Dear Dysruptors,
Fernando Santa Cruz here in the forty-sixth edition of Weekly Synapsis — where an AI independently disproved an open mathematical problem from 1946, a Chinese model worked uninterrupted for 35 hours without human intervention, and Google turned search and email into a team of agents operating while we sleep.
Writing from Toronto, wrapping up the week with clients in construction and real estate, while coordinating with partners in Mexico on the next phase of AI training programs for the Yucatán region.
This week changed the name of work.
Not the tool. The work itself.
For two years, we asked AI to do things. This week, AI started doing them on its own, for hours, without asking permission.
And the highest-paid role in the industry stopped being writing code.
It became directing the machines that write it.
This newsletter expands on the WhatsApp summaries (week of May 18–22) to understand why AI stopped being something we use and became something we supervise, why that’s good news for SMBs, and which skill suddenly became the most valuable one on your team.
Google Turns Search and Email into a Team of Agents: Gemini 3.5 Flash, 4x Faster and Much Cheaper
At Google I/O 2026, Google introduced Gemini 3.5 Flash, four times faster than its predecessor at $1.50 per million tokens.
It’s not the smartest model. It’s the fastest and cheapest.
And that’s exactly the strategy.
Gemini 3.5 Flash is now the default engine powering agents inside Search and Workspace. Google also introduced Omni Flash, which generates and edits video conversationally using text, audio, or images. And Antigravity 2.0 evolved from a code editor into a platform that orchestrates sub-agents.
The industry game changed.
It’s no longer about who gets the best score on a theoretical benchmark. It’s about who’s cheaper and faster when an agent has to work for hours.
It’s like moving from an intern you dictate every sentence to, to an art director who grabs the mouse and finishes the project for you.
For SMBs, the practical story is pricing. When token costs become fractional, an agent running all day stops being a luxury and becomes economically viable. What was too expensive to automate last year suddenly returns to the table.
Raw intelligence is no longer the frontier.
Token economics is.
Question for your operation: What repetitive task did you avoid automating last year because it was too expensive — and which one makes sense again if token costs drop to a fraction?
An AI Disproves an Erdős Problem Open Since 1946: The Day Machines Stopped Copying and Started Discovering
An OpenAI reasoning model autonomously disproved the planar unit distance conjecture proposed by Paul Erdős in 1946.
Eighty years unresolved. Solved without human guidance.
The result was verified by elite mathematicians, including Fields Medal winner Tim Gowers.
For years, we repeated that AI was just an advanced parrot. That it merely predicted the next word based on existing internet knowledge.
This week, that idea broke.
AI didn’t summarize old knowledge. It generated new knowledge — mathematics no human mind had managed to uncover in eight decades.
It’s the difference between a librarian finding the right book and a researcher writing the missing chapter.
Let’s be honest about the nuance. For now, this is pure mathematics without immediate commercial application. You won’t invoice with this on Monday.
But the point most people are missing is this: the same reasoning architecture that unlocked Erdős is already running, in a more modest form, inside the tools your business uses today.
If it can tackle a 1946 problem, it can tackle the operational knot your team has postponed for three years.
Question for your strategy: What complex business problem has nobody had time to solve — and what would happen if this week you gave it to an AI as a reasoning challenge instead of a quick query?
Glasswing Finds 10,000 Vulnerabilities in One Month and Nobody Can Patch Them Fast Enough: The Bottleneck Moved
Anthropic audited 50 partner companies in one month using its Claude Mythos Preview model. The Glasswing project found more than 10,000 critical vulnerabilities.
Two thousand in Cloudflare. Two hundred seventy-one in Firefox.
With fewer false positives than human experts.
And here’s the twist: the problem is no longer finding vulnerabilities.
The problem is patching them at the speed AI discovers them.
Finding bugs went from taking months to taking seconds. But writing patches, testing them, and deploying them still depends on human hands that can’t keep up.
It’s like a smoke detector finding ten thousand gas leaks in a building with only one plumber available to fix them all.
For SMBs, this has both a technical and broader implication.
The technical one: traditional antivirus solutions are increasingly insufficient, and serious cybersecurity providers will include defensive AI in their packages — at higher prices.
The broader lesson is more useful. AI will uncover more opportunities, more mistakes, and more pending issues than your team can handle. Just like in cybersecurity, your bottleneck will stop being detection and become prioritization.
Knowing what to address first becomes the scarce skill.
Question for your team: If an AI delivered a list of fifty possible improvements for your business tomorrow, would you have a clear framework to choose the top three — or would you freeze in front of the list?
Cursor Pays $1.1 Million to Engineers Who No Longer Write Code, and a Chinese Model Works Alone for 35 Hours: The Profession of Herding Machines Is Born
Cursor launched Composer 2.5. It matches giant models on demanding engineering benchmarks — nearly 80% on SWE-Bench — at less than one dollar per task, compared to eleven dollars from competitors.
But the statistic that stopped the industry was another one.
Cursor pays engineers $1.1 million annually for jobs that no longer involve writing code.
Their job is coordinating fleets of agents that write it for them.
One human using these tools can produce the output of one hundred traditional programmers combined.
And it’s not just the West. Alibaba introduced Qwen3.7-Max, a model that operated autonomously for 35 consecutive hours, with one million context tokens and more than one thousand tool calls, without any human intervention.
Think about that for a second.
If a machine works alone for 35 hours, supervision stops meaning checking every click.
The industry is moving from step-by-step supervision to trajectory supervision. You no longer approve every move an agent makes. You define boundaries — “work freely, but if you’re about to move money or delete data, stop and ask for permission” — and evaluate the final outcome.
You stop being the operator. You become the manager.
And this isn’t just for software companies. It’s the new relationship every professional will have with every AI tool. Your value no longer comes from executing the task. It comes from framing it correctly, setting boundaries, and verifying the result.
Fast typing no longer wins.
Delegating and reviewing does.
Question for your talent: Is your team practicing the skill of directing and verifying AI work — or still competing to do by hand what machines already do alone?
Musk Loses in 90 Minutes, Wall Street Invests $5 Billion in Chips, and Meta Cuts 7,800 Jobs: AI Has Someone Paying the Bill
Ninety minutes of deliberation. A jury unanimously dismissed Elon Musk’s lawsuit against OpenAI on statute-of-limitations grounds.
With that, OpenAI clears its path toward a potential IPO valued at $300 billion.
That same week, Google and Blackstone formed a $5 billion joint venture to build data centers powered by proprietary TPU chips. NVIDIA began shipping Vera, its first processor designed specifically to coordinate swarms of agents.
Wall Street no longer sees AI compute as a tech expense.
It sees it as an asset class, on par with real estate or energy infrastructure.
It’s a gold rush — except this time capital is funding entire refineries, not miners.
And every gold rush has a darker side. Meta announced 7,800 layoffs — 10% of its workforce — to free budget for that same infrastructure.
Billions to servers. Nearly eight thousand people out of work.
For SMBs, this contrast hides a rare advantage. Large corporations are painfully restructuring themselves to become “AI-first”: layoffs, friction, internal politics.
A small business carries none of that baggage. It can become AI-first by design from day one, without dismantling anything.
Big companies have the capital. Small companies have the speed that comes from not needing to destroy their past to build their future.
Question for your plan: What processes could you design today “AI-first” that a large competitor would need years — and expensive restructuring — to adopt?
Karpathy Leaves His Projects to Join Anthropic: When the Best Chef Closes His Restaurant, Pay Attention to Which Kitchen He Enters
Andrej Karpathy, former Tesla AI director and OpenAI co-founder, paused his educational projects to lead frontier-model research at Anthropic.
In his announcement, he said the coming years would be especially formative.
The sector read it as a signal.
There are only a handful of people on Earth capable of building frontier models. When one of them pauses their own ventures to join a lab, it’s not a career decision. It’s a bet.
A bet that breakthroughs not yet public are about to happen inside those walls — probably in reasoning and interpretability.
It’s like the city’s most celebrated chef closing his own restaurant to join another kitchen. He’s not telling you what’s cooking. He’s telling you where to look.
That same week, Anthropic added 28 new enterprise security integrations and acquired the startup Stainless to better connect its models to real-world workflows.
For SMBs, the lesson is not to wait for the next big model. It’s the opposite.
The tools already in your hands will keep improving on their own — and quickly. The people building the habit of using them today will be ready when the next leap arrives. Those who wait will start from zero alongside everyone else.
Question for your team: Are you building your team’s AI muscle today, or waiting for a “final model” that, when it arrives, will find you without practice?
Spotify and Universal Legalize AI Remixes: From Looting to the Toll Booth
Spotify and Universal Music Group launched a tool allowing Premium users to generate licensed AI-powered covers and remixes.
The system automatically pays royalties to original artists.
Three pillars: consent, credit, and compensation.
The contrast is obvious. Last year, platforms like Suno faced massive lawsuits for using music without permission. This week, the same activity became a revenue channel for artists.
The music industry has seen this movie before. Napster was the looting. Spotify became the toll booth that turned piracy into business.
This is chapter two of the same story — now with AI.
It’s the first mature model turning the fear of “they’re going to copy me” into a revenue stream. And it will likely become the template Hollywood and publishing follow.
For SMBs, the lesson is broader than it appears. If your business owns any creative asset — a brand voice, a photo catalog, designs, a recognizable style — the emerging path isn’t lawsuits. It’s licensing.
The question stops being how to prevent AI from using your work.
It becomes how to get paid when it does.
Question for your brand: What creative asset in your business — images, text, voice, or style — could become something you license instead of defend?
SynthID Marks 100 Billion Images: The Internet Changes Its Default Assumption
OpenAI integrated SynthID, Google DeepMind’s invisible watermarking technology, into all images generated by ChatGPT and Codex. It also adopted the C2PA standard and a public authenticity verification tool.
Google has already marked more than 100 billion images and videos.
Digital authenticity is becoming an internet standard.
And the deeper shift is subtle but enormous. Until now, the internet assumed everything was real and asked us to detect what was fake.
From now on, that reverses. The internet will assume everything is synthetic, and the valuable thing will be proving what was made by a human.
It’s like a “handmade” seal in a market flooded with mass production. It doesn’t disqualify the machine. It gives human work a premium.
For SMBs, this turns trust into a product feature.
The business that verifies and authenticates its own content — real photos of its store, real customer testimonials, real products — will earn a trust premium over businesses flooding their feeds with generic AI imagery.
Authenticity will not become the opposite of technology.
Technology will finally help certify it.
Question for your adoption: What part of your brand’s content is genuinely real and verifiable — and how are you signaling that to customers who already assume everything online is synthetic?
Tools You Can Use on Monday
Google Pomelli – A Google Labs tool for SMBs. Using your documents and photos, an agent generates a full brand manual — including colors and typography — and builds a website in one click. Digitizing a business goes from weeks to minutes.
Granola – Its new “AI Meeting Briefs” feature prepares you before every video call. It reviews previous notes and shared emails with the client, then delivers a concise summary with context and priorities.
Krea 2 – The beta “LoRAs” feature lets you train AI on a few images of your visual style or products, ensuring all generated social media and advertising content maintains consistent branding.
Google Universal Cart – An AI-powered universal shopping cart introduced at Google I/O. Customers add products while browsing the web, searching Google, or watching videos, and checkout works through Shopify-based merchants. For retailers, it’s a frictionless sales channel worth preparing for now.
My Invitation This Week: The Hour of the Manager
This week, a machine worked alone for 35 hours. And the world’s highest-paid engineers stopped typing to start directing.
This week’s exercise is about practicing that shift. In just one hour.
We’re not filling out a worksheet. We’re not making a list.
We’re going to delegate for real — and resist the urge to supervise.
Pick one real task. One you would normally do yourself step by step, watching every detail. Writing a proposal for a client. Analyzing monthly sales. Preparing the week’s content. Reviewing a contract. Anything that genuinely matters.
Now do something uncomfortable. Instead of asking AI for one little piece, reviewing it, then another little piece, reviewing again — hand over the entire task at once.
With one condition. Before you let go, define the boundaries exactly as you would for a new employee. “Do the full task. But if you need information I didn’t provide, if you’re about to assume pricing, or if something feels risky, stop and ask before continuing.”
And here’s the difficult part.
Let go. Walk away.
Do something else for a while. Don’t look over the machine’s shoulder. Don’t correct halfway through. Trust the boundaries you defined.
When you come back, evaluate only one thing: the final result. Does it work? What’s missing? Where did it fail?
That’s it. One task, a few boundaries, one hour.
What’s interesting is rarely the AI’s result. It’s what we feel during the process.
For almost everyone, letting go of control is the hardest part. And that discomfort is the valuable information. It measures the distance between the operator we used to be and the manager this week is asking us to become.
The operator executes the task. The manager frames it, defines boundaries, and verifies.
The good news is that this promotion doesn’t require anyone’s permission.
It starts Monday, with one task and one hour.
Closing
This wasn’t the week of a new model.
It was the week work changed its name.
AI stopped being something we use step by step. It became something we frame, constrain, and verify.
The highest-paid role no longer types.
It directs.
And for SMBs, that isn’t a threat. It’s a promotion.
Because the promotion from operator to manager isn’t reserved for companies with thousands of agents and GPU clusters.
This very week, it’s available to anyone willing to delegate a real task — and trust the boundaries they set.
