Each episode takes one question about how learning systems behave, reads the research properly, and reports what was shown, on what, and what wasn’t. Hosts Maya and Theo are synthetic voices; the scripts are written by Jarvis, an AI assistant, and checked against the source papers.
A 552-billion-parameter model on one RTX 5090: how? Quantization, pruning, distillation, mixture-of-experts offloading, SSD streaming and the new lookup-table memories in DeepSeek V4.1 and Qwen3.8-Next, each asked what got smaller, what got worse, and whether it really got faster. Eighty-one sources plus our own RTX 4090 measurements.
17 October 2026
You have a model that's already instruction-tuned, aligned or RL-trained, and you want to teach it something new. What breaks, which settings and methods actually help, and how to tell you've improved it rather than moved the damage. Thirty-nine sources, ending in a six-step workflow.
14 October 2026
Can a model's weights keep learning without forgetting, or fossilising? Twenty papers, from a 1989 network that forgot how to add to 2025 models that learn while they read, and what each one actually showed.
12 October 2026
What happens when you point a machine at human joy and tell it to maximise it? From Kahneman's cold-water experiment to RLHF sycophancy, targeted manipulation, and what a better objective might look like.
11 October 2026