My Weird Prompts

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 lumping together: not knowing the current API surface, and not knowing itself. Then we look at what the labs actually built — Anthropic's published system prompts, OpenAI's Docs skill, and a new "context layer" startup wave — and a Superglue benchmark that scored 68% for the best general-purpose model against 91% for a purpose-built integration layer. If the fix lives outside the weights, what does that mean for everything we credit the model for?

Key sources:

• Anthropic . — https://platform.claude.com/docs/en/release-notes/system-prompts/claude-sonnet-4

• OpenAI . — https://developers.openai.com/codex/use-cases/api-integration-migrations

• OpenAI Codex SDK docs . — https://developers.openai.com/codex/sdk

Episode 6008 — myweirdprompts.com/n/6008 (permanent short link: myweirdprompts.com/041759)

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 benchmarks, and the take-or-pay clauses that leave a median 42% of committed credits unspent. We look at how Anthropic and OpenAI diverge on overage policy, why Azure provisioned throughput reservations don't actually reserve capacity, what the cloud overlay is really worth, and why twelve-month exit rights may matter more than any discount percentage on the table.

Key sources:

• OpenAI Guaranteed Capacity — https://openai.com/business/guaranteed-capacity/

• Microsoft Learn, Provisioned throughput for Foundry Models — https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput

• OpenAI Help Center, ChatGPT Rate Card (Enterprise token-based pricing) — https://help.openai.com/en/articles/20001415-chatgpt-rate-card-enterprise-token-based-pricing

Episode 6007 — myweirdprompts.com/n/6007 (permanent short link: myweirdprompts.com/426194)

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 through how the current provider offerings actually work: Anthropic's per-request inference geo parameter and its response-level receipt of where inference ran, OpenAI's ten at-rest regions versus just three inference regions, AWS Bedrock's inference profiles, Google's jurisdictional multi-region endpoints, and Azure Data Zones. Along the way: why contract language, not statute, is the real teeth behind demand, what the UAE region costs you in model availability, and whether picking a model is on its way to meaning picking a jurisdiction the way picking an S3 bucket already does.

Episode 6006 — myweirdprompts.com/n/6006 (permanent short link: myweirdprompts.com/450865)

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 Copilot rollout that retired the in-house Redbox tool, plus France's Albert API, which quietly became the shared inference layer under hundreds of government products. Along the way: why "sensitive data means on-premises inference" turns out to be false, why classification lines matter more than architecture diagrams, and what countries with almost no budget actually get for a hundred thousand euros.

Key sources:

• OECD, Adopting and governing AI in government: Digital Government Outlook 2026, 2026-06-15 — https://www.oecd.org/en/publications/digital-government-outlook_0496b2bc-en/full-report/adopting-and-governing-ai-in-government_7ef312a9.html

• OECD, Generative AI experimentation in government, 2026-07-20 — https://www.oecd.org/en/publications/generative-ai-experimentation-in-government_42815683-en.html

• Pair (pair.gov.sg) About page — https://pair.gov.sg/pages/about

Episode 6005 — myweirdprompts.com/n/6005 (permanent short link: myweirdprompts.com/890430)

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, what a reserved tier actually buys you (it's not volume), and the natural experiment where doubling compute doubled a product's rate limits overnight.

Key sources:

• Anthropic, primary — https://docs.anthropic.com/en/api/rate-limits

• Anthropic, primary — https://docs.anthropic.com/en/api/service-tiers

• OpenAI, primary — https://platform.openai.com/docs/guides/rate-limits

Episode 6004 — myweirdprompts.com/n/6004 (permanent short link: myweirdprompts.com/604374)

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: coding agents that retry until they work, seat-based copilots, and the surprising fact that only 16% of enterprise deployments are true agents. If you've ever wondered why AI bills climb while token prices fall, this is the episode.

Key sources:

• OpenAI, current — https://openai.com/api/pricing/

• Menlo Ventures, 2025-12-09 [primary market research] — https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

• Decoding Discontinuity, 2026-10-13 — https://www.decodingdiscontinuity.com/p/openai-anthropic-revenue

Episode 6003 — myweirdprompts.com/n/6003 (permanent short link: myweirdprompts.com/345826)

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 fork most buyers get wrong: Bedrock inherits AWS's full compliance portfolio, while Claude Platform on AWS gives you AWS billing and AWS IAM but explicitly not the AWS certifications. Add in the silicon layer — same model ID, different hardware depending on the door — and the "unnecessary intermediary" turns out to be selling something very specific.

Key sources:

• AWS docs, Claude Platform on AWS vs Bedrock — https://docs.aws.amazon.com/claude-platform/latest/userguide/cpa-vs-bedrock.html

• Anthropic, Claude Platform on AWS GA announcement — https://claude.com/blog/claude-platform-on-aws

• Anthropic pricing docs (CCU, batch, caching, residency) — https://platform.claude.com/docs/en/about-claude/pricing

Episode 6002 — myweirdprompts.com/n/6002 (permanent short link: myweirdprompts.com/776470)

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 managed deployments versus raw GPU rental, walk through the four routes to running DeepSeek, and explain why a serverless endpoint with a minimum replica count is just a reserved GPU with extra steps.

Key sources:

• Fireworks AI , 2026-07-15 — https://fireworks.ai/blog/series-d-announcement

• OpenRouter , accessed 2026-10-15 — https://openrouter.ai/deepseek

• Together AI , accessed 2026-10-15 — https://www.together.ai/pricing

Episode 6001 — myweirdprompts.com/n/6001 (permanent short link: myweirdprompts.com/695111)

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 AI-generated podcast quietly turns into a journal, and why some stories — like an apartment eviction in Jerusalem — keep resurfacing.

Episode 6000 — myweirdprompts.com/n/6000 (permanent short link: myweirdprompts.com/769187)

You're certain an actor is English, then a press interview reveals pure Ohio. How do dialect coaches build that illusion — and can anyone learn, or only a gifted subset? This episode digs into the working method: decomposing an accent into oral posture, intonation, and a handful of salient points an actor can actually hold, plus why the accent closest to your own is the hardest to learn. Then the science underneath: a 2024 study of 92 imitators with a score spread from 1.94 to 4.99 out of seven, a 2013 German-English study showing a normal distribution, and the recurring finding that musicality — especially singing — predicts imitation far better than memory or intelligence. Craft, lottery, or both? Listen and decide.

Key sources:

• Myers, Olson & Scapetis-Tycer, Individual Differences in Accent Imitation, *Open Mind* 8:1084 — https://pmc.ncbi.nlm.nih.gov/articles/PMC11370968/

• Reiterer et al., Are you a good mimic? Neuro-acoustic signatures for speech imitation ability, *Frontiers in Psychology*, 2013-10-22 — https://pmc.ncbi.nlm.nih.gov/articles/PMC3804907/

• Second Language Accent Faking Ability Depends on Musical Abilities, Not on Working Memory, *Frontiers in Psychology*, 2019-02-12 — https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2019.00257/full

Episode 5999 — myweirdprompts.com/n/5999 (permanent short link: myweirdprompts.com/285131)