Training Data
322 episodes · Page 5 of 14
#5946: Inside Humanity's Last Exam: 2,500 Questions AI Can't Ace
2,500 questions, ~1,000 experts, and a top score that jumped from 2.7% to 54.8% in 18 months. What is HLE actually measuring?
#5945: How a Reward Model Actually Trains an LLM
One number, no gradients, four models in the room. What a reward model actually does inside the RLHF loop.
#5944: Keeping the Rollout Engine Fed: Inside RLHF Pipelines
Most of an RLHF run is spent waiting on generation, not gradients. Here's how TRL, OpenRLHF, and verl actually differ.
#5937: The Fine-Tuning Career: Titles, Skills, and What It Pays
The job above the labeling layer exists — it just doesn't have one name. What it takes, what it pays, and how to sell it.
#5936: Training a Model When You Don't Know the Labels Yet
Daniel wants to extract adapter specs from his shelf — but he can't name the labels until he's already annotating.
#5922: Teaching a Model Your Voice: SFT vs DPO
Three ways to teach a model your writing voice — and the evidence says one of them quietly makes things worse.
#5868: Fine-Tuning Parakeet for Um-Free, Hebrish Speech
Parakeet runs great on a phone — but fine-tuning it happens on a GPU. Here's the actual recipe for dropping ums and catching Hebrew.
#5860: Who Can Actually Train an LLM From Scratch?
Only a handful of organizations can truly train a model from random weights. Here's why nobody can count them.
#5853: Who Actually Competes With Hugging Face?
2.96 million model repos, and 85.6% have under 200 downloads. What ModelScope, Kaggle, and Ollama actually do differently.
#5805: Freeze the Model, Move the Labels
Your classifier isn't stale — your label set moved. Why retraining is cheap, and why not retraining might be smarter.
#5771: MCP Connectors: Talk to Your Data vs. Open Data
Same protocol, two patterns: pointing agents at your own database, and pointing them at public data — with very different results.
#5743: Why Claude Keeps Saying "Prior Art
Daniel noticed "prior art" creeping into his Claude reports. Is it wrong — or just a prestige term the model learned to love?
#5626: What Multimodal Models Actually Learn From Video
Late fusion, early fusion, and vision laziness — what actually happens when models train on video, audio, and text together.
#5625: Omni and the Marble That Rolls Uphill
Google's Omni claims an intuitive grasp of gravity and momentum. Can a video model really learn physics, or just remember what physics looks like?
#5506: AI Is Way More Than Chatbots
The word "AI" quietly became a synonym for "chatbot." Here's what the full model ecosystem actually contains.
#5485: Fine-Tuning Parakeet for Hebrew and Your Own Jargon
NVIDIA's Parakeet beats Whisper on Android — but can you teach it Hebrew, or just your own jargon? Two answers, one much happier.
#5478: When AI Hands Out Your Phone Number
A chatbot gave out a stranger's real phone number. What happens when the model can't forget it?
#5447: The Models That Never Talk Back
Some models read text, score it, and return a number. No chat, no reasoning, just decisions — and they're running under every router you use.
#5443: Heads vs Layers: How Model Merging Actually Works
Heads aren't the Lego bricks of model merging — layers are. Here's what heads really do and how frankenmerges get built.
#5442: When Do LLMs Suddenly Get Good at Things?
New capabilities seem to appear out of nowhere as models scale. Whether that's real — or just how we measure — is still an open fight.
#5438: Why Linguists Left the AI Room
Large language models grew out of linguistics — so why aren't linguists in the room where they're built?
#5437: Hunting the Tells That Vanish When Named
There's no tool for finding an LLM's verbal tics — you have to build one. Here's how keyness analysis works.
#5416: Xet, Buckets, and Auto-Pulling Model Weights
Hugging Face swapped its storage backend to Xet with barely a ripple. So why can't a bucket auto-pull new upstream weights?
#5412: Editing vs. Note-Taking for AI Fine-Tunes
Hand-editing a model's output gives three training signals at once. Writing notes gives one — and a weaker one at that.