Artificial Intelligence
Episodes whose main subject is AI itself: building with, training, evaluating, or debating AI systems. AI used only as a tool for another topic goes to that topic.
#6038: The Model You Pay For Isn't the Best Deal
Eight Chinese labs, two tiers each, and a pricing spreadsheet that says your default model is mid-pack.
#6037: Gemini 4 Argon: Google's Million-Token Flagship
A close read of Gemini 4 Argon's model card — a million-token window, thinking levels, implicit-only caching, and a lot of blanks.
#6036: GPT-6 Astra: A Million Tokens, No Fine-Tuning
A million-token window, ten built-in tools, a 272K pricing cliff — and no fine-tuning. What GPT-6 Astra actually offers builders.
#6035: Who Else Is Making AI Podcasts? A Survey
A survey of AI-hosted podcasts: one full match, a 730-episode run, a home NAS rig, and a graveyard of good tooling.
#6034: Claude Opus 5.5: What Anthropic Actually Shipped
Anthropic's new flagship is cheaper than the model it replaces — and won't let you turn thinking off.
#6033: Progressive Skills Discovery: When Bigger Context Windows Aren't Enough
Bigger context windows were supposed to make skill discovery moot. Here's why the listing budget still bites at scale.
#6032: Blending Datasets vs Merging Weights for Fine-Tuning
Merging weights happens in parameter space. Blending datasets happens in data space — and the messy part is worth it.
#6030: Who Do You Actually Hire to Train a Model?
Research engineers and data engineers are diverging. Here's who you actually hire to train a model — and why the org chart keeps changing.
#6029: One Model, Two Modes: Does AI Consolidation Hold Up?
OpenAI hides a router, DeepSeek merged two modes, Anthropic removed the knob. Is the model landscape really consolidating?
#6028: ASR vs Multimodal LLMs: What Actually Won
Dedicated ASR plus a cleanup model is quietly beating end-to-end dictation on the hardware you already own.
#6027: Can You Snap Two Fine-Tunes Together Like LEGO?
Two fine-tunes, one base model, and a question from Jerusalem: can you merge Hebrew code-switching with technical vocab — without retraining?
#6025: When Should Claude Code Start a New Session?
Long context windows don't fix stale context — they hide it. Why task boundaries, not token counts, are the real problem.
#6024: What's Actually Inside the KV Cache?
890 bytes per token: why the KV cache is both what makes LLMs practical and what makes long context expensive.
#6023: How Much Voice Data Is Enough?
Two hours of phone recordings took Hebrew code-switching from 20% to 60%. But when does more data stop mattering?
#6022: Fine-Tuning for Code-Switched Hebrew on 2 Hours of Audio
Two hours of one man's voice beat models trained on 120,000 hours — but only for code-switched Hebrew.
#6021: Why Chat Templates Break Models Silently
Every chat API flattens your messages into special tokens. Get one delimiter wrong and the model just gets quietly dumber.
#6016: What 70 Billion Parameters Actually Means
A single artificial neuron, built from scratch — then the hard part: does that neuron analogy still describe a real transformer?
#6014: Why Every AI Model Is Secretly a Decoder
Encoder, decoder, or both? The three transformer shapes, why one won, and why BERT still refuses to die.
#6013: What Happens Inside One LLM Forward Pass
One trip through the network, layer by layer — and why a classifier needs one pass while an LLM needs hundreds.
#6012: Inside the Decode Loop: How LLMs Generate One Token at a Time
The model doesn't choose tokens or decide when to stop — the engine does. A step-by-step tour of one decode iteration.