English · 17 released episodes
Four ways into AI: quick concepts, connected terms, deep dives and the week's news. New episodes in English. Andy's Café is also available in Deutsch, Italiano, Español and Français. Brewed by machines. Tasted by humans.
Eight consequential AI stories from 17–29 August 2026: the Hugging Face intrusion, automated alignment research, Qwen releases, OpenAI's inference chip, cryptographic evaluations, classroom evidence, Google's travel tools, and books acquired for training data.
A philosophical and empirical tour through meaning, grounding, generalisation, tool use, and the limits of behaviour as evidence of understanding.
A practical guide to what AI benchmarks measure, how test sets leak into training data, and why contamination can make progress look larger than it is.
How smaller numerical representations shrink language models, why calibration and software kernels matter, and why quantisation is an engineering trade-off rather than free compression.
Why language models do not simply choose the next word: a concise guide to temperature, probability distributions, and sampling controls such as top-k and top-p.
How a personal listening tool became a multilingual public podcast: the formats, editorial choices, speech pipeline, open-web distribution, privacy principles, growth strategy and questions that come next.
How fluent next-token prediction can produce precise falsehoods, what confidence and calibration really mean, and how to verify an answer without dismissing everything.
Why training and inference require so much arithmetic, data movement and coordination—and why more compute brings real gains without guaranteeing truth or value.
Where a trained model ends, what makes a system agentic, what the surrounding harness does—and why more autonomy is not always better.
How pretraining builds a model’s broad base, how fine-tuning adapts an existing checkpoint, what post-training adds—and why these labels overlap.
Claude’s possible mathematical advance, the race for faster and more persistent models, Meta’s open weights, Google’s reported billion-user scale for Gemini, ads in ChatGPT, simulated clinical video visits and Z.ai’s delayed weights.
How next-token prediction changes a model’s weights, what post-training adds, and why context and retrieval are not the same as permanent learning.
Why a long chat can remember your preferences and still forget an early instruction—and how tokens, context, memory and retrieval fit together.
Why language models count text in tokens, how tokenization works, and why tokens shape context, speed and cost.
The practical difference between training a model and using it for inference, including prompts, memory and retrieval.
What exactly is an AI model, how is it different from the surrounding product, and what does training produce?
This week's Morning Paper: cyber evaluations, OpenAI's Astra, mathematical research, Google leadership, new models, data-centre power and autonomous agents.