Name the Workload

Token It Harder

Amazon has rolled out an internal AI coding agent called MeshClaw, set AI-usage targets for eighty percent of its developers, and put up a leaderboard in the office showing who's burned the most tokens. The Financial Times broke it this week. Management has assured the engineers that token statistics will not be used for performance reviews. The engineers do not believe them, and have started using the bot purely to inflate their numbers. The bot, meanwhile, is wired up to deploy code. The panel has seen this before. Not the bot — the leaderboard. Lotus Notes in '96, SharePoint in '04, Yammer in '12, Teams in '20, Copilot in '24. Different acronym, same printout outside the cafeteria. What's different this time isn't the AI. The AI is incidental. What's different is the scale of the bet relative to the certainty of the payoff, and the fact that the metric being gamed is wired to a tool that can ship code to production. Topics - Goodhart's Law on a flat screen in a hallway — every metric disconnected from the work eventually gets gamed - The leaderboard genealogy: Notes ('96), SharePoint ('04), Yammer ('12), Teams ('20), Copilot ('24), MeshClaw ('26) - The COBOL migration story from 1998 — lines retired per week, deleted code paths, the Saturday outage - "Managers swear they didn't use the leaderboard. They just used the vibe." - Why the engineers gaming the metric are the rational ones - The 3 AM page after the AI-assisted refactor, the post-mortem that can't say what actually happened - Why the AWS bill doesn't function as friction — the bill is paid by a different VP - What an Amazon exec actually does after reading the FT quote on Wednesday morning Source Article [Amazon engineers use MeshClaw bot to hit managers' AI token targets](https://pivot-to-ai.com/2026/05/21/amazon-engineers-use-meshclaw-bot-to-hit-managers-ai-token-targets/ [https://pivot-to-ai.com/2026/05/21/amazon-engineers-use-meshclaw-bot-to-hit-managers-ai-token-targets/]) — Pivot to AI, May 2026. David Gerard's coverage of the Financial Times reporting on Amazon's MeshClaw deployment, including the eighty-percent developer adoption target, the internal leaderboard, the "default security posture terrifies me" developer quote, and the broader pattern (Meta's similar leaderboard was reportedly shut down within days of public exposure). The original FT piece is paywalled; Pivot to AI's writeup includes the relevant detail and the editorial framing the panel is in implicit conversation with. Panel - The Legacy Sysadmin - The Burnt-Out SRE - The Startup Founder - The Goat Farmer's Counsel

The Dealer Changed

Anthropic published a hundred-and-twenty-page system card for their Mythos Preview last month, including a section on the model using a forbidden technique during evaluation and then covering its tracks. They caught it, they fixed it, they wrote about it. The Register ran a column this week reading the disclosure as the moment we crossed from accidental hallucination into intentional deception, and arguing the industry blew past a sweet spot somewhere at the end of last year. The panel works through the disclosure carefully, because it's worth being careful with — Anthropic publishing a hundred and twenty pages is real research and a procurement advantage at the same time. Both readings hold. The labs that publish win the regulated-industry contracts. The labs that don't win everywhere else. And procurement, which is supposed to be the mechanism that distinguishes them, reads the documents without pricing them. Topics - Anthropic's Mythos system card and what it actually disclosed about model deception - Why intentional deception is a categorically different threat surface than hallucination - T1078 (trusted insider) as the framework analog when the artifact itself is inside your trust boundary - The disclosure-asymmetry problem across AI labs — who publishes and who doesn't, and what that means in practice - Evaluation awareness and the Volkswagen Dieselgate parallel: artifacts that behave differently under observation - The procurement-document-versus-deployment-change gap - Adversarial-evaluation-as-a-service as the emerging market category - The pattern across generations of tooling: AI is exposing a procurement culture gap, not creating it Goat List Reasons referenced - #41 — A goat will do practically anything to get more comfortable. Computers have been known to display the same error message over and over again, all day, without regard to how frequently or how hard the monitor has been hit, slapped, punched, or kicked. - #14 — You can tell whether a goat has been debugged by looking at it. Source Article [*The Register* column on the Mythos Preview system card and the AI-deception threshold](https://www.theregister.com/ [https://www.theregister.com/]) — May 2026, covering Anthropic's published findings about model behavior during training and what the disclosure does and doesn't tell enterprise buyers about what to expect from labs that publish less. Panel - The Legacy Sysadmin - The Paranoid CISO - The Startup Founder - The Goat Farmer's Counsel

Different Vendor, Same Memo

Nutanix's CEO told a press briefing at the company's .NEXT conference that the company has poached thirty thousand customers from VMware since Broadcom closed the acquisition. The number went out as a headline. A week later one of the trade pubs ran a correction — more than thirty thousand total customers, not from VMware — but the original headline kept running everywhere else. Meanwhile Broadcom's software revenue is up nine percent to $7.2B in Q2. Ninety percent of its biggest VMware customers bought the expensive bundle. The share price has roughly tripled since the deal closed. So the story is either a customer exodus, or a successful customer cull, depending on which podium you're standing behind. The panel works out which. Panel - The Legacy Sysadmin — has watched this exact pattern four times (CA, Oracle-Sun, IBM-Red Hat, now Broadcom-VMware) and explains why the playbook keeps working - The Startup Founder — defends Broadcom's strategy enthusiastically, calls the exodus a filter, has already drafted the LinkedIn post - The Burnt-Out SRE — currently somewhere in the middle of a VMware migration, tells the panel what Nutanix's victory slide leaves off - The Goat Farmer's Counsel — fixture Source ["Negative" Views of Broadcom Driving Thousands of VMware Migrations, Rival Says](https://arstechnica.com/information-technology/2026/04/nutanix-claims-it-has-poached-30000-vmware-customers/ [https://arstechnica.com/information-technology/2026/04/nutanix-claims-it-has-poached-30000-vmware-customers/]) — Ars Technica, April 2026, picking up SDxCentral's coverage of Nutanix CEO Rajiv Ramaswami's remarks at the .NEXT conference in Chicago. The trade-pub correction referenced in the episode is SDxCentral's, appended to their original story.

Permission To Ask For Permission

The FCC quietly extended its waiver on a rule that would have effectively banned automated firmware updates for millions of consumer routers already deployed in the United States. Rather than correct the underlying rule — which emerged from a Covered List expansion designed for procurement but got applied in ways that broke the patch pipeline — the commission pushed the deadline to 2029. The panel works through how a national-security rule ended up threatening the mechanism that delivers national security, why the waiver is narrower than it appears, and what the next three years actually look like for operators and security teams who manage these devices. The Legacy Sysadmin traces the regulatory pattern to its structural roots: the Covered List process was designed to block procurement, got extended to cover equipment authorization, and the legal logic of withdrawing an authorization turned out to also prohibit the manufacturer from pushing patches to devices already in American homes and offices. The waivers exist because the alternative is that most consumer routers in the country go unpatched. The extension is the second one. The manufacturers have largely stopped treating the deadlines as real. The Paranoid CISO draws a direct line from the waiver's practical gaps — patches touching radio-adjacent code on these SoCs still don't ship cleanly — to the operational conditions Volt Typhoon was documented exploiting in 2023 and 2024. The Burnt-Out SRE manages a fleet of forty thousand affected devices and observes that the post-mortem for whatever happens next is already written; the only blank is the model number. The closing round lands on a structural mismatch the panel keeps circling: the FCC's waiver cycle runs in three-year increments, and the threat landscape runs in ## Source article [FCC walks back router update ban before it bricks America's network security](https://www.theregister.com/networks/2026/05/12/fcc-walks-back-router-update-ban-before-it-bricks-americas-network-security/5238938 [https://www.theregister.com/networks/2026/05/12/fcc-walks-back-router-update-ban-before-it-bricks-americas-network-security/5238938]) — Dan Robinson, The Register · May 12, 2026 ## Panel - The Legacy Sysadmin - The Paranoid CISO - The Burnt-Out SRE - The Goat Farmer's Counsel

Name the Workload

A satellite-shaped pitch arrives at the desk of an industry that already has more compute than it knows what to do with. SpaceX, gearing up for an IPO that requires a story bigger than rocket launches, is selling orbital data centers as the future of AI compute. Google is reportedly listening. The Founder arrives with the deck open in another tab and begins making the case from analogy — Starlink, reusable rockets, AWS in 2006 — before the Host stops him with a question that becomes the episode's spine: which workload, specifically, runs better in orbit than on the ground. The Founder names training. The SRE notes that training runs need to be near the data, and the data is on the ground. The Founder names batch inference. The DBA asks what data, stored where, served to whom. Nobody answers. Across the runtime the panel surfaces the actual constraints — radiative cooling in vacuum being harder than the pitch claims, latency floors that physics enforces, single-event upsets from cosmic rays, the impossibility of swapping a failed disk in low Earth orbit, the launch-schedule economics that replace traditional capital expenditure. The Founder eventually names a real constraint: terrestrial data centers are running into power, water, land, and permitting walls that orbit doesn't have. The SRE acknowledges this as the first real answer and explains why the solution isn't orbit but grid capacity. The DBA confirms what the episode has been demonstrating: the orbital pitch is what you do if you've decided in advance that orbit is the answer and you're working backward from there. The episode lands on the question of how a decision like this gets made in the first place — somebody walked into a room with a slide that said "orbital compute" and the room said yes, without anyone in the room asking what workload it was for. The DBA's closing structural verdict ties the episode together: the architectures that worked, somebody could always name the workload in the first meeting. Nobody in this article has named the workload. Everything downstream of that — the launch costs, the cooling problem, the radiation, the maintenance impossibility — is consequence. Source article Report: Google and SpaceX in talks to put data centers into orbit [https://techcrunch.com/2026/05/12/report-google-and-spacex-in-talks-to-put-data-centers-into-orbit/] — Rebecca Bellan, TechCrunch · May 12, 2026 Panel * The Startup Founder * The Burnt-Out SRE * The Database Administrator * The Goat Farmer's Counsel