My Weird Prompts

Daniel has been fine-tuning NVIDIA's 600M-parameter Parakeet on his own read-aloud recordings, English with Hebrew mixed in. Each run costs about eighty cents and twenty-five minutes on a single L40S, so retraining from the stock model is cheap — but that raises a tangle of questions: what's actually inside a 6.9 GB training checkpoint versus a 2.5 GB finished weights file, why resuming a finished...

A listener has 150 hand-labelled dictated prompts and wants to know if that's a start or a third of what he needs — and his architect wife has the same question about photographs of buildings. This episode works through the real answer: count examples per label, not in total; treat vendor minimums as proof-of-concept ceilings rather than floors; and measure your own learning curve instead of trust...

Mixture of experts is sold as the trick that makes giant models cheap. It doesn't — it makes them bigger. This episode unpacks what the active parameter count actually measures, why a trillion-parameter checkpoint can run like a 3.6B model, and why the router fires a committee of experts per token instead of one specialist. We walk through the design paper arguing you should maximize total paramet...

Dictation has quietly reshaped how many of us work, but starting and stopping it well is its own engineering problem. This episode digs into the physical controls that make voice input dependable: desktop pedals like the PCsensor and Stream Deck Pedal, handheld buttons, and wearables, plus the assistive-technology history that made all of it possible. Then it moves to Android, where open-source ge...

Off-the-shelf "Humanizer" models flatten everyone into the same average voice — the method works, but the corpus is the problem. This episode walks through building a personal style adapter: manufacturing flat/neutral training pairs from your own posts, choosing a base model over an instruct one, curating hard, and knowing when it's overfit. Then we cover the practical part: running the adapter wi...

Daniel fine-tuned a speech model on his own mixed Hebrew-and-English recordings for the price of a coffee. So what happens if you point that same move at an Opus 5.5-class frontier model — raw, full precision, no practical reason at all? This episode walks through the arithmetic: the sixteen bytes per parameter that Adam demands, why a one-trillion-parameter model needs roughly 200 H100s just to h...

Every lab has a middle rung — Anthropic's Sonnet, OpenAI's Sol, Google's Flash, Mistral's Medium — but there's no agreed name for it, and the price lists have quietly converged into a ladder. This episode digs into why the tier most people use all day is getting squeezed by cheap open-weight models below and flagship promotions above. It's a conversation about cost per finished task rather than co...

Two Chatterbox hosts are talking to you right now, and this audio carries an inaudible watermark — can you hear it? That is the design goal, and it is also the problem. This episode digs into the real state of synthetic speech detection: listeners judge voice clones as human about as often as they judge real people, while off-the-shelf AI voices still give themselves away. Detectors that score nea...

There's a shelf of uncensored models, but it's stocked with fine-tunes of other labs' bases, not from-scratch builds. This episode surveys what actually exists — Nous Research's Hermes with its written-down guardrail spec, Perplexity's re-alignment of DeepSeek R1 across hundreds of censored topics, Venice's privacy-first model shop, and xAI's Grok, the one big lab that markets against the agreeabl...

What happens when researchers need to measure machine intelligence and end up grading puns, emoji, and a joke with the punchline surgically removed? This episode counts down ten of the strangest AI evaluations ever published — from BIG-bench's "ruin names" task to a high schooler's Minecraft model leaderboard, from a canary string that detects cheating to Claude stuck in Mount Moon for eighty hour...

Hugging Face hosts nearly three million model repos, and 85.6% have under 200 downloads. This episode wanders the long tail of AI: auto-generated "woolly sedate snail" swarms, a hotdog-not-hotdog classifier, a Trump-tweet device detector, and a Tiny Pirate that beats the Yoda. Then it goes back further — Janelle Shane's paint-name network producing "Stanky Bean" and "Turdly," a guinea pig rescue n...

Nine months after our quadratic bottleneck episode, the post-transformer world has actually shipped — just not in pure form. This episode is a census of the hybrid models now in the wild: Qwen 3.5's three-to-one Gated DeltaNet blocks, Nvidia's Nemotron 3, Kimi Linear's 75% cache cut, IBM's Granite 4, Ant Group's trillion-parameter Ling and Ring, and Falcon H1's side-by-side design. Then the counte...

There's no office anywhere that stamps an election genuine, yet most people assume one exists. This episode digs into the numbers behind sham elections — V-Dem counts 92 autocracies, 57 of them "electoral autocracies" holding multiparty votes that aren't real contests — and asks why the three big democracy indices disagree so often. It traces the actual minimum bar in international law, from Artic...

Elections are the one moment a government publishes its intentions in advance — which is exactly why foreign services, embassies, and trading desks want the unpublished version. This episode sorts the documented record from the movie version: leaked CIA tasking orders on the 2012 French presidential race, the very open apparatus of accredited embassy observers in Kenya and Colombia, the press acco...

Israel elects 120 seats by pure proportional representation, nobody wins a majority, and a party at 3.25% gets nothing at all. So what would it take to forecast an Israeli election properly — with real staff and a real budget? This episode walks through the staffing plan: running the actual Bader-Ofer rules in every simulated draw, cleaning twenty contradictory polls with a Bayesian model, using b...

You fine-tune a Whisper-style model or Parakeet on your own voice, run it twice, and get two word error rates several points apart. Neither number is a bug. This episode pulls apart the three layers of ASR evaluation — the scorer, the normaliser, and the harness — and shows why almost every WER you've ever read is really a statement about someone's regular expressions. We walk through jiwer and th...

Monte Carlo methods turn chance into answers, and they were born from a card game at Los Alamos in 1946. This episode traces the method from Stanislaw Ulam's solitaire problem through John von Neumann's neutron diffusion work, the ENIAC calculation, and Buffon's needle — then digs into the practical question everyone hits: how many runs do you actually need? The square-root rule, pilot runs, seque...

Colab is the default for every fine-tuning tutorial, but free has fine print: fluctuating GPUs, 12-hour runtime caps, and idle timeouts. This episode walks through five real options — Colab, Kaggle, local Jupyter, marimo, and pay-per-use GPU notebooks like Modal and Lightning — with honest tradeoffs for each. Then it tackles the question everyone almost doesn't ask: when you share a notebook so ot...

Five days before the vote, with polls putting Netanyahu's bloc anywhere from the mid-forties to sixty-two depending on who you read, this episode throws out the forecasting models and does it by hand. Two hosts build complete 120-seat tables independently — Corn nudging the mainstream average, Herman betting the pollsters are missing the same voters they missed in 2022 — and then argue over every...

Two Claude Code sessions on the same machine can now talk to each other natively — but the design has sharp edges. Names are the address, messages carry text and nothing else, and a "sent" receipt proves less than you'd think. We walk through ListAgents discovery, the name flag, the notify-when-idle subscription that replaces polling, and the file-on-disk pattern that makes handoffs trustworthy. E...

What happens when you freeze the evidence, hand two decision models the same 39 questions about Israel's election, and then add a table of Kalshi and Polymarket prices? Clef (open weights, Qwen backbone) and Jev (closed, API-only) both picked Likud and Netanyahu — until the market section went in, when both flipped to Yashar and Eisenkot, overshooting the market's own prices. The whole run cost on...

Cloudflare's Clef is a 27-billion-parameter model with a vow of silence: it reads text, images, and video, and returns nothing but a table of probabilities. Underneath, it's a post-trained Qwen 3.8 with a frozen backbone and a small routing head — and that architecture turns out to matter a lot. This episode opens the hood on Clef, Clef-flash, and Clef-omni, compares them with Jev, TypeSafe, and O...

Israeli election headlines report tidy seat counts, but those integers hide enormous uncertainty. SekerKolshehu, a small open source project by developer Isgav Saranga, takes a poll's vote shares, draws 100,000 plausible variations, and runs each one through the actual Israeli allocation rules — threshold, surplus agreements and all. The result is a distribution: average seats, a 90% range, and th...

Daniel wants to run a proper LLM-driven forecast on Israel's election on the twenty-seventh — not another loose script. So what do experiment platforms actually do? This episode sorts the vocabulary into three families: product testing tools like Statsig and Eppo, bioinformatics workflow managers like Snakemake and Nextflow, and run trackers like MLflow, Weights and Biases and ClearML. Then it tur...

Daniel has fine-tuned Qwen3 TTS 1.7B into a voice clone called Meorav, and it currently takes 2.6 seconds of compute for every second of audio. This episode digs into how to actually benchmark speech-model inference: what real-time factor means (and how people invert it), why the first run is always slow, how to measure peak allocated versus peak reserved memory, and why shrinking weights to 8 bit...