My Weird Prompts

A listener's question about video understanding gets the full treatment: what actually changed in the past year, why Google's Omni models and agentic video navigation matter, and why uploading a clip to ChatGPT still feels like sneaking past a bouncer. The conversation works through the real arithmetic — roughly 100 tokens per second of video at default settings — and shows how that number drives every product decision, from Google's 88% token reduction to OpenAI's decision to leave video off the supported upload list entirely. Then it turns to the local question: what does it actually take to run Qwen3-VL on your own hardware for a short, high-stakes clinical clip? The answer isn't the model weights. It's the KV cache, and the math gets uncomfortable fast. Along the way: specialist token-compression research, a hand-perfusion study with a brutal generalization gap, and why the life-or-death use case turns out to be the affordable one.

Key sources:

• Gemini API — https://ai.google.dev/gemini-api/docs/video-understanding

• Google blog — https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-agentic-video-in-gemini/

• Google blog — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/

Episode 5903 — myweirdprompts.com/n/5903 (permanent short link: myweirdprompts.com/537327)

Israel just held its first full-term Knesset in decades — and got thirty-eight party lists for its trouble. This episode digs into why closed-list proportional representation with a 3.25% threshold produces a system where nobody campaigns to win, only to be the kingmaker. Then it goes deeper: four groups — Haredi, Tel Aviv, National Religious, and Arab Israeli — each carrying a fundamentally different picture of what the social contract should be. The left's collapse gets dated to a single brutal frame (Labor's 44 seats in 1992 to 4 in 2022), and the demographic argument behind it gets taken seriously. The question at the center: is the Western democratic model simply ill-fitting to this part of the world, and if so, what arrangements — states within states, secession, anything — have held incompatible visions together elsewhere?

Episode 5902 — myweirdprompts.com/n/5902 (permanent short link: myweirdprompts.com/741251)

The word "Haredi" comes from the Hebrew root for trembling — awe before the divine, drawn from Isaiah. Its common English translation, "ultra-Orthodox," carries a very different charge. This episode traces how a descriptor became an identity, why the label stuck, and how the community it names was deliberately built with state funding within living memory — from Rabbi Shach's 1977 coalition deal to Shas and the Mizrahi expansion. With over a million Haredim in Israel today and one of the highest growth rates in the developed world, the vocabulary debate is inseparable from the demographic and political reality. We look at what the numbers mean for military service, education, and Israel's economy — and why the draft fight is really a fight about what the country looks like in 2050.

Key sources:

• Moment Magazine — https://momentmag.com/jewish-word-haredi/

• Religion News Service — https://religionnews.com/2014/02/06/ultra-orthodox-want-different-name/

• Wikipedia — https://en.wikipedia.org/wiki/Haredi_Judaism

Episode 5901 — myweirdprompts.com/n/5901 (permanent short link: myweirdprompts.com/130207)

You've got the weights, you've got the rack, you've got a GPU bolted in that you're not replacing. So which engines can actually serve DeepSeek-V4.1-Flash — and how much of your "engine choice" is really a hardware decision wearing a software costume? This episode walks the five-name shortlist (vLLM, SGLang, TensorRT-LLM, llama.cpp, ONNX Runtime) against the model card's supported paths, then runs the two-server thought experiment: one NVIDIA box, one AMD box, same budget line, wildly different menus. Along the way: why the most portable engine is the one that can't load the model, why "supports AMD" and "supports your AMD card" are two different sentences, and why the quantization method printed on the checkpoint can decide whether your inference runs at all.

Key sources:

• DeepSeek API Change Log (2026-09-10 V4.1-Flash release; full history) — https://api-docs.deepseek.com/updates

• DeepSeek-V4.1-Flash model card (architecture, benchmarks, local-app paths) — https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

• DeepSeek-V4-Flash model card — https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash

Episode 5900 — myweirdprompts.com/n/5900 (permanent short link: myweirdprompts.com/915386)

Two servers, same GPU, same weights, same prompt — different output. This episode digs into Daniel's thought experiment: three providers running vLLM, SGLang, and TensorRT-LLM on identical hardware, where the customer can't choose the engine. Does the inference engine actually change speed, cost, and the answers themselves? We walk through the architecture of all three — PagedAttention, RadixAttention, and the TensorRT-LLM backend overhaul that removed TensorRT — then look at the benchmark data, where the fastest engine changes with the load, and a 2026 CISPA paper showing that swapping only the backend moved GSM8K scores by up to 16.6 percentage points on the same weights. If you rent endpoints and never see the middle layer, this one's for you.

Key sources:

• Thinking Machines Lab (Horace He), 2025-09-10 — https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/

• SGLang Team, 2025-09-22 (updated 2025-09-24) — https://www.lmsys.org/blog/2025-09-22-sglang-deterministic/

• vLLM repo — https://github.com/vllm-project/vllm

Episode 5899 — myweirdprompts.com/n/5899 (permanent short link: myweirdprompts.com/616043)

AMD crossed a trillion dollars in market cap and its software stack is finally getting the attention. We dig into three ROCm releases — 7.14, 10, and 10.1 — and ask the honest question: is ROCm actually production-viable for inference now, or just for the data center? From TheRock's reproducible builds to Hyperloom's autonomous optimizer to NUMA-aware memory allocation, we trace what changed, what's still marketing, and why the 7900 XTX owner is still the person the roadmap forgot.

Key sources:

• ROCm GitHub README , commit 63fd839, accessed 2026-10-12. — https://github.com/ROCm/ROCm

• AMD ROCm Blogs, ROCm 7.14: TheRock Goes Production, 2026-07-15 (updated 2026-07-17) . — https://rocm.blogs.amd.com/ecosystems-and-partners/rocm-7.14-blog/README.html

• AMD, ROCm 10: A Simpler Path to Production AI, 2026-08-27 . — https://www.amd.com/en/blogs/2026/amd-rocm-10-a-simpler-path-to-production-ai-on-amd.html

Episode 5898 — myweirdprompts.com/n/5898 (permanent short link: myweirdprompts.com/597264)

Daniel terminated his first IEC 60309 connector — the blue 16A industrial plug known as commando, pin and sleeve, or CeeForm — and it worked. But two questions remain: what cable actually belongs on a 16A connector feeding a device that would run fine on a lamp cord, and why can't you buy a bench strip or PDU with these outlets? The cable answer is counterintuitive: you size the cable to the connector, not the appliance, because the circuit is defined by the plug and its breaker, not by whatever you happen to plug in. The market answer is stranger — the multi-outlet product Daniel went looking for doesn't appear to exist, and the reason is baked into what IEC 60309 was designed to do.

Key sources:

• Farnell/pro-POWER, H07RN-F BS EN 50525-2-21 datasheet (current capacity, voltage drop, de-rating, construction). — https://www.farnell.com/datasheets/1929043.pdf

• Nexans H07RN-F 3G2.5 (max OD 14.0 mm, standards). — https://www.nexans.de/en/products/Building/Rubber-cables/H07RN-F4943/product~10056140~.html

• Lapp H07RN-F 3G2.5 (450/750 V, HAR-certified). — https://e.lapp.com/apac/p/cables-for-harsh-conditions/h07rn-f-enhanced-version-3g2-5-4533031

Episode 5897 — myweirdprompts.com/n/5897 (permanent short link: myweirdprompts.com/108116)

Israel is projected to grow by three million people in fourteen years, and most of that new fabric will be laid down after the car. What went wrong in Modiin — the best-documented Israeli new city that was explicitly designed to be urban and became a bedroom community — and what would it take to build differently? This episode digs into Safdie's own planning guidelines, the code that produced coherence across dozens of developers, and why the form was right while the life was wrong. Then: land, approvals, financing, transit, and the regulatory reforms moving through the Knesset right now, plus the case for a brand-new mid-sized Israeli city of fifty to a hundred thousand people, walkable and rail-connected, built on traditional urban principles with a contemporary neo-Levantine language.

Key sources:

• Safdie Architects — https://www.safdiearchitects.com/projects/new-city-of-modiin

• Classic Planning Institute — https://www.classicplanning.org/bio

• Harvard University Press — https://www.hup.harvard.edu/books/9780674919242

Episode 5896 — myweirdprompts.com/n/5896 (permanent short link: myweirdprompts.com/875238)

Cities are the most efficient habitat humans have ever built — and the place people most want to leave once they have kids. This episode works both halves of that problem: what families are actually buying when they move out, and what premodern cities like Çatalhöyük got right about density and livability that we've forgotten. From the Urban Institute's "families as infrastructure" framing to Jane Jacobs' eyes on the street to the fifteen-minute city and the mobile-phone data that complicates it, we ask whether a dense place can also be a quiet one.

Key sources:

• UNESCO World Heritage Centre — https://whc.unesco.org/en/list/1405/

• Wikipedia — https://en.wikipedia.org/wiki/Insula

• Wikipedia — https://en.wikipedia.org/wiki/Teotihuacan

Episode 5895 — myweirdprompts.com/n/5895 (permanent short link: myweirdprompts.com/513920)

Every year the Fed orders billions of new banknotes, and every year headlines read it as money printing. But the print order is a maintenance schedule, not monetary policy — and the split between replacing worn-out notes and actually growing the money supply is a number the Fed doesn't publish. This episode walks through the whole physical life of cash: how the Bureau of Engraving and Printing and the US Mint split the work, why a one dollar bill lasts 5.8 years while a hundred lasts 15, what happens to the 5,287 tons of shredded currency a year, and why Israel — with a fully sovereign currency — prints its shekels abroad. Also: what the 2027 Fed order actually says, and why "cash is dying" doesn't show up in the numbers.

Key sources:

• Federal Reserve Board — https://federalreserve.gov/paymentsystems/files/currency_print_orders_2027.pdf

• Federal Reserve Board — https://www.federalreserve.gov/paymentsystems/coin_currency_orders.htm

• Federal Reserve Board — https://www.federalreserve.gov/paymentsystems/files/currency_print_orders_2026.pdf

Episode 5894 — myweirdprompts.com/n/5894 (permanent short link: myweirdprompts.com/334920)