General Financial Intelligence

One of the most consequential decisions that shape most people’s lives, how to manage their money, still has no real intelligence behind it. People routinely ask general-purpose AI for financial advice. How should I position my portfolio given rising inflation? Should I maximize my 401(k) or pay down my mortgage? What’s the tax treatment on RSUs? Can I afford this home?

These models provide articulate, broadly accurate, and generic responses, but their utility stops there. The models can talk, but they do not carry the burden of the problem. There is no follow-through, no continuity, no comprehension of your specific circumstances. They disappear the moment the tab closes.

Solving this doesn’t require an AI that excels at writing poetry, coding, or generating images. The push toward artificial general intelligence (AGI) focuses on systems proficient across diverse, general domains. Financial management is not a general task. It is a profoundly specialized field with its own rules and physics, demanding a specialized form of intelligence.

The need for specialization isn’t obvious. Finance looks like a generalist’s domain. Any frontier model can explain the efficient market hypothesis, explain derivatives, or outline the tax advantages of a Roth IRA. The textbook view is neat: rational participants, equilibrium pricing, risk-return trade-offs governed by mathematical laws.

The reality is dramatically different. Markets are reflexive systems driven by irrational actors on incomplete information, susceptible to herd dynamics and motivated by misaligned incentives. Regimes shift without warning. Correlations that held for decades reverse overnight. An asset classified as “safe” can become lethal in a single quarter. The chasm between financial theory and practice is where every generalist AI approach fails.

What this domain requires is a generally intelligent agent purpose-built for these complexities: one that reasons under deep uncertainty, plans across long time horizons, and acts on your behalf continuously. Not a chatbot. Not a robo-advisor. Not a human advisor with a copilot.

I call this General Financial Intelligence.

The hard problem

Your income, your debt, your tax situation, your equity comp, your goals, your risk capacity, your time horizons, the current market regime, the current tax code, and how all of these interact. Given all of that, what should you do?

This is a planning problem in a multi-dimensional, non-stationary environment. The state space is enormous. The objectives compete. The environment shifts while you’re optimizing. And the person at the center of it has preferences that are partially latent, inconsistent, and evolving. The correct comparison class isn’t fintech. It’s autonomous systems, robotics, self-driving, any domain where an intelligent system must maintain a model of a complex environment and operate under distributional shift.

The industry’s response to this complexity has been to ignore it: sort people into five risk buckets, hand them a static allocation, rebalance quarterly. And the industry’s AI response has been even less ambitious: use AI to help advisors search research reports, use LLMs to summarize documents, train a 50-billion-parameter model to sound financially literate. Incumbents are structuring AI around existing human roles (advisor, analyst, portfolio manager) rather than asking whether the intermediary should exist at all.

Nobody is building the intelligence itself.

What happens when rules meet reality

For decades, the canonical advice for conservative investors (retirees, people nearing retirement, anyone flagged as “low risk tolerance”) was to hold a large allocation to bonds. Not a suggestion. An axiom. As you age, shift to fixed income. Safety. Stability.

Then 2022 happened. The Fed hiked rates at the fastest pace in 40 years and the “safe” part of the portfolio got destroyed. The Bloomberg Aggregate Bond Index fell 13%. Long-duration Treasuries dropped over 30%. The 60/40 portfolio had its worst year since 1937. Millions of conservative investors, people who did exactly what they were told, watched the assets intended to protect them crater.

This wasn’t a black swan. Inflation was running hot, the Fed was telegraphing aggressive tightening, and bond math is bond math. Anyone reasoning about the macro environment would have seen that long-duration bonds were the highest-risk asset in that portfolio, not the lowest.

But the systems that manage people’s money don’t reason. They don’t model how monetary policy transmits into asset prices. They don’t understand that risk is conditional on the regime, not a fixed property. They apply templates to risk questionnaires.

The entire system runs on brittle heuristics. If you’re 60, hold 60% bonds, regardless of the rate environment, your tax situation, your other income sources, or what’s actually happening in the world. The machinery can’t adapt to you or to reality. It classifies, allocates, and hopes nothing changes. That’s what the 1-2% advisory fee pays for.

Toward world models

The relevant question in markets is almost never “what’s going to happen”? It is: “what does the market believe is going to happen, and where is it wrong?” Summarizing financial news is journalism. Comparing implied narratives to what’s actually priced in, identifying divergences, reasoning about why they exist, and constructing probabilistic views about how they resolve. That is the actual problem.

This requires a world model, a causal representation of how the global economy works. Not a statistical model that finds correlations in historical data, but a system that understands mechanism: when semiconductor export controls tighten, what happens to chip production, capex cycles, AI infrastructure buildout, and the revenue exposure of every company in your portfolio? When a central bank signals a pivot, what are the second-order effects on mortgage activity, housing supply, and bank earnings? When geopolitical tensions disrupt crude supply, how does that transmit through energy prices into consumer spending into inflation expectations?

You don’t need a model that sounds financially literate. You need one that thinks. And you can’t engineer a system this complex. The causal graph of the global economy is too dense, too dynamic, too context-dependent. It must be learned. The richer the model, the better the downstream reasoning. This is where scale matters, and where the problem becomes a genuine research challenge, not an engineering exercise.

The bitter lesson, applied

Rich Sutton’s famous essay describes a pattern across 70 years of AI: approaches that leverage human-engineered knowledge work for a while, then general methods that leverage computation and scale blow past them. Every time. Chess, Go, vision, language, protein structure.

The entire financial planning industry runs on exactly these kinds of hand-engineered approximations.

Replace heuristics with search over the full action space, conditioned on a person’s actual situation, the current environment, and a world model with real causal structure, and you get a phase shift. The system finds strategies no template would surface, because the space of possible moves across tax timing, account selection, asset location, contribution sequencing, withdrawal ordering, estate planning, and their interactions is vast. Heuristic reasoning explores almost none of it.

This is what happened when AlphaGo played moves no human had seen in 3,000 years of Go. Not better heuristics. Search over a space that heuristic play had never explored.

Architecture

There is a lazy version of this future where incumbents keep the same architecture and sprinkle AI on top. A brokerage adds an assistant. A robo-advisor adds conversational onboarding. An RIA hands its advisors better tools and rebrands the stack as intelligent. But once you accept that the core problem is persistent state, world modeling, inference, and action, the intelligence cannot sit at the edges. It has to sit at the center: informing what gets held, what gets sold, what gets deferred, what gets harvested, what gets surfaced, and when the system should do nothing. A decorative AI layer is a facelift. An AI-native system changes what the software is.

In an agentic system, the AI is the system. The agent architecture is the research function. The reasoning engine is the portfolio manager. The world model informs every decision at every level, continuously. There is no expensive and high latency intermediary between the intelligence and you.

Build from first principles and you arrive at an agentic harness that orchestrates domain-specialized agents across rates, equities, commodities, and geopolitical risk. Each maintains its own model of the world. Each is equipped with the tools and data access to do real work, not summarize someone else’s. These agents don’t just describe what’s happening. They detect where market pricing diverges from what the evidence supports. The output isn’t a report. It’s a view: a probabilistic claim about a mispricing, with a causal argument and a scenario distribution.

The harness manages what each agent sees, how questions get decomposed into parallel subtasks, when to call tools or sub agents and when to reason internally, and how results get synthesized across domains. It maintains memory, tracks calibration, and learns which of its own components to trust. Agents challenge each other. Disagreements are resolved with explicit uncertainty, not consensus. Confidence is earned through prediction, not assertion. It learns.

General Financial Intelligence

Financial services is usually discussed as a market-size story: large, fee-heavy, fragmented, a legacy industry ripe for disruption. All true, but not the most interesting part.

It is one of the most consequential domains where real intelligence is still scarce. People make decisions here that determine where they live, when they retire, how they absorb shocks, whether they can start companies, what happens to their family if something goes wrong, whether a decade of work compounds or gets quietly taxed, diluted, misallocated, or stranded in the wrong wrapper. The decisions are nonlinear. The interactions are ugly. The constraints are real. And most of the systems touching those decisions are either too shallow to reason or too human to scale.

General Financial Intelligence, built properly, is not better investing software. It is a new layer in the stack: a system that carries the state of a financial life, forms and updates views about the world, searches over a large decision space, and converts all of that into continuous, personalized action.

That is the category I think should exist. And it is the one I am building at ATG.