The Goalposts Have Left The Stadium

Good morning,

I've been doing this newsletter for three years now. 156 weeks of trying to make sense of the AI chaos. And I'll be honest—I'm getting tired of sounding like a broken record. One week it's the apocalypse. The next week we'll all be fine.

About 1,5 years ago I was asked to give a keynote speech about AI. Just speaking before me was a lawyer talking about AI law , after me an AI startup founder. They both didn’t think it would go this fast. My talk got me strange looks because I painted a picture that was at the time unimaginable. The end of work. The end of privacy. Afterwards I had the impression they thought I was crazy and haven’t been invited back.

I DID say that AI law would be useless once the singularity is here and the lawyer got quite “salty”.

With every week that passes - we’re edging closer to this vision.

But this week no edges but a giant leap. Something ACTUALLY shifted.

Claude Code dropped about a week ago. And if you were out there touching grass or watching Netflix or whatever normal people do, you might have missed it. The rest of us are sitting here staring at our screens trying to figure out what the fuck just happened. (Strong language is warranted in this case). Believe you me.

The One-Hour Year

A principal engineer at Google (not some random startup founder, not some AI hype person, but someone who actually builds this stuff at one of the biggest tech companies in the world) she fed Claude Code a description of a “distributed agent orchestrator”.

Her team spent a year building it. At Google. With the smartest people in the world in the team.

Claude Code replicated it in one hour.

ONE HOUR !!!!

She said it wasn't perfect. That she had to iterate. But she also said "I'm not joking and this isn't funny."

When the people at Google start getting nervous about the velocity of what they're building, you should probably pay attention.

The Automated Money Printer

Then there's Ethan Mollick. He’s one of the rare sources I follow religiously. He asked Claude to build him a business. Not a business plan. An actual functioning business with code and a website and payment processing. The whole thing.

It took 74 minutes.

The machine wrote hundreds of files. Deployed a live site. Even wrote the marketing copy (sketchy "get rich quick" stuff, but still). Mollick pulled the plug before taking anyone's money, but the gun was loaded.

Five years ago this was science fiction.

Truth is—I don't know what to do with this information.

The Goalposts Are Gone

Remember when the Turing test was "can it talk like a human for ten minutes?"

Mustafa Suleyman at Microsoft says says the new test should be: "Can an agent turn $100k into $1 million legally?"

Two weeks ago in this newsletter I told you that task time is doubling every seven months. New data suggests it's happening every four months.

What If You're Already Behind

David Shapiro - who I talked about last week - compares this to jet pilots. When you're flying at Mach 2, your brain can't process the inputs fast enough. You have to trust the machine.

That's where we are now. All of us. Fighter pilots in jets we didn't build, flying into something we can't quite see.

The interface for Claude Code is a command line. Looks like something from 1985. Ugly. Raw. Orange or Green. This little bit of technical friction keeps most people away.

But don't let that fool you.

We're entering the era of the Agent. And if you're not figuring out how to use these things—or at least understanding what they can do—you're bringing a knife to a nuclear battle.

To be honest: I don't know if Elon's right when he says we've entered the singularity. But I know the machine just did a month's worth of engineering work before lunch. And I know that should probably worry you more than it does.

So what about you? What's your plan?

Because it's happening. Every single day something new drops. And the gap between "this is impossible" and "this is Tuesday" keeps getting shorter.

Truth is—most people are still asleep. They have no idea this is coming. They think AI is ChatGPT writing emails and Midjourney making weird art.

They don't know the machine can build businesses now. They don't know it can replicate a year of Google engineering in an hour.

And even if you're reading this and you're wide awake instead of asleep, there might be nothing you can do about it.

But at least you'll see it coming.

AI News

• Anthropic has reportedly blocked Elon Musk’s xAI from using its Claude models after discovering the lab was using the rival system to help build and train its own competing technology. This enforcement action follows reports that xAI developers were using Claude through the coding platform Cursor, which directly violates Anthropic’s policy against leveraging its models for competitive development. While the move is expected to temporarily slow down xAI’s internal work, the company is using the setback as motivation to speed up the launch of its own independent coding tools.

• Google has introduced the Universal Commerce Protocol (UCP), a new open-source framework co-developed with retail giants like Shopify and Walmart to allow AI agents to manage the entire shopping journey. The protocol enables AI assistants like Gemini to handle product discovery, apply discounts, and complete secure checkouts using native payment tools directly within a chat interface. By establishing these universal standards, Google aims to turn conversational AI into a primary storefront where users can buy items from multiple brands without ever leaving their assistant.

• Apple and Google have officially entered a multi-year partnership to use Google’s Gemini technology as the primary brain for the upcoming Siri upgrade and other core AI features on the iPhone. While Apple is paying an estimated $1 billion annually for this integration, the company confirmed it will still keep its existing deal with OpenAI intact to provide users with a variety of advanced tools. This landmark deal validates Google's rapid rise in the AI space and helps Apple quickly modernize its digital assistant by outsourcing its most complex AI challenges to a major rival.

• Anthropic has released Cowork, a new tool for Mac users that allows the Claude AI to function as a hands-on office assistant by autonomously organizing files, writing reports, and managing professional expenses. Unlike a standard chatbot that just answers questions, Cowork can be assigned several background tasks at once and even integrates with work apps like Notion and Asana to complete projects on a user's behalf. This release brings advanced automation to everyday office workers, moving powerful technology out of the world of computer programming and into mainstream daily productivity.

• A new report from Microsoft shows that global AI adoption has reached over 16%, but reveals a surprising trend where the United States ranks only 24th in widespread use despite being the leader in developing the technology. The data highlights a growing "digital divide" between wealthy and developing nations, though the Chinese AI model DeepSeek is successfully bridging the gap in underserved markets like Africa through free, accessible software. These findings suggest that while the race to build the best AI is happening in the West, the race to actually integrate it into daily life is being won by nations like the UAE, which currently leads the world in adoption.

• Meta has launched "Meta Compute," a massive new division dedicated to building AI infrastructure on a scale that could eventually reach hundreds of gigawatts of power. This initiative is backed by a $600 billion investment plan and long-term nuclear energy deals to ensure its future data centers have a reliable and sustainable energy source. The announcement coincides with a 10% workforce reduction in the company’s Reality Labs division, signaling a major strategic pivot from the metaverse toward the physical hardware required for advanced AI.

• Microsoft has unveiled its "Community-First AI Infrastructure" plan, a series of commitments designed to ensure its massive data centers are better neighbors to the local residents surrounding them. The company has pledged to pay higher utility rates so that residential electricity prices do not spike and aims to replenish more water than it consumes through new, high-efficiency cooling technologies. Additionally, the tech giant will forgo traditional tax breaks and invest in local job training, addressing growing public concerns over the environmental and economic impact of the global AI buildout.

• UK startup Basecamp Research, in partnership with Nvidia, has introduced Eden, a family of AI models trained on evolutionary data from over one million species to design cures for genetic diseases. This technology has successfully created new gene-editing tools that proved functional in 63% of tested cases and antibiotic candidates that are 97% effective against dangerous, drug-resistant "superbugs." By learning from billions of years of biological history, the system provides a safer and faster way to develop treatments for conditions like muscular dystrophy that currently have no cure.

• Google has launched Personal Intelligence, a new beta feature that allows Gemini to securely search across your Gmail, Photos, YouTube, and Search history to provide highly personalized answers. By connecting these apps, the assistant can understand your real-world context—like finding a specific receipt in your inbox or identifying a car part from your photos—without you needing to switch between different apps manually. The tool is currently available to paid subscribers in the U.S. and is disabled by default, with Google promising that this private data will not be used to train its AI models.

• Actor Matthew McConaughey has secured eight federal trademarks for his voice, likeness, and famous catchphrases to protect himself against the rising threat of unauthorized AI deepfakes. These trademarks, which include his iconic "Alright, alright, alright" line, provide a clear legal framework to pursue companies or individuals who use his identity without permission in federal court. As an investor in AI voice technology himself, McConaughey’s move marks a significant attempt to establish "consent and attribution" as the global standard for digital ownership in the AI era.

• Chinese AI startup Zhipu AI has released GLM-Image, a major open-source image generator that was trained entirely on domestic Huawei hardware instead of U.S.-made microchips. While the model has achieved high scores on technical tests for rendering text within images, early user reviews suggest its overall artistic quality still lags behind top-tier rivals like Nano Banana Pro. Despite these early performance gaps, the release demonstrates that China is successfully building its own independent AI infrastructure to bypass global hardware restrictions and trade bans.

Quickfire News

• OpenAI launched a new version of ChatGPT designed specifically for hospitals and doctors. This platform follows strict "HIPAA" privacy laws, meaning it can safely handle private patient records to help medical staff write summaries, check research, and organize their daily work.

• Meta, the company that owns Instagram and Facebook, signed a huge deal to use nuclear energy to power its computer systems. They plan to buy enough electricity to power millions of homes by 2035, making them one of the biggest buyers of nuclear power in the United States.

• The head of Alibaba’s AI team, Justin Lin, said that Chinese companies have a less than 20% chance of catching up to top U.S. companies like OpenAI in the next few years. He explained that U.S. labs have much more money and computer power to spend on new research, while Chinese teams are busy just trying to keep up with daily tasks.

• OpenAI’s GPT-5.2 Pro model solved three famous, unsolved math problems called "Erdos problems." With help from world-famous mathematician Terence Tao, the AI used a new logic tool called Aristotle to write proofs that were checked and confirmed to be correct.

• The research group Epoch AI found that the world’s total amount of computer power for AI is doubling every seven months. Their report shows that chips made by the company Nvidia have made up more than 60% of all the AI hardware produced since 2022.

• ElevenLabs released a new tool called Scribe v2 that turns speech into text. The company claims it is the most accurate system ever made for this task, meaning it makes fewer mistakes than any other program when transcribing long recordings or difficult accents.

• A special AI system called AxiomProver achieved a perfect score of 12 out of 12 on the 2025 Putnam exam, which is known as the most difficult college-level math test in the world. Instead of just guessing, the AI used a formal logic language to write out and double-check every step of its math proofs.

• Anthropic launched "Claude for Healthcare," a new set of tools designed to help doctors, hospitals, and insurance companies. These tools are built to follow strict privacy laws and can connect directly to major medical databases to help organize patient records, check insurance coverage, and summarize scientific research.

• The computer chip company Nvidia and the pharmaceutical giant Eli Lilly are spending $1 billion to build a new AI research lab in San Francisco. This joint project will use a massive supercomputer and advanced AI models to try and discover new life-saving medicines much faster than human scientists can on their own.

• OpenAI bought a small startup called Torch for $100 million to help improve its health features. The Torch team, which has only four employees, built technology that acts like a "medical memory" by pulling together a person’s scattered health records from different labs, doctor visits, and wearable devices into one place.

• The AI company Manus introduced a new feature called "Meeting Minutes" that records in-person conversations and turns them into notes. The system is smart enough to recognize which person is speaking and can automatically create to-do lists and summaries so teams can start working on the project right after the meeting ends.

• Google released an update for its medical AI called MedGemma that can now read 3D medical images. The system is able to look at complex scans like CTs and MRIs to help find health problems, and it also includes a new tool that turns a doctor's spoken notes into written text.

• The Japanese research lab Sakana AI made history when its computer program, ALE-Agent, won first place in a major coding competition called the AtCoder Heuristic Contest. This was the first time an AI ever beat all the human programmers in this specific event by inventing a new way to solve a math problem that the contest creators hadn't even thought of.

• The head of the consulting firm McKinsey said that about 25,000 of the company's "workers" are actually AI agents. The company plans to give every one of its 40,000 human employees their own AI assistant to help them with tasks like making charts, doing research, and writing reports.

• U.S. Defense Secretary Pete Hegseth announced that the military will start using the Grok AI from Elon Musk’s company xAI. By the end of this month, the AI will be added to the Pentagon’s computer networks to help the military process information and make decisions more quickly.

• Anthropic started a new department called Anthropic Labs to build and test early versions of new AI tools. The team is led by Mike Krieger, who was one of the people who originally started Instagram before joining the AI industry.

• A leader at Anthropic revealed that their new "Claude Cowork" tool, which helps people manage their computer files and folders, was built almost entirely by an AI. The AI coding tool finished writing the software for the new feature in only about ten days.

• OpenAI teamed up with the chip company Cerebras to use 750 megawatts of specialized computer power. This new system uses massive, plate-sized chips that are designed to make AI responses like chat and coding up to 15 times faster than the chips most companies use today.

• ElevenLabs is working with Deutsche Telekom to create AI voice assistants for customer service. These digital workers can answer phone calls and app messages 24/7 in a realistic human voice, helping to solve about 80% of customer problems without making people wait on hold.

• Microsoft has become one of the biggest customers for the AI company Anthropic, spending nearly $500 million a year to use their models. Microsoft is using these AI tools inside its own products, like GitHub and Office, to give users more options and to avoid relying only on OpenAI.

• Slack updated its built-in Slackbot to act like a personal office assistant that understands the "context" of your work. The bot can read through your past messages, files, and channels to answer questions about projects, summarize meetings you missed, and even draft emails in your specific writing style.

• OpenAI made its GPT-5.2-Codex model available for outside developers to use in their own apps through a new connection called the Responses API. This model is specifically trained to be a "coding agent" that can handle complex software engineering tasks and long-term projects better than standard AI.

Closing Thoughts

That’s it for us this week. Please like and subscribe :)