Sunday, 15 November 2026
12 episodes · 6596 in all
#6618: How Label Studio's XML Configs and Storage Work
Label Studio doesn't import your media—it leaves files in your cloud bucket and loads them via pre-signed URLs or a storage proxy.
#6617: Why CVAT Beats Generalists at Video Annotation
Born inside Intel in 2017, CVAT uses keyframe interpolation and track primitives to skip frames that generalist tools make you label by hand.
#6616: Small Labeling Tools That Outship Big Platforms
LabelImg is archived with 25,000 GitHub stars, while live single-machine tools like labelme and X-AnyLabeling ship AI updates weekly.
#6615: Quantization Quietly Deletes Fine-Tuned Skills
Three-bit GPTQ drops Llama-3.1-8B-Instruct by 14.5 points while the 70B version loses just 2, proving bit-width alone cannot predict damage.
#6614: How Humans Actually Grade Generated Video
While computer vision tools track objects across 100,000 frames, AI video evaluation still relies on two buttons and a five-second timer.
#6613: Where RLHF Preference Annotation Tools Stand
Open-source preference annotation specialist Argilla is in maintenance mode, leaving commercial platforms to ship the newest RLHF UIs.
#6612: How the Interface Shapes Data Annotation Tools
Audio and image annotation tools need heavy viewers, while LLM preference data collapses the UI to two text panes and a radio button.
#6611: Image Annotation When Experts Are the Bottleneck
The WikiChurches art historian annotated just 631 prototypical bounding boxes, narrowing 117 style classes down to four.
#6610: What Goes Into a Character-Voice LoRA Adapter
A character-voice LoRA adapter can be just twelve megabytes, yet it still requires a reference audio clip on every single generation.
#6609: Merging LoRA Voice Adapters vs Keeping Them Separate
Keeping fifty voice adapters separate takes 24 gigabytes of memory instead of 800 gigabytes for fifty merged models.
#6608: How Multi-Speaker TTS Models Keep Voices From Blending
Adding speaker conditioning to a vocoder can boost unseen-speaker quality by up to 95%, stopping the model from averaging every voice together.
#6607: How Water Dowsing Works Without Finding Water
In a German test, dowsers set their own 83% success threshold and scored 52.3%, barely above a coin flip.