My Weird Prompts

Every token an LLM generates has to attend to every token before it, and the KV cache is what keeps that from being recomputed from scratch — while quietly becoming the thing that caps your context length, your batch size, and your latency. This episode walks through what's actually stored (keys and values, never queries), why prefill is compute-bound and decode is memory-bandwidth-bound, and the...

Daniel fine-tuned NVIDIA Parakeet V2 on two hours of LLM-generated sentences recorded on his phone, and Hebrew-English code-switching recognition jumped from roughly 20% to 60%. Now he wants to record fifteen minutes every morning for weeks — so the real question is where the fine-tuning curve flattens. The literature disagrees: a 2023 Singapore study on Mandarin-English code-switching suggests sa...

Daniel fine-tuned a speech recognition model on two hours of his own voice and hit 60% accuracy on code-switched Hebrew, up from under 20% across every leading model he tested. The catch: Parakeet V2 is English-only, and Hebrew isn't in its 25-language sibling either. This episode digs into what a fine-tune actually captures — new tokens versus new accents, why control sentences matter, and whethe...

Your JSON messages don't reach the GPU as JSON. Somewhere in between, a Jinja chat template flattens them into a string of reserved special tokens — the exact format each model family was trained on. Llama 3 brackets roles with start_header and end_header; OpenAI's ChatML embedded the role in a single opening token; gpt-oss uses Harmony with separate stop tokens for "done talking" and "calling a t...

We spend over 95% of our lives indoors, yet almost nobody asks what the building is doing to them. This episode digs into the neuroarchitecture research — controlled studies using EEG, cortisol, and heart rate variability — to see what traditional and modernist architecture actually do to the human nervous system. From a 73% cortisol spike in windowless rooms to the fractal dimensions of Piranesi'...

When a city rezones land, it creates value out of nothing — so who should capture it, and who decides what gets built? This episode digs into Hannah's question about developers' proper place in shaping a city, starting from her architecture school project in Jerusalem's Ir Ganim neighborhood. A look at Toronto's politically negotiated Section 37 deals, Vancouver's staff-driven Community Amenity Co...

Jerusalem's statutory land-use plan, the Taba, is a legal object that no human mind has ever fully contained. It carries a British mandatory skeleton, Ottoman land law in the basement, and 171 amendments stacked on top. On this episode, we trace how a single Rehavia lot can be governed by four overlapping plans, why 60% of local plans between 2016 and 2022 were authorized as spot-planning, and wha...

Israel's average time to a building permit hit 319 days in 2019 — roughly three times the Western average of 114 days. But where does that year actually go? This episode walks the full pipeline for private houses, apartment buildings, and high-rises: the five strictly sequential approval stages, the relief track that costs architects a year, and the automatic-rejection rule that quietly resets the...

Seventy billion parameters gets repeated on every earnings call, but almost nobody who says it could tell you what one parameter is. This episode builds the artificial neuron from the ground up — weights, bias, activation — then works through the math objects underneath it: scalars, vectors, matrices, and tensors. Along the way, the conversation lands on why stacked linear layers collapse into a s...

What actually happens when someone wires you a million dollars? The answer splits in two: the number is trivial, but the plumbing is a whole industry. This episode walks through three transfers — a hundred dollars, a million, and a billion — and shows where the paths diverge. Along the way: why money never physically moves, how CHIPS netting meant one large payment never had to move as a large pay...

Encoder-only, encoder-decoder, decoder-only — three shapes, one skeleton. This episode pulls apart the architectural fork at the center of modern language models: how attention masking and training objectives separate BERT from GPT, why the decoder-only design swallowed the field, and why encoder-only models still quietly run classification and retrieval at a fraction of the cost. Along the way: t...

A listener asked a deceptively simple question: what actually happens during an LLM forward pass? This episode walks through the whole thing — input tensors, linear layers, why nonlinearities are the only reason depth means anything, and how logits become a prediction. Then it gets to the part that trips people up: a classifier produces its answer in a single forward pass, but an autoregressive mo...

The model scores; the engine chooses. That split is the whole story of autoregressive decoding, and this episode walks through it token by token. Starting from a warm KV cache, we trace a single decode iteration — the forward pass, the query-key-value projections, the logits row that actually gets used, and the sampling step that feeds the next token back in. Along the way we work a concrete examp...

A single request to an LLM isn't one computation—it's a stack of stages wearing a trench coat. This episode walks the full pipeline: flattening a chat transcript into one flat document with special role tokens, tokenizing text into IDs, the massively parallel prefill pass that produces your first token, and the strictly sequential decode loop that produces every token after. Along the way: why for...

Wearable tech promised to change how we live. It gave us step counters, a haptic buzz when someone likes your photo, and a robot that feeds you tomatoes mid-marathon. This episode tours ten of the strangest wearables ever manufactured — not failures, exactly, but objects built with complete sincerity that look absurd on a human body. From Necomimi's brainwave-controlled cat ears (a real product, t...

Vertex AI was Google's 2021 bet on unifying ML tooling — pipelines, feature stores, model registry — into one platform. In April 2026, Google renamed it the Gemini Enterprise Agent Platform and said all Vertex services would be delivered exclusively through it. So what actually changed? This episode traces the lineage from AutoML and AI Platform through the 2021 launch to today's agent-first rebra...

A coding model confidently hands you a deprecated SDK. Claude denies a newer Claude exists. These aren't hallucinations — they're a knowledge cutoff problem wearing a costume. This episode digs into the "Know Thyself" study from the University of Chicago, where ten frontier models guessed who wrote a piece of text and landed at essentially random accuracy. We trace the two failures people keep lum...

Enterprise AI inference isn't bought like seats or like cloud compute — it's a hybrid of both, and the gap between the published rate card and the signed contract is where the entire market lives. This episode walks through what actually happens once you cross the threshold: the commit floors that get a provider to pick up the phone, the discount curves that disagree by twenty points between bench...

Everyone cites GDPR and the EU AI Act as the reason AI inference has to run in-country. Neither text names a location. This episode untangles the three layers that get used interchangeably — data residency, inference residency, and data sovereignty — and shows why a Frankfurt data center owned by an American cloud provider can satisfy residency and fail sovereignty at the same time. We walk throug...

Thirty-five of thirty-six OECD countries now use AI somewhere in government — but the interesting part is what sits behind the chat box: whose model, on whose hardware, bought through whose procurement process, and governed by whom. This episode traces the two opposite archetypes, Singapore's open-weight Pair running in a Government-on-Commercial-Cloud environment and the UK's commercial-first Cop...

Can a Fortune 500 company buy unlimited AI inference at any price? The answer is no — and the reason why is more interesting than the answer. This episode walks through the published rate-limit tiers at Anthropic and OpenAI, the reserved-capacity arrangements above them, and the one vendor document that admits the real constraint out loud. Along the way: why "uncapped scale" still returns a 429, w...

Enterprise spending on frontier model APIs hit $12.5 billion — yet no lab publishes who actually pays what. This episode digs into the brackets, from self-service startups spending hundreds a year to multinationals with nine-figure commitments, and explains why the scariest numbers in circulation are compute purchases, not customer bills. Then it turns to what generates the bill in the first place...

Claude's per-token rate is identical whether you go direct to Anthropic, through Amazon Bedrock, or through Claude Platform on AWS. So the entire buying decision lives somewhere other than price. This episode walks through the three AWS doors into Claude, why procurement mechanics — not discounts — drive the choice, and how a $100 CCU can retire an existing cloud commitment. Then it gets to the fo...

Open weights are free, so why does the same DeepSeek model cost four and a half times more at one host than another? This episode breaks down what specialist inference providers — Fireworks, Together, Baseten, DeepInfra, Novita, Hyperbolic, Nebius, Groq, Cerebras — actually sell, and why "serverless GPU" hides three completely different products. We run the real arithmetic on token APIs versus man...

My Weird Prompts just crossed 6,000 episodes — and the human behind it, Daniel Rosehill, is stepping out from behind the pipeline to talk about how it actually runs. From the TTS struggles that nearly derailed the whole thing to the shift to DeepSeek 4.1 Flash as the script agent, this is a rare look at the machinery and the person operating it. It's also a reflection on what it means when a daily...