My Weird Prompts

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 behind at roughly eight times the speed, and a new third layer has appeared: specialized post-processing models like Superwhisper's S1-mini that turn raw transcripts into formatted text. This episode works through the three-layer stack, why the latency gap is an architectural fork rather than a maturity gap, why word error rate is under active attack as a metric, and why no benchmark yet compares audio-to-formatted-text against ASR-plus-normalization head to head.

Key sources:

• Open ASR Leaderboard paper (v2, 2025-10-08 data) — https://arxiv.org/html/2510.06961v2

• NVIDIA Parakeet TDT 0.6B v3 model card — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3

• S1-mini model card — https://huggingface.co/superwhisper/s1-mini

Episode 6028 — myweirdprompts.com/n/6028 (permanent short link: myweirdprompts.com/351585)

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 through what merging actually does to weights — task vectors, weight averaging, TIES-Merging, DARE, and LoRA adapter merging — and ask whether transcription's demand for exactness makes unified training the only honest path. Along the way: why word error rate can hide the exact failures you fine-tuned to fix, and why fine-tunes from the same base checkpoint come pre-aligned.

Key sources:

• TIES-Merging: Resolving Interference When Merging Models — https://arxiv.org/html/2306.01708v2

• An Introduction to Model Merging for LLMs — https://developer.nvidia.com/blog/an-introduction-to-model-merging-for-llms/

• NeMo Speech Fine-Tuning docs — https://docs.nvidia.com/nemo/speech/latest/asr/fine_tuning.html

Episode 6027 — myweirdprompts.com/n/6027 (permanent short link: myweirdprompts.com/991040)

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 fine-tunes and merging them later a real, evidence-supported path to better ASR? Whisper merges are showing twelve percent relative WER drops, and TIES with correction cut Fair-Speech WER from 7.38 to 5.13. But tokenizer mismatch and non-local merging may be the actual blockers.

Key sources:

• nvidia/parakeet-tdt-0.6b-v3 model card, Hugging Face, released 2025-08-14 — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3

• NeMo-Speech Fine-Tuning docs — https://docs.nvidia.com/nemo/speech/latest/asr/fine_tuning.html

• arcee-ai/mergekit merge_methods docs — https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md

Episode 6026 — myweirdprompts.com/n/6026 (permanent short link: myweirdprompts.com/757014)

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-boundary hook would actually need to do, what already exists in sub-agents and handoff utilities, and what you give up when the interface makes the call for you.

Key sources:

• Claude Code Hooks reference — https://docs.claude.com/en/docs/claude-code/hooks

• Create custom subagents — https://docs.claude.com/en/docs/claude-code/sub-agents

• Manage costs effectively / context management — https://docs.claude.com/en/docs/claude-code/costs

Episode 6025 — myweirdprompts.com/n/6025 (permanent short link: myweirdprompts.com/194035)

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 formula that turns a single Llama 2 conversation into two gigabytes of cache. Then it maps the whole lineage of fixes onto three multiplicative dimensions — entry size, sequence length, and layers — from Multi-Query and Grouped-Query Attention through DeepSeek's Multi-head Latent Attention, DeepSeek Sparse Attention, cross-layer sharing, and the inference-only tricks like PagedAttention and KV quantization that need no retraining at all.

Episode 6024 — myweirdprompts.com/n/6024 (permanent short link: myweirdprompts.com/640482)

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 saturation can arrive as early as one to ten hours, while a Maltese study that trained thirty models found no diminishing returns for Whisper. A single-speaker Swedish study shows most of the gain captured by four hours, with a long shallow tail after. This episode digs into the contested research, what extra hours actually buy, why transcription noise may matter more than fresh clips, and whether a dataset built from LLM-written sentences is a diet version of how you really speak.

Key sources:

• NVIDIA Riva — https://docs.nvidia.com/deeplearning/riva/user-guide/docs/asr/asr-customizing.html

• Rosehill — https://huggingface.co/blog/danielrosehill/whisper-hebrish

• thewh1teagle/parakeet-tdt-0.6b-v3-he — https://huggingface.co/thewh1teagle/parakeet-tdt-0.6b-v3-he

Episode 6023 — myweirdprompts.com/n/6023 (permanent short link: myweirdprompts.com/872707)

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 whether a 30% English ratio is the right number. Plus: why fine-tuning helps where a model is ignorant and hurts where it's merely imprecise, and the hallucination risk that comes with over-controlling.

Key sources:

• NVIDIA Parakeet TDT 0.6B V2 model card, release 2025-05-01 — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2

• Daniel's Whisper fine-tune benchmark — https://github.com/danielrosehill/Whisper-Fine-Tune-Accuracy-Eval

• Choosing Epochs for ASR Fine-Tuning, 2025-11-25 — https://gist.github.com/danielrosehill/0fd5579dd554596b9c4a73fedf917697

Episode 6022 — myweirdprompts.com/n/6022 (permanent short link: myweirdprompts.com/443386)

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 tool"; Gemma has no system role at all; DeepSeek uses full-width CJK pipes. This episode walks through what special tokens actually are, how tokenizers register them, why the add_generation_prompt flag matters more than anyone thinks, and why a format mismatch never throws an error — it just produces a model that's a bit dimmer than it should be.

Key sources:

• Hugging Face, Chat templates — https://huggingface.co/docs/transformers/chat_templating

• Matthew Carrigan / HF, Chat Templates: An End to the Silent Performance Killer, 2023-10-03 — https://huggingface.co/blog/chat-templates

• OpenAI, Harmony Response Format, 2025-08-05 — https://cookbook.openai.com/articles/openai-harmony

Episode 6021 — myweirdprompts.com/n/6021 (permanent short link: myweirdprompts.com/765230)

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's prisons and the face-like patterns in traditional facades, we take a building apart piece by piece and examine what each design system is really doing to the person standing in front of it. Plus: the practical question of how to design something new that doesn't look historic but keeps the proportions that make traditional architecture feel humane.

Key sources:

• Brielmann, Buras, Salingaros & Taylor, What Happens in Your Brain When You Walk Down the Street?, *Urban Science* 6(1):3, 2022-01-07 — https://www.mdpi.com/2413-8851/6/1/3

• Institute of Traditional Architecture, Design Guidelines — https://www.institute-of-traditional-architecture.org/self-study/design-guidelines/

• ITA, Openings — https://www.institute-of-traditional-architecture.org/openings/

Episode 6020 — myweirdprompts.com/n/6020 (permanent short link: myweirdprompts.com/686908)

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 Contributions, and Israel's pinui-binui program reveals three completely different answers — and three very different outcomes. The through-line: cities that decide what they want before a developer walks in the door capture far more public value than cities that improvise deal by deal.

Key sources:

• Jerusalem Post — https://www.jpost.com/business-and-innovation/real-estate/article-826268

• Times of Israel — https://www.timesofisrael.com/how-to-make-urban-renewal-smoother-keep-developers-and-homeowners-on-the-same-side/

• Al Jazeera — https://www.aljazeera.com/features/2012/12/18/demolition-fears-haunt-israeli-neighbourhoods

Episode 6019 — myweirdprompts.com/n/6019 (permanent short link: myweirdprompts.com/147695)