Training Data
322 episodes · Page 7 of 14
#4670: Why AI Weights Are Indecipherable
Can you actually read an AI model's weights like a book? We explore why the answer is mostly no — and what researchers are doing about it.
#4669: What "Distribution" Really Means in AI Models
Unpacking what "distribution" actually means under the hood — and why training data variety shapes model behavior.
#4666: Why 30 Seconds of Audio Beats 3 Minutes for Voice Cloning
Why adding more audio made Daniel's voice clones worse — and what it reveals about how voice embeddings actually work.
#4505: What DeepSeek's Training Data Reveals About Model Voice
Why Chinese AI models trained on 60% English still produce dialogue that feels distinctively different.
#4492: What Makes a Container Robot-Ready?
The hidden specs that turn a plastic box into something a robot can pick, read, and move at industrial speed.
#4412: When Hacking Becomes Disclosure: The Suno Breach
A hacker stole Suno's internal training code and handed it to a journalist. Was it a crime or whistleblowing?
#4268: When AI Trains on AI: The Model Collapse Problem
What happens when AI trains only on AI-generated content? The answer is model collapse — and it's already happening.
#4169: Change Detection AI for Construction Sites
How AI catches what experts miss in construction site photos — from Siamese networks to transformer models.
#4106: Embedding Models vs LLMs: What Actually Connects?
Can you mix any embedding model with any LLM? And why are new embedding models still dropping if they're "solved"?
#3874: How to Tag 4000 Episodes Without Losing Your Mind
Why tagging breaks at scale, and how a two-stage AI pipeline fixes it for good.
#3852: The Hidden Workforce Behind AI's Intelligence
Behind every "intelligent" AI system are millions of workers in Kenya, India, and the Philippines doing repetitive tasks for poverty wages.
#3767: How LLMs Actually Learn: Stages or Slurry?
Do large language models learn grammar first, then facts? The honest answer is messier and more fascinating.
#3713: How a Real PI Manages Thousands of Photos
Phone camera rolls don't cut it. Here's how real PIs organize, tag, and store thousands of evidence photos per month.
#3649: When Wikipedia Feels Less Reliable Than AI
One reader explains why he now trusts AI more than Wikipedia on contested topics like Israel and Zionism.
#3596: Why an AI Model Kept Calling Itself Sonnet 4.6
When a Chinese model insists it's "Sonnet 4.6," is it theft, sloppy training, or something stranger?
#3595: How DeepSeek Feels More Open Than Western AI
Why Chinese AI models sometimes feel less censored on American political topics than American models do.
#3393: Ireland's Moral Cost Accounting Problem
When moral stances meet economic reality—examining Ireland's pattern of avoiding costs for its stated principles.
#3261: The Hidden Zoo of Drug Testing
Mice dominate headlines, but drug validation relies on dogs, pigs, ferrets, and macaques — each chosen for a specific human system.
#3170: Pharmacokinetics vs Neural Nets: Two Meanings of "Model
Two things called "models" that work completely differently — and why the confusion matters for patient safety.
#3157: Opus 4.8: What Actually Changed Under the Hood
Anthropic dropped Opus 4.8 with no fanfare. New training data, faster inference, and smarter refusals — here's what changed.
#2982: Why Your TTS Model Nails "Shabbat" but Not "Keren Hishtalmut
Why multilingual TTS models handle loanwords but fail at niche vocabulary — and what you can do about it.
#2854: What Our Analytics Dashboard Reveals About Hidden Audiences
Hilbert uncovers suspicious spikes in podcast data. Are they covert ops or just university students?
#2842: Fixing Your New Apartment: The Israeli Tool Kit
The eight essential tools and hardware every Israeli apartment needs — with Hebrew names and where to buy them.
#2836: Can ANC Handle Real City Noise Now?
ANC has gotten smarter, not just stronger. Can it handle construction noise? And what about cancelling noise in a whole room?