Quick Hits

• @ericwilliamrea says Podium went from engineers averaging about 5 PRs a week to several shipping over 100, with AI reviewing 100% of PRs and auto-approving 60%, after a full CI rebuild cut median cycle time from roughly 1,000 minutes to 200.
• @JoshuaSWarren reports that an AI agent pointed at his omarchy-apple-dev repo built and notarized a macOS app on a Linux machine in 10 minutes and passed Apple review, no Mac required.
• Emulation went viral: @Silverhand83 says Sword's open-source PS5SX2, which boots PS2 discs on a PS5, is drawing millions of views, while @pablowasserman asks whether emulators are now "over" given @aisearchio's "huge database of decompiled games."
• @BayouTD predicts "At some point Slopcore will be nominated for a Grammy," reacting to @AstrablueAI's "KILL SWITCH" video ("She passed every AI safety test. She was lying."), made entirely from prompting with zero human edits.
• @JasonBassler1 asks Grok whether Zuckerberg, Thiel, and Sam Altman are buying apocalypse insurance and building bunkers. It is a question posed to a chatbot, not reporting, and the post offers no evidence.

Agents Ship More, and the Pipeline Becomes the Constraint

The day's most detailed engineering claim is that AI moved the bottleneck off engineers and onto infrastructure. @ericwilliamrea describes a three-year effort at Podium: standardized AI coding workflows across the org, AI review on every PR with 60% auto-approval, and WALL-E, an internal agent that diagnoses CI failures, reruns jobs, and applies fixes without an engineer. The stated results are a median PR cycle time down from about 1,000 minutes to 200, a change failure rate below 1%, and a goal of shipping 10x more code this year than last. All of it is self-reported by the company, but the shape of the story, throughput outrunning the pipeline, is the interesting part.

@bibryam shares what he calls Anthropic's AI-native SDLC playbook, a governance counterpart to the same problem: versioned intent, specs, and plans; skills for policy with hooks for enforcement; continuous evals with gated deployment. Read together, the two posts describe the same shift from different ends, one from volume, one from process control.

@CXCarroll adds the market framing, arguing "the vast, vast majority of investors have no idea what just happened to software in the last five days" and that it will "hit like a tsunami." He never says what happened in those five days, so this is assertion without receipts, useful mainly as a sentiment data point.

An Agent Passes Apple Review Without a Mac

@JoshuaSWarren's post is the day's most concrete agent demo: someone handed their agent his omarchy-apple-dev repo, and ten minutes later it had built and notarized a macOS app on a Linux host, then passed Apple review. @oomathias, whose post JoshuaSWarren quotes, attributes the run to "Sol" and calls it a sign of "amazing times." If accurate, it removes a long-standing hardware requirement from the Mac release path.

Take the fleet-size claims with more salt. @supremaxbt promotes a workshop quoting a SpaceXAI engineer, Lauren Tan, who says she runs 21+ GrokBot agents with /loop and /goal, that 85% of engineers there already run teams of 10+ agents, and that a Chief of Staff agent coordinates everything. The post is marketing a free course, attributes its numbers to a single quoted engineer, and offers no verification. Treat it as aspiration, not evidence.

Interviews Compress to System Design Plus an Agent Build

@poteto (lauren) argues SWE interviews could collapse into two technical rounds: system design, to test whether you can articulate and engineer a solution, and an onsite project building "a real thing with agents," to test whether you can turn intent into outcomes. In her framing, "everything else we used to ask is no longer necessary." She is quoting @GeoffreyHuntley's post that AI use "is no longer optional if you wish to be employed," which also promises advice on hiring AI-first candidates in 2026. @badlogicgames separately flags Huntley's piece as recommended reading, so it circulated twice today.

@jacobgold undercuts the agent-first persona with a joke, posting that "@poteto is a fraud" because "she's reading all of the code." The tease lands on a real tension in the hiring debate: the people advocating agent-driven workflows still audit what the agents produce.

@MakeAI_CEO quotes @satyanadella's post titled "Models as Insider Risks in the Super Intelligence Era" and claims, in Japanese, that anyone who understands it will stay employable for 30 years. The post is even tagged as a hype claim in its own metadata, which is about as honest as hype gets.

Local AI Reaches Laptops, Phones, and ESP32s

The on-device thread runs from workstations to microcontrollers. @ClementDelangue amplified a report that clef-flash was distilled into a 0.6B model with an Apple Neural Engine option, with 1,000 support tickets triaged on an M5 Pro "in 57" (the unit is cut off mid-sentence in the repost). The direction matters more than the truncated benchmark: real triage work on a laptop.

@0xSero points Omarchy users to @IbrahimSait_'s "Omarchy x Local AI" demo, captioned "It's never been this easy." Omarchy shows up twice more today: @dhh says he built a dedicated monologue-recording app for it and will drop OBS as a default, since recording monologues was all he used OBS for. His post quotes @LundukeJournal's older story about OBS declining to support Rumble, but dhh's stated reason is scope, not politics.

Two hardware posts are enthusiasm without artifacts: @QingQ77 says (in Chinese) that the ESP32 feels like it is becoming a "magic tool," with video thumbnails but no details, and @derekmross declares smartphones "will never be the same again" across three screenshots, also without specifics. Interesting to watch, unverifiable as written.

Practical Takeaway

If your team is scaling agent-assisted coding, the Podium post suggests measuring where your constraint actually sits now: PR cycle time, review capacity, and CI queue length, not lines written. Before copying anyone's numbers, whether Podium's 60% auto-approval or a promo thread's agent fleet, instrument your own pipeline and track change failure rate as the counterweight. The open question worth investigating is whether your review and CI setup can absorb a 5x to 10x increase in merge volume, because that is the failure mode these posts describe.