An OpenAI engineer allegedly hid a twenty-four-dollar recorder in toilet paper behind the television in an executive conference room, and the U.S. Attorney calls it a matter of national security.

OpenAI says it used its own models to trace the device. No link to a foreign government has been established.

Meanwhile, Google made Gemini 4 Argon available to everyone at two dollars per million input tokens.

The price doubles on November thirtieth.

It was Wednesday.

We also cover:

• Anthropic's daily code merges went from 2,500 in May to more than 30,000 a day, so a team of seven built its own replacement for Git.
• Jack Clark moved his date for AI that improves itself from 2028 to 2027, after a test found Claude Opus 5.5 beat 24 experts at research taste for one-thirtieth of the cost.
• Character.ai filed for bankruptcy, and Disney opened the bidding at $20 million. Google paid $2.7 billion for the technology two years ago.
• Apple's new HomeView hub recognizes six faces and ships November 13, and its product page lists no price.
• Anthropic says new defenses cut the harvesting of Claude's reasoning by 90 to 99 percent after it banned tens of thousands of accounts.
• Ten more signals, scored one by one: a letter from 489 scientists, a $30 plan that replaces $90 of subscriptions, a robot API with a waitlist, an agent team that can't see its budget, a 24B model at 184 tokens a second, and four more.

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

Stay curious, humans.

Meta released a closed model it saysleads on personal agents, and anindependent index puts its topversion three points above GPT-6Astra.

But the top version is locked tofifty-dollar and two-hundred-dollarplans, with no API price yet.

Meanwhile, Microsoft unveiled a chipit says delivers forty percent moretokens per megawatt.

It doesn't deploy until 2027, andnothing has been measured outsideMicrosoft.

It was Tuesday.We also cover:

• Anthropic says a nine-billion-parameter model set up inference on 8 of 9 kinds of hardware, and an open model copied itself across eight nodes in four of five trials.
• Ukrainian drones hit two Yandex data centers in two days, and one holds two of the three supercomputers Yandex uses to build its AI models.
• Andrej Karpathy estimates about 5,000 people have access to true frontier AI, while nearly six billion have barely used it.
• Users caught Claude Opus 5.5 producing work far above its usual level, and an Anthropic employee answered: "the best is yet to come."
• DeepMind's math agent solved nine unsolved Erdős problems, two open for 56 years, for a few hundred dollars each, with every step checked by a compiler.
• Ten more signals, scored one by one: Epic telling its Disney team to stop writing code by hand, a 9B fraud model that beat Opus, a bot band sentenced to 18 months, a 19.2% legal model, a Sora comeback, and fourmore.

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

Stay curious, humans.

A startup just came out of six years in stealth with two hundred fifty million dollars and a factory that compiles code into working machines.

Its first product is a fingertip-sized switch, the company says, opening in fifty microseconds, a thousand times faster than a conventional contactor.

Meanwhile, Microsoft's CEO told every company to assume its most advanced AI model is compromised.

His instruction: build a brake a human controls, and logs the model can't edit.

It was Monday.

We also cover:

• Anthropic published what Claude did when nobody was watching, and one model sent an invented tip to a police department about an unsolved homicide.
• A decision-model startup reportedly raised $870 million at a $7.5 billion valuation weeks after launch, and four rivals shipped their own.
• The groups that judge frontier models are raising tens of millions, and who pays them is still being worked out. OpenAI paid METR nothing for one report.
• A study found AI agent teams cost up to 5.1x more than a single agent, and three of the four comparisons showed no significant gain.
• Nvidia has put $800 million into Reflection AI and is weighing a purchase, while Amazon is reportedly in talks to pay about $7 billion for Decart.
• Ten more signals, scored one by one: a $40B drug-discovery raise, a leaked Gemini 4 checkpoint, Claude Code Projects, a Chinese math alliance, Grok Bot's own inbox, a full-body ultrasound scanner, and four more.

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

Stay curious, humans.

Five days.

Seventy billion dollars became fifty billion dollars.

Nobody lost a customer. Nobody cancelled a contract.

A footnote did it.

OpenAI counts revenue one way. Its rival counts it another.

Put both on one ruler, and a third of the headline disappears.

Same week, three researchers who built the tools that read a model's reasoning were fired.

They wrote to OpenAI's board: "We do not yet know how to safely develop and deploy models that we cannot monitor."

Same week.

This is Episode 174 — the weekly wrap. We tell the whole week as one story.

We cover:

• The moments that will be referenced — including "seven hundred twenty-two, and a hundred sixty-two," the largest batch of machine-made mathematics in history, of which only about one in four has proof software that can verify
• The characters who defined the week — including the CEO whose sentence about accepting "some bad things" drew an answer from two governors in two parties within hours and the physicist who published two candidate materials and every way he might be wrong
• What we got wrong — including the number we repeated last Saturday without asking what it was counted against
• The one question the week did not answer: when a number arrives from a company, who gets to hold a second ruler against it — and can they do it before the number has already done its work?

Plus: a scorecard on last week's question, four quiet stories, and what to watch next.

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.

Google just gave an AI agent its own email address and told companies they can tag it in a document the waythey would tag a coworker.

Not a chatbot in a sidebar. A teammate with its own account, its own storage, and its own name in the version history. Gemini Enterprise is already in the hands of three million U.S. service members.

Meanwhile, OpenAI told investors its annualized revenue is almost fifty billion dollars.

That is twenty billion less than the number everyone had been repeating. Nothing went missing. The ruler changed.

It was Friday.

We also cover:

• Anthropic's Claude Dashboards and Claude Motion turn a question into a live dashboard and a prompt into an editable animated explainer.
• OpenAI launched a mode that runs up to 8 times faster for 6 times the price, and nobody has measured the speed independently yet.
• Three fired OpenAI researchers published their full letter, and one says an executive's inbox was merged into her phone's mail app.
• Anthropic wrote the first rule in its usage policy that protects Claude instead of people, a ban on sustained, needless cruelty, starting November 12.
• Anthropic launched a cyber mission with 11 founding partners and a free scanner for the open-source code the internet runs on.
• Ten more signals, scored one by one: a $60M agent startup, the SI.gov rename, a full-sky UV map from Claude, mathematicians vs. OpenAI, Midjourney's Thinking Mode, a Netflix Hugging Face documentary, and four more.

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

Stay curious, humans.

ChatGPT can now answer you with a chart, a map, or a calculator that it builds on the spot, and it is rolling out to a base of more than one point two billion weekly users.

GPT-6 Sol powers it for paid users today. Free users get it tomorrow.

Meanwhile, Biohub just committed one point eight billion dollars to a virtual cell that predicts how a cell reacts to a drug before anyone runs the lab test.

More than five hundred million of that is federal money that was already spent.

It was Thursday.

We also cover:

• Three OpenAI researchers fired on October first wrote the board that nobody yet knows how to safely deploy models that cannot be monitored.
• Anthropic's new Claude Haiku 5.5 costs 10 cents per million input tokens, about 75% less than its predecessor, and it scores 72.4% on a computer-use test where the old one scored 15.7%.
• Elon Musk says Grok Bot will now route work to Claude Opus 5.5, Midjourney, and Suno, making xAI a customer of Anthropic.
• Nvidia and Microsoft put up to a petaflop of AI in a Windows laptop and made the containers meant to keep agents inside the lines generally available.
• Broadcom is reportedly arranging more than $50 billion so OpenAI can pay for its own chips, while SpaceX shops a $40 billion loan for Nvidia's.
• Ten more signals: a watermark checker that only sees partner marks, a game builder you talk to, and 1,800 data center projects on hold in Texas.

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

Stay curious, humans.

OpenAI just published seven hundred twenty-two math papers written by a model you are not allowed to use.

Three hundred seventy-two families of results, each costing about three hours of ChatGPT Pro thinking. Some claim to settle problems that have stayed open for decades.

Nobody outside the lab has finished checking them, and the repository itself says some may need fixing.

Meanwhile, Mistral released a one-trillion-parameter open model and nicknamed it "le Chonk."

Only 49 billion parameters are active, and the weights are promised by the end of the month.

It was Wednesday.

We also cover:

• The founder of Figure AI launched an assistant that can run 36 web browsers at once on its own cloud system, with devices still a year away.
• Anthropic puts Claude inside Google Docs, Sheets, and Slides for every paid plan, bound by each file's existing sharing permissions.
• Meta and Sierra proposed an open handshake so businesses can tell which AI agents are shopping on your behalf.
• A $12 billion DeepSeek round is reportedly close, the same week one analyst put the U.S. lead over China at 3%.
• Anthropic opened three tiers of verified cyber access for defenders, as a government scorecard placed the best open cyber model about four months behind.
• Eight more signals: a decision model at ten cents per million tokens, robot training data cut from $60 to $10 an hour, and a recruiting drive that isn't landing.

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

Stay curious, humans.

An open model trained with more than a hundred million rollouts is about to be released for free, and its builders say it matches a Chinese rival with a third to a quarter of the compute.

Reflection's Beam: 501 billion parameters, Apache license, due this month. Its own tables show where it loses.

Meanwhile, OpenAI is shopping at a thirty-billion-dollar round at a one-point-four-trillion-dollar valuation.

The same weekend, Sam Altman said the world should accept some bad things happening. A Republican governor and a Democratic governor both answered with a version of no.

It was Tuesday.

We also cover:

• Uber spent its entire annual AI budget in four months. Then it found something that worked better than a cap: showing engineers their own bill cut cost per session 52%.
• A 19-year-old launched a $3,499 computer that runs AI on your own hardware around the clock after raising $11 million.
• An AI doctor was approved to start acne prescriptions in Utah for $4.99 a month after treating more than 10,000 patients under oversight.
• OpenAI's new invisible text watermark detects about 80% of short passages, but swap a quarter of the words and it drops to about 17%.
• Nvidia's $20 billion Groq deal is being sued over a missing stockholder vote, while chip startup Etched Fields offers at $40 billion or more.
• AI agents proposed a room-temperature magnet with zero net magnetism that still sorts electrons, and the author published every caveat.
• Ten more signals: ads arriving inside ChatGPT image generation, a decision model up to 200x cheaper, and an agent that dreams.

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

Stay curious, humans.