Kevin Kelly: No expert can see through the future of AI (full speech)

Today, we’ll focus on future trends, the upcoming industrial shifts, and brand-new possibilities.

First, I want to emphasize this: at the moment, there is no AI expert who can truly see through the future of AI.

We still can’t be sure whether Artificial General Intelligence (AGI) will be truly achievable; we don’t know whether AI development will become highly centralized, or split into countless small, decentralized intelligent units; we don’t know whether AI will replace human jobs at large scale; and we can’t judge whether the future of AI will be dominated by open-source models or closed-source models.

We have countless guesses and scenario-building exercises, but no one can provide definitive answers. This is the front-line uncertainty of the AI industry: development speeds vary widely across different areas—some tracks iterate rapidly, while others grow slowly and deeply.

Moreover, this kind of uncertainty will not disappear in the short term. Within one year, two years, or even five years, these core questions still can’t be fully resolved. After five years, we’ll still face these unresolved industry challenges. So learning to coexist with uncertainty, and to adapt to this unknown, is a mandatory lesson for all of us.

But at the same time, uncertainty itself is the biggest opportunity. All the trends I’m about to share—though full of unknowns—also open a door for everyone here, giving you a chance to create something unprecedentedly new.

In the next five years or so, opportunities and uncertainties in the AI field are mainly concentrated in three cutting-edge frontiers: humanoid robotics, emotional intelligence, and the smart agent ecosystem.

Let’s start with humanoid robots. I firmly believe that humanoid robots will be the most complex creations humans have ever built—on today’s Earth, besides human beings themselves, they are the products with the highest difficulty in structure, logic, and technology.

Their R&D difficulty is extremely high, requiring thousands of top talents to come together to tackle it. Because humanoid robots aren’t just built on the most complex AI systems available today—they also fuse three core components: precise hardware, core energy supply, and supporting software algorithms. Each of these areas is highly challenging on its own; integrating and compressing them into small humanoid devices makes the difficulty rise exponentially.

In other words, developing a humanoid robot that can walk into homes, enter classrooms, and serve humans routinely will be the hardest engineering task in human technology history—and we are only just getting started right now.

At the current stage, large language models (LLMs) only replicate a very small portion of the many cognitive abilities of the human brain, and they aren’t even the core way humans think. The AI we’re building today is merely a single, one-sided simulated form within humanity’s complex system of thought.

Existing large language models have two major core weaknesses: they can’t learn autonomously, and their long-term memory is weak. But autonomous iteration, continuous learning, and dynamic adaptation to environments are the essential capabilities for intelligent agents to accompany humans and keep evolving. This also means we can’t confine ourselves to large language models—we must explore entirely new forms of AI cognition.

Large language models are indeed powerful and irreplaceable. They can be deployed across a massive range of scenarios. But their training foundation is massive text and textual knowledge—they understand the world in books, but they don’t understand the real physical world.

Future AI needs to break away from “talking about things on paper,” and truly have real-world cognitive capability. It must be able to sense three-dimensional space, recognize object orientations, and understand the three-dimensional physical world—having real scene-judgment ability. This is what the many “world models” being tackled by labs around the globe aim to achieve.

In the future, AI won’t be only language models. It will also have physical models, chemical models, and biological models. It won’t rely on textual cognition to understand the world; instead, it will perceive reality through real physical laws and natural rules, understanding natural phenomena like object collisions and liquid flow.

This type of technology is also called spatial intelligence, embedded AI, or world models. The core idea is to take abstract knowledge and cognition and ground them into the three-dimensional physical world. This is the most essential cutting-edge track right now, and it contains a vast number of innovation opportunities.

And the humanoid robot’s core “must-have” requirement is precisely this kind of spatial intelligence. Language ability is important, but far from enough. Robots need to go beyond textual cognition and truly read the physical world. Breakthroughs in the future robotics field will inevitably rely on the miniaturization and high-precision of this spatial intelligence—meaning it must be integrated into smartphones and small robot devices. This is a huge technical challenge, and also an excellent opportunity for entrepreneurship and innovation.

Besides that, robotic hands are another major technical barrier. Human hands are exquisitely precise in structure—they can sense pressure and identify temperature. Replicating a mechanical structure comparable to a human hand is extremely difficult. We don’t just need to develop simple mechanical grippers—we also need to build bio-inspired robotic hands with multi-dimensional sensing. That requires countless engineers and AI researchers to keep pushing through; it’s a major engineering challenge for the future, and it will also create a huge number of jobs and research opportunities.

Finally, there’s the issue of robot energy. Although related technologies have made many breakthroughs, compared to human biological energy efficiency, the gap is still enormous. As a biological individual, a human brain’s supercomputing power requires only 25 watts, and the highest power consumption for the entire body is only 300 watts, yet we can work continuously and stay active for 12 hours. For humanoid robots at this stage, their energy efficiency can’t even reach half of that of humans. Improving energy efficiency is an urgent core engineering problem.

Even more stringent: to integrate into human life—caring for family, deeply interacting with people—humanoid robots must achieve 99.999% ultra-high reliability. But currently, robots are nowhere near this standard.

This is the industry’s “nine-step precision improvement rule”: from 99% to 99.9%, then to 99.99% and 99.999%—for every additional “9” of precision, the amount of R&D work and technical difficulty required is equal to the total across all previous stages. The higher you go in precision, the harder it is to keep improving. Even the most top-tier global robotic factories have only 91% automation coverage—showing that we still have a very long road ahead.

Of course, non-humanoid robots have already been deployed at scale. For example, precision agriculture robots can precisely monitor the growth status of each lettuce plant—enabling accurate water control and measured fertilizer application for individual crops. This is a precision agriculture model that traditional farmers have been unable to achieve. With agricultural robots, this approach has already become a reality. Beyond that, specialized robots like smart milking robots are also continuously iterating and optimizing with AI-enabled empowerment.

San Francisco’s self-driving cars are a benchmark product for robotics technology, but even then, their maturity reaches only 99.9%. The remaining 0.001% of extreme scenarios still require remote human takeover and backup guarantees. To completely eliminate this last trace of human intervention would require time, funding, and technical costs equivalent to the entire accumulated effort of the self-driving industry over the past four decades.

So I always emphasize: humanoid robots are the most complex technological creations in human history. Conservatively, it will still take about ten years for them to truly mature and be deployed in the real world. This isn’t just my personal judgment—market prediction data also confirms it. The difficulty of R&D and deployment is far beyond what most people imagine.

After talking about robots, let’s move to the second frontier track: emotional intelligence. This will be the biggest surprise in future AI development—to give AI emotional capabilities, and making that emotional layer the core next step in AI iteration.

The reason we want AI to have emotions is to match human interaction habits. Humans are naturally able to perceive and understand emotions without needing extra training or adaptation. Now that AI has visual cameras, it can already precisely recognize human emotions like happiness, surprise, and fear, and then provide corresponding responses.

You can imagine: an AI smart toy for a child that can sense the child’s low mood, proactively accompany them and listen; in the future, human pet cats and dogs can truly speak up and communicate with their owners. These kinds of scenarios already have conditions for real deployment today.

I firmly believe that AI and robots with emotions will form real, deep emotional connections with humans. Even if they are products of artificial intelligence, the bonds and emotional relationships between us and them will be incredibly real.

The last core track is the smart agent ecosystem—also the main direction that the industry is heavily discussing right now.

In the future, everyone will have a dedicated personal smart agent, while behind the scenes many invisible secondary smart agents will coordinate and work together. The biggest feature of the core smart agent you use up close is that it’s permanently online and can respond at any time—it can be integrated into devices like smart glasses and handheld terminals, serving users around the clock.

With long-term companionship, these smart agents will keep learning your habits, and understanding your needs; ultimately, they will know you even better than you know yourself. I define this as an “external self” (an external intelligence agent): it isn’t another independent person, and it’s not fully you either; it’s a dedicated intelligent individual that fits you, adapts to you, and extends you.

In the future, a complete AI smart agent economic ecosystem will form: countless smart agents will autonomously connect, assign tasks, coordinate work, and carry out all kinds of affairs. More importantly, this ecosystem will generate a dedicated transaction system—smart agents will autonomously settle, issue credit, and conduct trades with each other.

This also finally gives cryptocurrencies and stablecoins real-world deployment scenarios. They won’t be just speculative tools anymore—they can become dedicated circulating currencies for the AI smart agent ecosystem. Multiple projects like Strike have already started testing this model in live deployments.

But after smart agents become widespread, a series of entirely new problems will also arise: who owns smart agents in the end? Who will they ultimately serve—development companies, the users who use them, or some third-party entity?

Among these, the most core and most opportunity-filled proposition is building a smart trust framework: how can smart agents that are strangers and haven’t undergone security verification achieve safe, trustworthy interaction and collaboration with each other? This need will give rise to a whole new set of trust technology systems, and it will be the core foundation for the future real-world deployment of the smart agent ecosystem.

To sum up, future uncertainty in the AI industry and its core opportunities all focus on three areas: technical breakthroughs in humanoid robots, the deployment of AI emotional intelligence, and building an all-domain smart agent ecosystem.

Finally, I want to emphasize this: looking back from a ten-year perspective, we’ll find that right now, there’s no one who can truly be called an AI expert. And that also means that entering the field now is never too late. I’m looking forward immensely to seeing everyone here create results that will overturn the era in this brand-new track.