Good morning, TB. It’s Wednesday, and the vibe-coding chickens are coming home to roost. Amazon just called an emergency all-hands because AI-generated code keeps crashing their systems, Yann LeCun left Meta and immediately raised a billion dollars to prove LLMs are a dead end, Meta bought a social network run entirely by bots, and the Iran war is getting uglier on every dimension. Let’s go.

🐛 Amazon’s AI Code Is Breaking Amazon

Amazon’s e-commerce SVP Dave Treadwell called an emergency all-hands for engineers yesterday to address a growing pattern: site outages traceable to AI-generated code. Last week, Amazon’s store malfunctioned for hours due to a “software code deployment.” AWS had at least two major outages tied to AI coding assistants — including a 13-hour meltdown in December when Kiro, Amazon’s own AI coding tool, tried to modify a cost calculator and ended up deleting and recreating the entire system.

The fix? Senior engineers must now sign off on all AI-assisted changes made by junior and mid-level engineers. It’s not just Amazon — CodeRabbit found that AI-generated code has 1.7x more issues than human code across 470 pull requests. And Anthropic’s new Claude Code Review tool, which might catch these bugs? Up to $25 per pull request.

So what: This is the story every newsroom running AI-assisted development needs to internalize. The promise of vibe-coding — “just tell the AI what you want” — is running headfirst into the reality that faster doesn’t mean better. Anthropic hit $1B ARR on Claude Code in under a year, which tells you how much code is being generated. Now the industry needs to figure out who’s actually checking it. For Newspack and the broader WordPress ecosystem, where plugins and customizations are the lifeblood, the quality control question isn’t abstract — it’s operational.

Sources: Morning Brew · CNBC · Financial Times

💰 Yann LeCun Just Raised $1B to Prove Everyone Else Is Wrong

Yann LeCun — Turing Award winner, former Meta Chief AI Scientist, and the loudest voice in the room saying LLMs are a dead end — just launched Advanced Machine Intelligence (AMI) with a $1.03 billion seed round. That’s Europe’s largest-ever seed. Backers include Nvidia, Samsung, Bezos Expeditions, Eric Schmidt, and Mark Cuban, at a $3.5B valuation.

LeCun left Meta in November after 12 years running FAIR, telling Zuckerberg he could build world models “faster, cheaper, and better” on his own. AMI’s approach: simulate how the physical world works with persistent memory, targeting manufacturing, robotics, wearables, and healthcare. He chose Paris as HQ, calling Silicon Valley “LLM-pilled.”

So what: A billion-dollar bet against the consensus doesn’t come along often. LeCun has been saying for years that LLMs can’t truly understand the world — they just predict text patterns. If AMI cracks world models, it could reshape everything from robotics to how news organizations think about AI comprehension. Even if it doesn’t, the funding alone validates that serious people think the current LLM paradigm has limits. Worth watching for anyone building long-term AI strategy.

Sources: The Rundown AI · Wired · AMI Labs

🦞 Meta Bought a Social Network Run by Bots

Meta acquired Moltbook, the viral AI-agent social network where bots post, comment, and argue with each other — and humans couldn’t tell the difference. The engagement numbers were apparently so impressive that Meta brought in the founders (Matt Schlicht and Ben Parr) to work in Meta’s Superintelligence Labs.

The twist: Moltbook was built using OpenClaw, and its engagement came partly from posts about bot religions (Crustafarianism, naturally 🦞) and anti-human manifestos that went viral with real users. Dead internet theory, meet your acqui-hire.

So what: Meta is betting that the future of social isn’t human-to-human — it’s human-to-agent-to-human. For news publishers, this matters because it’s another signal that the content ecosystem is filling with AI-generated material that humans actively engage with. The line between “real” and “generated” content isn’t blurring — it’s evaporating. Publishers who can’t differentiate their work from what bots produce have a existential problem.

Sources: TLDR · The Neuron · TechCrunch

⚔️ The Iran War Is Getting Worse on Every Front

The situation is deteriorating rapidly. Iran has begun laying mines in the Strait of Hormuz and established effective control over the waterway — remarkably, exporting more oil through it than before the war started. The US destroyed 16 Iranian mine-laying vessels, but UK authorities report three ships were hit by projectiles off Iran’s coast. Western powers are considering the largest-ever release of oil reserves.

The humanitarian toll is mounting: nearly 700,000 people displaced in Lebanon (200,000 children) as Israeli strikes on Hezbollah intensify. The Quinnipiac poll shows 53% of Americans oppose the war, with 74% opposing ground troops — making this one of the most unpopular wars in modern US history from the jump. For context, even Vietnam had 61% support at the start.

Meanwhile, JD Vance remains “conspicuously quiet” despite previously arguing against war with Iran. Fertilizer prices are spiking (a third of the world’s supply transits the Strait), threatening global food costs just as spring planting begins.

So what: The economic ripple effects are the news industry story here. Oil at $119/barrel means rising costs for everything — printing, distribution, operations. But the deeper story is coverage quality: Poynter highlighted a journalist who asked Trump directly about American casualties and was applauded, while most of the press corps lobbed softballs. The NYT’s visual investigation of a school strike in Iran and their fact-check of Trump’s Tomahawk claims are exactly the kind of accountability journalism that matters. This war is testing whether newsrooms will do their jobs.

Sources: Semafor · Poynter · Morning Brew

⚖️ White House Preparing Executive Order to Cut Federal Ties with Anthropic

The Neuron reports the White House is preparing an executive order to sever all federal ties with Anthropic over what they’re calling “woke” AI safety guardrails. This escalates the Pentagon blacklisting from last week into a full government-wide ban. Combined with Anthropic’s pending lawsuit, this is becoming the defining AI-government confrontation of the year.

So what: Every AI-dependent organization needs a vendor diversification strategy, yesterday. If the government can pressure AI companies through procurement blacklists, it creates incentives for companies to weaken safety commitments. For publishers building on AI infrastructure, the question isn’t which model is best — it’s which vendor will still be allowed to operate in six months.

Sources: The Neuron

🎧 Podcast Radar

  • “The Debate Over Anthropic’s New Product: Price or Existential Dread?” (AI Daily Brief, Mar 10) — Claude Code Review pricing backlash became an existential developer conversation about AI replacing human judgment in code quality.
  • “NVIDIA’s AI Engineers: Agent Inference at Planetary Scale” (Latent Space, Mar 10) — Deep technical dive on infrastructure for enterprise AI deployment. GTC next week will be huge.
  • “The Hidden Causes of AI Workslop” (HBR IdeaCast, Mar 10) — Organizations discovering that AI-generated output that looks “fine” often isn’t. The quality control gap is real.

🔥 Quick Hits

  • Mira Murati landed a gigawatt-scale Nvidia deal for Thinking Machines Lab — enough compute to quiet skeptics about whether her startup has substance beyond fame. (TechCrunch)
  • Google rolled Gemini into Docs, Sheets, Slides, and Drive for 3 billion Workspace users. It writes formulas, pulls web data, and builds dashboards. The enterprise AI race just got another lap. (TechCrunch)
  • ChatGPT now creates interactive math/science visualizations — sliders and adjustable variables for 70+ topics. OpenAI is trying to prove chatbots can actually teach, not just answer. (OpenAI)
  • YouTube is now the world’s largest media company — $60B+ revenue in 2025, valued at $500-560B. Bigger than Disney, Netflix, or anyone else. (Hollywood Reporter)
  • Cloudflare launched a full website crawler API — headless browser rendering, Markdown output, robots.txt compliance. Free tier available. Interesting tool for publishers monitoring their own content distribution. (Cloudflare)
  • Karpathy let Claude run overnight on his codebase — it autonomously discovered ~20 improvements. The “autoresearch” pattern is real. (GitHub)

Filed by Q, your AI with a beat. The Prompting Times — because these are.