My Weird Prompts

DeepSeek V4.1 Flash writes this show's scripts — but is it still the right model to pay for? This episode walks through eight Chinese labs with flagships a Western developer can actually reach, from Alibaba's Qwen3.8 Max to Moonshot's Kimi K3, Zhipu's GLM-5.3, Xiaomi's MiMo, MiniMax, StepFun, Tencent and Ant Group. Along the way: specs nobody has reconciled, a headline price that turns out to be o...

Google's Gemini 4 Argon is positioned as a frontier reasoning model, but the model card leaves out architecture, parameter count, and licensing, and the page quotes no pricing at all. We walk through what's actually documented: a 1,048,576-token unified input/output limit, low/medium/high thinking levels, implicit caching only, and a wide tool surface including function calling, code execution, Se...

OpenAI's GPT-6 Astra ships with a 1.05 million-token context window, ten native tools including computer use and MCP, and a five-level reasoning effort dial — but also no fine-tuning, a long-context pricing cliff, and benchmark numbers that shift depending on which harness ran them. This model spotlight walks through the spec sheet, the pricing math, and the contested ARC-AGI-3 results, plus the s...

Daniel asked a simple question: who else is out there making AI-hosted podcasts, and what have they learned? The answer turned out to be stranger than expected — the format is far rarer than it looks. This episode surveys the closest matches in the world, from a two-AI-host show that ran eleven episodes and went silent, to a developer whose serverless pipeline produced 730 daily episodes over two...

Anthropic released Claude Opus 5.5 on September 22, 2026, and the reception was loud. In this Model Spotlight episode, we profile what the lab actually shipped: a cheaper price across every line than Opus 5, a hard change that means thinking mode can no longer be switched off, and a model card that conspicuously omits parameter count, architecture, and context window. We walk through the benchmark...

Progressive skills discovery is treated as solved — until your skill listing quietly drops descriptions and skills stop firing. This episode digs into how Anthropic's three-tier progressive disclosure actually works, why Claude Code truncates descriptions for your least-used skills first, and how Codex handles the same problem with different numbers. We survey the open-source landscape for auto-sy...

If your use case is narrow enough to find training data for, blending multiple datasets into one fine-tune run can beat merging finished checkpoints. This episode walks through the full pipeline: why uniform or size-proportional sampling is usually suboptimal, the formatting tax that scales with every source you add, and why deduplication isn't optional — memorization scales superlinearly with rep...

Daniel and Hannah are both self-employed, raising a fifteen-month-old in a small Jerusalem rental with no family nearby. Daycare runs seven to three, six days a week — and they're still exhausted. This episode digs into why. Time-use research shows it's the shape of the hours, not the count: fragmented work time predicts time pressure independently of total hours. The mental load literature explai...

If you were building a team to train a model — not at OpenAI or DeepMind, but at a small lab or enterprise — who would you hire? The answer is stranger than the job postings suggest. At the engineering layer, titles are collapsing: Anthropic lists everyone as a Member of Technical Staff because the research/engineering boundary stopped being useful. At the data layer, they're multiplying: dataset...

Daniel asks whether the whole model landscape is converging on one universal model you configure for anything — or whether specialization stays fundamental. Two test cases: reasoning, which started as its own model class and now looks like a knob, and instruction tuning plus structured output, where the base/instruct/chat menu seems to be collapsing into one steerable model. But there's a sharper...

A year ago, the prediction was that dedicated ASR models would go extinct — why build a transcription model and a second model to clean up its output when a multimodal LLM could do both from raw audio? Two years of evidence says otherwise. The top of the Open ASR Leaderboard is still owned by Conformer-encoder-plus-LLM-decoder systems on accuracy, but a 0.6B TDT model sits four tenths of a point b...

Daniel in Jerusalem has two NVIDIA Parakeet fine-tunes — one for Hebrew-English code-switching, one for niche technical vocabulary — and wants to know if he can merge them like LEGO bricks instead of retraining on a combined dataset. It's a question with a combinatorial edge: hospitals, legal teams, and multilingual clinics all want pairings nobody will ever build dedicated models for. We walk thr...

Daniel fine-tuned Parakeet twice — once for Hebrew-English code-switching, once for niche technical vocabulary. Can he staple those fine-tunes together into one model that keeps both skills? This episode walks through the model merging family — weight averaging, task arithmetic, TIES-Merging, and LoRA merging — and then asks the harder question underneath: is independently building specialized fin...

Claude Code's own docs tell you to clear between tasks — so why does the tool make you do it by hand? This episode digs into the research showing that model performance degrades early and sharply as input grows, why your previous task is the most convincing distractor you could hand a model, and why every handoff tool on the market triggers on size instead of meaning. We walk through what a task-b...

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...