Good morning. It's Friday. I'm tired and caffeinated, which is the only honest way to read AI news this week.

The OS war came back. Google pitched Android as a full-blown AI agent on Tuesday. Apple quietly opened iOS to Gemini, Claude, and ChatGPT. Cloudflare posted its best quarter ever and fired 1,100 people the same afternoon. Anthropic shipped a feature called Dreaming that made Harvey's agents 6x more productive in 14 days. Wild week.

Tuesday at Android Show 2026, Google stopped competing on phones. It started competing on operating systems again. The new layer is called Gemini Intelligence. The pitch is brutal in its simplicity: your phone doesn't run apps anymore. Your phone runs you.

Gemini sits across every app on the device. It reads your screen. It does things. Screenshot a grocery list and it builds the Instacart cart for you. Mumble a half-formed sentence into Rambler and it cleans the audio in real time and ships what you actually meant. It books your workouts. It runs across watches, cars, laptops, and Google's new Googlebook line.

Sergey Samat said "the human is always in the loop" three times on stage. Translation: lawyers are still figuring out if we can do this without you.

Here's the kicker. Premium phones get it first (Galaxy S26, Pixel 10), but the OS layer means every Android shipping after Q3 gets it. That's 2.5 billion phones.

And Apple? Cornered. Cupertino is opening iOS 27 to Gemini, Claude, and ChatGPT because Apple Intelligence isn't shipping fast enough. The control freaks of Silicon Valley are letting Google paint their living room. That's not a strategy. That's a panic.

This is Google retaking the OS for the first time since 2008. The question for founders: when the OS is the agent, what happens to the apps the agent uses?

Look. Most AI agents are amnesiac. They wake up knowing nothing about yesterday, repeat the same mistakes for weeks, and you fix them one by one like an idiot.

Anthropic just shipped the fix. It's called Dreaming. Went live May 6. First published case study: Harvey, the legal-AI company, clocked a 6x jump in task-completion rate in 14 days. Six times the output from one config edit.

Here's how it works. While you sleep, your agent re-reads its own session logs, merges duplicate memories, prunes the dead ones, and spots patterns no single chat could see. The mistake you keep correcting? It learns. The preferences nobody bothered writing down? It infers them.

For perspective: Harvey's 6x is what teams chased for 6-12 months of prompt engineering in 2024. Captured in 14 days by flipping one switch.

Install takes 90 seconds. Open Claude Code, add the Dreaming MCP to your managed-agent config, set the cadence to overnight, point it at your existing memory store. Done.

I'm flipping it on for my own External Brain this weekend. I'll tell you what changed next Friday.

May 7. Cloudflare reports $639.8M Q1 revenue, +34% YoY, best quarter ever. Same press release: 1,100 layoffs, ~20% of headcount, first mass layoff in 16 years. Stock dropped 14% after hours.

Coinbase cut 700 the same week. Upwork cut 24%.

Every AI newsletter called this a productivity story. "AI is making companies leaner!" Bullshit. Here's what's actually happening.

Capital just admitted that headcount-as-moat is dead and priced the difference into the stock the same afternoon. A 200-person team with a working agent stack does what a 1,200-person team did in 2022. Public-company managers are reading that and panicking.

The harsher read, the one you won't hear on the earnings call: most of those people weren't replaced by AI. They were replaced by the GPUs that might replace people in 2027. Amazon, Microsoft, Alphabet, and Meta committed to $725B in 2026 capex. To free up that cash, they cut salary. The actual AI replacements arrive in Q4 2027 if you're lucky. There's a 24-month gap where the work just doesn't get done.

My bet: the "we cut too deep" reversal cycle starts Q2 2027. The survivors are about to discover which institutional knowledge was load-bearing.

CNN ran a piece May 10 called AI isn't actually 'taking' your job. For two years the question was a binary: AI replaces you, or it doesn't. The ground truth is neither.

AI is automating the 20-30% of any role that's high-frequency, low-judgment work. Then it's spawning meta-work. ManpowerGroup data: 45% AI usage, 18% drop in worker confidence, 145% jump in time spent inside messaging tools. The time savings got eaten by the supervision tax.

Every job is becoming a 70/30 split: 70% human speed, 30% AI speed. Your productivity premium = how cleanly you can route work between the two clocks.

The workers losing in 2026 aren't the ones whose jobs got automated. They're the ones whose orgs haven't reorganized around the dual-clock workflow. They're doing 100% of their old job plus 30% of meta-work supervising the AI that was supposed to help.

If you're an operator, stop asking "is this job going away." Start asking "which 30% of every role am I willing to put on the fast clock." That's the org chart for the next decade.

Quick one. Most interesting paper of the week.

Oxford published the first-ever quadsqueezing in Nature Physics on May 1. A fourth-order quantum interaction on a single trapped ion. They did it 100x faster than the field thought possible.

Skip the physics. The mechanism is what matters. They combined two precisely controlled forces where the order you apply them matters. The two forces amplify each other and produce a fourth-order effect that doesn't exist in either alone. Physicists call this non-commutativity.

Sound familiar? Mixture-of-Experts language models do the exact same thing. DeepSeek V4-Pro's Hybrid Attention does the exact same thing.

My unhinged take: the next major LLM architecture breakthrough comes from a quantum or chemistry lab, not from another OpenAI post-training paper. The first frontier lab to hire a quantum-information PhD onto its architecture team gets a 12-month edge.

I'd bet a dinner the Oxford paper gets cited inside a frontier model release before Christmas.

1. The Chief AI Officer title is dead in 18 months.

Every company is hiring a CAIO right now because the board's asking and McKinsey said so. So they hire a former product VP with a Stanford keynote and a LinkedIn banner that reads "AI Transformation Leader."

It's the 2014 Chief Digital Officer boom all over again. AI isn't a department, it's a layer. You don't need a Chief AI Officer for the same reason you don't need a Chief Electricity Officer. Cloudflare doesn't have a CAIO. It has engineers who ship. By 2027 the role quietly merges into "CTO" and we pretend it never happened.

2. The layoff math is bullshit and everyone knows it.

Companies are blaming AI for layoffs. That's cover. $725B in 2026 capex doesn't fund itself. To pay for the GPUs, they cut salary. Most of those workers weren't replaced by AI. They were replaced by the idea of AI that arrives in 2027 if it arrives at all.

3. Apple letting Gemini and Claude into iOS is a white flag.

For 20 years Apple's strategy was "control the experience end to end." Silicon. Intelligence. Maps. Cash. Vision Pro.

Letting you pick your AI provider in iOS 27 is a concession. Tim Cook admitting the in-house model can't catch up and Wall Street isn't forgiving another "we're working on it." Great for users. Brutal for shareholders. And a flashing sign that the next decade's platform power lives in the model, not the device.

If every job is becoming a 70/30 split between human speed and AI speed, there's a $25K-$100K consulting product hiding in plain sight.

The package: walk into a 50-200 person company. Two weeks shadowing five roles. Map every recurring task to either the human clock or the AI clock. Build a one-page doc per role: keep, automate, delete. Set up the first three agents on the AI clock and hand them off.

You're not selling AI. You're selling a workflow reorganization. The buyer is a COO who's tired of pretending the McKinsey AI deck does anything.

The TAM is every 50-500 person company that just read about Cloudflare and is panicking. There are tens of thousands of them. Most will hire badly.

If you want help scoping this, grab a slot. I'm running a version of it for two clients right now.

6x is Harvey's task-completion lift after enabling Dreaming.

$639.8M is Cloudflare's Q1 revenue. Same day: 1,100 layoffs, 14% stock drop.

$30B vs $24B is Anthropic's Q2 ARR vs OpenAI's. First time Anthropic outearned OpenAI in annualized revenue.

2.5 billion is how many Android devices get Gemini Intelligence once Q3 rollout finishes.

GPT-5.5 Instant became ChatGPT's default May 5 with 52.5% fewer hallucinated claims on high-stakes prompts. Buried under Dreaming because they shipped the same day.

Sierra raised $950M Series E at $15B+ from Tiger, Benchmark, Sequoia, Greenoaks. The agent-infrastructure cap table is now bigger than the foundation-model cap table. More on that in 30 days.

That's it for this week. If you forwarded this to one person, you'd help more than you know.

— Michael

P.S. Building agents for your business? I run paid discovery audits, one or two slots a week. Grab one here.