My Weird Prompts

Most speech recognition coverage leads with accuracy — word error rate, character error rate, how close the transcript got. But that's not the number that decides whether a system is usable. This episode digs into the computational side of ASR benchmarking: real-time factor, why an RTF above one means broken rather than merely slow, and why the mean latency on your slide describes a request that never happened. From there it moves through throughput under concurrency, the memory envelope on-device, and the three environments where speech models actually run. It closes on a question worth asking out loud: is tokens per second even a meaningful metric for speech recognition, or is it borrowed from somewhere it doesn't belong?

Key sources:

• ASR Leaderboard paper (arXiv 2510.06961v4, v4, data as of 27 March 2026) — https://arxiv.org/html/2510.06961v4

• NVIDIA ASR NIM Performance docs (v26.07.0) — https://docs.nvidia.com/nim/speech/26.07.0/reference/performances/asr/performance.html

• Keen Research, Evaluating ASR Systems, Part 3: Latency and Responsiveness (2026-04-23) — https://keenresearch.com/blog/evaluation/evaluating-asr-systems-part-3-latency-responsiveness.html

Episode 5959 — myweirdprompts.com/n/5959 (permanent short link: myweirdprompts.com/077812)

Freedom of information laws stretch back to Sweden in 1766, and roughly 140 countries now have one. But a 2025 study filed identical requests in 76 countries and found that 38% drew no substantive reply — not a denial, just silence. This episode traces the history from Peter Forsskål's 1766 Freedom of the Press Act through the post-1990 adoption wave, compares how laws differ on who can request, which bodies are covered, and what exemptions exist, and asks why the strongest laws on paper — Afghanistan, Mexico, Serbia — sit above Sweden, the UK, and the US on the Global Right to Information Rating. Israel's 1998 law anchors the comparison, including its section 9A/9B exemption split and the institutional carve-out for Shin Bet, Mossad, and the Atomic Energy Commission.

Key sources:

• Global Right to Information Rating, by country — https://www.rti-rating.org/country-data/

• CLD/UNESCO, Global Comparative Testing of Responses to Requests for Information, September 2025 — https://www.law-democracy.org/wp-content/uploads/2026/02/IDUAI.report.25-09-26.pdf

• Israel Freedom of Information Law, 5758-1998 (full text) — http://www.freedominfo.org/documents/Israel--FOIL1998.pdf

Episode 5958 — myweirdprompts.com/n/5958 (permanent short link: myweirdprompts.com/774962)

ChatGPT's plugin directory now lists 5,723 plugins with 238,000 callable tools, yet only 52% have useful descriptions. The leaderboard rewards commercial SaaS with clean schemas — Ashby, Supabase, Miro, Stripe — while scientific tools like Wolfram and Consensus don't crack the top rankings. Meanwhile, the genuinely capability-expanding drug discovery work — LIDDIA, MADDIE, DrugPilot — lives in arXiv papers that haven't migrated into the directory at all. This episode explores the split between the visible commercial layer and the deeper research layer, and asks whether general-purpose assistants are quietly becoming the access layer for serious scientific infrastructure. We dig into measurable impact: LEADS showing 20.8% time savings in study selection versus a systematic review finding GenAI missed 68-96% of studies during searching. The evidence cuts both ways, and the honest numbers are more interesting than the marketing.

Key sources:

• OpenAI, 2025-10-06 (updated 2025-11-13) — https://openai.com/index/introducing-apps-in-chatgpt/

• ChatGPT Learn, current — https://learn.chatgpt.com/docs/plugins

• OpenAI Developers, current — https://developers.openai.com/apps-sdk/quickstart

Episode 5957 — myweirdprompts.com/n/5957 (permanent short link: myweirdprompts.com/381131)

MCP turned the AI assistant into a universal front end, and OpenAI's Apps SDK is built on it. So why is a grocery list the battleground? This episode works through a prediction that conversational AI becomes the new interface for everything — and that vendors, not your employer, end up selling you the bolt-on. We trace the MCP timeline from Anthropic's open-source release through the ChatGPT App Directory, weigh the interface argument against the distribution argument, and look at what a brutal year for SaaS valuations is actually pricing in. Then we find the gap: the mechanism is fully built, but nobody has shipped the first-party version yet.

Key sources:

• OpenAI — https://openai.com/index/introducing-apps-in-chatgpt/

• OpenAI — https://openai.com/index/developers-can-now-submit-apps-to-chatgpt/

• OpenAI Help Center — https://help.openai.com/en/articles/12515353-build-with-the-apps-sdk

Episode 5956 — myweirdprompts.com/n/5956 (permanent short link: myweirdprompts.com/005169)

If you've ever bought a work light on its lumen rating and been disappointed, you're not alone — lumens measure light leaving the fixture, not light landing on your work surface. This episode works through the specs that actually predict whether a light will do the job: lux and illuminance at working distance, beam angle, runtime per setting, thermal step-down, CRI, colour temperature, and the IP and IK durability codes most listings quietly omit. Along the way: why wattage tells you almost nothing, why a quoted beam distance is a warning label, and how to tell a flood from a spotlight using nothing but a product photo. If you've ever stood in an aisle staring at four form factors under one label, this is the buyer's guide that sorts them out.

Key sources:

• Milwaukee Tool (ROVER 2366-20 dual power, 4,000 lm) — https://www.milwaukeetool.com/products/details/m18-rover-dual-power-flood-light/2366-20

• Milwaukee Tool (magnetic + clamp mounting) — https://www.milwaukeetool.com/products/details/m18-rover-mounting-flood-light/2365-20

• Milwaukee Tool (3 clamp, 1 — https://www.milwaukeetool.com/products/details/m18-rover-clamping-flood-light/2358-20

Episode 5955 — myweirdprompts.com/n/5955 (permanent short link: myweirdprompts.com/853818)

Three safeguards that each work ninety percent of the time should fail together only once in a thousand tries. That arithmetic is correct, and it's also where the trouble starts. This episode works through the real mathematics of layered redundancy, why the independence assumption behind it so often breaks, and what reliability engineers call common cause failure. From the beta-factor model and the Humphreys checklist to the FAA's one-in-a-billion flight hour standard, from dissimilar redundancy on the Airbus A320 to physically separating critical systems, the episode explores why adding more of the same thing is not the same as designing genuinely independent layers—and why knowing when to stop adding layers may be the hardest question in the field.

Key sources:

• Rausand & Lundteigen, Chapter 10. Common Cause Failures (CCFs), NTNU SIS textbook, Version 0.1 — https://lundteig.folk.ntnu.no/Presentations/SIS-book/chapt10-1-CCFs-new.pdf

• World Nuclear Association CORDEL, Defence-in-Depth and Diversity: Challenges Related to I&C Architecture, Report No. 2018/003, April 2018 — https://world-nuclear.org/images/articles/CORDEL-Defence-in-Depth-Report-10-April.pdf

• FAA Advisory Circular 25.1309-1B — https://www.faa.gov/documentLibrary/media/Advisory_Circular/AC_25.1309-1B.pdf

Episode 5954 — myweirdprompts.com/n/5954 (permanent short link: myweirdprompts.com/387838)

A billion shekels for twelve Apache helicopters, funded partly by cutting a five-year plan for Arab communities, eleven days before an election. But the real question isn't why twelve Apaches — it's how any military decides what it needs, in what quantity, and how urgently. This episode uses Israel's October 2026 approval as a doorway into the American requirements system: JCIDS, the Joint Requirements Oversight Council, and the Capabilities-Based Assessment that replaced systems with capabilities. Along the way: a deal that died after three deadline extensions and got revived anyway, a two-year Air Force staff study, a squadron-sized buy shaped by a Boeing production slot, and a genuine doctrine fight about whether attack helicopters still have a role that remotely piloted aircraft can't fill.

Key sources:

• CRS In Focus IF12817 — https://www.congress.gov/crs_external_products/IF/HTML/IF12817.html

• JCS J-8 — https://www.acqnotes.com/Attachments/Capabilities-Based%20Assessment%20

• IDA — https://www.dmi-ida.org/download-pdf/pdf/NS%20D-10369.pdf

Episode 5953 — myweirdprompts.com/n/5953 (permanent short link: myweirdprompts.com/788505)

Most people know Mossad, Shin Bet, and Aman — but there's a fourth body inside Israel's Defense Ministry that almost no one can name. Malmab, the Director of Security of the Defense Establishment, protects the country's most sensitive nuclear and biological facilities, and its practices reveal how governments actually organize information security when a breach could be catastrophic. From clearances that expire in practice to compartments that can't be acknowledged, this episode explores the gap between cybersecurity and true information security — and asks which of these extreme measures could ever make sense for an ordinary business.

Key sources:

• Israel Ministry of Defense (mod.gov.il) — https://mod.gov.il/.../ראש-מלמ-ב-חדש-במשרד-הביטחון-גיל-ריי

• Israel State Comptroller — https://library.mevaker.gov.il/sites/DigitalLibrary/Documents/2024/2024.11-Cyber/EN/Cyber-106-Refael-Taktzir-EN.pdf

• ODNI — https://www.odni.gov/files/documents/ICD/ICD-705-SCIFs.pdf

Episode 5952 — myweirdprompts.com/n/5952 (permanent short link: myweirdprompts.com/011187)

Reasonableness is the load-bearing wall of the legal system — negligence, self-defence, harassment law, and constitutional equality all run on it. But the "reasonable person" is a fiction, and what counts as reasonable is shaped by class, culture, gender, and professional habit. Drawing on Mayo Moran's work on the reasonable person and Victoria Nourse's argument that the standard was never a person at all, this episode asks how much subjective judgment hides inside supposedly objective tests, why courts adjust the baseline for children and physical disability but not mental illness, and what a 19,000-ruling study of French child-custody decisions says about whether you can predict a case better by knowing the judge than by knowing the law.

Key sources:

• Mayo Moran, The Reasonable Person: A Conceptual Biography in Comparative Perspective, Lewis & Clark Law Review Vol. 14:4 (2010) — https://www.lclark.edu/live/files/7230-lcb144art1moranpdf

• Victoria Nourse, After the Reasonable Man: Getting Over the Subjectivity/Objectivity Question, 11 New Crim. L. Rev. 33 — https://scholarship.law.georgetown.edu/cgi/viewcontent.cgi?article=2133&context=facpub

• Spamann, Klöhn, Jamin, Khanna, Liu, Mamidi, Morell & Reidel, Judges in the Lab: No Precedent Effects, No Common/Civil Law Differences, Journal of Legal Analysis Vol. 13, Issue 1 (published 2021-03-23) — https://academic.oup.com/jla/article/13/1/110/6180994

Episode 5951 — myweirdprompts.com/n/5951 (permanent short link: myweirdprompts.com/564687)

Daniel wants to know if you can bolt together a search-and-grounding stack from open-source parts and never rent a search API again. On the table: SearXNG for discovery, Crawl4AI for extraction, and a TypeScript aggregation library for grounding — plus how you'd actually run the whole thing on Modal. The honest answer is yes, with one large asterisk: the bill you save is not the bill you pay. You stop paying Exa and start paying a server, a cron job, and about fifteen hours of your own time. This episode walks through all three layers, the one-line settings change that trips up every self-hoster, the reciprocal rank fusion trick that turns six mediocre result lists into one decent one, and the latency numbers that undercut the whole purity argument.

Key sources:

• SearXNG repo (commit 22c64b8, README, AI_POLICY.rst, engine list incl. `exaapi.py`) — https://github.com/searxng/searxng

• SearXNG Search API docs (2026.10.13+22c64b8d3) — https://docs.searxng.org/dev/search_api.html

• Crawl4AI repo (commit aadd82b, README, v0.9.4, 2026-09-23) — https://github.com/unclecode/crawl4ai

Episode 5950 — myweirdprompts.com/n/5950 (permanent short link: myweirdprompts.com/819604)