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.
#6169: What It's Like to Be a Token Inside an LLM
A token's-eye view of an LLM's forward pass — no cone, no squeezing, just a hallway where every layer scribbles on you.
#6168: Hebrish: Hebrew-Aware Dictation That Actually Works
Two new speech models fix English-Hebrew code-switching dictation — and the whole training run cost about fifty cents of GPU time.
#6167: Streaming Browser Cookies to a Headless Server
Daniel wants his logged-in browser on the home server. Chrome bolted the front door — so what's left?
#6165: Watching 20 Batch Jobs at Once
Batch APIs give you a discount, then hand you a babysitting job. Here's what actually watches a batch run — and what doesn't.
#6164: We Gave Each Host Its Own AI Agent
We split our two hosts into separate AI agents with private context windows. The bill came to 78 cents an episode.
#6163: Forking a Keyboard for One Voice Feature
Daniel built a Hebrew-English dictation engine. Now he wants to fuse it into someone else's keyboard — without inheriting its entire future.
#6162: Why Dictated Words Rewrite Themselves
Hold-to-talk dictation shows dimmed text that silently changes when you release the key. Here's the architecture behind that rewrite.
#6161: Turning Words Into Sound: G2P, IPA, and the Hebrew Vowel Problem
Text-to-phonetics tools exist, but the neural ones guess wrong on exactly the words you care about — names, brands, and Hebrew.
#6160: Where Do Tiny Speech Rules Live? A Ladder From Regex to LoRA
A regex notebook, a WFST grammar, or a fine-tune? Where do tiny speech-rewrite rules actually belong in a 600M-parameter model?
#6159: When a Speech Model Can't Decide How to Write Dollars
One dataset says "$20," another says "twenty dollars." Fine-tuning a speech model means finding out which convention wins — and why it matters.
#6158: Naming the ASR Sub-Tasks Nobody Mapped
A 2001 linguistics paper already named Daniel's problem — but it points the other way. So what do you call a fine-tune set that expands "F Y I" to ...
#6157: Where to Sit When You're Fine-Tuning Your Own Voice
Your phone's built-in mic is the point — the setup around it is what ruins the takes. Rooms, remotes, and the app setting that deletes your data.
#6154: Who Checks the Labels? Weak vs Strong Supervision
Human labels, rule-based labels, or a teacher model writing the exam — what actually makes supervision "strong," and who verifies it?
#6153: Decision Models: Typed Answers in One Forward Pass
Clef, Jev, and OpenAI's Decisions API return typed answers with probabilities in a single forward pass — no generated text to parse.
#6149: 4 Hard Fails in 84 Answers: Testing AI Bias on Israel-Palestine
A 19-prompt, 6-model eval tested whether AI models push back evenly on false claims about Israel and Palestine. One model failed four times.
#6148: The Tiny Hebrew Models Behind the Show's Voice
Dicta's 300M-parameter nikudizer, Phonikud's adaptors, and the ktiv male model hiding inside a vowel adder.
#6147: Hebrew AI: Ears, Mouth, and Brain
Open Hebrew speech recognition beats the paid services. Hebrew text-to-speech is a different story — and the language models are a third.
#6146: Grounding Is Not Solved: DIY vs Gateway vs Native Search
Bolt-on search, gateway search, or native grounding? The plumbing didn't vanish — it just moved to someone else's bill.
#6145: Do LLMs Nod Along With Conspiracy Theories?
A 98-prompt eval asks one question: when a believer brings a conspiracy theory, does the model push back, or help them dig?
#6143: Propaganda Bait: Testing 7 Models in 5 Languages
Five prompts, seven models, five languages — a first look at a benchmark that doesn't exist yet, testing what LLMs do with government talking points.