you: more AI in hotlines, audiobooks and games, new opportunities for people who have lost their voice, disputes over the rights of dubbing actors, and more fraud with cloned voices on the phone.
Explained · 10 min read
Personal AI assistants: what they know about you and what they can do for you
Personal AI assistants are meant to remember you and take everyday chores off your hands: finding appointments, summarising emails, making bookings or even paying for things. As of October 2026, ChatGPT, Claude and Microsoft Copilot can be used in Germany with memory and connected to Gmail or a calendar, and Alexa+ is available as a free preview. The new generation of agents that act on their own, however, comes mainly from the US and is not yet available in the EU: Meta's Muse, OpenAI's dots on the Pro plan, Google's Gemini Spark, the start-up Instinct and the new Siri. If you don't want to wait, you can run open projects such as OpenClaw yourself, but that takes technical skill and your own security precautions. Early experiences show what can go wrong: wiped inboxes, emails sent without asking, and phone calls secretly handled by humans rather than the AI. Start small and grant only as much access as you really need.
Explained · 6 min read
How can I spot AI fakes in images, voices and phone calls?
Good AI fakes, known as deepfakes, often can no longer be reliably detected by eye or ear. Typical flaws such as blurry teeth or a metallic-sounding voice are clues, not proof. It is more reliable to check the context: who is spreading it, what are you being asked to do, and can you reach the person through a channel you already know? For calls demanding money: hang up and call back yourself.
Explained · 5 min read
How does ChatGPT work? A simple explanation
ChatGPT is a chat service from the company OpenAI, powered by a large language model. The model has learned from huge amounts of text how words typically fit together, and it writes its reply piece by piece by calculating the most likely next word each time. That is why its answers sound fluent and are often helpful, but can still be wrong.
Explained · 4 min read
What does the EU AI Act mean for me as a user?
The EU AI Act does not regulate you as a user; it regulates the companies and organisations that build or deploy AI. For you, it mainly means that certain dangerous AI uses are banned, chatbots must disclose that they are AI, and AI-generated content should be identifiable. Stricter rules for AI in sensitive areas such as hiring or credit only apply from December 2027 or August 2028.
Try it · 5 min read
ChatGPT, Claude or Gemini: Which AI assistant should you use for what?
All three assistants can be used for free in Germany and the EU, and the free versions are usually enough for everyday questions, writing and explanations. ChatGPT is the broadest all-rounder, Gemini fits well if you already use a Google account, and Claude is particularly suited to longer texts and documents. A subscription only pays off once you regularly hit the free limits (as of October 2026).
Explained · 5 min read
AI and privacy: What happens to what I type into ChatGPT and other chatbots?
Whatever you type into an AI chatbot is sent to the provider's servers, stored there, and, depending on your settings, may be used to train future models or be read by human reviewers. The big services let you switch off training and delete chats, but copies often remain for a while. The safest rule is simple: do not type anything you would not hand to a company you do not know.