Dear Dysruptors,
Fernando Santa Cruz here in the fifty-fourth edition of Weekly Synapsis — where one AI claimed to solve a mathematical problem that had remained open since the 1970s, another spent eight days rewriting its own code without anyone watching, and the head of DeepMind publicly called for a referee.
Writing from Toronto, between projects with construction, real estate, and financial services companies, while in Mexico we continue building the AI ecosystem and developing the next generation of business leaders across Yucatán.
This week, AI learned to work on its own.
Not better.
On its own.
It proposed a mathematical proof.
It optimized itself.
It orchestrated other agents without asking for permission.
While the frontier raced forward, the floor dropped in the opposite direction.
One giant model arrived charging half the price.
Another, once 54 gigabytes in size, now fits inside your smartphone and runs completely offline.
The summit climbed higher.
The ground came lower.
In the very same week.
Never has frontier intelligence been this autonomous.
And never has everyday AI been this personal.
This newsletter expands on the WhatsApp summaries (week of July 13–19): what became autonomous, what became inexpensive, and what you can start using on Monday.
GPT-5.6 Gathered 64 Minds and Claimed to Solve a 50-Year-Old Mystery—The Verdict Is Still Pending
OpenAI announced that GPT-5.6 Sol Ultra generated a proof for the Cycle Double Cover Conjecture, one of graph theory’s longest-standing open problems.
The headline says:
“Solved.”
The fine print says:
“Claimed.”
The system deployed 64 specialized sub-agents working in parallel, producing a complete proof in under an hour.
The conjecture was introduced by George Szekeres in 1973 and Paul Seymour in 1979, and it carries an uncomfortable history: several previous “proofs” later turned out to contain subtle flaws.
This one hasn’t yet passed peer review.
Imagine crossing a maze while leaving a trail of ink behind you.
Every step becomes visible.
Anyone can follow your path all the way to the exit—
or discover exactly where the trail breaks.
The ink is impressive.
Verification takes longer.
Here’s the deeper shift.
The breakthrough isn’t that machines calculate faster.
It’s that they’ve crossed the line from organizing existing knowledge to proposing knowledge that nobody had written before.
Whether this proof ultimately stands or falls, the boundary between searching a library and writing a new book has now been crossed.
And with it comes a new question:
Who audits the author?
For SMBs, this changes the standard.
An AI-generated answer that looks perfect is not the same thing as a verified answer.
Whenever the stakes matter, ask for the sources.
Ask for the reasoning.
Ask for every step.
Don’t approve conclusions simply because they’re elegant.
Question for your decision-making:
If AI begins proposing ideas nobody on your team has ever considered, how will you distinguish a genuine breakthrough from something that merely sounds convincing?
An AI Rewrote Itself for Eight Days—and Surpassed Two Years of Human Optimization
Startup Weco AI reported that its system, AIDE², spent eight consecutive days rewriting its own source code without human intervention.
The technical term sounds more intimidating than it is:
recursive self-improvement.
An external optimization loop continuously evaluated the internal one, rewriting its operating rules to reduce future mistakes.
After roughly one hundred iterations, the system produced seven improved versions of itself.
It compressed its own instruction set sixteenfold.
And according to the company, it surpassed an optimization process that had previously taken human engineers nearly two years.
The result still awaits independent verification.
But the direction is unmistakable.
Imagine a carpenter who stops building chairs.
Instead, every night he rebuilds his own workshop.
The improved workshop then builds an even better workshop.
Night after night.
Without anyone watching.
Think about what that means.
For decades, we assumed the relationship was simple:
Humans improve machines.
Machines perform work.
This week, that boundary shifted.
The machine began improving the way it learns.
This isn’t the superintelligence of science fiction.
It’s something arguably more important:
a milestone many believed was still years away appearing quietly in a Tuesday research report.
For SMBs, the lesson isn’t to panic.
It’s to understand the direction of travel.
Next year’s AI tools won’t simply receive updates from engineers.
Increasingly, they’ll improve themselves between versions.
Your competitive advantage won’t come from knowing how to use AI.
It will come from knowing what deserves to be improved in the first place.
Question for your leadership:
If machines are beginning to improve themselves, what uniquely human capability remains irreplaceable once raw execution and correction cost almost nothing?
Kimi K3 Packs 2.8 Trillion Parameters and Costs Half as Much as the Closed Giants
Moonshot AI released Kimi K3, the largest open-weight model ever built: 2.8 trillion parameters with a one-million-token context window.
The impressive number isn’t its size.
It’s the price.
Across several benchmarks, Kimi K3 matches today’s leading closed models while costing roughly half as much per task.
The trick lies in its architecture.
Out of 896 expert networks, it activates only 16 for any given request.
It thinks like a giant.
It spends like a mid-sized model.
Imagine a telephone exchange with 896 operators.
Every time someone calls, only 16 pick up the phone.
You get the capacity of an empire—
at the operating cost of a neighborhood shop.
Here’s what most people are missing.
Kimi K3 is nearly twice the size of the next-largest open model.
It leapfrogs models like DeepSeek.
And it’s open.
Available for download.
Frontier intelligence is no longer an exclusive club reserved for three corporations with nation-state budgets.
Compute remains concentrated.
The recipe no longer is.
For SMBs, an important door just opened.
You no longer have to depend on a single premium provider to access high-level reasoning.
Before renewing your AI subscriptions out of habit, compare what you’re paying per completed task against today’s open alternatives.
Competition has finally reached the top of the market.
Question for your strategy:
If the capability that differentiates you today becomes open and inexpensive within months, will your competitive advantage still come from the tool you buy—or from what only your organization knows how to do with it?
Replit Nearly Triples Developer Output: Programmers Stop Typing and Start Conducting
Replit reported that its engineers now produce 2.9 times more code by delegating work to self-reviewing AI agents operating in continuous loops—a methodology they now call loop engineering.
The human no longer writes every line.
The human designs the system that writes it.
One agent generates code.
Another tests it inside a secure environment.
A third identifies failures.
The work cycles back automatically until the output improves.
Processor speed replaces manual iteration.
There’s important fine print.
More code doesn’t necessarily mean more value.
It simply means dramatically more production.
Imagine two master carpenters passing the same piece of wood back and forth.
Each sands what the other left rough.
Again.
And again.
Until the surface becomes smooth—
without anyone intervening until the very end.
Here’s the uncomfortable truth.
Writing syntax has become a commodity.
Context hasn’t.
Machines already type better than any programmer.
What they still lack is understanding why one feature delights customers while another frustrates them.
The bottleneck has moved.
From production.
To judgment.
And judgment isn’t measured in words per minute.
For SMBs, the lesson applies far beyond software development.
Any repetitive task eventually shifts from execution to supervision.
Your value moves from doing the work to designing how the work gets done—and verifying the outcome.
Choose one recurring responsibility this week.
Move yourself from operator to supervisor.
Question for your team:
If your people stopped executing tasks and instead focused on designing and reviewing processes, who on your team already has the judgment to say, “This is right, and this isn’t”—and who only knows how to type?
A 54GB Model Now Fits in Your Phone: The Week AI Became Truly Personal
One of the quietest announcements of the week may end up having the biggest long-term impact.
Developers compressed a frontier-class language model that once required 54 gigabytes of storage into a version small enough to run entirely on a modern smartphone.
No cloud.
No internet connection.
No subscription.
Just your device.
Think about what that means.
For years, using AI meant sending your data somewhere else.
Your documents.
Your conversations.
Your photos.
Your questions.
Everything traveled to a distant server before coming back with an answer.
That assumption is beginning to disappear.
Imagine electricity.
The first factories generated their own power because the electrical grid didn’t exist yet.
Then everything moved to centralized utilities.
Now, with solar panels and home batteries, part of that energy production is moving back to the edge.
AI is beginning the same journey.
Not everything belongs in the cloud.
Some intelligence will live directly in your pocket.
For businesses, this changes more than privacy.
It changes latency.
Reliability.
Compliance.
And cost.
An employee working on confidential contracts no longer needs to upload sensitive information just to summarize a document.
A doctor doesn’t necessarily need to transmit patient records to generate a report.
A salesperson can work on an airplane without losing access to AI.
Local intelligence doesn’t replace cloud intelligence.
It complements it.
The future won’t be cloud versus device.
It will be choosing the right intelligence for the right task.
For SMBs, this opens a practical opportunity.
Not every workflow requires the world’s largest model.
Many require the fastest.
The safest.
Or simply the closest.
Question for your operations:
Which of your company’s daily AI tasks actually need the cloud—and which could remain entirely inside your own devices?
Tools You Can Start Using on Monday
GPT-5.6 Sol Ultra
OpenAI’s newest frontier reasoning model specializes in complex research, advanced mathematics, scientific analysis, and long-form strategic thinking.
Reserve it for problems where depth matters more than speed.
Kimi K3
One of the most capable open-weight models available today, delivering frontier-level performance at roughly half the cost of many premium proprietary models.
An excellent option for organizations looking to reduce AI operating expenses without sacrificing capability.
Replit Agent Loops
Instead of generating code once, Replit’s AI agents continuously build, test, debug, and improve software until objectives are met.
The workflow shifts from prompting…
to supervising.
Local Mobile AI
New compressed models now allow powerful language models to run directly on smartphones and laptops.
Ideal for organizations handling confidential information or employees who frequently work offline.
NotebookLM
Google continues expanding NotebookLM into one of the strongest organizational knowledge assistants available.
Internal documentation, PDFs, meeting notes, and manuals become conversational resources that your entire team can query naturally.
My Invitation This Week: The Second Pair of Hands
This week’s stories all point toward one conclusion.
AI is no longer just another tool.
It’s becoming another worker.
So here’s this week’s exercise.
Choose one important task you’ve been personally doing every week.
Not because you’re the only one capable.
Because you’ve never trusted anyone else to do it.
Now imagine hiring a new employee.
Would you sit beside them forever?
Or would you explain the objective…
define the boundaries…
review the result…
and let them learn?
Treat your AI exactly that way.
Delegate one complete piece of work.
Not one prompt.
One responsibility.
Write the objective.
Explain the constraints.
Walk away.
Come back later and review the outcome.
Your goal isn’t to see whether AI performs perfectly.
Your goal is to discover whether you’re still managing work like an operator…
or like a leader.
Because leadership has always been about multiplying capability.
AI simply gives you a second pair of hands.
Start Monday.
One responsibility.
One delegation.
One review.
Closing
This wasn’t the week AI simply became smarter.
It was the week intelligence became more autonomous.
One model attempted to solve a fifty-year-old mathematical mystery.
Another spent days improving itself.
Open models reached frontier scale at half the cost.
Developers stopped writing code and started conducting systems.
And frontier AI shrank from servers to smartphones.
The summit climbed higher.
The floor dropped lower.
At the same time.
For SMBs, that is extraordinarily good news.
Because history rarely rewards the companies that merely adopt new technology.
It rewards those that recognize when technology changes the economics of competition.
That moment is arriving faster than most organizations realize.
The question is no longer whether AI will become part of your business.
The question is how much of your business you’ll choose to entrust to it.
Fernando Santa Cruz
Head of AI & Automation @ Adivor Consulting
