#fine-tuning
54 episodes
#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?
#3406: LoRA Isn’t Just for Image Generation
LoRA lets you fine-tune an LLM’s behavior with a 50MB file. Here’s how it works and why it matters.
#3283: Fine-Tuning DeepSeek for One Podcast
Can a purpose-specific fine-tune fix a model's stubborn writing tics? We explore the practical engineering behind it.
#3171: How to Break an LLM's Bad Verbal Habits
Blacklists fail and regex inverts meaning. Here's what actually works to clean up AI writing tics.
#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.
#2693: When AI Ignores Your Style Guide
Why your AI ignores formatting instructions and how to fix it with pipeline architecture, not model swaps.
#2664: Can You Trust an LLM's Raw Knowledge?
Why pre-trained knowledge isn't reliable for facts — and what actually makes models useful.
#2650: How to Catch an LLM's Bad Writing Habits
A practical guide to analyzing podcast transcripts for repetitive language and dialogue patterns — from Python word counts to embedding clustering.
#2517: How Unsloth Makes LLM Fine-Tuning 2x Faster
Unsloth cuts memory usage by 50-70% and speeds up training 2.2x for models like Llama 3 and Mistral.
#2516: Overfitting Is Not a Binary Condition
Overfitting isn't binary. Learn the real triggers, the bias-variance tradeoff, and modern techniques to prevent it.
#2495: How to Bake Personality Into an LLM in 15 Minutes
Fine-tune a model's personality with ~300 examples and a consumer GPU. SFT + DPO explained.
#2470: Where Intelligence Should Live in Your Pipeline
When should you fine-tune a tiny model for prompt enhancement instead of prompting a large one? The answer depends on latency, precision, and domain.
#2426: Why DeepSeek V4's Prose Feels More Vivid Than Claude or GPT
A million-token context window at 2% the KV-cache cost — and prose that actually breathes. Here's what makes V4 different.
#2334: How AI Flattens Your Voice in Emails
Why AI-generated emails feel impersonal and how to reclaim your authentic voice in professional communication.
#2316: Who’s Building AI’s Next Training Data?
How boutique dataset firms are reshaping AI training, from rights-cleared content to domain-specific precision.
#2315: How to Update AI Models Without Starting Over
Exploring the challenge of updating AI models with new knowledge without costly full retraining.
#2307: Inside Frontier LLM Training: Stages, Costs, and Checkpoints
Discover the multi-stage process of training frontier large language models, from pretraining to post-training, and why checkpoints are the key to ...
#2196: The Invisible Workforce Behind AI
Annotation is the invisible foundation of AI—and a $17B industry by 2030. Here's what dataset curators actually need to know about the tools, platf...
#2187: Why Claude Writes Like a Person (and Gemini Doesn't)
Claude produces prose that sounds human. Gemini reads like Wikipedia. The difference isn't capability—it's how they were trained to think about wri...
#2177: Skip Fine-Tuning: Shape LLMs With Alignment Alone
Can you build a personalized LLM by skipping traditional fine-tuning and using only post-training alignment methods like DPO and GRPO? We break dow...
#2067: MoE vs. Dense: The VRAM Nightmare
MoE models promise giant brains on a budget, but why are engineers fleeing back to dense transformers? The answer is memory.
#1907: Why We Still Fine-Tune in 2026
Despite million-token context windows, fine-tuning remains essential. Here’s why behavior, not just facts, matters.