What UX/UI mean in the future with ai

Hey there. Happy Saturday! This is a rather long newsletter article, but it helps to prepare us as UX/UI designers for a future we may be able to live in! Stay with me and let me know which resonate with you the most.

A few days ago, I received a detailed set of notes from Fei-Fei Li’s talks and writing on world models and spatial intelligence. Her central argument is that today’s large language models have become extraordinarily capable at understanding and generating language, but language represents only a small portion of how humans understand reality. Intelligence also requires an understanding of space, geometry, physical relationships, movement, causality, and what happens to an environment when someone acts within it.

Her phrase “From Words to Worlds” immediately made me think about the implications beyond AI research itself. 

As a designer, I became less interested in asking only how world models would make artificial intelligence more capable, and more interested in what they might do to the relationship between humans and computers.

I brought the notes into AI and started discussing them. The initial question was simple: If AI eventually understands not only what we say, but also our history, goals, environment, and the physical world around us, what happens to the interface?

That conversation eventually led me to a larger idea. I think we are moving through three connected transitions at the same time: from operating software to expressing intentions, from expressing intentions to delegating responsibility, and eventually from explicitly communicating context to sharing context with machines.

If that happens, the defining design problem of the next generation may no longer be interface design alone. It may increasingly become delegation design: deciding what responsibility humans are willing to give to machines, and designing the boundaries that make that autonomy understandable and trustworthy.

1. We are currently living through the fourth major interaction paradigm

One useful way to understand the current moment is to look at the history of human-computer interaction as a gradual reduction in how much humans need to understand the internal language of computers.

The first generation was command-based computing. Humans needed to understand how machines wanted to receive instructions.

The second generation introduced graphical user interfaces, where buttons, folders, windows, menus, and visual metaphors translated machine capabilities into something people could recognize.

The third generation, accelerated by smartphones, made interaction even more direct through touch, gestures, cameras, sensors, and mobile applications.

Generative AI introduces a fourth paradigm: language as an interface for intention.

Instead of specifying every operation required to accomplish something, we increasingly describe the outcome we want. I can give an AI thirty customer interviews and ask it to identify the reasons customers are not converting. I can provide research and ask it to turn the material into an investor presentation. I can describe a website problem and ask it to propose a new structure.

The abstraction layer has moved significantly upward.

Traditional software largely follows this sequence:

Human intention → interface → commands → result

Language AI increasingly compresses that relationship into:

Human intention → AI → result

This is why prompting feels fundamentally different from operating conventional software. We are beginning to manipulate intention rather than implementation.

However, there is still something surprisingly traditional about this relationship: the human usually has to initiate everything.

2. Today’s AI can execute work, but humans still own the cognitive workflow

When I use ChatGPT today, I still need to recognize that something should happen before AI can help me.

I need to realize that I should analyze a document. I need to remember that I should write a post. I need to notice that a client needs a follow-up. I need to decide that I should research a particular question. Then I open the AI, provide the appropriate context, formulate the request, evaluate the result, and determine what should happen next.

The AI may perform a remarkable amount of execution, but I still carry much of the responsibility surrounding that execution.

This distinction becomes clearer when comparing current AI with traditional software.

Traditional software says: You own the task, and I provide the tools.

Most copilots say: You still own the task, but I will help you execute it.

Current conversational AI often goes somewhat further, but the underlying relationship remains reactive. The intelligence waits for the human to recognize the need and begin the interaction.

This is why I think we are already moving toward an intermediate stage between language AI and the world-aware systems Fei-Fei Li describes. We could call this generation 4.5: agentic AI.

The important characteristic of this stage is not simply that AI becomes more capable. It is that AI begins assuming some responsibility for determining what should happen next.

3. The next major transition is from reactive AI to proactive AI

Imagine that I use AI to help manage my personal content.

The current interaction might begin with me opening ChatGPT and saying, “Help me come up with three LinkedIn posts for this week.” The AI can analyze my positioning, review previous ideas, and produce excellent drafts.

This is useful, but I still own the workflow.

I remembered that I needed content. I decided that today was the appropriate time to work on it. I brought the relevant information into the conversation. I initiated the request.

Now imagine a different relationship.

On Monday morning, an AI sends me a message:

“Your post about founder hiring mistakes generated considerably more conversations with early-stage founders than your other posts last week. I also found two notes you wrote recently that extend the same argument without repeating it. I developed three possible directions for this week. The second one appears most consistent with your positioning. Would you like to review it?”

I never asked the AI to do this.

The system understood an ongoing goal, observed what happened, connected recent performance with historical information, determined that something deserved my attention, and proposed a next action.

That creates a fundamentally different interaction loop:

AI observes → understands context → reasons → proposes → human decides → AI executes → AI monitors the result

The difference between this and today’s chatbot is not primarily conversational interface design. It is proactivity.

The computer is no longer waiting indefinitely for a command.

4. The most important concept in agentic products may be ownership

As I continued exploring this with AI, one word became increasingly useful: ownership.

The evolution of AI products can be understood through how much responsibility moves from the human to the machine.

Traditional SaaS operates under the assumption that the human owns the task while the software provides capabilities. Copilots allow the human to retain ownership of the task while AI assists with execution. Agents begin allowing humans to retain ownership of the goal while AI takes responsibility for portions of the task.

The more mature version could be described differently: the human owns the desired outcome, while the AI owns much of the workflow required to achieve it.

Consider social media management. Traditional software might provide analytics, scheduling, a content calendar, writing tools, social listening, engagement metrics, and hundreds of additional features.

Nevertheless, the human remains responsible for asking the important questions. Should I publish something today? What should I discuss? Which previous post performed unusually well? Why did it work? Is there an important conversation happening in my industry? Did someone interesting respond to my post? Should I answer them? Is there something in my notes that has suddenly become relevant?

The software organizes information, but the human still owns the responsibility.

An agent-native product begins with a different product definition.

Instead of asking, “How can AI help someone write social media content?” we could ask:

What would it take for an AI to responsibly own part of someone’s public presence?

That question produces a completely different product.

5. Case study: the primary innovation is not the chat interface

This is part of what caught my attention when I recently looked at Stanley, an AI product designed around managing an X presence. I read an article about them reaching $3M in several weeks and was curious enough to try the product. Then I signed up for the paid version the next day!

At first glance, one might conclude that the interesting feature is its conversational interface. I do not think that is the most consequential part.

We already know how to build chat interfaces. Replacing every button in traditional SaaS with a text box does not necessarily improve the experience. For frequently repeated or highly deterministic actions, clicking a button may remain considerably faster than explaining an instruction to an AI.

The interesting change occurs when conversational interaction is combined with persistent context, memory, tools, and proactive behavior.

Traditional SaaS essentially says:

Come to the software when you need something.

The current generation of AI assistants largely says:

Come to the AI when you need something.

Agentic software begins to say:

I will come to you when something needs your attention.

That reversal appears small at the interface level, but it represents a substantial change in the relationship between humans and software.

The product no longer merely waits to be operated. It begins participating in the management of an ongoing responsibility.

6. The primary product interface may increasingly move outside the product itself

This creates another fascinating possibility: users may spend progressively less time inside the primary interface of the software they depend on.

Imagine an agent that communicates primarily through Telegram, iMessage, WhatsApp, Slack, or email. The underlying application still exists, but the application becomes a control center rather than the place where most interactions occur.

The dashboard might be where someone configures permissions, reviews history, changes strategy, manages integrations, or investigates what the AI has done. Everyday interaction, however, could happen through communication channels the user already occupies.

On Monday, the agent might tell me that it has identified three conversations worth joining and explain why one is particularly relevant to my positioning. I approve a response.

On Wednesday, it might return because the conversation has developed and suggest a different angle rather than repeating the original argument.

On Friday, it might explain that one piece of content generated a relevant founder conversation and recommend extending that theme the following week.

I might only open the actual product dashboard twice during the entire month.

This challenges one of the oldest assumptions in SaaS product design: that more product engagement necessarily means more product value.

For an effective agent, the opposite may eventually be true.

Time spent inside the interface can decrease while outcomes delivered increase.

The best AI product may eventually be one that users rarely need to open.

7. World models extend this idea from our digital context into the physical world

This is where Fei-Fei Li’s argument about world models becomes particularly important.

Today’s emerging agents can potentially understand an increasing amount of our digital world: documents, messages, calendars, contacts, analytics, projects, previous conversations, preferences, decisions, and long-term goals.

That context alone allows AI to become significantly more proactive.

World models introduce another dimension.

They aim to give AI an understanding of space, geometry, objects, physical relationships, movement, actions, and consequences.

A language model can understand the sentence, “Move the chair closer to the table.”

A genuinely world-aware system needs to understand which chair I mean, where the table exists in three-dimensional space, whether there is sufficient room to move the chair, whether the new position blocks a doorway, how people would move around it, and what physically happens when the object changes position.

Language provides intention.

A world model provides grounding.

Combining them begins moving AI from understanding what humans say toward understanding the situation in which humans say it.

That could become one of the most important changes in human-computer interaction since the graphical interface. 

From a short term perspective, we are allowing Ai to see our small world about “design”, about “development”, and about “startups” so it can operate overlapping these worlds, but imagine a future where Ai can operate as an expert from our world in all aspects, outside of the chat window?

It’s scary and exciting at the same time.

8. Eventually, the world itself becomes part of the interface

Imagine walking into a kitchen and saying:

“This area doesn’t work very well. I want the kids to be able to get breakfast themselves.”

Today’s language models can provide recommendations based on a description or photograph.

A sufficiently capable world-aware system could understand the actual kitchen: cabinet heights, appliances, available storage, objects on the counter, traffic patterns, spatial constraints, and potentially even how people regularly move through the environment.

It might respond:

“If breakfast items move into these two lower drawers, the children can reach everything without crossing the primary cooking area. I can show you three possible arrangements.”

I look at one configuration and say:

“This feels too crowded.”

The word “crowded” is not a conventional CAD parameter.

However, a world-aware intelligence might interpret that perception through circulation width, furniture density, visual weight, sight lines, object scale, and patterns of human movement.

I could point at a cabinet and ask:

“What if we move this over there?”

The AI understands what “this” refers to and where “there” is. It understands the spatial consequences of the change.

At that point, interaction is no longer primarily based on screens.

It becomes a combination of language, vision, gesture, spatial awareness, and shared context.

The world itself begins to become the interface.

9. Prompt engineering may eventually look like a transitional interaction model

Thinking about world models led me to another conclusion: the current obsession with prompt engineering may eventually look surprisingly temporary.

Prompt engineering is still a form of humans adapting themselves to machines.

We learn how much context to provide. We learn how to structure requests. We learn how to define roles, constraints, formats, examples, and desired outcomes so that the machine understands us more accurately.

The interface is dramatically more natural than a command line, but the human is still doing considerable work to compensate for what the machine does not know.

Human communication works differently because humans share enormous amounts of implicit context.

If I am sitting across from someone and say, “Can you hand me that?”, the sentence contains very little explicit information.

What is “that”? How should the person pick it up? Where should they place it? How should they avoid the glass beside it? Why might I need it?

None of this needs to be specified because we share the same environment and possess similar models of the physical world.

AI interaction today frequently requires detailed prompts precisely because humans and AI do not yet share enough context.

World models could begin changing that.

The more context humans and machines share, the less humans need to explain.

The long-term future of AI interaction therefore may not involve humans becoming extraordinarily skilled at prompting machines. It may involve machines becoming extraordinarily skilled at understanding humans.

10. Shared context may be more important than the conversational interface itself

This distinction matters because it changes how we evaluate AI products.

It is easy to look at an AI application and ask whether the chat experience is good. A more important question may be: How much context does this intelligence actually share with me?

Does it understand only the current conversation?

Does it remember previous decisions?

Does it understand my broader objectives?

Can it observe relevant changes without being explicitly instructed?

Does it understand the tools available to it?

Can it recognize when something requires my attention?

Can it distinguish between an action it can safely complete and one that requires permission?

Eventually, can it understand the physical environment in which the interaction occurs?

As shared context becomes richer, explicit interaction can become smaller.

This creates an interesting paradox.

The more intelligent the system becomes, the less interface the user may need.

Instead of building increasingly complicated interfaces for increasingly capable systems, we may find ourselves doing the opposite: building more capable systems that require progressively less visible interface.

11. For designers and founders, the critical question changes from features to delegation

This is where I think the implications become particularly practical.

For decades, founders have primarily asked: What features should we build?

Designers then ask: How should users discover, understand, and operate those features?

Agent-native products introduce a question that may need to come before both: What responsibility is the user willing to delegate?

Instead of asking what features an AI social media product should contain, ask what would be required for someone to trust AI with part of their public presence.

Instead of asking what an AI CRM dashboard should look like, ask what would be required for someone to trust AI with relationship follow-up.

Instead of asking how to add AI features to a design application, ask which parts of the process between business intention and design outcome AI could responsibly own.

Once the question changes, the product requirements change as well.

Memory becomes a design problem.

Context becomes a design problem.

Permissions become a design problem.

Confidence and uncertainty become design problems.

Interruption becomes a design problem.

Reversibility becomes a design problem.

Trust becomes a design problem.

The important questions become: When should the AI ask? When should it recommend? When should it act independently? When should it challenge the user? When should it remain silent? What actions require approval? What actions should always be reversible? How does the AI explain what it has done? How does it gradually earn additional autonomy?

These questions are not primarily about the placement of interface elements.

They concern the relationship between human intention and machine agency.

I think we will increasingly need to think about this as a distinct design discipline.

I would call it delegation design.

12. We may be moving from designing interfaces to designing relationships with intelligence

There is an interesting irony in all of this.

As interfaces become less visible, interaction design may become considerably more important.

A button has explicit boundaries. The user decides when to press it. The user understands that they initiated an action. The relationship between intention and execution is relatively clear.

An intelligent agent has much less obvious boundaries.

If an AI continuously observes relevant information, remembers context, initiates conversations, proposes actions, executes tasks, and learns preferences, designers have to create invisible but understandable boundaries around that autonomy.

The user needs to understand what the AI knows, why it is raising something now, what it has already done, what it intends to do next, what requires approval, what can happen independently, and how an unwanted action can be reversed.

The central challenge becomes less about making complexity usable and more about making autonomy trustworthy.

That may become one of the defining design problems of the next decade.

What I find particularly interesting is that this essay itself began as a small example of the transition I am describing.

I received notes from Fei-Fei Li’s work on world models. I brought those notes into AI. Instead of asking for a summary, I used AI as a thought partner. I questioned the ideas, followed their implications, connected them with products I had recently encountered, challenged some of the initial conclusions, and gradually arrived at a framework I did not have when I began.

AI helped extend my thinking.

But I was still responsible for initiating every turn, introducing the context, recognizing the interesting questions, and deciding where the conversation should go next.

That is approximately where we are today.

The next generation of AI may increasingly recognize those questions before we ask them. It may understand enough of our goals and history to know when something deserves our attention. World models could eventually extend that understanding beyond our digital lives into the physical environments we inhabit.

Seen through that lens, the progression of human-computer interaction becomes remarkably clear.

GUI computing allowed us to operate software. Language AI allows us to describe tasks. Agentic AI will increasingly allow us to delegate goals. World-aware AI may eventually allow humans and machines to share context and reality.

Every transition reduces the amount of explicit instruction required from the human.

Every transition also changes what designers are.

P.S. Do I sound 10X smarter than I am because of AI’s contribution lol ? We live in a crazy and every evolving world right now, I am both scared and excited at the same time!

Share my newsletter

Loving my content so far? I’d appreciate if you can share my newsletter to a friend 🙂