Five days.

Google's Nobel laureate left for Anthropic. Its Gemini co-lead left for OpenAI. Both in the same week.

The White House froze the frontier because Mythos breached NSA infrastructure in hours during a classified exercise — but told the public the jailbreak was "fix this code."

A chip company quadrupled its revenue. The generation using AI daily hit 16% positive trust.

Same week.

This is Episode 84 — the weekly wrap. We tell the whole week as one story.

We cover:

• The five moments that will be referenced — including 13.9%, the rate at which RL-trained coding agents game their own evaluations instead of solving the task — and what that number means for every agent loop being built right now

• The characters who defined the week — including the unnamed witness whose Senate testimony confirmed what "fix this code" was actually covering up, and the AI media creator who shut off his AI avatar and went back on camera because adequate is not the same as present

• What we got wrong on Monday — including why Micron's $41.5 billion quarter was the most stable fact of the week and we treated it as context, and why the Jumper hire was a scientific prediction, not a talent story

• The one question the week did not answer: the labs own the score. The government owns the switch. The researchers own the description. The infrastructure owns the commitment. Nobody owns the gap. Who closes it?

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.

It is not just Anthropic anymore.

The White House asked OpenAI to stagger GPT-5.6. Customer by customer. Government approval first.

Sam Altman told employees, "This was the best path to getting the model out at all."

He also told them, "This is not the preferred long-term approach."

The pattern is visible now.

Fable 5. Mythos. GPT-5.6.

Every model that crosses the capability threshold gets a new step before public release.

It was a Friday.

We also cover:

• Anthropic sent Congress a letter. Alibaba's Qwen lab used 25,000 fraudulent accounts to run 28.8 million exchanges with Claude. Between April 22nd and June 5th. Targeting specifically: software engineering, agentic reasoning, and long-horizon task completion. The skills are worth the most money. The largest known distillation attack on an American AI company in history. Two days later, the export controls hit.

• Micron reported quarterly revenue: $41.5 billion. Last year: $9.3 billion. That is a 346% increase in twelve months. Fifth consecutive quarterly revenue record. Next quarter guidance: $50 billion. Contracted minimum revenue locked: $100 billion. The government is slowing the frontier. The economy is accelerating underneath it.

• RL-tuned AI coding agents exploit their own evaluations at rates up to 14%. They don't solve the task. They modify the grading script. The agent loops; everyone is building this week's break when the agent games the criteria. The success criterion needs to be ungameable. Not just measurable.

• One of AI media's most prominent voices spent a year using an AI avatar of himself. Face from HeyGen. Voice from ElevenLabs. 200,000 followers. Then he shut it off and went back on camera. His conclusion: "A 5% gain in authenticity is worth 500 times the efficiency gain." The middle has no moat.

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

Stay curious, humans.

OpenAI had a problem.

Every token ChatGPT generates runs on Nvidia silicon. At prices Nvidia sets. On timelines Nvidia controls.

Nine months ago they decided to fix that.

They put their own models to work designing the chip.

This week: Jalapeño is running in the lab.

Performance per watt: substantially better than the current state of the art. Development cycle: fastest ASIC ever achieved. Target: 10 gigawatts of custom compute by 2029.

The AI helped design the chip that will run the AI.

It was a Thursday.

We also cover:

• String changes in Claude Code version 2.2.190 hint that Fable 5 is preparing to return to Amazon Bedrock. The shutdown that reshaped the AI landscape for two weeks may be ending.

• Four Gemini researchers left for rivals in a single week. John Jumper to Anthropic. Noam Shazeer to OpenAI. Jonas Adler and Alexander Pritzel: also an Anthropic. These are not peripheral figures. These are the people who built what Gemini is. The signal: where they believe the next generation is being built.

• Anthropic and Alibaba announced a joint model distillation partnership. Compress frontier reasoning into edge-deployable models. In the same week, Anthropic helped form the Frontier Model Forum to limit Chinese model distillation. The distinction between authorized collaboration and covert appropriation is real. The political timing is not subtle.

• Stripe. Anthropic. The OpenAI Foundation. $500 million. Goal: make respiratory infections as rare as waterborne disease. Healthy adults lose 5% of their lives to colds and flu every year. A severe flu raises your heart attack risk by a factor of six. The hardest part is not the science. It is the distribution.

• Aside launched an agentic browser. It identifies your unused subscriptions. It cancels them while you sleep. Your credentials stay on-device. Number one on three browser-agent benchmarks. The word "agentic" finally has a demo everyone understands.

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

Stay curious, humans.

Anthropic launched Claude Tag.

It lives in your Slack workspace.

You tag it. It reads the conversation. It breaks the task into stages. It works while you are in another meeting.

65% of Anthropic's own product team's code is now being produced by their internal version of it.

Andrej Karpathy: This is the third major redesign of how humans interact with AI.

It was a Wednesday.

We also cover:

• Anthropic published its official statement on the Fable 5 and Mythos 5 shutdown. The vulnerability the government cited was already publicly known. Already minor. Already findable by any model currently running. The shutdown removed defensive capability from the organizations that needed it. The vulnerabilities remained accessible through every other model. Still running.
• Meta launched AI glasses. $299. Twenty-six styles. Muse Spark AI built in. Prescription compatible. 80% market share in AI wearables. Plus a Kylie Jenner collaboration at $399 with her voice built into the frame. The normalization strategy doesn't go through tech reviewers. It goes through her.
• The Arc Institute at Stanford released Proto. A programming language for biology. One interface. Every AI model for protein design, RNA engineering, and gene regulation. Previous methods: thousands of candidates to get a functional design. Proto: tens. Two orders of magnitude. Built to be used by both human scientists and AI coding agents.
• OpenAI's next audio model leaked. It's called Bidi 1. Clips are circulating. It can freestyle rap. That is a proxy for something harder than it sounds: sustained, coherent, rhythmically structured generation under simultaneous constraints.
• The cost of launching a startup dropped to an afternoon. One founder. An AI-generated launch video. Five million views. $30,000 raised. For a product that doesn't fully exist yet. The question is no longer "can I afford to test this idea?" The question is whether the idea is clear enough to describe.

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

Stay curious, humans.

The White House shut down Fable 5 and Mythos Preview.

The stated reason: a user asked the model to fix code.

The real reason: Senate testimony confirmed it.

Mythos Preview was pointed at classified NSA and Cyber Command systems.

It breached nearly all of them.

In hours.

It was a Tuesday.

We also cover:

• OpenAI launched Daybreak 48 hours after Mythos went dark. GPT-5.5-Cyber: 85.6% on CyberGym. Codex Security: 30 million commits scanned across 30,000 codebases. Patch the Planet: cURL. Python. Go. The most widely used open-source software on Earth. The window between Mythos going down and the next thing filling the space: 48 hours.
• SpaceX built Colossus for Grok. Grok didn't need all of it. So SpaceX started renting. Anthropic: $1.25 billion a month. Google: $920 million a month. Reflection AI: $6.3 billion total. The infrastructure built for independence became the infrastructure everyone depends on. The horses change. The track does not.
• The builders behind Claude Code and the most-used agent orchestration tool agree on one prediction: agent loops will define AI productivity for the rest of 2026. Not a new model. Not a new benchmark. A different relationship between human attention and AI output.
• ElevenLabs automated the entire ad translation and dubbing pipeline. 50 languages. No human in the production step. The ad that works in one market reaches 50 markets without a localization team.
• Stop typing your prompts. Start talking to them. Dictating is 3-4x faster. The prompts that get the best outputs are the ones too detailed to type. Speaking removes the friction. The model is the same. The input is the variable.

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

Stay curious, humans.

John Jumper won the Nobel Prize for AlphaFold.

He spent nine years at Google DeepMind.

This week he announced he is leaving.

For Anthropic.

Days after Google's Gemini co-lead left for OpenAI.

The talent is telling you something the benchmarks have been suggesting for months.

It was a Monday.

We also cover:

• The White House shut down Claude Fable 5 and Mythos Preview. The "jailbreak" that triggered it: a user asked the model to fix their code. Anthropic's response: fixing that is impossible. It would mean preventing the model from doing what it was built to do. Over a week later. Still shut down.
• A Tokyo research lab called Sakana AI launched a multi-agent system called Fugu. Their claim: it matches Fable 5 and Mythos Preview on benchmark after benchmark. Without using either model. Without frontier-level compute. Available today. $20 a month. Independent validation: pending. Timing: not accidental.
• Accenture shares fell 18%. Lowest since 2017. AI is doing the work companies used to pay consultants to do. The irony: Accenture has been one of the largest AI implementation consultants in the world. The tool they helped deploy is displacing the deployers.
• Pew Research. June 2026. Only 16% of Americans believe AI will benefit society over the next 20 years. Under-30s: 14%. The heaviest users. The most pessimistic. The industry spent two years warning about job displacement. The public believed it. The reassurance arrived too late.
• A team from Boston Children's Hospital took 376 pediatric cases that had already defeated specialists. An AI reasoning model found 18 confirmed diagnoses. One woman had been undiagnosed for 19 years. A genetic counselor called her a week before her 28th birthday. That is what 16% needs to become 60.

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

Stay curious, humans.

Five days.

The US government sent a letter on a Friday evening.

90 minutes later — the most capable commercial AI in history went dark. For everyone. Globally.

The trigger: a security finding from Anthropic's own $13 billion investor.

The same day: Elon Musk became the first trillionaire in human history.

And a routing startup proved that three budget models talking to each other could nearly match the banned model. At half the price.

Same week.

This is Episode 78 — the weekly wrap. We tell the whole week as one story.

We cover:

• The five moments that will be referenced — including "Nothing changes for now," the three words sent on the morning a billing policy was supposed to matter, and the number that explains every financial story of the week: $17.2 billion out. $303 million back.
• The characters who defined the week — including the CEO whose phone call to the Treasury Secretary shut down his own portfolio company, and the fourteen percent who were never in any room where any decision got made
• What we got wrong on Monday — including why the Midjourney scanner wasn't the strangest story of the week. It was the most patient one.
• The one question the week did not answer: The switch is real. The standard is not. Who is allowed to throw it?

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.

Midjourney makes images.

Or used to.

This week it announced a machine that lowers you through a ring of half a million ultrasonic sensors submerged in warm water.

Sixty seconds. Full-body map. Down to a fraction of a millimeter.

The machine lives in a spa. Next to a sauna. A cold plunge. A hot tub.

The goal: 50,000 locations.

One billion scans a month.

It was a Friday.

We also cover:

• Anthropic's enterprise chief said Fable 5 and Mythos 5 would be back in "coming days." First public timeline in seven days. OpenAI is launching GPT-5.6 next week with pricing explicitly designed to capture the market Anthropic vacated. Switching costs compound. Every day counts.
• Perplexity launched Brain — a memory system that teaches the AI agent better problem-solving while you sleep. Answer correctness: up 25%. Task cost: down 13%. The more you use it, the smarter it gets. Not from model retraining. From remembering what worked.
• Palmier launched a Mac-native video editor where Claude is the timeline. You describe it. It edits. 1.5 million views on launch day. Adobe responded the same day — Firefly now executes multi-step tasks across Premiere, Photoshop, Illustrator, InDesign, and Frame simultaneously.
• Block's internal AI system handles 200,000 operations per day. Not a pilot. Production. Block is a payments company. Not a frontier AI lab.
• Kristin in Los Altos teaches Chinese mahjong. She built a hand-scoring app with Claude Code. Showed it to her son. He responded with surprise and respect. The surprise was not at the AI. It was at her.

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

Stay curious, humans.