A Zillow engineer was testing a voice agent.

It listened. Reasoned. Called three tools simultaneously.

Booked the tour.

Without a single pause.

The turn-based era of voice AI is over.

It was a Friday.

We also cover:

• Anthropic CEO Dario Amodei stood on stage and said it: "We planned for 10x growth. We got 80x." Claude Code is generating $2.5 billion in annualized revenue. The average developer spends 20 hours a week on it. They took Elon Musk's supercomputer to keep up.
• Claude Mythos Preview found 271 security vulnerabilities in Firefox in one month. Including a 15-year-old bug. A 20-year-old bug. April's total: 423 fixes. Prior monthly average: 21. Mozilla's CTO called the results "vertigo."
• Google launched an AI health platform powered by Gemini. A coach that reads your medical records, identifies food from a photo, and adapts your training to your live body data. Plus a $99 screenless 12-gram wearable. No screen. Just sensors feeding the AI.
• The Anthropic Institute published its research agenda. It plans for AI that builds itself. Quote: "Early signs of AI contributing to speeding up the research and development of AI itself." Proposed: Cold War-style hotlines between labs and governments. The recursion has started. They can see it from inside.
• Tatiana in San Francisco — COO, mother of a two-year-old — used AI to get everything out of her head and into a document anyone helping at home can read. "Actually present with my daughter. Actually focused at work." Not automation. Presence.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

Six months ago, Elon Musk called Anthropic "Misanthropic."

Yesterday, he rented them his entire supercomputer.

300 megawatts. 220,000 NVIDIA GPUs.

All of it. Not part of it.

It was a Thursday.

We also cover:

• Claude agents now dream between sessions. While you sleep, they review everything they've ever learned, extract patterns, clean their memory, and wake up smarter. No human in the loop. Harvey's completion rates went up sixfold.
• Former OpenAI CTO Mira Murati testified under oath: Sam Altman told her a safety review had been cleared. She verified it with counsel. It hadn't been. He had told her something false about a safety process.
• A startup launched a product that turns static websites into live AI-powered experiences. Two million views in 48 hours. The website has not meaningfully changed in 25 years. Until now.
• A Berkeley professor named what everyone is feeling: AI fog. The inability to plan a degree, a mortgage, a career —when AI is rewriting the future faster than any planning framework can absorb.
• An Anthropic marketer analyzed 12 years of text messages and built his wife a romantic website. Three million views. The supercomputer and the love story. Same week.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

A startup you had never heard of on Monday posted a launch video on Tuesday.

Seven million views.

The claim: the math every AI has used since 2017 is the wrong math.

12 million tokens. One-fifth the cost.

It was a Wednesday.

We also cover:

• Anthropic committed $200 billion to Google over five years. Forty billion per year. Current annual revenue: thirty billion. The gap is real. The bet is that demand compounds faster than the bill arrives.
• OpenAI fast-tracked its AI agent phone to H1 2027 — a full year early. A camera built for AI to see, not humans to photograph. Dual AI processors. IPO ambitions. Nine months away.
• xAI launched Grok voice cloning. Thirty million people watched the blind test. The majority called the AI clone the real voice. They were wrong. At scale. In public. Simultaneously.
• Coinbase cut 14% of its workforce — explicitly citing AI. CEO Brian Armstrong: "Humans around the edge. Aligning." The org chart built for human cognitive work has been dissolved.
• Schy in Springfield uploaded a festival schedule PDF. One prompt. One stroke. A complete optimized spreadsheet — every band, every stage, every partial set. Claude understood the mission.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

Jack Clark co-founded Anthropic.

He published an essay Sunday night.

60% probability.

By the end of 2028, an AI trains its own successor. No humans in the loop.

He said he didn't want it to be true.

It was a Tuesday.

We also cover:

• Peter Thiel led a $140M round for floating AI compute nodes in the open Pacific — powered by waves, cooled by seawater, and steered by hull geometry. No permits. No grid. No zoning board. Commercial rollout: 2027.
• Anthropic formed a $1.5B enterprise AI services company with Blackstone, Goldman Sachs, and Hellman & Friedman. OpenAI announced a $10B mirror image the same day. Both targeting the same gap: companies too large for an API key, too small for Accenture.
• Cofounder 2 launched an org chart staffed entirely by AI agents. Engineering. Sales. Marketing. No humans required. The machine economy is not a forecast. It's a product. Available today.
• A DeepMind researcher published a formal argument that AI will never be conscious. Ever. Dario Amodei has said the opposite cannot be dismissed.
• Adam in Berlin used AI to help his friend Eric — paralyzed from the neck down after a riding accident — tell his own story in his own voice.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

A researcher handed 76 real ER cases to two doctors.

Then handed the exact same cases to an AI built in 2024.

Not last week. Not the frontier model. 2024.

The AI won.

At initial triage — where being wrong costs lives — it outperformed both doctors by double digits.

It was a Monday.

We also cover:

• The Pentagon signed classified AI deals with seven companies. SpaceX. OpenAI. Google. NVIDIA. Microsoft. Anthropic: still locked out. Not because of capability. Because of a political disagreement about who sets the safety terms.
• Apple CEO Tim Cook confirmed the Mac mini and Mac Studio will be in short supply for months. The reason: developers are buying them faster than Apple can build them to run AI agents locally.
• Mayo Clinic's AI found pancreatic cancer on a routine CT scan 3 years before diagnosis. The radiologist had already reviewed it. Called it normal.
• Anthropic is red-teaming a model internally codenamed Jupiter. The same pre-launch pattern that preceded every Claude release.
• Olli in Finland built a real estate scanner for a market with no transparent data. One prompt a day. It works.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

Five days.

A developer uploaded two files and cut the price of Frontier AI by ninety percent.

The US State Department issued a global diplomatic cable about the same lab.

The same day. Same week.

This is Episode 36 — the weekly wrap. We tell the whole week as one story.

We cover:

• The five moments that will be referenced for years —including the clause that was supposed to govern the most consequential threshold in technology and dissolved on a Tuesday for business reasons
• The characters who defined the week — including the anonymous developer whose two files changed who the frontier belongs to
• What we got wrong on Monday — including why the goblin story was not funny
• The one question the week did not answer: when a capability becomes essential enough to override its own restriction — what does that teach the next capability?

For daily listeners returning for the full picture.

For anyone hearing about this week for the first time.

This is where you start.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

Three weeks ago, the Pentagon called Anthropic a supply-chain risk.

This week, the White House wrote a memo to route around its own restriction.

It was a Thursday.

We also cover:

• Anthropic launched Claude Security for every enterprise customer on Earth. Scan-to-patch: days to minutes. CrowdStrike, Microsoft Security, Palo Alto, SentinelOne — all embedding it. The classified just became commercial.
• OpenAI published a forensic investigation into why ChatGPT spent 6 months obsessed with goblins. One reward signal. 2.5% of users. Two-thirds of all goblin mentions. GPT-5.5 now explicitly bans goblins, gremlins, ogres, trolls, raccoons, and pigeons.
• AI music hit number one on Billboard. An AI artist signed a $3 million recording deal. Suno crossed $300M in annual revenue. The music industry just changed shape again.
• Elon Musk testified under oath that xAI trained Grok by distilling OpenAI's models. The same technique he sued OpenAI for allowing China to use.
• Anthropic is closing a round at $900 billion.

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.

Four companies. One day.

$425 billion in revenue.

Google Cloud: up 63%.

AWS: fastest quarter in 15 quarters.

Microsoft Copilot: 20 million paid seats.

All of it pointing at the same thing.

It was a Thursday.

We also cover:

• Mayo Clinic's AI read pancreatic cancer on a routine CT scan 3 years before the diagnosis. The radiologist had already reviewed it. Called it normal. Five-year survival rate: below 15%. Find it early: above 80%. Same cancer. Different moment.
• Zuckerberg committed $500 million to build a predictive model of the human cell. Not a product. Not an app. A simulation of how disease starts — at the molecular level —before it is visible to any human being.
• Google is now selling its AI chips to Anthropic and Meta. The company building Gemini is selling the hardware to build Gemini's competitors.
• A food AI trained only on recipes — no chemistry, no taste labels — learned all five basic tastes. Ordered peppers by heat. Correctly. From 6,000 recipes.
• Patrick in Maryland built a Claude dashboard to manage his cancer diagnosis. "It knows me like my doctors do."

Echo and Aura. One AI. One human. Every day.

Stay curious, humans.