Nils Durner's Blog Ahas, Breadcrumbs, Coding Epiphanies

AI Agent efforts underway

WirtschaftsWoche, a German business magazine and key source to decision makers, features German-led (but US-based) AI startup uiAgent. Having prototyped something somewhat similar and with top AI Labs preparing launches (1, 2, 3, and more), their core tech appears unremarkable to me. But their overall narrative is: replacing up to 50% of offi... Read more

Google Search AI Summary: ficticious Encanto 2

Simon Willison has surfaced renewed cases where Google Search shows so-called “featured snippets” (🇩🇪 “hervorgehobene Snippets”) that answer the user search flat-out wrong. Simon writes on Bluesky: This is a particularly bad case-study in how badly AI summarization can go when its exposed to the wilds of the internet It’s not “AI Summarizat... Read more

2024 Quantum Computing Review

With end-of-year roundups and predictions being in season, it’s also worthwhile to consider what hasn’t, despite ceaseless cheerling, been achieved or announced: Quantum advantage in breaking real-world encryption. While classical computing has cracked very weak RSA (trivially small key size of perhaps 1/10^535 the strength of today’s d... Read more

Document-to-Markdown Converters for LLM Use

Recently, a few open-source tools for converting PDFs, Office documents, and other formats into Markdown have drawn attention. Among these are MarkItDown from Microsoft, Docling from IBM Research, PyMuPDF4LLM, and the Jina AI Reader API. They aim to provide text suitable for downstream tasks, including LLM-driven analysis, without requiring manu... Read more

o1 Pro Mode & Llama 3.3

Quick notes on last week’s foundation model releases: OpenAI o1 o1 was released through ChatGPT: it’s a stark improvement over the o1-preview available through API; o1-preview basically not representative. The new “o1 Pro Mode” is a class of its own: it aces through almost all of the subject tasks in a survey paper I have under submission, and ... Read more