Dear Disruptors,
Fernando Santa Cruz here in the forty-eighth edition of Synapsis Weekly, where AI started writing 80% of its own code, Anthropic prepared what could be the largest IPO in tech history, and New York said “enough” to data centres.
Writing from Toronto, where Canada just launched “AI for All,” a national strategy to move AI adoption from 12% to 60%, with this city’s Vector Institute as one of its anchors. In Mexico, we continue advancing with partners on AI training for leaders and the ecosystem in Yucatán.
This week, AI stopped being just software.
It became heavy industry. It wrote itself. The world started drawing lines.
One model is already building its successor. One AI company is worth nearly a trillion dollars. At the same time: a state banned its data centres, two universities proved that automating without measure empties everyone’s pockets, and a CEO vetoed his own people for trying to make it addictive.
That’s the nerve of today’s edition: a technology that no longer has internal limits, colliding with external ones. Water. Demand. Ethics.
This newsletter goes deeper than the WhatsApp summaries (week of June 1 to 6) to understand what it means for a machine to build itself, where the walls it’s hitting actually are, and what you can use starting Monday.
The Printing Press That Prints Better Printing Presses: AI Is Writing 80% of Its Own Code
Anthropic revealed that its Claude Mythos and SWE-1.6 models already write more than 80% of the company’s production code.
Its human engineers now deploy 8 times more code than before.
Because the figures come from financial filings ahead of its IPO, they are legally binding. This isn’t marketing. It’s real.
This has a name: recursive self-improvement.
AI building the tools that will train the next AI.
It’s like watching a printing press print the manuals for manufacturing better printing presses.
Think about this for a moment. The bottleneck in technology was always the human brain: it gets tired, sleeps, makes mistakes. This week that bottleneck shifted. The limit is no longer intelligence. It’s the speed at which servers compile what the machine itself writes.
A practical warning buried in the fine print. The new multi-agent Claude Code can spiral: a broad instruction like “research this thoroughly” can spawn up to 199 parallel sub-agents, burning through your budget before delivering anything useful.
The engineer stopped typing. Now they direct.
Directing starts with setting limits for what executes.
Question for your role: If the machine is already handling 80% of the execution, what percentage of your week is still trapped in executing, and how much do you dedicate to deciding what’s worth executing in the first place?
Anthropic Is Worth $965 Billion, and Sanders Wants Half of It to Belong to Everyone
Anthropic filed a confidential S-1 with the SEC to go public at a valuation of $965 billion.
Following a $65 billion Series H round, it officially overtook OpenAI in the private market.
Annualised revenue: $47 billion.
This is the largest financial event in the history of AI.
It marks the end of an era. The romantic era of research labs is over. AI has formally become heavy, institutional industry.
It’s like when the garage band signs with the label: the money comes in, and so do the rules.
Because the same week, Senator Bernie Sanders proposed the American AI Sovereign Wealth Fund Act: the public would receive 50% of the shares and dividends from these mega-corporations. His argument: the models were trained on the data, the art, and the knowledge of all of humanity, without compensation.
Today it’s politically unlikely.
But the question it installs isn’t going away: if the raw material of your business is society’s digital footprint, whose is the profit?
For an SMB, the useful signal is in the $47 billion. Wall Street is no longer paying for AGI promises. It’s paying for real revenue from companies that use these tools every day. Adoption stopped being an experiment. It’s the economic mainstream.
Question for your vision: If AI was trained on everyone’s knowledge, what share of the value it generates in your business do you plan to return to your team, your clients, your community?
1 Gigawatt Against a Veto: the Wall AI Is Hitting Is Made of Water and Steel
OpenAI broke ground on Stargate in Michigan: a data campus that will consume 1 gigawatt.
The energy of an entire small city. For a single facility.
That same week, across the map, New York approved a one-year moratorium banning new data centres larger than 20MW.
The reason? The electrical grid and water supply can no longer keep up.
Public opposition sits at 70% across the United States. Google had to spend $17 million just to replace 7 billion gallons of fresh water.
For years, we assumed the obstacle to the next generation of AI would be the algorithm, or the data.
Turns out it’s steel, building permits, and water rights.
The competitive advantage moved from the lab to the construction site.
For an SMB, this collision explains a trend that does affect you: if the cloud hits physical walls, the industry pushes intelligence toward your laptop and your office. We get to that below, and it’s the best news of the week for a small business.
Question for your operations: What physical resource (energy, space, water, suppliers) does your growth plan take for granted, and what happens to that plan the day it gets more expensive or runs out?
The Layoff Trap: 192,000 Cuts and the Monopoly Nobody Wins
Economists from Wharton and Boston University published a mathematical model with an uncomfortable conclusion: mass automation can collapse the very economy financing it.
There are already 192,000 recorded tech layoffs between 2025 and 2026.
The logic is a prisoner’s dilemma at global scale. For any individual executive, replacing 500 analysts with agents is perfectly rational. But those 500 analysts are also someone else’s customers.
If everyone automates at once, they produce without limit for customers who can no longer buy.
Infinite production. Zero demand.
It’s like playing Monopoly: if one player accumulates all the money, nobody wins the game. It ends, because nobody can land on your properties and pay.
The outcome is paradoxical. Every company acts with perfect rationality, and the whole system destroys itself. The only correction with mathematical backing in the study: a Pigouvian Automation Tax, a charge for every task that replaces a human, to force companies to internalise the social cost they generate.
Here’s the reading I’ll defend for an SMB, and it’s an optimistic one. The trap is automating to cut. The way out is automating to grow: same people, triple the capacity, new services you couldn’t offer before. The study doesn’t condemn AI. It condemns using it only with scissors.
Question for your leadership: When AI frees up hours on your team, do you already know what they’ll build with them, or is your only plan that there’ll be people left over?
Scout Lives Inside Windows, and Nadella Vetoed Addictive AI
Microsoft introduced Scout at Build 2026: an always-on agent integrated at the deepest level of Windows.
It orchestrates Teams, Outlook, and OneDrive in the background. It summarises, organises, and executes while you work on something else.
AI stopped living in a browser tab.
Now it lives in the operating system.
But the most important news from Build wasn’t technical. It was ethical.
A vice president proposed in an internal memo to design Scout to be “addictive,” using social media tactics. CEO Satya Nadella publicly overruled him: the goal is productivity, not hijacking attention.
Picture the TikTok algorithm embedded inside your spreadsheet and your inbox.
That’s exactly what Nadella stopped.
Not everything in this industry is a blind race. This week, the line was drawn by someone on the inside.
For an SMB, the practical point: agents are coming to the operating system your business already pays for, with native security isolation. It doesn’t require buying another platform. It requires deciding, calmly, which workflows you hand over first.
Question for your attention: Of the tools your team uses every day, which ones are designed to save you time and which ones are designed to take it from you? And can you tell the difference?
ChatGPT Stops Forgetting You: Memory at 82.8% and Emails That Send Themselves from Chat
OpenAI rolled out Dreaming V3, a system that consolidates user memory in the background.
It remembers your business context with 82.8% accuracy. Without you repeating the same prompt every session.
It can also send emails directly from the chat. No copying. No pasting. No switching tabs.
It was like training a new employee every morning, because they arrived with amnesia.
Today they arrive having already read your file.
The connection most people aren’t making: persistent memory plus direct execution is the combination that ends the chatbot. ChatGPT stops being an oracle waiting for questions. It becomes an engine that learns from you, synthesises your history while you’re away, and acts.
AI no longer just gives you answers so you can do the work.
It does the work while you’re not watching.
For an SMB, the savings are direct: the hours spent “let me explain what we do again” disappear. But the coin has another side, and in Europe it’s already triggering privacy alerts: an automated, persistent profile of you now lives on a server you don’t own. What you share, it remembers. Choose carefully what deserves to be in that memory.
Question for your workflow: How much time does your team lose every week re-explaining context to AI, and what should you write once so you never have to repeat it?
AI Leaves the Cloud: Gemma 4 Runs on Your Laptop and Nobody Else Has the Key
Google launched Gemma 4 QAT, a family of models that reduces RAM consumption by 72%.
The result? Full multimodal AI running on a standard 16GB laptop.
No internet. No subscription. Not a single byte of your data leaves your machine.
Process your internal documents, your contracts, your client base, with complete privacy.
It’s like going from renting office space to owning the building: nobody else has the key.
The trend is bigger than Google. Perplexity introduced a hybrid orchestrator that handles simple tasks on your chip and only sends complex ones to the cloud. NVIDIA launched Nemotron 3 Ultra, 550 billion parameters, activating only 55 billion per task: 30% lower operating cost and responses in under 100 milliseconds.
Sending everything to the cloud was a transitional phase.
The future of enterprise AI is split between the server and your desk.
For an SMB, this is the quietest, most valuable news of the week. The barrier is no longer the monthly cost or the fear of uploading sensitive information to someone else’s servers. The intelligence fits on the hardware you probably already have.
Question for your data: What information about your business have you never wanted to upload to the cloud, and what would you do with it if you could process it with AI without it leaving your office?
The Robot’s Brain Now Downloads: MolmoAct 2 Runs 37 Times Faster
The Allen Institute released MolmoAct 2, an open-source robotic brain 37 times faster than its predecessor.
It already automates CRISPR gene-editing experiments at Stanford.
In parallel, startup Generalist AI raised $400 million to scale generalised industrial robotics, reducing total cost of ownership by 90%.
Ninety percent less.
For decades, the robot came with the brain welded in: you bought the hardware and were married to the manufacturer’s proprietary software, licences included.
That broke this week.
The brain now downloads. The hardware is just the body.
Intelligence became universal. Hardware became replaceable.
For an SMB in manufacturing or logistics, humanoids are still far off. But the underlying equation changed sign: when the software is open and the cost drops 90%, physical automation stops being exclusive territory for large corporations. It’s the same story we lived through with computers: first banks had them, then everyone did. It’s worth starting to browse the catalogue, even if you’re not ready to sign yet.
Question for your plan: What’s the most repetitive physical task in your operation, and at what monthly price would it stop making sense to do it by hand?
Tools You Can Use Starting Monday
- Codex Plugins by Role – OpenAI launched six plugins that turn Codex into a data analyst, sales assistant, or creative producer, connected to tools like Salesforce, HubSpot, and Tableau. Sites also builds internal websites and apps for your team from a prompt. Agentic automation is no longer just for developers.
- Newo Vibe Mode – If you use voice agents to handle calls (restaurants, clinics, services), you can now configure your “AI employee” by talking to it in plain language: change schedules, promotions, or service rules instantly, no developers needed, with the option to revert any change. The step-by-step guide is from IONOS, Newo’s global partner, which has already integrated it for its 6 million customers.
- Ideogram 4.0 – Image generator with accurate, legible text, native 2K resolution, and transparent backgrounds. Ideal for logos, posters, and social media assets with prices and dates, no spelling errors. Open version available.
- Reve 2.0 – Solves the biggest problem with AI-generated images: control. Define with bounding boxes the exact position of each element, text, or logo, in 4K resolution. Your assets respect your brand’s proportions and guidelines.
- Runway Aleph 2.0 and Edit Studio – Upload an existing video and modify only what you specify: the product colour, the lighting, a garment. Everything else stays intact, in clips up to 30 seconds at 1080p. Campaign variations without reshooting.
- Miso One – Open 8-billion-parameter voice model that interprets real emotions (sarcasm, whispers, hesitation) with 110-millisecond latency. Marketing videos that sound human, or natural phone support, without a voiceover budget. Currently English only, making it ideal if your content targets that market.
My Invitation This Week: The Onboarding Manual for Your Digital Employee
This week, AI learned to remember.
Dreaming V3 stores your business context with 82.8% accuracy. But there’s a catch: it can only remember what you’ve told it clearly.
This week’s exercise is to write, just once, what you’ve been repeating in every chat for months.
We’re going to draft the onboarding manual for your digital employee. One page. Forty minutes. Like the document you’d hand someone new on their first day.
Open a blank document and fill in five sections, three to five lines each.
- Who we are: what your business sells, to whom, and what makes you different.
- How we speak: your brand’s tone, two phrases you’d say, two you’d never say.
- Our sacred numbers: prices, margins, or policies the AI must never change or promise.
- Our typical client: describe them as you’d introduce them to a new colleague.
- What I always repeat: that context you type in every new conversation. It ends here.
Now for the interesting part: the onboarding test.
Load the document into ChatGPT’s memory, a Claude project, or the start of your favourite tool. Give it the exam you’d give a new employee in their first week: ask it to describe your business in its own words, and to draft an email to a typical client.
Score both responses from 1 to 10.
What comes out wrong isn’t the machine’s error. It’s a gap in your manual. Fix it and test again.
At the end of the exercise you have something no tool gives you: your business context, distilled to one page, ready to onboard any AI, present or future, in minutes.
Agents will change every six months.
Your onboarding manual gets written once.
Start on Monday. One page, five sections, two tests.
Closing
This wasn’t the week of a launch.
It was the week of limits.
AI no longer has an internal brake: it writes itself, improves itself, remembers itself.
The brakes now come from outside. From water. From demand. From a CEO who said no. From a senator who asked whose profit it is.
For an SMB, that realignment is a door. Intelligence is dropping in price, leaving the cloud, and going open, just as the giants are hitting their walls.
Their limit is physical.
Yours is only deciding where to start.
Start on Monday, with one page.
Fernando SantaCruz Head of AI & Automation @ Adivor Consulting
