The asymmetry is getting uncomfortable.
Andrew Lo at MIT mapped out a century of investment innovation. Every major paradigm shift—Graham and Buffett’s fundamental analysis in the 1920s, Markowitz’s modern portfolio theory in the 1950s, Simons and D. E. Shaw’s quant revolution in the 1980s—had one thing in common: the core resource of investing changed. Experience gave way to math. Math gave way to compute. Now compute is giving way to something else entirely.
Here’s what’s happening, and why most of the industry is still looking the wrong way.
The three models that dominate today
Discretionary investing relies on human judgment. Buffett and Lynch built legends this way. But one person can only cover so many companies, and cognitive biases—FOMO, ego, anchoring—are features of the hardware, not bugs to patch. Even the best fund managers hit a cognitive ceiling.
Quant investing solved the discipline problem. Systematic, data-driven, backtestable. It can cover hundreds of stocks simultaneously. But alpha decays as strategy capacity fills up, and the signals that actually matter often don’t look like clean structured factors. A pause in an earnings call. A subtle shift in regulatory language. A cross-industry supply chain connection. Traditional quant doesn’t miss these entirely—it just processes them at high cost and low fidelity.
Index investing is efficient market plumbing. Low cost, transparent, diversified. It doesn’t judge direction—it just captures the result. Like a reliable vehicle that doesn’t decide which road to take.
All three share one assumption: the core cognitive unit in investment decisions is a human. Even in quant investing, the strategy design and factor selection are still done by people. That assumption is loosening.
Three signs the shift is real
This isn’t gradual improvement. It’s a structural change underneath the entire industry.
First, pre-trained models are replacing parts of expert experience. Traditional quant relies on feature engineering: researchers manually build factors based on market intuition, then combine them into signals. Factor quality and quantity remain bottlenecked by the researcher’s experience and imagination.
A pre-trained model trained on massive financial text corpora has already internalized enormous implicit knowledge: the relationship between industry cycles and stock prices, the correlation between management tone shifts and earnings warnings, and the mapping between regulatory wording changes and sector trends. These patterns don’t need to be explicitly coded as factors. The model remembers them as distributed representations.
Tacit knowledge that once lived only in the heads of senior researchers can now be scaled and reused. A model can accumulate experience across dozens of industries and hundreds of sub-domains simultaneously. No human team can do that.
AlphaEvolve-style iteration creates a new research paradigm
Pre-training has a natural limitation: training data is always historical. But financial markets keep changing. What worked yesterday may not work tomorrow, and the definition of a good company in one cycle doesn’t necessarily hold in the next.
The next critical capability isn’t just learning from history—it’s continuously improving analytical methods in real tasks. The insight behind AlphaEvolve-style approaches is not to learn everything at once, but to build a closed loop: propose candidate approaches, evaluate them, retain what works, and evolve from there.
Google DeepMind describes AlphaEvolve as an agent system that evolves codebases and improves algorithm design. The parallel to investment research is direct.
Multi-agent systems complete the execution layer
If the first two signals are brain upgrades, this one is the hands. A system that could truly replace an investment analyst can’t just understand information and make judgments—it needs execution capability.
That capability spans real-time search and retrieval across news, filings, patents, and supply-chain data; reasoning and synthesis that turn fragmented information into investment theses; tool use for valuation, backtesting, and risk calculations; and decision execution through trade orders, risk parameters, and dynamic position sizing.
This isn’t automation. Automation does repetitive things for you. Agentic systems do things that require judgment.
When multiple agents—research, trading, and risk—work together through structured protocols, the human role shifts from executing every step to setting objectives, assigning authority, monitoring the system, and handling exceptions. We call this Agentic Investing.
What this means for the industry
If asset management is shifting from brainpower-intensive to compute-intensive, it changes the entire competitive logic.
For investors, the core competency shifts from individual cognition and industry understanding toward building stronger systems: better data pipelines, more market-adaptive models, and more robust agent architectures. It’s no longer only about who knows more. It’s about who can systematize knowledge better.
For institutions, the organizational model may need rewriting. The traditional portfolio-manager-plus-analyst-team structure organizes decisions around humans. As more research, analysis, and execution moves to systems, the human role shifts toward objective-setting, governance, and exception handling.
For the competitive landscape, this may become a value-chain restructuring. Resources that historically concentrated around the smartest individuals may increasingly flow toward teams with stronger compute infrastructure, more complete technology stacks, and more effective human-machine collaboration.
The pattern has happened before. Spreadsheets began as a faster way to do accounting, then changed how companies budgeted, analyzed, and managed. AI entering asset management may follow the same trajectory: starting as a research-efficiency tool, then rewiring how investment research is produced and how humans and machines divide decisions.
GIM’s bet
We don’t think this paradigm shift happens overnight. Financial markets have their own complexity—regulation, trust, inertia, and irrationality—and these factors will keep traditional models effective for a long time. But the direction is clear.
At GIM, we’re building the investment system as an AI-native system from day one. Our goal isn’t to pick sides among the three traditional models. It’s to build a new kind of capability: model like quant, pick stocks like fundamental, and scale like index.
The technical stack requires three layers: one foundation model purpose-built for investment research and financial reasoning; multiple agents that divide labor, collaborate, validate, and continuously evolve; and one infrastructure layer connecting data, toolchains, risk constraints, and execution feedback into a real closed loop.
All three are required. With any layer missing, you’re plugging AI into old processes. Only when all three work together can a true AI-native investing system emerge.
Finance has never been an industry where technological optimism can simply bulldoze through. Regulation, trust, risk constraints, and real trading environments will make this path slower than many expect.
Even so, we’re increasingly convinced: the next generation of investment firms won’t just use AI. They’ll rebuild around models, agents, and infrastructure.
It’s early. Far from mature. But we believe this will be one of the most important bets in asset management over the next decade.