The Metaverse Didn’t Die. It Became Intelligent.

Five years ago, I defined the metaverse as “Digital Life + Physical World.” AI agents, digital humans, and synthetic media are making that convergence more likely—and much more complicated.

Almost five years ago, after serving as CTO of a metaverse startup, I wrote an article reflecting on the industry and proposed a simple definition:

Metaverse = Digital Life + Physical World

At the time, “metaverse” was one of the most overused words in technology.

Facebook had renamed itself Meta. Companies were buying virtual land, launching NFTs, opening virtual stores, and organizing events inside digital worlds. Investors and consultants were predicting that we would soon spend a substantial part of our lives inside immersive 3D environments.

It was difficult to tell where the long-term idea ended and the short-term marketing began.

Since then, much of the excitement has disappeared. The crypto market went through another cycle. Many virtual-land projects lost momentum. Companies reduced or redirected their metaverse investments. Artificial intelligence replaced the metaverse as the dominant technology story.

I also moved on and have not worked directly on metaverse projects for several years.

But after watching what has happened with generative AI, AI agents, digital humans, smart glasses, deepfakes, and synthetic media, I have started thinking about the metaverse again.

My conclusion is slightly unexpected:

The word “metaverse” may have lost its momentum, but the underlying idea is more likely today than it was five years ago.

The next version, however, may look very different from what people imagined in 2021.

It may not begin with virtual real estate, cartoon avatars, or a headset.

It may begin with everyone having a digital agent.

What the first metaverse wave got wrong

The first metaverse wave often treated the metaverse as a place.

It was described as a persistent 3D environment that we would enter through a VR headset. Inside it, we would work, shop, socialize, attend concerts, play games, own digital property, and represent ourselves through avatars.

This was not necessarily wrong. Gaming and virtual worlds had already demonstrated that people could form relationships, build communities, and create real economic value inside digital environments.

But it was too narrow.

It assumed that the metaverse had to be somewhere we went. It also implied that the physical world would gradually be replaced by a digital alternative.

I have never found that version completely convincing.

I do not believe people generally want to replace physical life. A digital bottle of water cannot satisfy physical thirst. A virtual meal cannot provide nutrition. A digital home cannot protect a physical body from heat, cold, or rain.

At the same time, digital resources clearly have real value. Cloud storage is not a physical hard drive sitting in front of us, but we still pay for it. A digital reputation can affect a physical career. A navigation app can determine which physical road we take. A dating profile can lead to a real relationship. A digital payment can result in a physical product arriving at our door.

Digital life and physical life are already connected.

That was the reason for my original definition:

Metaverse = Digital Life + Physical World

The important part was never the replacement of the physical world. It was the merging of digital and physical life.

The question was not how much time we would spend inside a virtual world.

The better question was:

How much of our physical life would be understood, influenced, assisted, and represented by digital systems?

Seen this way, the metaverse did not fail.

One particular commercial interpretation of it failed to become universal.

AI supplies the missing intelligence

The metaverse envisioned five years ago had environments, avatars, digital assets, currencies, and social interactions.

What it did not have was much intelligence.

An avatar could represent how you looked—or how you wanted to look—but it did not really know you. It could not understand your intentions, learn your preferences, communicate on your behalf, or complete meaningful tasks while you were away.

It was a digital body without a digital mind.

Generative AI changes this.

We now have systems that can communicate in natural language, analyze documents, generate images and video, use software tools, browse websites, work across different sources of information, and remember aspects of a user’s preferences.

The industry is also moving from AI that only answers questions to AI that can take actions.

For example, AI agents can already use virtual computers to research information, navigate websites, create documents, work with spreadsheets, and carry out multi-step tasks. These systems remain imperfect and require supervision, but the direction is clear.

The progression might look something like this:

  1. Digital profile — stores information about you.
  2. Avatar — visually represents you.
  3. Assistant — helps when you ask.
  4. Agent — completes tasks under your direction.
  5. Digital representative — acts within defined limits based on your identity, preferences, and authority.

That fifth stage is much closer to the real metaverse than a virtual character walking through a digital shopping mall.

A digital representative could exist across your phone, computer, glasses, car, home, workplace, and online accounts. It could move between digital and physical contexts without requiring you to enter a separate virtual world.

The metaverse, therefore, may not become a destination.

It may become a persistent layer of identity and intelligence that follows us between environments.

From avatar to digital delegate

It is useful to distinguish between three concepts that are often mixed together.

An avatar represents how I appear.

A digital twin represents information about my condition or current state.

A digital agent represents what I want done.

When these are combined, we get something new: a digital delegate.

A digital delegate might know my schedule, communication style, recurring preferences, professional responsibilities, spending limits, and relationships. It might be able to speak using a synthetic version of my voice or appear through a realistic digital representation of my face.

Technologies such as Meta’s Codec Avatars point toward increasingly realistic remote presence. Synthetic voice and video systems can already produce representations that would have required specialized studios only a few years ago.

Combined with an AI agent, this creates the possibility of a digital entity that can look like me, sound like me, understand some of my intentions, and take authorized actions for me.

But it is important to be precise:

A digital representative is not the same thing as the person it represents.

Human beings are inconsistent. We change our minds. We make exceptions. We behave differently depending on our mood, relationships, health, environment, and information available at that moment.

An AI trained on my previous decisions might predict what I normally do. It may still fail to understand why today is different.

It may know that I usually decline meetings after 6 p.m. It may not know that this particular meeting is important because an old friend is involved.

It may know which flights I normally prefer. It may not understand that I am willing to accept a longer journey this time to travel with someone.

It may learn my patterns without fully understanding my reasons.

Will an AI make most of our decisions?

It is tempting to imagine an AI that makes decisions with 70, 80, or 90 percent of the accuracy of the person it represents.

I still think something in this direction is possible, but I would now frame it more carefully.

The percentage depends entirely on the kind of decision.

For repetitive, bounded, and low-risk decisions, a digital agent may eventually handle a large majority of the work:

  • arranging routine appointments;
  • filtering messages;
  • preparing standard replies;
  • renewing subscriptions;
  • comparing common purchases;
  • selecting travel options according to known preferences;
  • coordinating calendars;
  • organizing information;
  • handling basic customer-service requests;
  • monitoring recurring bills or deliveries.

But accuracy in familiar situations does not guarantee good judgment in unfamiliar ones.

Research into AI-agent performance shows that capability is improving but remains uneven. An agent may perform well on one multi-step task and fail unexpectedly on another. METR’s work on agent task-completion horizons illustrates why reliability needs to be measured against the complexity and duration of a task rather than described by one universal intelligence score.

For financial, medical, legal, emotional, or irreversible decisions, human involvement will remain much more important.

A better prediction is:

AI agents may handle a large majority of recurring, low-risk decisions while escalating unfamiliar, consequential, or irreversible decisions back to their human users.

This is less dramatic than saying an AI will become a perfect copy of us. It is also more useful.

The goal should not be an unrestricted duplicate of a person.

It should be a bounded, accountable, and revocable representative.

Every agent may need a personal constitution

A personal agent should not have to infer all its instructions from previous behavior.

People should be able to give their agents explicit rules—a kind of personal constitution.

For example:

  • Never spend more than a defined amount without approval.
  • Never disclose medical or financial information unless I confirm.
  • Never agree to legal terms for me.
  • Ask before making a commitment involving another person.
  • Prioritize time with family over optional meetings.
  • Do not imitate my voice without clearly identifying yourself as an AI.
  • Explain any decision that falls outside my normal preferences.
  • Require stronger confirmation for irreversible actions.
  • Keep work, health, financial, and personal information separate.
  • Allow me to inspect and delete what you remember.

This may become more important than trying to create a system that understands everything about us.

A person may also have several agents instead of one universal agent: one for work, another for travel, another for healthcare, and another for finances.

Each agent would receive only the information and authority required for its role.

This separation would reduce risk. It would also reflect how people behave in real life. We do not give every colleague, doctor, bank, shop, and friend access to the same information.

Why should we give one AI system access to everything?

Representation creates an identity problem

Once an agent can act for a person, identity becomes more complicated.

In the earlier internet, we often asked:

Is this information real?

In the next internet, we will also need to ask:

Is this really the person, an authorized agent of the person, or an unauthorized imitation?

A realistic face or voice will no longer be sufficient evidence.

Deepfakes are usually discussed as a problem involving false videos of politicians or celebrities. But the more common risks may be personal: fraudulent calls, fake instructions from a manager, fabricated family emergencies, unauthorized endorsements, fake meetings, and synthetic versions of ordinary people.

The same technology that creates a useful digital representative can create a malicious impersonator.

This is why the next metaverse needs more than visual realism. It needs a trust layer.

A legitimate digital representative should be able to prove:

  • which person or organization authorized it;
  • what it is allowed to do;
  • which information it can access;
  • when its authority began;
  • whether that authority is still valid;
  • whether a specific action was within its permission;
  • whether its output was edited;
  • and who is responsible if something goes wrong.

NIST is already exploring standards-based approaches to identity and authorization for software and AI agents. This may sound like a technical infrastructure problem, but it is central to whether people will trust agents to operate in everyday life.

Deepfakes are both an enabling technology and a warning

Synthetic media has two sides.

On one side, it can improve accessibility, translation, entertainment, education, and remote communication. People may communicate across languages while preserving aspects of their own voice. Someone unable to attend an event could send an authorized digital representative. People with disabilities could use new forms of expression and interaction.

On the other side, synthetic media can separate appearance from identity.

A video may show my face without my involvement. A voice may sound like mine without my consent. An agent may communicate in my style without having permission to represent me.

Efforts such as the C2PA Content Credentials standard aim to provide information about the origin and editing history of digital content. Regulators, including the European Union, are also developing transparency requirements for AI-generated and manipulated media.

These are important steps, but a label saying “AI-generated” will not answer every question.

If a video uses my authorized digital likeness, it is AI-generated but may still be legitimate.

If someone creates an unauthorized imitation of me, it is also AI-generated—but represents a very different situation.

The more important questions are:

  • Was the person’s likeness used with consent?
  • What exactly did the person authorize?
  • Can that consent be withdrawn?
  • Is the agent clearly identifying itself?
  • Is the content authentic to the stated source?
  • Can the authorization be verified independently?

We will need to distinguish not only between real and fake, but also between authorized and unauthorized synthetic reality.

We may need the right not to be simulated

The development of digital representatives creates a new category of personal rights.

People may need:

  • the right to control commercial use of their face and voice;
  • the right to know when they are interacting with an AI;
  • the right to revoke an agent’s authority;
  • the right to inspect actions taken in their name;
  • the right to correct or delete an agent’s memory;
  • the right to prevent an agent from exceeding its permissions;
  • the right to decide what happens to a digital representative after death;
  • and perhaps the right not to be simulated at all.

The question of death is particularly difficult.

A sufficiently detailed digital representative could continue communicating after the human it represents has died. It might preserve memories, speech patterns, images, opinions, and stories. It could provide comfort or preserve family history.

It could also become a disturbing commercial product that the original person never intended.

Would the agent continue to learn after the person’s death? Who would own it? Could family members change its personality? Could a company keep charging a subscription to preserve it? Could it make statements the person never made?

A digital life may outlast the physical one, but persistence is not the same as consent.

The interface may be surprisingly ordinary

The early metaverse was closely associated with VR headsets.

Headsets will continue to matter, particularly for games, training, design, simulation, therapy, and remote presence. But the wider integration of digital and physical life may happen through more ordinary devices:

  • phones;
  • cameras;
  • earbuds;
  • smart glasses;
  • watches and health sensors;
  • cars;
  • payment systems;
  • maps and location services;
  • smart homes;
  • robots;
  • and connected workplaces.

Smart glasses are particularly interesting because they allow digital intelligence to share some of the user’s physical context.

An agent could potentially see what the user sees, hear what the user hears, recognize objects, recall names, translate conversations, give directions, or provide information without requiring the user to look down at a phone.

This form of augmented reality may be less about projecting animated objects into the room and more about giving an intelligent system situational awareness of the room.

That changes the role of the interface.

The most important metaverse device may not be the one that blocks out physical reality. It may be the one that understands physical reality well enough to assist us within it.

The physical world may also gain digital agents

The integration will not only happen on the human side.

Physical places, objects, and businesses may also have agents.

A hotel could have an agent that negotiates with a traveler’s agent. A car could schedule its own maintenance. A building could manage access, energy, repairs, and deliveries. A shop’s agent could answer questions about inventory, compatibility, and delivery without requiring a customer to browse through pages of products.

In this model, commerce could increasingly involve agents communicating with other agents.

My travel agent might tell an airline agent:

  • my preferred departure times;
  • acceptable connections;
  • accessibility requirements;
  • budget limits;
  • loyalty memberships;
  • and the situations in which it should ask me for approval.

The airline agent might then return options tailored to those constraints.

This could be more efficient than searching through dozens of near-identical listings. It could also create new risks if agents manipulate one another, prioritize hidden commissions, expose personal data, or make decisions their users do not understand.

Once again, convenience depends on trust.

Interoperability matters more than one platform

Another weakness of the first metaverse wave was the expectation that one platform would become the metaverse.

But our digital lives already span many systems: messaging, work, shopping, entertainment, health, finance, travel, education, and social networks.

A useful personal agent must cross at least some of these boundaries.

That requires interoperability—not necessarily one universal virtual world, but compatible methods for handling:

  • identity;
  • permissions;
  • credentials;
  • payments;
  • data portability;
  • digital assets;
  • spatial information;
  • provenance;
  • and agent-to-agent communication.

The Metaverse Standards Forum continues to work across areas such as avatars, identity, privacy, IoT, digital twins, geospatial systems, and 3D interoperability.

Blockchain may play a role in some of these systems, but it is not a requirement for the metaverse. The first wave sometimes treated blockchain, NFTs, and the metaverse as if they were the same thing.

They are not.

A decentralized ledger may be useful for certain ownership, credential, or transaction problems. It does not automatically solve identity, privacy, governance, interoperability, or trust.

The difficult question is not whether an item is on a blockchain.

It is whether people have meaningful control over their identity, information, assets, and agents.

A digital life that can exist only inside one company’s ecosystem is not a second life.

It is a rented account.

The business model will shape the result

Who pays for a personal agent?

If users pay directly, the agent may be more likely to work in their interests—but access could be unequal.

If advertisers pay, the agent might influence the user rather than represent them.

If retailers pay commissions, recommendations may be commercially biased.

If employers provide the agent, the boundary between assistance and workplace surveillance could become unclear.

An agent that knows our schedule, relationships, purchases, location, preferences, conversations, and health could become the most intimate technology product ever created.

That makes the business model a design decision, not an afterthought.

We should ask:

Is the agent working for me, learning from me, selling to me, monitoring me—or doing all four?

The most dangerous version of the metaverse may not be a dystopian virtual world.

It may be an invisible layer of personalized persuasion operating continuously across physical and digital life.

A new definition

My original definition still works:

Metaverse = Digital Life + Physical World

But I would now expand it:

Metaverse 2.0 = Digital Life + Physical World + Agentic Intelligence + Trust

Digital life includes our data, relationships, reputation, assets, history, and identity.

The physical world includes our bodies, homes, workplaces, devices, locations, and material needs.

Agentic intelligence allows digital systems to understand context, make plans, communicate, and take actions.

Trust determines whether those actions are authorized, transparent, accountable, and accepted by others.

Without intelligence, the metaverse is mostly a collection of environments, assets, and avatars.

Without trust, it becomes a world of impersonation, surveillance, manipulation, and unaccountable automation.

Both layers are necessary.

What I would watch now

Five years ago, my metaverse watchlist focused largely on virtual worlds, avatars, VR, AR, blockchain, digital twins, and devices.

I would still watch those areas, but I would now add:

  • personal AI agents;
  • agent identity and authorization;
  • smart glasses and ambient interfaces;
  • synthetic voice and video;
  • lifelike digital humans;
  • content provenance;
  • digital-likeness rights;
  • agent-to-agent commerce;
  • memory and preference systems;
  • robotics and embodied AI;
  • data portability;
  • and digital inheritance.

I would also watch the social behavior around the technology.

Will people permit agents to answer personal messages? Will meetings accept AI representatives? Will organizations allow agents to negotiate? Will a person feel insulted if someone sends an agent instead of attending? Will children grow up treating personal agents as normal extensions of identity?

Technology makes a new behavior possible.

Social acceptance determines whether it becomes part of everyday life.

An updated cultural watchlist

Several films and television series are especially relevant to this newer version of the metaverse:

  • Her — personal AI, intimacy, and the possibility that an assistant may become more than a tool.
  • Black Mirror: “Be Right Back” — reconstructing a person from their digital traces.
  • Marjorie Prime — memory, identity, and digital representations after death.
  • Upload — digital continuity, ownership, subscriptions, and inequality.
  • The Congress — licensing a person’s identity and synthetic likeness.
  • Pantheon — uploaded intelligence and the boundary between a person and a copy.
  • The Creator — embodied AI and the moral status of artificial beings.
  • Mission: Impossible – Dead Reckoning — AI, surveillance, digital identity, and uncertainty about what is real.

These stories are not predictions, but they are useful thought experiments. They explore what happens when identity, memory, intelligence, appearance, and physical presence no longer belong to one inseparable human body.

The word may never return—and that does not matter

The technology industry may not bring back the word “metaverse.”

It may prefer terms such as spatial computing, ambient computing, digital twins, embodied AI, agentic systems, personal intelligence, or intelligent assistants.

That does not matter very much.

The internet did not become significant because everyone agreed on a perfect definition. It became significant because separate technologies gradually connected into infrastructure that people used every day.

The same thing may happen here.

The next metaverse may not announce itself with virtual land, a token launch, or a headset.

It may emerge quietly as our digital identities gain memory, intelligence, appearance, voice, permissions, and the ability to act.

We may notice it when our agents begin interacting with companies.

We may notice it when they represent us in routine situations.

We may notice it when smart glasses connect intelligence to what we see.

We may notice it when we need proof that a video, voice, or digital person is authorized.

And eventually, we may realize that the boundary between digital life and the physical world has become difficult to identify.

The word became less popular.

The idea did not disappear.

It became intelligent.

Research and technology references

Cultural watchlist references

Related posts:

  1. It’s not just about platform, but big platform for #bigdata and #AI ready future. forbes.com/sites/heatherh…
  2. ⚡️ “AI, Machine Learning, Automation” by @EvolvingViews https://twitter.com/i/moments/909063667802546176
  3. My AI Notes – Sharing tips and tricks from my AI playground
  4. From Ollama to MLX: Achieving 2-3x Performance on Apple Silicon