Original Title:AI in 2026: 3 Trends
Original Author: a16z crypto
Translated by Ken, Chaincatcher
1. This year, AI will take on more substantial research work.
As a mathematical economist, in January 2025, I even had great difficulty getting consumer-grade AI models to understand my workflow; but by November, I was already able to give the models abstract instructions in the way I would guide a Ph.D. student—and sometimes they could provide novel and correct answers. Beyond my personal experience, AI is being increasingly applied in research fields—especially at the reasoning level, where models are directly assisting in the discovery of innovations, and can even solve difficult problems from the Putnam Mathematical Competition (one of the most challenging university-level math exams in the world).
It is currently unclear in which fields such research assistance is most effective, or exactly how it exerts its effects. However, I anticipate that AI research this year will foster and reward a new type of "polymath-style" research approach—one that values the ability to speculate on connections between concepts and to rapidly draw conclusions from answers that are not yet certain.
These answers may not be entirely accurate, but under a certain topological structure, they can point in the right direction. Ironically, this is somewhat akin to harnessing the power of "model hallucination": when models are intelligent enough, allowing them abstract space for exploration may sometimes result in nonsense, but it can also tear open a window of discovery—just as humans are often most creative in nonlinear, ambiguous states without clear instructions.
This mode of reasoning requires a completely new AI workflow—not just interactions between agents, but a pattern of "agents encapsulating agents," in which multi-layered models can help researchers evaluate the outputs of earlier models, filtering them layer by layer to separate truth from falsehood. I have been using this method to write papers, while others use it for patent searches, inventing new art forms, or (unfortunately) uncovering vulnerabilities in new types of smart contracts.
However, running such a "packaged reasoning agent cluster" for research requires better interoperability between models, as well as a mechanism to identify and fairly compensate each model's contribution—cryptocurrency could help address both of these issues.
2. Moving from "Know Your Customer" to "Know Your Agent"
The bottleneck of the agent economy is shifting from "intelligence" to "identity." In the financial services sector, the number of "non-human identities" is already 96 times that of human employees—yet these identities remain ghost-like entities without bank recognition or coverage.
The key fundamental element missing here is KYA: Know Your Agent.
Just as humans need credit scores to obtain loans, agents also require cryptographically signed credentials to conduct transactions—binding the agent to its principal, constraints, and accountability. Until such a credential system is established, merchants will continue to block agents behind firewalls.
The industry that took decades to build its KYC infrastructure now has only a few months to get KYA sorted out.
3. We must address the "hidden taxes" on the open internet.
The rise of AI agents is imposing an invisible tax on the open web, fundamentally undermining its economic foundation. This erosion stems from an increasingly pronounced misalignment between the internet's "content layer" and "execution layer": current AI agents extract data from websites that rely on advertising for survival (the content layer), providing convenience to users while systematically bypassing the revenue sources that support content creation (such as advertising and subscriptions).
To prevent open networks from being eroded (and to protect the diverse content that fuels AI development), we need to deploy large-scale technical and economic solutions. These could include next-generation sponsored content, micro-attribution systems, or other innovative funding models. Existing AI licensing agreements have proven to be unsustainable stopgap measures, typically compensating content providers for only a small fraction of the revenue lost due to AI-driven traffic erosion.
The network requires a completely new "techno-economic" model that enables value to flow automatically. A key transition in the coming year will be shifting from static licensing to real-time, usage-based compensation mechanisms. This means we need to test and scale new systems—potentially leveraging blockchain-enabled micro-payments and sophisticated attribution standards—to automatically reward every entity that contributes information toward an agent's successful task execution.
