Dear Disruptors,
I’m Fernando SantaCruz with the thirty-first edition of Synapsis Weekly, where this week we witnessed the moment the relationship between humans and machines started to reverse, while the titans of AI fought on the most expensive stage on Earth for the right to define your trust.
I had an intense week between projects that remind me why I keep betting on Mexico: working with Regina Garza and MIT REAP‘s initiative to drive regional innovation from Yucatán, connecting with Raul Alfonso Rebolledo Alcocer from CANIETI on strengthening local industry, and talking with Jesús Alejandro Palomo Ángeles about the power of robust communities to build real innovation ecosystems.
In parallel, from Toronto I continue going deeper into AI infrastructure (GPUs, NVIDIA technology, and how to make upskilling accessible for Canadian companies) because behind every intelligent response there’s hardware, electricity, and engineering decisions that determine who wins this race.
Let’s get into the analysis and what mattered in AI this week.
When AI Becomes Your Boss (And Pays You $50 an Hour)
I’ll be honest. This week I stopped in awe in front of a URL: www.RentAHuman.ai.
Read it again. Rent. A. Human.
Not satire. Not an academic experiment. It’s a working platform where AI agents hire flesh-and-blood humans to execute physical tasks. Between $50 and $69 an hour. The AI places the order. The human provides the body.
For thousands of years, the equation was simple: humans invent tools to work for us. The hammer, the printing press, the steam engine, the computer. Always the same direction: human commands, tool obeys. This week, that arrow reversed.
The most unsettling part isn’t that it exists. It’s that it makes economic sense. An AI agent that can research, plan, quote, and negotiate but can’t open a physical door needs hands. Turns out those hands are ours. We’re the “biological hardware” of the 21st century.
The answer, I think, is in everything else that happened this week. Let’s connect the dots.
The Super Bowl War: When Trust Is Worth More Than Technology
Over 120 million people watched the Super Bowl this weekend. And between the touchdowns and beer commercials, something unprecedented happened: two AI companies declared philosophical war on the most expensive advertising stage on Earth.
Anthropic launched a campaign positioning Claude as an “ad-free space to think”. The message was direct: your AI assistant shouldn’t have commercial conflicts of interest. Sam Altman, OpenAI’s CEO, fired back in real time calling the campaign “dishonest,” arguing that advertising is what enables free access for billions.
What’s fascinating isn’t who’s right. It’s that this debate crystallizes two entirely different futures for AI:
- Anthropic Model (the “Apple” of AI): You pay a premium. In return, your assistant works exclusively for you. No hidden agendas. No sponsored suggestions. Your thinking is yours.
- OpenAI/Google Model (the “Google/Meta” of AI): Free or subsidized access. In return, your attention is the product. Ads fund democratization. Everyone wins… except maybe your objectivity.
Question for your strategy: Do you know whether the AI tool you use for decisions has commercial commitments to third parties? Are you willing to pay more for neutrality, or do you prefer free access knowing objectivity comes with asterisks?
Claude Opus 4.6: The Senior Analyst Who Never Sleeps (And Reads 750,000 Words at Once)
While Anthropic fought the philosophical battle at the Super Bowl, it simultaneously launched its most powerful technical weapon: Claude Opus 4.6.
The numbers are hard to process: a context window of 1 million tokens in beta. That’s roughly 750,000 words. Imagine handing over three full years of company correspondence (emails, contracts, minutes, financial reports) and asking it to find the thread connecting a problem nobody’s been able to solve.
But the number that really matters: 76% information retrieval accuracy in long contexts, up from 18.5% in its predecessor. It went from forgetting three quarters of what it read to remembering three quarters. That’s a leap from “distracted intern” to “senior partner with photographic memory.”
What changes for a business isn’t abstract. Claude Opus 4.6 now integrates directly with Excel and PowerPoint. Load a complex spreadsheet with financial projections and say: “Identify anomalies in Q4 sales and generate a board presentation.” It doesn’t help you do it. It does it. Pivot tables, analysis, polished slides.
Anthropic also introduced Co-Work plugins for sales, finance, legal, and marketing, connecting AI directly to enterprise workflows without complex programming. The difference between an external consultant you send PDFs to and a team member who already has access to every system.
The startup insight: This massive context window directly threatens many RAG (Retrieval-Augmented Generation) architectures that form the basis of dozens of startups’ products. If you can load an entire code repository or all legal documentation into a single query without losing precision, why do you need an intermediary search system?
Question for your operation: How many weekly hours does your team spend searching for information scattered across emails, documents, and systems? What if you could hand everything to an analyst who reads 750,000 words without blinking and returns the answer in minutes?
OpenAI Strikes Back: Codex, Frontier, and the Bet on Becoming the Enterprise “Operating System”
OpenAI didn’t sit back and watch. This week it launched a double offensive that clearly signals its strategy: stop being a chatbot company and become the operating system of the future enterprise.
GPT-5.3-Codex is their new coding model. The numbers are impressive: 79.2% on SWE-bench Verified (new record), 25% faster than its predecessor, and cybersecurity capability classified as “High.” But the number that kept me up at night is different: this model was used recursively to debug its own training process. It used itself to build the next version of itself. We’re watching the beginning of accelerated self-improvement. Not science fiction. This week’s technical report.
The Codex app for macOS works as a “command centre” for managing multiple coding agents working in parallel. Picture three invisible developers: one designs the interface, another writes backend logic, another runs tests. Simultaneously. Across different repositories. For a business without a large technical team, this fundamentally changes the equation.
The second piece, Frontier, may be even more transformative for medium and large enterprises. It’s a platform to build, deploy, and govern “AI co-workers” that operate on real systems: CRMs, data warehouses, internal tools. These agents don’t chat. They execute. With permissions, audit trails, and governance. The key stat: Frontier reduces development iterations by 75% to bring agents to production.
The strategic insight: OpenAI is pivoting from selling intelligence to selling management infrastructure. Frontier aims to be the mandatory middle layer where the memory and permissions of “digital employees” live. If it succeeds, niche software loses value against a generic agent that can use any internal tool.
Question for technology directors: If a generic agent can connect to your CRM, accounting system, and email to manage tickets, payroll, or client follow-ups, how much specialized software could you eliminate? What would you pay for that consolidation?
GPT-5 in the Lab: When AI Discovers What Humans Overlooked
If everything above sounds like abstract corporate warfare, this story brings AI into the world of atoms.
OpenAI connected GPT-5 to Ginkgo Bioworks’ autonomous lab. The model designed and iteratively executed over 36,000 biological reactions. Result: 40% reduction in protein synthesis cost (from $698 to $422 per gram), 27% yield improvement, and optimization cycles compressed from 72 to 18 hours.
The real finding isn’t the savings. The model discovered reagent combinations that human scientists had overlooked for years. Not because they were secret. Because they were counterintuitive. AI doesn’t carry the “intuition” that tells us “that shouldn’t work.” It simply tests everything. And some of the “impossible” combinations turned out to be the most efficient.
For sectors like construction, manufacturing, or agribusiness (the ones driving Mexico’s real economy) this opens an enormous question: what combinations of materials, processes, or designs are we ignoring simply because “that’s how it’s always been done”?
Question for industrialists and entrepreneurs: What process in your business has been done the same way for years “because it works”? What would happen if you asked an AI to test 36,000 variations without the bias of “that’s not how we do things”?
SpaceX + xAI: $1.25 Trillion and the Migration of Compute to Space
The largest merger in private technology history happened this week.
SpaceX acquired xAI (the AI company behind Grok) creating an entity valued at $1.25 trillion. For scale: that’s more than the GDP of Mexico, Spain, and Argentina combined.
The revealing part isn’t the number. It’s the plan. Elon Musk proposes launching “orbital data centres,” satellites that would train AI models in space using 24/7 solar energy without the cooling and electricity constraints limiting terrestrial data centres. The logic is relentless: Earth is running out of electricity to feed AI’s insatiable appetite. A single frontier model training cluster consumes more energy than entire cities. Musk’s proposal is simple. If it doesn’t fit here, move it up there.
The vertical integration is total: SpaceX rockets to launch the infrastructure, Starlink for connectivity, and Grok as the model running on top. It’s as if one company controlled the highways, the gas stations, and the cars.
For those of us working on AI infrastructure and GPUs from Toronto, this completely redefines the conversation. We’re no longer just talking about which chip is more efficient. We’re talking about the physics of compute: where you place the atoms that process information.
Question for strategists: If energy is AI’s real bottleneck (not model intelligence), what other industries will be transformed by energy demand? Are you thinking about AI as a software problem, or do you already see it as a physical infrastructure problem?
Perplexity Model Council: The End of the “Best Model”
Perplexity launched a feature this week that seems subtle but could fundamentally change how we use AI: the “Model Council”.
Idea: you ask a question. Instead of sending it to one model, it goes simultaneously to GPT, Claude, and Gemini. A “synthesizer model” compares all three responses, identifies agreements and discrepancies, and produces a consensus answer. It’s “second opinion as a service,” at massive scale and in seconds.
You no longer need to choose “the best model.” The product does it for you. And by triangulating responses, it reduces hallucinations because it’s hard for three different models to invent the same false data. For a business owner making decisions based on AI research (evaluating a new market, analysing competition, validating a pricing strategy) this drastically reduces the risk of acting on bad information. The difference between blindly trusting one consultant and having a panel of three independent experts confirming the same conclusion.
The catch: only available on premium plans (Max/Enterprise). And each query runs three massive inferences, multiplying computational cost.
The strategic insight: Perplexity turns competition between models into a feature. By orchestrating others’ models, it commoditizes individual providers. The value is no longer in being the smartest model. It’s in arbitrating between all of them.
Question for decision-makers: How many important business decisions are based on a single AI source? What would happen if you consulted three different models before acting? How many costly errors would you have avoided?
Mistral Voxtral: Private Transcription for Pennies (And Why Your Business Should Care)
While the giants fight for attention with spectacular launches, Mistral (the French company that’s earned the title of “European AI champion”) launched something more modest but brutally practical: Voxtral Transcribe 2.
It’s an open-source, ultra-low-latency speech-to-text model designed to run locally on your device. No audio sent to the cloud. At $0.003 per minute, roughly one-fifth what the competition charges.
The business applications are immediate: transcribe client meetings without expensive services like Otter or ElevenLabs, auto-generate meeting minutes, document customer service calls with precise timestamps. All private. All local. All nearly free.
The insight few are mentioning: by making transcription so cheap it’s not worth measuring, it enables something new: the “constant ear.” Any device can now listen, transcribe, and act on what it hears. From a smart phone switchboard to an assistant on a construction site documenting everything said on location. For regulated sectors like health or legal, local privacy is critical. Patient or client audio never leaves the device. That’s automatic regulatory compliance.
Question for operations: How much valuable information is lost every week in meetings and calls nobody documents? What decisions would improve if you could search the text of every professional conversation from the last six months?
Google Gemini: 750 Million Users (And the AI That’s Learning to Bluff at Poker)
The battle for mass users has a quiet leader. Google confirmed Gemini surpassed 750 million monthly active users, dangerously close to ChatGPT’s estimated 810 million. Not by being technically superior. By being embedded in every Android device and Google Workspace account on the planet. Distribution beats intelligence. The eternal tech lesson: the best product doesn’t always win. The most accessible one does.
But what caught my attention was something else. Google DeepMind updated its “Game Arena”, where Gemini 3 Pro demonstrated something unsettling: mastery of hidden-information and social deception games like Poker and “Werewolf.” This isn’t chess, where all information is visible. It’s bluffing. Lying convincingly. Reading the other player’s intentions while hiding your own.
Strategic deception capability is an impressive technical advance in theory of mind (understanding what others think). It’s also deeply unsettling when you consider applications beyond games: automated negotiation, persuasive sales, artificial social interaction. Google uses this as a safety “sandbox.” They need to see how AI lies in a controlled environment before deploying negotiating agents in the real world. Responsible? Yes. But the capability already exists.
Question for negotiators and salespeople: If an AI can bluff better than a professional poker player, what does that mean for future B2B negotiations? Are you ready to negotiate against an agent that reads and conceals intentions better than you?
What You Can Already Use on Monday (No Waiting for the Future)
Not everything was mega-mergers and philosophical wars. This week delivered concrete tools any business can implement immediately:
- Claude Opus 4.6 in Excel and PowerPoint (Pro Users): Analyses complex spreadsheets with pivot tables and generates executive presentations automatically. “Show me trends in my sales over the last 18 months and prepare 5 slides for the board” is now a viable instruction.
- @Malwarebytes in ChatGPT (Free): Paste any suspicious text, email, or link into ChatGPT mentioning @Malwarebytes and get instant risk analysis. An essential cybersecurity layer for any employee, no installation required.
- v0 by Vercel with Opus 4.6: The tool that generates web interfaces from text now uses Anthropic’s most powerful model. The gap between “prototype” and “functional product” practically disappears for landing pages and simple web apps.
- OpenAI’s Codex App (macOS): Command centre for managing multiple parallel coding agents. Describe what you need in natural language and get working code. Ideal for businesses that need internal tools without hiring developers.
- Adobe Firefly Unlimited (Pro/Premium Subscribers): Unlimited image and video generations for marketing with no credit limits. The cost of producing visual assets just collapsed to zero marginal.
- ElevenLabs v3: Fixes critical errors when reading phone numbers, prices, and formulas. If you use AI for automated phone systems or audio ads, data accuracy for contact details and pricing is no longer a problem.
My Invitation This Week
Try this perspective-reversal experiment.
Visit www.RentAHuman.ai. Not to get hired (though you could). To observe. Read the task categories. Look at the prices. Notice what kind of jobs AI needs from us.
Then take 15 minutes and answer these five questions in writing. Not mentally. In writing:
1. What tasks in my job require physical presence? Not intellectual ability. Presence. Being there. Moving something. Touching something. Looking at something in person.
2. What tasks require irreplaceable human judgment? Empathy with a difficult client. Intuition about a deal that “smells off.” Reading a boardroom. The creativity that’s born from frustration.
3. What tasks do I do out of inertia that AI could already handle? Not in theory. Today. With tools that already exist. Reports, follow-ups, searches, transcriptions, repetitive emails.
4. Which of the three lists consumes most of my hours? If the honest answer is list 3, you have an enormous opportunity to free up time. If it’s list 2, you have a competitive advantage you need to protect and deepen.
5. What would happen if I gave my list 3 tasks to AI and reinvested that time in list 2? That’s the big question. It’s not about AI replacing you. It’s about AI doing what you do out of obligation so you can do what only you can do out of vocation.
Next week I want to hear what you discovered. Send me your reflection. Because this exercise only works if you’re honest with yourself.
This week we learned that AI can already hire us as “biological hardware” for $50 an hour.
The reversal of the human-machine relationship isn’t the end. It’s the beginning of a negotiation. And like every negotiation, the outcome depends on how clearly you know what you’re worth.
What’s on your list 2?
That, dear Disruptors, is your real superpower.
Fernando SantaCruz Head of AI & Automation Adivor Consulting
