THE GROUND MOVES THIS FAST. THE ROOM MOVES WITH IT.

September 2026 is a snapshot of how quickly the family-office operating model is changing. 

At the frontier, Claude Fable 5.1 and GPT-6 Astra give a principal access to levels of reasoning, research, coding, and computer use that only recently would have required an extraordinary team. The opportunity is not simply to ask better questions. It is to put multiple intelligences against the same problem at once — one developing the thesis, another attacking it, another searching the family’s history, another building the downside case — while the human remains responsible for the decision.

At the same time, the AI Chief of Staff is becoming real.

With Grok Bot, Muse, and Instinct, the model is moving beyond a chat window. You can increasingly create persistent digital workers with defined responsibilities: a research analyst monitoring holdings, a diligence agent reviewing opportunities, an operating assistant tracking unresolved commitments, or a chief of staff coordinating the work and surfacing only what requires human judgment.

Instinct makes the idea especially easy to understand: rather than learning another complicated AI interface, imagine texting or calling an intelligent assistant that works across the systems you already use and increasingly notices what has not been finished before you have to remember it yourself.

That changes the economics of a small family office.

You do not need thirty people to begin building thirty capabilities.

And the intelligence no longer has to live entirely in somebody else’s cloud. Qwen3.8-27B is an example of increasingly capable open-weight intelligence that can run on hardware many families already own or can buy inexpensively. That means private documents, historical decisions, investment memoranda, entity information, and institutional memory can increasingly be processed inside infrastructure the family controls.

Then comes the next question: once agents begin acting, how do you govern them? That is why work such as Buzz, from Block, matters. The family increasingly needs to know which human authorized which agent, what it did, and whether that history can be reconstructed later. AI governance is becoming family-office governance.

Now put the pieces together.

Imagine an investment opportunity arrives Thursday afternoon. Before Friday morning, your system has read the material, compared it with previous family decisions, researched the counterparty, challenged the thesis from several directions, built a downside case, and identified what still cannot be verified.

Or imagine a principal saying on Monday morning:

“Watch these ten holdings, these three counterparties, and these two jurisdictions. Do the work continuously. Bring me only what changes the decision.”

That is the family office we are beginning to build.

And this is the point of the AI Family Office Masterclass.

You are not coming to New York merely to hear about Fable, Astra, Grok, Muse, Instinct, Qwen, or whatever surpasses them before November. My books can explain the technology.

The room is where we build with it.

Skills. Agents. Investment red teams. Institutional memory. Private intelligence. Operating workflows. Governance. Tools that remain useful after the two days end.

At the August New York City Masterclass, faculty and members developed 27 working Skills around the broader two-day program. November will not repeat August. By November 4–5, the frontier will have moved again, and the room will begin with what is possible then.

That is why I run these Masterclasses multiple times each year.

What can your family office build now that it could not build three months ago?

That is the question we will answer in New York.

SFO Continuity membership includes your seat for November 4–5.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

THE TRADITIONAL FAMILY OFFICE WAS BUILT FOR A WORLD THAT NO LONGER EXISTS

The traditional family office was designed for a world in which intelligence was expensive, most execution had to be human, information moved more slowly, and the safest answer was often to build committees, hire more people, and rely on a trusted network of outside institutions. 

That architecture made sense.

The environment changed underneath it.

Intelligence is becoming dramatically cheaper. Markets, regulation, geopolitics, and technology move faster. AI can now monitor, research, prepare, compare, remember, and execute continuously in ways that previously required layers of people and outside providers.

That does not mean the great family office replaces people with machines.

It means the best people stop spending their time on work that no longer requires them.

It means the family’s judgment is preserved instead of walking out the door when someone retires. It means important changes are surfaced before the next quarterly meeting. It means the principal can bring far more intelligence against a decision without building a larger bureaucracy around it.

For an established family office, the opportunity is to rebuild selectively around this new reality.

For a family creating its first office, the opportunity may be even greater: you can begin without installing the legacy architecture in the first place.

That is what I mean by an AI-native, sovereign family office: the family owns the information, memory, judgment, and critical operating architecture that should remain its own — while using the best external intelligence in the world where it creates an advantage.

The objective is not a more technological family office. It is a more capable one.

And that is what the two days in New York are designed to begin building.

"A high-performing family office is not measured by how busy it is. It is measured by the quality of its decisions, the speed of its execution, the strength of its people, the clarity of its purpose, the adaptability of its systems, and the degree to which the family becomes more capable because the family office exists." — Angelo Robles

---

TWO IMMERSIVE DAYS. ZERO THEORY. ALL OPERATIONAL EDGE.

November 4–5 · New York City 

This is not a conference about what AI may eventually do to the family office. It is a working room built around a much more practical question:

What can your family office build now that materially changes how it operates when you return home?

Over two days, we work from the frontier as it actually exists in November — not from a syllabus written six months earlier. The models will have moved. The agents will have improved. New tools will exist. Some ideas that looked experimental in August will already be practical.

That is exactly why the room exists.

You will not finish a sovereign AI family office in forty-eight hours, and we do not pretend otherwise. What you will do is begin building the architecture with your own hands, alongside faculty and members who are already doing the work: Skills, agents, investment red teams, institutional memory, private intelligence, governance, and workflows that continue to matter after the event is over.

The real danger is not that another family knows more about AI than you do.

It is that they started earlier, and their systems, judgment, memory, and operating fluency have been compounding while yours has not.

That gap cannot be closed later simply by buying better software.

SFO Continuity runs these Masterclasses four times each year because the frontier moves too quickly for an annual event — and because building capability is a process, not a presentation.

Your seat for November 4–5 is included with SFO Continuity membership.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

WHAT THE TWO DAYS ACTUALLY PROVE

The Masterclass exists to prove something in the only way that really matters: by letting you build it yourself. 

You see, on your own machine and against real family-office problems, that AI can move beyond conversation into execution — agents, Skills, institutional memory, investment red teams, private intelligence, and workflows that continue operating after the room closes.

The revelation is rarely simply that AI is powerful. Most people know that before they arrive.

It is realizing how much more capable their own family office could already be — and that the distance between where they are and where they want to go is buildable.

The two days begin that build.

SFO Continuity is what keeps it moving after New York.

---

YOU DON'T NEED TO BE TECHNICAL. NO PRIOR AI EXPERIENCE REQUIRED. NONE.

You do not need to become an engineer to build an AI-native family office. 

You need to understand enough to direct the engineers, the models, and the agents — and to know what should be built in the first place.

That distinction matters.

The highest-value skill in this new operating model is not coding. It is judgment: knowing which problem is worth solving, what a good answer looks like, where human authority must remain, what the family should own, and when the machine is wrong.

At our August Masterclass, one of the family principals in the room was 83 years old. He attended with his family-office executive. He did not arrive trying to prove that he understood the technology better than everyone else. He listened, challenged assumptions, and kept asking the questions that forced the room to think more deeply.

By the end of the two days, my faculty broadly agreed that some of the best questions in the room had come from him.

That is the point.

AI does not reward youth as much as people assume. It rewards curiosity, judgment, pattern recognition, and the willingness to engage with something new before you feel completely comfortable with it.

I am in my early sixties and came into this with no technical background. Over the last two years I have become deeply fluent because I kept building, testing, asking questions, and refusing to outsource my understanding of something this consequential.

You can do the same.

The objective is not to turn the principal into a programmer. It is to turn the principal into a far more capable orchestrator — someone who can sit above extraordinary human and machine intelligence and know how to direct both.

That is why the Masterclass is hands-on.

You will be surrounded by faculty and members who can help with the technical layer, but you will not sit passively while someone demonstrates it from a stage. You will work with the systems yourself, see what they can do, understand where they fail, and leave far more capable of directing what comes next.

You do not need a technical background. You need to begin.

---

WHAT YOU ACTUALLY WALK OUT WITH

Not a binder. Not two days of impressive demonstrations. Not a list of AI tools you will forget by the following week. 

You leave with the beginnings of an operating capability.

The first shift is conceptual. Most family offices using AI today are still chatting: open a window, ask a question, close the window. It feels productive, but almost nothing persists.

The Masterclass moves you from chatting to building.

You begin to understand the simple architecture underneath an AI-native family office: how instructions are written, how judgment becomes reusable, how agents receive standing responsibilities, how institutional memory is organized, how private intelligence can stay under family control, and how different models are used for different jobs rather than treating one chatbot as the entire strategy.

Then you begin building those pieces yourself.

YOUR FIRST AI-NATIVE FILES

A surprising amount of the architecture begins with ordinary Markdown — plain-text files the family owns permanently and that can travel across Claude Code, Codex, Hermes, and whatever system becomes best next.

You begin creating the files that tell the intelligence how your office actually operates: its priorities, rules, decision standards, standing mandates, and repeatable methods.

This is where a Skill becomes much more important than a prompt.

A prompt asks the machine to do something once. A Skill captures how an exceptional person performs a recurring task so that the method can be used repeatedly, improved over time, and carried forward even when the person who originally held the judgment is no longer in the room.

That may be a due-diligence process. A premortem. A red-team review. A manager evaluation. An archive interrogation. A repeatable way of turning a complex problem into a decision-ready answer.

The family stops renting the process every time it needs the answer.

It begins owning the method.

YOUR SECOND BRAIN BEGINS

The most valuable information inside a family office is rarely the document itself. It is the reasoning behind what the family did.

Why did we pass on that investment? Why did we trust that partner? What concerned the founder in 2018? What happened afterward? Which assumptions turned out to be wrong?

Normally, that knowledge lives in people’s heads until it disappears.

You begin learning how to organize that history into a searchable, reasoning memory for the family — what I call the Second Brain — so that the institution can increasingly retrieve not only what happened, but why.

That is not simply an AI feature. It is continuity.

YOUR FIRST WORKING TOOLS

This is also where the abstract becomes tangible.

In June, one of our faculty members converted years of investment judgment into the Family Office Investment Thesis Advisor. At the August New York Masterclass, faculty and members developed 27 working Skills around the broader two-day experience — usable resources that participants could continue building on after the room ended.

November will create its own body of work against the technology and problems that exist on November 4.

And remember the rule: you must be a member, and you must be in that particular room, to receive what that room builds.

What gets built there stays with the members who were there contributing to it.

YOU BECOME THE ORCHESTRATOR

The goal is not to make you technical for the sake of being technical.

It is to make you fluent enough that you can sit above the architecture and direct it intelligently.

You should understand what a Skill is, what a second brain does, why some work belongs in the cloud and some should stay private, what an agent can and cannot be trusted to do, and how plain-English intent can increasingly become working software.

That level of fluency is more attainable than most principals assume.

The benchmark inside family offices is still extraordinarily low. Someone who begins seriously now, spends time with the books, works through the Masterclass, learns beside peers doing the same thing, and keeps building afterward can progress remarkably quickly because so much of the field is still standing at the starting line.

You do not need ten years of technical experience.

You need thirty days of serious engagement to stop being a spectator.

And then the learning compounds.

THE PEOPLE BUILDING BESIDE YOU

This may ultimately be the most valuable asset of the two days.

You are not trying to solve this alone.

You are sitting beside principals, executives, next-generation family members, and faculty who are confronting the same questions: what to build, what to own, what failed, what worked, what became obsolete since the last room, and what suddenly became possible.

That is where the learning accelerates.

Brendan Boyle, CIO of a single family office, described it well:

“Angelo’s AI Family Office Masterclass takes what’s coming and makes it tactical — not theory, but what to build now. He challenged me, and every challenge evolved the family’s thinking. He is a large part of why our office is as AI-forward as it is today. In SFO Continuity, he has built something I haven’t found anywhere else in the SFO space: cutting-edge perspective on the AI-native family office, at a level you can actually implement.”

That is the outcome I care about.

Not that you leave impressed by artificial intelligence.

That you leave more capable of building with it.

And because the finished operation is built in the weeks and months after New York, the Masterclass is not the end of the process. SFO Continuity is the environment that keeps the building moving after the room closes.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

WHAT THIS ACTUALLY DOES FOR YOUR FAMILY OFFICE

Real Operations. Real Workflows. Real Results. 

Picture a Tuesday.

You have a decision to pressure-test — a deal, a manager, a counterparty, a jurisdiction. Instead of opening a chat window and beginning another conversation with AI, you brief the office the way you would brief an exceptional chief of staff: by voice, from your phone, perhaps from the car.

The system divides the work. One agent researches the counterparty and its history. Another attacks the investment thesis. Another models the downside. Another checks the structure and jurisdictions involved. Another searches the family’s own past decisions for anything relevant. The work happens in parallel, and by the time you are ready to decide, the analysis has been assembled around the question rather than around whichever analyst happened to have time.

The principal still decides. The architecture makes sure the principal is deciding from a radically better position.

That is the practical shift.

In investment research and due diligence, the office can interrogate an opportunity from several independent directions before capital moves. Instead of asking one analyst to prove the thesis, you can deliberately build adversaries whose job is to find the hole, challenge the assumptions, identify what cannot be verified, and surface the one issue the entire investment may turn on.

In macro, geopolitical, and counterparty intelligence, the office no longer has to wait for someone to remember to look. Agents can monitor the specific countries, regulations, central banks, companies, counterparties, and risks that matter to your family across the full 168-hour week, then surface only the developments that cross a threshold you defined in advance.

In portfolio and operating oversight, the architecture can be shaped around your actual family rather than around somebody else’s software. A dashboard built for your entities. Monitoring built for your concentrations. Reporting built for the decisions you actually make. The increasing ability to describe what you need in plain English and have software built around it changes the economics of customization completely.

And across legal, tax, regulatory, and risk work, AI does not replace excellent counsel. It changes what you ask counsel to spend expensive human judgment on. The surveillance, document review, first-pass analysis, comparison, and preparation can increasingly happen before the call, so the professional enters where their judgment is actually valuable rather than spending the first hour reconstructing information your own systems could already have assembled.

This is what makes the opportunity larger than “AI productivity.”

A family office has historically been constrained by how many intelligent human hours it could bring against a problem at one time. That ceiling is lifting.

The office can increasingly monitor continuously, research in parallel, remember what humans forget, challenge its own conclusions, and produce working software around the family’s particular needs — while the principal and the best people remain focused on judgment, relationships, and the decisions that genuinely require them.

That is what SFO Continuity members are learning to build — Masterclass by Masterclass, capability by capability.

The next section goes one level deeper into what may be the most valuable capability of all: making sure the judgment that built the family’s wealth does not disappear with the people who created it.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

YOUR FAMILY'S MOST VALUABLE ASSET IS STORED IN THE ONE PLACE GUARANTEED TO FAIL

The most valuable asset in a great family office is not the portfolio. 

It is the judgment that built it.

Why did the family pass on that investment in 2009? Why did the founder trust one counterparty and reject another? What did a failed deal teach the office that never made it into the final memo? Which rule was broken once — correctly — and why?

That intelligence usually lives in the most fragile place imaginable: people’s heads.

A founder retires. A CIO leaves. A trusted advisor dies. A new generation arrives. The capital transfers, but much of the reasoning that produced it does not.

That is one of the quietest failures in private wealth.

AI creates a different possibility.

A family can begin building what I call a Second Brain: not simply an archive of documents, but a searchable, reasoning memory of the family’s decisions, frameworks, mistakes, exceptions, and accumulated judgment.

Imagine asking:

“We are considering this investment. What have we done before that resembles it? What assumptions were wrong then? Which counterparties have we encountered before? When have we broken our own rules, and what happened?”

The system does not merely retrieve files. It brings the relevant history into the current decision.

That is where institutional memory becomes institutional intelligence.

And it compounds.

Every important decision captured today makes the system more useful tomorrow. Five years from now, the family that began building this memory will be reasoning against five years of accumulated context that the family beginning later simply does not possess.

The technology can be purchased later.

The lost years of captured judgment cannot.

That is why I consider the Second Brain one of the most important things a family office can begin building now.

Not because it is an impressive AI application.

Because it solves one of the oldest problems in wealth: how to transfer judgment, not merely assets.

You begin building yours in the Masterclass — against your own history, your own decisions, and your own operating reality.

---

CHAT ISN'T DEAD. IT'S JUST NOT ENOUGH

The Era of Agency Is Here. 

For the first few years of generative AI, the experience was mostly conversational. You opened a window, asked a question, received an answer, and then you did the work.

That remains enormously useful. I use chat constantly. But the more consequential shift is from AI that answers to AI that acts.

Agency means giving intelligence an outcome rather than a single prompt: research this opportunity, interrogate the documents, compare it with our history, attack the thesis, build the model, prepare the questions for the meeting, and come back when the work is ready for human judgment.

The system can divide that mandate among specialized agents, use software and tools, work for hours without continual supervision, and return something much closer to completed work than another conversation.

For a family office principal, imagine saying from your phone:

“Prepare me for Thursday’s manager meeting. Compare the opportunity with anything similar we have seen before, identify the three assumptions most likely to be wrong, red-team the investment case, and give me the questions I should ask.”

The principal does not become a programmer. The principal becomes the orchestrator.

That is the new skill: learning how to define the outcome, establish the boundaries, decide what the machine may do on its own, and reserve the final judgment for the human.

This is one of the central shifts we build around in the November 4–5 AI Family Office Masterclass. You are not only learning how to talk to AI more intelligently. You are learning how to put intelligence to work.

SFO Continuity membership includes your seat.

---

SIX AI ECOSYSTEMS. EACH WITH IRREPLACEABLE SUPERPOWERS.

The question is no longer "which AI do we use?" It is "how do we orchestrate all six — routing the right model to the right task automatically?" That is the new alpha. That is Session One, Day One.

OpenAI / ChatGPT. GPT-5.6 shipped July 9 in three levels — Sol, Terra, Luna — and Sol is now the lead orchestrator of this architecture: a million-token context, native computer use, agentic workflows across your applications. Codex turns plain language into working software; ChatGPT Work is OpenAI's head-on answer to Claude Cowork, and we give you the honest comparison in the room.

And on September 3, GPT-6 Astra — the successor to Sol, OpenAI's most capable model, the first at its Critical cybersecurity threshold, priced at Fable's ten and fifty. It does not take the orchestrator's seat today, and in the room you learn exactly why: the seat is decided by terms, cost and proven behaviour, never by which model is best this month.

Anthropic / Claude. Opus 5 is the deep-work standard and the second seat — the adversarial counterweight to the orchestrator, holding your institutional memory in a million-token context. Sonnet 5 is the fast, high-volume workhorse. 

Above them sits Claude Fable 5.1, released September 1 — the Mythos-class tier, the sharpest instrument any lab offers, summoned deliberately for the hardest decision and stepped back down from. For four editions of my books it was never trusted with the family's crown jewels, because there was no way to run it without a thirty-day retention hold. As of September 1, eligible accounts can run it with zero retention — so the second seat is now a test your own paper answers. We run that test in the room.

Google / Gemini. Gemini 3.1 Pro remains the deep multimodal engine for work where text, images, documents, financial statements, property imagery, and other unstructured material have to be understood together. But Google is moving quickly: Gemini 3.8 Flash, released September 2, is now its high-volume agentic workhorse, and Gemini 3.8 Live pushes the voice layer toward real-time reasoning and execution. Google’s superpower remains the breadth of the medium — text, vision, documents, audio, video, and live interaction increasingly inside one ecosystem.

xAI / Grok. Grok 4.6 delivers Opus-class reasoning at an affordable price — and it is the only frontier model natively wired into live market data and the social layer, where a counterparty's trouble surfaces days before it reaches a filing. A different information clock than everything else in the stack.

Meta / Muse Spark. Muse Spark 1.3, released September 2, is Meta's fourth in five months — and the independent indexes now score it level with Grok 4.6, Sol and Opus 5 at roughly half their cost per task, built for agentic tool use at a quarter of top-tier pricing, speaking both the OpenAI and Anthropic API formats. The cheapest uncorrelated fourth opinion in the stack.

Perplexity. No longer a search engine. Perplexity Computer orchestrates a fleet of frontier models on autonomous workflows, and its Model Council puts several in parallel on your most consequential decisions — forcing independent conclusions and showing exactly where they agree and where they break apart. Two hundred dollars a month.

---

THE RIGHT ANSWER IS NOT CLOUD OR SOVEREIGN. IT IS BOTH.

The most capable frontier models on earth are extraordinary, and I use them constantly. 

Claude Fable 5.1. GPT-6 Astra. Opus 5. Grok. Gemini.

For a family office just beginning its AI journey, that is often exactly where I would start. The frontier gives you immediate access to extraordinary reasoning with almost no infrastructure to build, and it teaches the principal very quickly what modern intelligence can actually do.

The mistake is not renting intelligence.

The mistake is becoming dependent on intelligence you do not own.

We saw that clearly when Fable was temporarily unavailable earlier this year. The larger lesson was not about Anthropic. It was about architecture. Any frontier provider can change pricing, usage limits, retention terms, product availability, or access. That is not a criticism of the labs. It is simply what it means to build on infrastructure somebody else controls.

A sophisticated family office therefore does not choose between the cloud and sovereignty.

It uses each for what it is best at.

The frontier is where you reach for the deepest reasoning available: the exceptionally hard investment question, the adversarial review, the strategic problem where another increment of intelligence may materially improve the decision.

But the family’s accumulated memory, private documents, repeatable workflows, institutional judgment, standing agents, and much of the day-to-day execution increasingly belong on infrastructure the family controls.

That is where open weights become important.

Models such as Qwen3.8-27B, Gemma, gpt-oss, Nemotron, DeepSeek, and whatever earns the seat next can increasingly run locally, on hardware the family owns, without sending every piece of institutional DNA to an outside provider and without paying by the token for work that may run continuously.

This is not an ideological preference for open source over the frontier.

It is architecture.

Rent extraordinary intelligence when extraordinary intelligence is worth renting. Own the information, memory, judgment, and operating capability that should never depend on somebody else’s permission.

For most serious family offices, that hybrid model is where this is going.

And it is one of the most important distinctions we build around in the Masterclass, because “sovereign AI” does not mean unplugging from the frontier. It means the family decides what leaves, what stays, which model gets which job, and how easily the entire system can move when something better arrives.

The models will change.

The family’s intelligence should remain the family’s.

---

THE SOVEREIGN INTELLIGENCE MANDATE

Stop Bleeding IP. Start Building Your Fortress.

Every prompt your team enters into a commercial AI without a sovereign architecture is an intelligence transfer. Your investment theses. Your tax strategies. Your succession architecture. Your deal terms. Processed, stored, and potentially training the models your competitors will use against you tomorrow.

Don't take it only from me. Alex Karp built Palantir on the thesis that serious institutions must own their compute, their models, and their data — that real safety means knowing you "own the means of production, and it's not being transferred to someone else." Not a vendor's promise. Ownership.

A frontier lab is not your advisor. It is a counterparty whose next product is built from what its customers teach it — you saw the proof above: the finance agents the largest banks now run were assembled from what those banks taught the frontier. So turn the questions on your own office. Who owns your data? Where is it cached? Are your prompts secure — or training material? If you can't answer cleanly, you don't have an AI strategy. You have an intellectual-property leak with a monthly invoice.

And the leak is only half the exposure. You cannot rent intelligence from the same place that rents it to your competitor. Even with perfect security, every office in your peer set feeding the same frontier models converges on the same analyses, the same conclusions, the same quarter. The rented model is the great equalizer — everyone paying a premium for the privilege of becoming more alike. The families in this room build the opposite: a sovereign corpus no other family holds, judgment no other family encoded. Everyone has the frontier. Only your family has your corpus. That is the edge.

The question is binary: Sovereign Intelligence — or Leaked IP. 

"True Sovereignty means owning your AI infrastructure like you own your family's gold." — Angelo Robles

---

SOVEREIGN IS NOT ONE THING. IT'S SIX.

Most family offices thinking about AI sovereignty fixate on data privacy — rightly. But data leakage is only the first layer. Miss any of the six and you are still exposed. Command all six and you have built something that cannot be taken from you.

01 — Data Sovereignty. Every prompt entered into a commercial AI is transmitted to servers you don't own, under retention policies written by lawyers whose clients are those companies, not you. Sovereign architecture routes that intelligence through locally hosted models on hardware you control. The data never leaves. The models never phone home.

02 — Terms Sovereignty. Commercial vendors update their Terms of Service unilaterally. What is permissible today can shift overnight, and you have no recourse — you clicked accept. With sovereign infrastructure, you set the terms. Permanently. There is no counterparty with the power to rewrite your operating environment.

03 — Operational Sovereignty. Vendors go down, throttle access, pivot their offerings — documented events, not hypotheticals. Your Vault runs whether or not a San Francisco server farm is having a bad day. No usage limits, no rate caps, no silent model swaps changing your most sensitive workflows without your knowledge.

04 — Audit Sovereignty. Regulators, trustees, and future generations will one day ask: what AI touched this decision? With sovereign architecture, every query and output is logged in infrastructure you control — full provenance, auditable by whoever you authorize. With commercial AI, you have a vendor's approximation of an audit trail, subject to their interests.

05 — Model Sovereignty. Vendors deprecate models; if your workflows are built on one that disappears, you rebuild on their timeline. Sovereign infrastructure gives you model continuity: you choose what runs, when it updates, and whether it updates at all. The decision is yours.

06 — Cost Sovereignty. You are renting intelligence at prices set by others, designed to capture value as your dependency deepens. The sovereign family office pays once for hardware and runs models at effectively zero marginal cost per query. No per-token rent. No subscription that can be repriced.

The sovereign family office does not ask permission to protect its intelligence. It builds infrastructure that makes the question irrelevant.

---

THE SOVEREIGN STACK: BUILT FOR YOUR FAMILY OFFICE. IN THE ROOM. IN TWO DAYS.

Two years ago, running frontier-capable models on local hardware — privately, at family-office cost — was not possible. That changed. The window to build sovereign infrastructure before it becomes table stakes is open now. Not indefinitely.

Layer One: The Cognitive Vault — a device you own, not a brand you rent. The Vault is a category, not a logo: a small machine in a cabinet you control, running frontier-capable open models with zero cloud exposure. There are three credible ways to build one, and we configure the right one for your family in the room — by ecosystem, by budget, by who maintains it.

The Apple path — the default for most families. The Mac Mini remains the simplest Vault: silent, briefcase-sized, built for continuous operation, at home in any office already living in Apple —refreshed August 25 onto the M6 and M5 Pro, shipping September 22, with the Mac Studio moving to the M5 Max and a new M5 Ultra at up to 512 gigabytes. One honest note: the memory shortage capped high-memory configurations for a period this summer; the August refresh lifted that cap, and a two-bit quantization of the leading open model now fits on a desk-side box. The window to build affordably is open now, not indefinitely.

NVIDIA DGX Spark — the Vault for the NVIDIA world. A Grace Blackwell supercomputer the size of a Mac Mini, 128 gigabytes of unified memory, the full NVIDIA stack preinstalled, running open models up to 200 billion parameters on your desk — and two of them link with a single cable into one 256-gigabyte pool.

AMD Ryzen AI Max+ "Strix Halo" — the most capacity for the least money. The first x86 chip to put 128 gigabytes of unified memory on a single package, in a mini PC running Linux or Windows, at the lowest entry price of the three. For the family that wants the largest sovereign models at the value end.

Three doors. One principle. The intelligence runs on hardware you own, on open weights answering to no vendor's release schedule, and nothing your family thinks ever leaves the building. Which door is right for you is settled in the room — not from a sales page, including this one.

On your Vault, your sovereign arsenal. The everyday sovereigns — Gemma 4 (Google's open weights, frontier-level reasoning small enough to run fast and local), DeepSeek V4.1-Flash (open frontier reasoning under MIT license),Qwen3.8-27B (Alibaba's small flagship — native multimodal, a hundred-plus languages, built for multi-jurisdictional families, and quantized aggressively it comes into range of a 16GB-class laptop), gpt-oss (OpenAI's own weights in your hands, Apache 2.0), and Nemotron 3 Nano (NVIDIA's compact American reasoner, tuned to fly on its own hardware) — carry the bulk of real family-office work on a Mac Mini-class machine.

Of those, watch the Qwen 27B. It is the model that changes who this is available to: quantized aggressively, genuinely capable open weights come into range of a 16GB-class laptop — a machine most people in this room already own. Context length, speed and quantization quality all still bite, and anyone telling you otherwise is selling something. But it means the sovereign floor no longer requires a purchase decision. It requires a download.

GLM-5.3 (released August 14, among the highest-ranked open models in the world — and read its licence, because the flagship's weights no longer carry the MIT terms its predecessor did), Nemotron 3 Ultra (the 550-billion-parameter American heavyweight), and Kimi K2.6 (a trillion-parameter agentic specialist). Be clear what giants mean in metal: these are server-class systems — not the Mini, not a maxed Studio. Running one clean means real GPU iron, or the honest alternative most families choose: the overnight batch run, for the heaviest questions only. We size that honestly in the room — including whether you need the giant at all.

The American lane now has a flagship: Inkling, from Mira Murati's Thinking Machines Lab — at 975 billion parameters, the largest American open weights ever released, Apache 2.0, multimodal, downloadable from day one. When provenance matters and you have the hardware, this is the U.S. flagship.

And Kimi K3 — the largest open-weight model ever released: 2.8 trillion parameters, shipped as public weights July 27. Genuinely competitive with the closed frontier flagships, as a download you own — and far beyond any hardware a family office runs today. Which makes this architecture's standing doctrine a live instruction: download and archive the weights on storage you own, because owning the artifact is sovereignty, even years before you own the machine to run it. Own it this quarter. Run it when the hardware curve catches up.

Why a Chinese model on your Vault never touches China. Every model above runs as open weights: you download the file once from a public registry and it executes entirely on your hardware. You never call the lab's API. A downloaded weight is inert with respect to its maker — no telemetry, no visibility, no way in. The exposure was never the model. It was the hosted API, and the sovereign architecture does not call one. The standing rule is absolute: a Chinese-hosted endpoint is never touched — for any model, at any ranking. What remains is provenance — whose research produced the weights — which is why the arsenal is deliberately mixed: the Chinese frontier for raw capability, the American options for the mandates where provenance is part of the decision. Owning the stack means you make that call yourself, model by model. We build that map with you in the room.

Layer Two: Frontier reasoning — secure pipelines, zero DNA exposure. Your Vault connects through encrypted pipelines to the frontier models — but your Institutional DNA never leaves. Before any query goes out, your sovereign layer strips the identifiers: the family name, the entity structure, the holding, the counterparty. What leaves is the question without the fingerprint. The frontier applies its full reasoning and returns the intelligence; your Vault reassembles the answer inside your walls, against your full proprietary data. Sovereign in. Intelligence back. Your DNA never exposed.

Layer Three: MCP — the architecture that makes everything one. Before MCP, every AI tool was an island. MCP is the universal standard that connects all of it into a single sovereign system: deal flow feeding due diligence agents, CRM feeding relationship context into every analysis, regulatory monitoring wired to your entity structure, your document vault searchable by every agent at once. We architect your MCP integration in the room, built to your specific infrastructure.

Layer Four: The autonomous workforce — 168 hours a week. A 40-hour work week is a sprint. Continuous operation is the new standard: risk and reputation surveillance catching a counterparty's deterioration at the signal stage, weeks before it is news; research swarms connecting macro signals to portfolio implications while the world sleeps; legal and regulatory agents tracking every jurisdiction in real time; perimeter defense against the intelligence threats that come with being a family worth targeting; capital optimization flagging the window that opens at 2am and closes before morning. Your people's time freed entirely for what only human judgment can do.

The phase transition your family office was not built for. Agrarian era: control of land. Industrial era: control of capital. Financial era: control of information asymmetry. AI era: speed and quality of decision-making under uncertainty. Each transition, the families who moved early dominated. Information is now a commodity — your competitors have the same data and the same feeds. The only remaining alpha is Decision Velocity: processing global complexity at machine speed, validating through adversarial multi-model analysis, executing before the window closes. November 4-5 is where you rebuild for it.

This is the architecture we begin building on November 4–5 in New York. 

You do not need to arrive knowing which machine, model, or stack is right for your family office. That is part of the work. You leave understanding the architecture well enough to make those decisions intelligently — and having begun the build yourself.

SFO Continuity membership includes your seat.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

"The SFO of the future isn't just a smaller version of today's model. It is a sovereign fortress of intelligence that produces superior results through elegant efficiency." — Angelo Robles

---

A YEAR AGO, THIS MASTERCLASS WOULD HAVE BEEN IMPOSSIBLE

The most remarkable thing about AI right now is not any single model release. It is the distance the entire operating environment has traveled in less than a year.

Late last year, even Andrej Karpathy — one of the people who understands this technology as deeply as anyone — was publicly skeptical that useful AI agents were close. The systems were too unreliable, too forgetful, too difficult to trust with sustained work. Within months, his own workflow had moved dramatically toward directing agents rather than doing all of the underlying work himself.

That reversal captures the moment better than another benchmark ever could.

The question moved from “Will agents actually work?” to “How much of my work should I now be giving them?”

And since then, the architecture around the models has matured just as quickly.

Take OpenClaw and Hermes. Some of the underlying ideas existed earlier, but what these systems could do then bears little resemblance to what they can do now. OpenClaw became one of the breakout agent platforms in the first part of this year. Hermes has since become, in my view and in the way I increasingly build, the more interesting system for the sovereign family office — particularly when the objective is persistent memory, reusable Skills, standing mandates, schedules, computer use, agent-to-agent work, and intelligence that gets better the longer it operates.  

For someone newer to this, a harness is simply the layer that turns an intelligent model into something closer to a worker.

The model is the brain. The harness gives that brain memory, tools, instructions, schedules, permissions, access to software, and the ability to keep working after the conversation ends.

That distinction is enormous.

A brilliant model sitting in a chat window can answer an extraordinary question. Put the same intelligence inside a properly designed harness and it can begin carrying responsibility.

At the frontier, Claude Fable 5.1 and GPT-6 Astra have pushed the available reasoning and computer-use capability forward again. And tools such as Claude Dynamic Workflows increasingly allow one difficult mandate to be divided across many specialized subagents — research here, modeling there, adversarial review somewhere else — with the work brought back together before the human decides.

At the same time, the idea of an AI staff is becoming accessible to people who will never care what a harness is.

Grok Bot lets you begin defining named digital workers and giving them responsibilities. Muse is moving toward a persistent personal agent that can continue working in the background and coordinate larger tasks. Instinct may be the most intuitive expression of the destination: instead of forcing the principal to learn another AI application, the intelligence increasingly fits around the way the principal already works — text it, call it, connect it to the systems already in use, and let it begin behaving more like a genuine chief of staff than a chatbot.

That is an extraordinary transition.

A family office can begin imagining a research analyst, diligence analyst, operating assistant, monitoring layer, and eventually a chief-of-staff function not as five software subscriptions, but as a coordinated digital workforce that can be stood up and refined far faster than a traditional staff could ever be assembled.

Then the sovereignty side moved just as dramatically.

Qwen3.8-27B is important because it demonstrates how capable open-weight intelligence has moved down the hardware curve. Serious AI can increasingly run on a machine a family may already own — with the model itself downloaded, private work remaining local, and no token meter running every time an agent reads another document.  

That changes the economics completely.

Use Fable 5.1, Astra, and whatever surpasses them next when the hardest problem deserves the deepest intelligence available.

But increasingly, the family can own the intelligence underneath: its documents, institutional memory, repeatable workflows, standing agents, encoded judgment, and a growing share of the execution.

And once those agents begin acting rather than simply advising, governance follows immediately. That is why developments such as Buzz, from Block, matter — not because every family office should install it tomorrow, but because the question it addresses is inevitable: when humans and agents increasingly work side by side, the institution needs to know who acted, who authorized the action, and whether the history can be reconstructed later.  

Look at how far the question has moved.

A year ago, we were debating whether agents were real.

Today we are deciding which agents should do what, which intelligence should remain private, what belongs on family-controlled hardware, how institutional memory should be captured, which frontier models deserve the hardest questions, and how an autonomous workforce should be governed.

That is why comparing the November 2026 Masterclass with what could have been taught in November 2025 is almost meaningless.

It is not the same curriculum with newer model names.

It is a different capability set.

And by November 4–5, this page will already be a snapshot of an earlier moment.

That is exactly why we convene four times a year.

The Masterclass is not designed to teach you a frozen technology stack. It is designed to bring the frontier into the family office as it actually exists when the room opens — then turn that frontier into something useful.

Skills. Agents. Red teams. Institutional memory. Private intelligence. Custom workflows. Sovereign infrastructure. Governance. The machinery that allows a family office to leave materially more capable than it arrived.

You are not coming to hear what changed. You are coming to build with what changed.

SFO Continuity membership includes your seat for November 4–5 in New York City.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

WHY ANGELO ROBLES ON AI

There are excellent technologists who understand artificial intelligence but have never operated inside a serious family office. And there are experienced family-office professionals who understand wealthy families but are still approaching AI primarily as a productivity tool. 

My work sits at the intersection.

For more than twenty years, I have worked inside the ecosystems of over 100 billionaire families across four continents, watching how consequential decisions are actually made, how family offices operate under pressure, where institutional knowledge disappears, and why some families build extraordinary capability while others simply build expensive organizations.

Over the last several years, I have applied that experience directly to artificial intelligence — not as commentary, but as architecture.

I am the author of 10 Things Every Family Office Should Be Doing Now in AI and The AI-Native Single Family Office, two living books that I continually revise because this field moves too quickly for static thinking. My work now spans frontier intelligence, open-weight models, sovereign hardware, agents, orchestration, institutional memory, reusable Skills, custom applications, AI governance, and the emerging digital workforce around the principal.

And multiple times each year, I bring that work into the AI Family Office Masterclass, where exceptional faculty, family principals, executives, and next-generation members move from understanding the technology to actually building with it.

At the August New York Masterclass, that broader process produced 27 working Skills alongside the other architecture and learning developed around the two days. November 4–5 begins from a frontier that has already moved materially since then.

The objective has never been to make family offices more impressed by AI.

It is to make them more capable because of it.

That is the work.

---

THE URGENCY IS REAL.

Don’t fight technology. In the end, it always wins. 

Every major technological shift creates the same temptation: wait until the tools are better, the standards are clearer, the risk is lower, and someone else has already figured out what works.

The problem is that by then, the advantage has moved from access to experience.

AI will be better a year from now. The agents will be better. The hardware will be cheaper. The interfaces will be easier.

That is not an argument for waiting.

It is an argument for beginning now, because the families that start building today will meet those better tools with accumulated judgment, working systems, institutional memory, trained people, and a year of practical fluency already behind them.

Technology gets easier. Catching up does not.

That is the urgency.

November 4–5 · New York City

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

JOIN SFO CONTINUITY

The Masterclass Is Two Days. The Membership Is the Year. 

November 4–5 is not a standalone event you purchase a ticket to attend. Your seat is part of SFO Continuity, my private 12-month membership for families building the next generation of family offices.

That distinction matters because nobody builds an AI-native family office in forty-eight hours.

The Masterclass creates acceleration. SFO Continuity provides the environment around it — the people, reference points, intellectual architecture, recurring rooms, and continued access that keep the work moving after everyone leaves New York.

Over the course of your membership, that includes four AI Family Office Masterclasses as the frontier continues to move. Each room begins from what is possible then, not from a curriculum written the year before. And remember the rule: membership gives you access to attend; what is built inside a particular Masterclass stays with the members who were actually there participating in that room.

But SFO Continuity is much broader than AI.

Throughout the year, members come together through private family-office forums, roundtables, dinners, soirées, and smaller working sessions around the issues consequential families are actually navigating — investing, macro and geopolitics, global structuring, jurisdiction, governance, succession, counterparty risk, family capability, and the accelerating impact of technology across all of them.

You also have direct access to me, a peer network of principals and family-office executives facing comparable complexity, the Continuity Index, my family-office frameworks and books, the two continuously evolving AI volumes, and the larger body of proprietary work built around creating a genuinely high-performing family office.

The advantage is not any single benefit.

It is the combination.

The Masterclass may show you what suddenly became possible. A conversation with another member may prevent you from spending two years learning an expensive lesson yourself. A forum may change how you think about a jurisdiction or counterparty. A new Skill may turn exceptional human judgment into a resource your office can use repeatedly. A private conversation may reveal that the problem you thought you had is not the problem at all.

That is why I do not think of SFO Continuity as a content subscription or an events membership.

It is the operating environment around the family office you are trying to build.

And because the membership includes the family and the executives serving it, the value does not have to stop with one principal. Bring the people who should become more capable with you — the executive running the office, the son or daughter who will inherit responsibility, the family member who should understand what is being built before they are eventually asked to lead it.

$10,000 · 12 Months

No tiers. No sponsors buying access to the room. No providers waiting for their opportunity to sell to the family.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

---

NOVEMBER 4–5. NEW YORK CITY. THIS IS WHERE YOU MOVE.

Two days will not turn you into an AI engineer. That is not the objective. 

What two days can do is change your reference point permanently.

You begin to understand the architecture well enough to direct it: frontier models for the hardest reasoning, open weights for work you want to own, the Sovereign Vault, agents and harnesses, Skills, institutional memory, MCP connecting previously isolated systems, and the governance required when machines begin doing rather than merely answering.

More importantly, you begin using it.

You should leave New York having touched the systems yourself, begun your own files, built against real family-office problems, and understood enough of the architecture that the following Monday is no longer another round of evaluating AI. It is the beginning of implementation.

And the amount of intelligence now available to one capable human is difficult to overstate.

With Fable 5.1, GPT-6 Astra, and the frontier systems that follow them, we are approaching something that, operationally, can feel like having multiple extraordinary minds available on demand. When difficult work is divided among agents and subagents — one researching, another modeling, another attacking the thesis, another searching history, another checking the others — the principal or executive can direct an intellectual bench that works across all 168 hours of the week.

That does not diminish human judgment.

It makes human judgment vastly more leveraged.

For a family-office executive, I struggle to think of a capability more valuable to build right now.

You may be outstanding at investments, operations, tax, governance, or relationships. Keep all of it. But becoming genuinely fluent at directing AI multiplies the value of nearly every capability you already possess. It allows one person to research further, prepare more deeply, build faster, challenge assumptions harder, and operate with a level of leverage that would previously have required an organization around them.

And because the family-office benchmark is still surprisingly low, the distance from novice to unusually capable is much shorter than most people imagine.

You do not need ten years.

You need to start building.

The same is true for the principal. The advantage is not becoming technical for its own sake. It is learning to command extraordinary intelligence without surrendering judgment, privacy, or control — and understanding enough of the architecture that nobody else gets to make those decisions for you.

That is why the Masterclass is hands-on.

The books can teach you what a Skill is. They can explain the Second Brain, sovereign infrastructure, frontier versus local intelligence, agent orchestration, MCP, red teams, and the architecture underneath all of it.

The room is where those ideas become yours.

And November 4–5 is not the finish line. It is the beginning of the twelve months around it.

You leave as an SFO Continuity member, with the faculty, the peer network, the books, the frameworks, the future Masterclasses, and the people traveling down the same road beside you. The technology will keep changing. The membership keeps you inside the environment where those changes are interpreted, tested, and turned into family-office capability.

Somewhere in the next five years, there is a version of you who moved and a version who did not.

By then, both versions will have access to better technology.

Only one will have five years of experience building with it.

That is the asymmetry.

NOVEMBER 4–5 · NEW YORK CITY

SFO Continuity membership includes your seat.

Request a Private Membership Briefing →

Ready for November 4–5? Join SFO Continuity →

Sovereign or Obsolete.

---

Sovereign or Obsolete. The AI Family Office Masterclass | November 4-5, 2026 | New York City The seat is available. The membership is open. The only question is whether your family is in the room.