From Abstract Intelligence to Real World Autonomy

We recently came across a post by Andrej Karpathy — a co-founder of OpenAI and the founder of Eureka Labs, who joined Anthropic’s pretraining team in May 2026.

Karpathy describes an interesting paradox: AI capabilities are growing rapidly, but people’s understanding of those capabilities is becoming increasingly uneven.

By his rough estimate:

• 6 billion people have barely interacted with an LLM.
• 1–2 billion use AI as an improved search engine or an everyday assistant.
• 20 million professionals are already watching AI agents complete in hours and days what used to take weeks and months.
• About 5 thousand people have inside access to frontier systems, where thousands of agents can jointly tackle the hardest engineering, scientific, and research problems.

It’s a striking situation: while most people debate how well ChatGPT writes emails, a small group of specialists is already witnessing an entirely different technological reality.

We largely agree with Karpathy here. But in our view, the most interesting part begins beyond his observations.

1. The AI revolution will unfold unevenly

Karpathy uses the term jagged intelligence — the unevenness of AI capabilities.

Modern models are remarkably strong at tasks where results can be verified quickly: programming, mathematics, data analysis, some areas of scientific research.

Yet there are domains where a correct answer cannot be produced or verified instantly: controlling physical objects, complex production processes, safety, strategic decisions, and interaction with unpredictable environments.

So autonomy will not spread as a single wave, but as separate islands.

And it’s at the borders of these islands, we believe, where the most interesting technological and commercial opportunities will emerge.

2. Code is ceasing to be the main competitive advantage

If AI can reproduce an application, an architecture, or an interface in a matter of hours, traditional technical moats get considerably weaker.

But it’s important to distinguish a digital product from a real system.

Creating a software clone of Uber is not the same as building Uber. Generating an industrial robot control system is not the same as keeping it running safely on a factory floor.

Physical infrastructure, regulatory approvals, operational reliability, real-world data, and complex operational processes cannot be copied with a single prompt.

Competitive advantage is gradually shifting from the ability to write code to systems design and the ability to reliably perform complex work in the real world.

3. The economics of software itself is changing

For decades, the industry has been selling tools used by people.

SaaS companies monetized seats, licenses, and subscriptions.

But if a software system performs the work of an analyst, a developer, or an operator on its own, a natural question arises: why pay for the number of users when you can pay directly for the result?

We expect outcome-based models to keep expanding: payment per completed operation, processed transaction, achieved savings, or created value.

At the same time, more and more agent-native companies will appear — organizations where AI doesn’t assist existing employees but originally performs a significant share of the operational work itself.

This is no longer just automation. It’s a redesign of the economic structure of business.

4. The next infrastructure layer — managed autonomy

Today the industry is mostly focused on how intelligent models can become.

But as multi-agent systems mature, a different question becomes ever more important:

How do you ensure predictability, coordination, verifiability, and safety for thousands of autonomously acting agents?

Asking an agent to write a program is one thing.

Letting a system autonomously run a production line, warehouse infrastructure, a robotic complex, or energy equipment is quite another.

In such scenarios, an intelligent answer is not enough. You need managed execution, constraint enforcement, reproducible results, audit trails, and safe recovery mechanisms.

We believe this infrastructure layer will become one of the most important directions in AI development.

5. The physical world may prove to be the most durable source of competitive advantage

There’s an interesting economic asymmetry.

The cheaper digital solutions become to create, the more relative value accrues to assets that cannot be instantly reproduced: factories, power grids, logistics infrastructure, robotic production lines, and licensable systems.

And AI can radically reduce the cost of operating them.

A potentially powerful combination emerges: capital-intensive physical infrastructure + intelligent autonomy with low operating costs.

This is where companies with fundamentally new economics may appear.

6. Why this matters to us

We look at this transformation from the perspective of autonomy execution infrastructure.

Our work on Holo Engine is about building a compute architecture that can support coordination, state management, and the execution of complex processes — including at the boundary between the digital and physical worlds.

We believe the next stage of AI will require not just more intelligent models, but fundamentally more efficient, controllable, and verifiable mechanisms for how they interact with the environment.

It’s also important to understand that deterministic execution alone does not guarantee the safety of the whole system. Real autonomy requires additional layers of verification, control, and handling of uncertainty.

That’s why, for us, this is not just a question of AI performance. It’s a question of the architecture of the next generation of autonomous systems.

Our main takeaway:

Today the industry competes over who builds the most intelligent AI.

But tomorrow, another race will matter just as much — who can turn that intelligence into reliable, scalable, and economically efficient autonomy.

We expect a significant share of long-term value to be created not only at the model level, but at the level of systems that let intelligence act safely in the real world.

Because building intelligence and learning to trust it are two completely different engineering problems. And the second one may well be the most interesting challenge of the coming decade.