My Weird Prompts

Google's Gemini audio docs promise emotion labels with timestamps. A benchmark called LISTEN tested whether the models actually use the sound of a voice or just transcribe and guess — and the numbers aren't flattering. This episode digs into the gap between what speech models claim to detect and what they really hear, what specialist open models like emotion2vec+ and SenseVoice offer instead, and...

Video-capable AI models can describe a room in detail but stumble when asked whether someone is moving slowly, which way they turned, or what happened in what order. Benchmarks like TOMATO, TimeBlind, FAVOR-Bench, and MotionBench all point the same direction: models are good at what's in the frame and shaky about what happens between frames. This episode digs into why — default sampling of one fra...

Can a single microphone tell where you're standing, where you're from, or which way you're walking? This episode sorts the real capabilities of audio AI from the claims that don't hold up. From language ID models that nail 107 languages but stumble on regional accents, to bioacoustic models that learned birds and turned out to work on whales, to industrial systems that detect failing machinery wit...

Your model is flattering users and you can't say why. This episode walks through the process a well-funded lab would actually follow: build an eval for the exact failure, bisect between the last good and first bad checkpoint, ablate the suspects, then reach for the research-grade tools — influence functions, persona vectors, sparse autoencoders. Along the way we look at what OpenAI's GPT-4o sycoph...

Ask an AI to warm up your photo and one of two things happens: it hands you a recipe of exact slider values you can read and adjust, or it repaints every pixel and hopes you don't notice the logo grew a letter. This episode digs into the recipe camp — JarvisArt's agent over 200+ Lightroom tools, PhotoArtAgent's render-look-revise loop, and VeraRetouch's fully differentiable renderer — and the rege...

Restoration tools have quietly become generative — they don't subtract noise, they rebuild the speech they think was underneath it. That works beautifully until it invents a phoneme you never said. This episode maps what actually ships in 2026: DeepFilterNet 3, ClearerVoice-Studio, Adobe Podcast's stem-splitting Enhance Speech update, ElevenLabs' voice isolator, and NVIDIA Maxine on the mature sid...

Daniel wanted an inference manager the way he already has a secrets manager, and what he described is basically a LiteLLM-style proxy. The failover machinery is genuinely built out: cooldowns, health probes, latency-based and least-busy routing. The gap is memory. Ollama will tell you a model's VRAM footprint, not the headroom left, so "is this box full?" isn't a question you can put to it — you w...

Most explanations of how a language model works show you a diagram. This one hands you a coat and walks you through the building. Starting from the token "is" in the sentence "The capital of the state containing Dallas is," the episode follows a single vector through eighty layers of a transformer — the Look Back stops where attention pulls in what earlier words are offering, the memory rooms wher...

If you dictate in English but live in Jerusalem, stock speech recognizers mangle words like arnona and makolet. Hebrish is a not-for-profit project that fixes this with two new models, Jachnun and Sabich, both built on NVIDIA's Parakeet. We walk through every evaluation, from term recall to never-seen Hebrew words, and explain why Sabich is the one the apps recommend. Plus: the Shakshuka streaming...

Daniel wants his agent-driven browser off his laptop and onto the home server, but the server has never logged into anything. Chrome closed the obvious path in March 2025 when version 136 started ignoring the remote debugging port on your default profile — infostealers had been using it to lift sessions. So what are the real options? This episode walks through the extension cookie API and CookieCl...

Every tutorial starts with an empty machine and a playbook you write first. But what if your Ubuntu box grew for years — packages, containers, midnight config edits — and keeps drifting? This episode works through Daniel's question: how do you turn a living machine into a recipe that stays true? The answer splits rebuild from restore, separates config from data from secrets, and accepts that no to...

Everyone talks about batch API discounts, but nobody talks about the babysitting. When you're running a speech fine-tune across thousands of clips or a personal classifier, the real problem isn't the price — it's knowing which job finished, which one expired, and where the results went. We walk through the actual state of batch monitoring: which providers now offer webhooks (OpenAI, Gemini, OpenRo...

What happens when you stop letting one model voice every character? This week the show rebuilt itself on Google DeepMind's Concordia framework — a separate agent for each host, each with its own context window, plus a showrunner deciding who speaks next. The result: real questions instead of filler, but a cost six times the normal price. We walk through what got better, what got worse, every rule...

Daniel has a working voice model that flips between English and Hebrew, shipping as a standalone Android dictation app. The obvious next move — fork an open source keyboard and wire the engine in — comes with a catch: the moment you fork, you inherit the upstream's entire future. This episode works through the patch-based alternative: keep your integration as a tiny replayable patch, let automatio...

You dictate into a voice prompt, watch a word appear in gray, and the moment you release the key it quietly becomes a different word. That flicker is not a bug — it is the seam between interim and final speech recognition, and the vocabulary for it exists across Google, Amazon, Microsoft, and Deepgram under different names. This episode unpacks what "dimmed until finalized" actually promises, why...

Daniel records himself reading prompts aloud to fine-tune a speech recognition model, and every brand name, Hebrew term, or Yiddish word he's only seen written down becomes a wall: he doesn't know how to say it, so whatever he says becomes the training label. The obvious fix is a pipeline — word in, predicted phonetics, a voice says it back. The pipeline exists, in parts. Grapheme-to-phoneme (G2P)...

Daniel wants his on-phone Parakeet model to write "FBI" instead of "F B I" and "fifty NIS" instead of "fifty shekels." That sounds like a rule, but speech models don't obey rules — they nudge probabilities. This episode walks a ladder from cheapest to heaviest: rule-based inverse text normalization with weighted finite state transducers, decode-time phrase boosting, shallow fusion with a KenLM lan...

Daniel is fine-tuning Parakeet v2, a 600-million-parameter speech model, on his own voice — English and Hebrew mixed in the same sentence. His training data is built as small, atomic datasets, one per phenomenon: currency, brand names, acronyms, self-corrections, code-switching. But when one module writes "$20" and another writes "twenty dollars," does the model pick a winner, split the difference...

Daniel fine-tunes Whisper-style models and keeps finding hundreds of tiny jobs hiding inside the big "ASR" label — like teaching a model that "F Y I" should be written out as "for your information." It turns out Sproat and five co-authors named this exact decision back in 2001, in a taxonomy built for text-to-speech front ends. The catch: their arrow points the other way. This episode digs into wh...

Fine-tuning a speech model on your own voice means recording hundreds of sentences a session, on a phone, twice a day, with no studio. This episode walks through the practical setup: why buying a better microphone actually works against you, which Bluetooth page turners send the right keycodes (and which quietly change your volume instead), where the microphone actually sits on a phone, and how to...

Göbekli Tepe was built before farming. Rome ran water-powered factories. The Greeks cut dozens of bronze gears by hand. So why did the industrial takeoff wait until around 1760 in Britain? This episode pulls the camera back from individual tools to ask what actually kept the climb going — energy, skills, markets, and reasons to replace the existing method, all reinforcing each other. Along the way...

Acupuncture works better than no treatment for chronic pain — but only slightly better than fake needling, and the gap is small enough that researchers still argue about what it means. This episode digs into the Vickers individual patient data analysis of 39 trials and 20,000+ patients, the Cochrane review's more cautious read on low back pain, and why sham design is the hardest problem in the fie...

Paying someone to label every example feels reassuring until you ask who checked the labels. This episode untangles the mess of strong supervision, weak supervision, self-supervision, and distillation — separating how a training target is produced from how trustworthy it actually is. From labeling functions and data programming to pseudo-labeling and the Llama 3.1 distillation work, we dig into wh...

A new category of "decision models" has gone from one launch to a crowded field in about a month. Clef, Jev, and OpenAI's Decisions API take a state plus a schema of typed questions and return choices, booleans, and scores with probabilities — no free-form text to generate and parse. The pitch is speed and cost: 70–500ms latencies and output priced at zero. But usable answers aren't the same as ac...

A banker published the first one-time pad in 1882 and told readers to erase the numbers after use. Shannon proved in 1949 that it's mathematically unbreakable. So why has almost every attempt to actually use one failed? This episode follows the pad's real lineage — Vernam's repeating tape, the Soviet duplicate pages that cracked Venona, SIGSALY's fifty tons of vinyl key records, the Moscow-Washing...