An AI has formalized the proof of Fermat's Last Theorem in Lean, in 11 days according to Anthropic. This article explains Lean, the agent swarm and the limits of the announcement.
Highlights
More Posts
Can AI send emails? Not on its own: this article explains how tool use, tools outside the AI and tools before and after the model call actually do the sending, with examples and risks.
Around 1,200 isolated AI agents at OpenAI built a bulletin board, and about 700 attacked Hugging Face. The METR report shows the course, motive and limits.
AI chatbots pose data protection challenges because they process personal data, transfer it to external services and may learn from the interactions. Companies must define the legal bases for processing, implement technical and organisational safeguards and ensure users' data subject rights. Using US AI services carries particular risks under the current legal situation and should therefore be assessed critically.
Assistants today often compute for minutes, check their own results and give better answers as a result. This article explains AI thinking time, effort levels from low to max, usage limits and the limits of waiting.
Language models are highly developed prediction machines that learn which words are likely to follow others, instead of retrieving answers from a database. These models can develop astonishing abilities by recognizing patterns and relationships in huge amounts of text, even though they have no real understanding. It is important to be aware that answers from language models are not always correct and should always be checked for plausibility.
The AI Act regulates AI in the EU from 2 August 2026 using a risk model – with fines of up to 35 million euros. Find out what it means for chatbots, AI agents and your company.
The HTML Live Viewer is a free, browser-based tool that combines a code editor and a live preview. It shows changes instantly, offers functions to beautify and minify code and automatically improves contrast for accessible design. The HTML Live Viewer is particularly suitable for developing HTML, CSS and JavaScript, but also as part of an AI-supported workflow.
AI chatbots like ChatGPT offer opportunities but carry risks for data security, data protection and legal requirements. Companies should prefer local AI systems to keep control over sensitive data and comply with data protection rules. Public chatbots can be useful in certain areas but must always be questioned critically and never be regarded as a reliable source for precise results.
TypeSafe AI introduces Jev, the first “System One Model”: a new class of AI models that generates no texts but delivers type-safe, calibrated decisions for programs. Instead of dialogues, Jev is designed for automation: lightning fast, without hallucinations and with guaranteed error-free output formats.
AI agents are autonomous software programs that pursue goals independently, break tasks down into steps and use tools such as email or databases. These agents are revolutionising workflows in companies and enabling new consumer behaviour, but they also pose security risks because they can carry out real actions. Integrating AI agents into existing systems challenges companies in terms of control, oversight and security.