Nat Rubio-Licht

If you thought AI was everywhere in 2025, that future is only going to accelerate in the next 12 months. There were developments we never could have anticipated a year ago. And while there will undoubtedly be plenty of surprises again this year — make sure you're subscribed to The Deep View to catch the biggest ones every day — there are plenty of future developments already in sight. Today, Nat Rubio-Licht and I break down the biggest ones to watch. Let's count them down.

1. Can AI glasses make consumers love AI?

The big tech companies — Meta, Google, Apple, Amazon, and others — are betting that AI glasses will convince consumers that AI is good for more than just chatbot searches and laughing at AI slop. Consumers will still need a lot of convincing. As much as the Meta Ray-Ban AI glasses have been a surprise hit, they've still only sold several million pairs in total over the past couple of years. And screenless AI hardware devices like the Humane AI Pin and the Rabbit R1 have been largely rejected by consumers. OpenAI has generated buzz for its screenless Jony Ive-designed device that sounds a lot like the doomed Humane AI Pin. And startups like Pickel have created heat around their forthcoming AI glasses, hyperbolized as an $800 "soul computer" that looks too good to be true from the slick marketing video. Still, I've tried the Meta Ray-Ban Display glasses, and there are some compelling everyday AI features, such as live captions and live language translation. I'm also interested in the forthcoming Brilliant Labs Halo glasses, which will feature a privacy-first, multimodal AI agent to help you remember things — similar to how meeting assistants take notes for you, but for real-life conversations. But make no mistake, companies large and small are coming for your face with AI in 2026.

2. Robotaxis will foreshadow AI's future

Self-driving taxis from Waymo, Tesla and others will likely grow to become one of the most common ways consumers interact with AI to get a taste of the future in terms of declining costs, societal impact on jobs, and automating previously manual tasks. During 2025, Waymo — owned by Google's parent company Alphabet — transformed from a quirky AI/robotics experiment into an emerging alternative to both Uber and taxis in five US cities. Those five were San Francisco, Los Angeles, Atlanta, Phoenix, and Austin. Near the end of the year, it also announced expansion into 5 more cities: Miami, Dallas, Houston, San Antonio, and Orlando. By the end of 2026, Waymo's expansion could extend up to another dozen US cities and its first international location in London. Meanwhile, Tesla's Robotaxi service launched in Austin, Texas and the San Francisco Bay Area during 2025, and the company has slated its next expansion for Las Vegas, Phoenix, Dallas, Houston, and Miami. Waymo rides are currently about 30% more expensive than standard Uber, Lyft, and taxi rides. I suspect that's because Waymo wants to avoid the negative publicity of undercutting human drivers and potentially putting people out of work, at least for now. Tesla appears to have fewer qualms about that — it's rides have been noted to be significantly cheaper than Uber at times. That feels like the inevitable reality here. I've been surprised at how many people have already told me they prefer robotaxis to human drivers, for a variety of reasons. It's a future that's about to take another big step forward in 2026.

3. We enter the post-LLM phase of generative AI

At an event in November, Hugging Face CEO Clem Delangue called it: We’re not in an AI bubble, we’re just in an “LLM bubble.” Major model providers have been on a frenetic mission to make their large language models more formidable, with many chasing after the elusive and ill-defined goal of artificial general intelligence. But some of AI’s foremost thinkers have started to question the validity of this chase, turning their attention beyond the LLM. World models and physical AI have become a major focal point in recent months, with voices like Yann LeCun and Fei-Fei Li calling it the next frontier. And as large language model inference only becomes more costly, enterprises may find more luck with small language models and domain-specific models, or as IBM Research Chief Scientist Ruchir Puri put it, “artificial useful intelligence.” 2026 might be the year that LLMs hit a plateau – not in capability, but in no longer being the sole driving force behind the AI boom.

4. Agents usher in the post-chatbot era

Tech companies went absolutely feral for agents in 2025 as the tech promised to break us out of the call-and-response requirements of conventional chatbots. And as companies figure out what exactly these autonomous little helpers are capable of, the hype will only continue. With that hype, new complexities will emerge. These agents are already creating cybersecurity hiccups for IT teams, and could present new challenges for HR as the “digital employee” impacts workforce operations and morale. But agents might also force some of the biggest players in AI to work together for once, such as with the launch of the open-source Agentic AI Foundation in December, bringing the industry one step closer to interoperable agents that unlock far more value working together than they could working alone.

5. New chips will redefine AI beyond Nvidia

The AI industry’s manic obsession with Nvidia reached such a fever pitch this year that there’s literally a name for it: Jensanity, dubbed after the company’s perpetual leather jacket-wearing CEO, Jensen Huang. But some are starting to recognize that there exist alternatives to Nvidia’s reign. Amazon unveiled upgrades to its Trainium chips at Re:Invent in December, and Google is spreading its TPUs, the company’s custom chips, far and wide, making its seventh-generation Ironwood chip generally available and discussing deals with Anthropic and Meta. Though it’d likely take some kind of catastrophic event to truly dethrone Nvidia at this juncture, AI innovators who want to move rapidly but are constrained by their Nvidia orders will have more alternatives in 2026.

6. Model developers will start to weigh the risks

Though voices like Geoffrey Hinton, Yoshua Bengio and (for some reason) Joseph Gordon-Levitt have been shouting from the rooftops about the risks that AI presents, model developers have been keenly optimistic so far. That, however, might change in 2026 as models become more powerful and new capabilities emerge. OpenAI might already be on guard, with CEO Sam Altman posting on X in late December that the company is seeking a “head of preparedness” to help implement “increasingly complex safeguards.” Given that the world’s most valuable startup has been treated as an industry bellwether thus far, it wouldn’t be surprising if other AI firms followed in its footsteps.

7. Government relationships with AI will get stickier

Last year saw a number of major AI firms seek to get cozy with the US government, from offering steep discounts to the General Services Administration to multi-billion-dollar investments that support the Trump Administration’s AI Action Plan. The Administration, too, is trying to make it easier for AI companies to have free rein by stymying state legislation. But not everyone is pleased. Sen. Bernie Sanders, for example, has pushed for a moratorium on data center development amid an “unregulated sprint to develop & deploy AI.” As AI becomes more and more politicized, don’t expect the regulatory landscape — or the discourse around how it should be governed — to become any less heated. With mid-term elections looming in the US, AI is on course to become a national issue rivaling health care, immigration, and the national debt.

8. We still won’t know what AI is going to do to us

To put it plainly, there’s still plenty of debate about whether AI is good or bad. There are the AI zealots who believe the tech is the key to unlocking a level of growth and prosperity that’s exponential and therefore incomprehensible to most of the human race. On the flip side, there are the doomers who believe AI and robots will diminish our humanity — if not literally rise up and kill us all (they nearly view The Terminator as a documentary). And there's abundant data available to support either side of the debate. Conflicting reports are constantly emerging. AI is creating jobs, and AI is killing jobs. It’s going to tank the stock market and make billionaires out of a lot more startup founders. It’s eroding our ability to forge human connections and providing people with an outlet for mental healthcare. It’s destroying our capacity to think critically, and it’s supercharging discovery. As these contradictory notions only continue to conflict, people are going to do what they’ve always done: Find the data that supports their preconceived biases and cling for dear life.

It's clear that 2025 was the year tech companies became obsessed with humans doing less. Practically every big tech firm went crazy over AI agents last year. The constant refrain at major conferences was agentic innovation, as these firms repeatedly touted the astonishing productivity gains that could result from installing digital coworkers alongside existing human workforces. 

Enterprise C-suites were all in on breaking agents out of the pilot phase and automating legacy processes. Agents even brought tech rivals like Anthropic, OpenAI, Google, and Microsoft together in a coalition dedicated to developing open-source standards for them. The excitement around the tech has reached such a fever pitch that Salesforce is even considering a total rebrand to Agentforce (à la Facebook-to-Meta in 2021). 

The big takeaway? We’ve built foundational models with formidable intelligence. Agents let us actually do something with them, Steve Zisk, principal data strategist for Redpoint Global, told The Deep View.

“We're finally at a point where the AI pattern recognition engines, if you will, start to actually resemble what people think of as human interactions,” said Zisk. “That has meant that a lot of people on both sides of the equation, the consumers, the big brands, big companies and so on, are reassessing what they can actually hand off to the machine.”

Agents are the natural evolution of where AI and technology broadly are going, Prasidh Srikanth, senior director of product management at Palo Alto Networks, told me. Search engines democratized information, chatbots democratized intelligence, and now agents democratize work, he said.

“It's behaving on behalf of a human being, thinking about our intelligence and actually making sense of what you need to do to fulfill an objective,” Srikanth said.

But despite all the buzz, anticipation and starry-eyed hopes, agents are far from ready to be our actual work companions, Neil Dhar, global managing partner at IBM Consulting, told me. As it stands, we’re still in the “first or second inning of the whole agent race,” he said. And while enterprises are aware that agents have the potential to upend the way we work, “people are just getting in tune with what an agent actually is.”

AI agents have trust issues we can't ignore

While tech companies talk a big game with agents, there are cracks in the foundation that can’t be ignored.

At their core, agents still face all the same problems as their conventional chatbot predecessors. Enterprises struggled to work around the fundamental issues they faced with generative AI, and a new form factor isn’t necessarily the solution. A study from Gartner estimates that more than 40% of agentic AI projects will be canceled by the end of 2027, citing high costs and an inability to control risks. 

The throughline problem is trust, said Dhar, in all of its “different flavors.”

• Security and privacy remain major complications. The more autonomy and agency we give these systems, the more access we have to give them, too, said Srikanth. That introduces a new level of risk that enterprises aren’t ready to handle. “Security often comes as an afterthought,” he said.
• On top of security, there’s the quality and accuracy issue, said Zisk. Enterprises must constantly weigh the “hallucination tax,” he said. What are the costs, financially, operationally, and reputationally, of one of these systems flying off the rails?

Finally, there’s a broader trust issue that employees and executives alike are facing, questioning how this tech will entirely shift their workflows, their processes and potentially their livelihoods. “There's just the fear of the unknown,” said Dhar.

If enterprises can’t trust these systems, they’re effectively stuck. The agents will remain in pilot mode, only being tasked with small and insignificant processes that have little to no effect on a company’s earnings – a far cry from the transformational powers that tech giants are claiming.

But foundational problems may require foundational solutions, said Srikanth. Just as with conventional AI systems, good data is key. One study from Capgemini found that less than 20% of enterprises have a high level of data readiness for AI adoption, and only 9% consider themselves fully prepared. 

While having agents explain themselves could help instill trust, said Zisk, “explainability is a cure for a disease that the AI agents have,” he said. “What you really want is to find a prevention for that disease. The prevention is to make sure that they have the right data, the right prompts, [and] the right controls in the first place.”

Can agents break out of the 'productivity lens'

With upwards of a trillion dollars being poured into the AI market broadly, the pressure is on to make the investment worth it.

Though cost forecasts vary, one estimate from Bank of America analysts found that agentic AI spending could reach $155 billion by 2030, and could deliver up to $1.9 trillion in value for enterprises as these systems take over more and more workflows. 

However, stakeholders are getting antsy, said Dhar.

“Boards are putting pressure on CEOs to actually deliver return on investment,” said Dhar. “And as such, stuff rolls downhill. CEOs are putting a lot more pressure on their CFOs, COOs and CTOs to deliver actual returns.”

As it stands, companies are keen to use agents to cut as much spending as possible, said Dhar. Enterprises are looking at ways to do more with fewer people, with sights set solely on returns from productivity gains. Things like headcount, efficiency, and “bending the cost curve” are top of mind. “I think most boards and CEOs right now are measuring AI through a productivity lens,” he noted.”

But the real value isn’t going to be derived from the bottom line, he said. Rather, it’s going to come from the top, rethinking not just what can be done by substituting agents for human workers, but also reimagining how processes can be transformed entirely, he said.

Actually getting there is going to take some trust from the people writing the checks, said Dhar. He pointed out that it's often easier to justify an investment by cutting something and saving money than to bet on something you expect to generate more revenue. While that may sound paradoxical, the fact is that it's more tangible to reduce spending than to believe in a key move that will make more money flow your way in the future.

Mike Clark has seen tech go from zero to one hundred over the years. But now, as the director of product management for AI agents at Google Cloud, “none of those step functions compare to the agentic step function,” Clark told The Deep View.

He said 2025 was the year of enterprises seeking to give agents more responsibility and control. But challenges remain in the market, which is still in its infancy. Clark sat down with The Deep View to discuss agent adoption challenges, the possibilities agents unlock, and whether we should be calling these systems our “digital coworkers.” This interview has been edited for brevity and clarity.

Nat Rubio-Licht: What do you think enterprises are getting wrong, or right, as they build and deploy AI agents?

Mike Clark: I’ll start with the in-between answer. For enterprises, or for anybody building agents today, it's the moment [where we're] trying to understand what agents really are. We don't have a clear definition of what agents are. The one thing that I'm watching a number of enterprises do is step back and actually define what is an agent to us, and what do we expect out of it.

For Google Cloud and a lot of our customers, it's the four key pieces: The LLM that wowed us three and a half years ago; connecting tools to that model and being able to do things; orchestration – and this is where agents truly come to be – how do we do these in a multi-step phase; and then finally, like a runtime to build them at scale.

One of the big misses that I see is a lot of companies are focused on purely the chat side of the world. “Agents” have this meaning in human language that doesn't have the same meaning technically. Everybody thinking about it as chat first really limited the capabilities and scope. Think about background agents and things that are solving these traditional background tasks, not just replacing workflow, but, really think about the objectives of what your company is trying to do.

Rubio-Licht: What challenges have you come across at Google Cloud as far as agentic adoption?

Clark: The number one thing that enterprises care about is quality. So the number one challenge is that security and governance don't matter if I can't have a quality product. Getting to a place of trust and risk mitigation has meant getting to a quality place.

I think as models continue to improve and that technological capability continues to grow exponentially – which Gemini 3 did for us. Watching the impact that those model improvements have on all the other pieces of agents was awesome. Watching Claude's advancements and OpenAI’s advancements have similar impacts on agents has been great for the whole industry. It's helped our enterprise customers find trust in the quality that they have. If people don’t trust it, they’re not going to scale it.

Rubio-Licht: How do privacy and trust play a factor in Google Cloud’s agentic strategy?

Clark: Google has a strong reputation from a privacy, security and trust perspective, in the products that we build and ship. We try to lead the charge on that. But I want to uplevel to the industry for a moment. As an industry, when I think about agents, it introduces some new potential vulnerabilities – the ability for prompt injection, the ability for tools to do some things like that, and we've invested in tools customers can use and leverage in the Vertex AI ecosystem to look for those patterns of prompt injection. 

On our agent platform that we released, we've taken agents from acting like a user and having the identity like a user to having this principal identity as an object in [Google Cloud] that you can attach security to. You don’t have the exposure of managing it as a user, instead you’re actually managing it as part of the infrastructure. We've tried to take the approach of making least privilege a core part of how agents get built.

Rubio-Licht: How do ethics and workforce comfort play a role in enterprise agentic adoption?

Clark: Personally and anecdotally in having conversations with enterprises, the folks that are having the most success, the employees are the ones that have helped unlock some of these agent capabilities. And it's not replacing them.

Some of the agents that companies have built that have some of the best interactions are those that users don't even interact with, but it helps inform them about their day. It helps inform them about decisions that are being made. It helps inform them about a number of other things in ways that they themselves wouldn't have had a mental context window big enough to handle. It's actually helping them get more done.

From an ethics perspective, the companies that are most successful are really transparent about the agent products and the impacts of those. The goals of those projects have very specific meaning about growing the business and unlocking new lines of business.

Rubio-Licht: Do you think agents should be treated as tools, or more so as digital coworkers, and why?

Clark: So I grew up on a farm, where you have animals that work versus animals that are pets. Even the things that work, sometimes, you anthropomorphize. You give this cow a name, or that horse a name. This is an interesting challenge that we've always had: How do you humanize the things that you interact with, and what's the impact of humanizing it in that way? That's why you see such a strong contrast in organizations being really strong one way versus the other, because I don't think there's a clear industry foundation on humanizing. We build agents that talk to you in very humanized voices, but at the same time, are merely AIs operating as a model in the back end through a series of API's to make that happen. I don't think this is unique to AI. I think it's just the moment in time that we happen to be in.

One clear observation is technological capabilities are happening on this exponential curve, but organizational changes are logarithmic changes. They happen on a much, much slower curve. For some organizations, humanizing the technology helps me get closer to it. It closes some of that gap from fear. For other organizations, dehumanizing it and keeping it just as an API running in the background has its same effect, because it's tied more to their culture as an organization than it is to agents or AI as a concept.

Rubio-Licht: What do you think is the biggest misconception about agents today?

Clark: The biggest misconception … is that they are there and ready to solve every single problem that we have. While they are very capable, and while a lot of the tools are happening, most of the tools, most of the interoperability, most of the interconnect between things, it’s very nascent. A2A, the first protocol to give interoperability between agents, was introduced by us eight months ago. MCP [Model Context Protocol] is only a year old. But also, I think everybody looks at the rate of change today, and a month in AI cycles is like five years in tech cycles 10 years ago. It's a misconception of the capabilities and where they are today.

The second misconception is that it just replaces the same thing we're doing with just automation. A lot of our processes and governance and other things that we have in enterprises are built around processes that date back to the [1940s], 50s and 60s, where people would type in triplicate on a typewriter and take one thing and hand it to one person. And we've just added technology on top of those processes over the years. The misconception is we're just replacing the same workflows by putting an agent where there may have been a typewriter or a person or a computer.

Rubio-Licht: What does the future look like for the agent market?

Clark: I think a lot of our interactions with AI are going to be driven through agents. It's going to become less obvious when we're interacting with AI [and] when we're not, or what work’s being done by AI or not, because it's just going to become a blend. We're going to see more and more things happening and being done with agents. We're going to start to see agents defining contracts, making economic transactions, making transactions around assets between organizations. What that's going to do is that's going to unlock more and more capabilities for businesses around trade, around how they interact with one another, both from a data trade perspective and even physical asset management. Those are going to be some of the creative things that we're going to start to see [during 2026].

I also think it changes all of our careers. Myself as a product manager, how much like the role of a PM is going to change, the role of designer, the role of everybody. I can now do a bunch of these cross functional pieces and start to interact on a much deeper level, solving deeper problems. I think we're going to solve a lot of problems that have plagued society, and plagued the world.

Enterprises can't wait to deploy AI agents, and AWS is dying to help.

Agents were front and center at Amazon Web Services’ re:Invent in Las Vegas this week. Two of the week’s keynotes were filled with agentic releases, aiming to help companies build, deploy, and keep track of their agentic coworkers more quickly and seamlessly.

“One of the biggest opportunities that is going to change everyone's business is agents,” said Matt Garman, CEO of AWS, in his Tuesday keynote.

Some of the highlights include:

• Tools for quality evaluations and policy controls for agents in Amazon Bedrock’s AgentCore platform;
• Systems that allow users to build agents that learn from experience and are easily customizable;
• And a new offering called "frontier agents,” which go beyond conventional agentic capabilities in autonomy, scalability and how long they can run. Debuting three of these frontier agents in his keynote, Garman called these systems a “step function change more capable than what we have today.”

Amid the mad dash to get these agents into the workforce, whether enterprises and their workforces are ready for them remains unclear. In a panel at re:Invent on Tuesday, May Habib, co-founder and CEO of agentic AI platform Writer, said that agents could fundamentally change the concept of career progress.

“You are no longer going to be promoted because you can execute tasks effectively,” Habib said. “You're going to be promoted because you can build systems of agentic orchestration that make these tasks happen.”

And with all the promises of what agents could be capable of, human employees may be getting uncomfortable. A survey of more than 1,000 workers by EY found that, while 84% of employees are eager about the prospect of agents, 56% worry about their job security. The anxiety comes amid companies such as Salesforce to Klarna to Accenture slashing staff as they pour more cash into AI.

“Just looking at this with the realist point of view, the vast majority of the CFOs (and) the C-suite out there are not mincing words around trying to be the most efficient, leanest, fastest company,” Habib said.

AI agents might finally be breaking out of the pilot phase.

Full AI implementation is up year over year, from 11% to 42%, according to a survey of 200 CIOs from Salesforce released today. Budgets dedicated to AI have nearly doubled, and 61% of CIOs reported feeling ahead of competitors on their deployments.

And amid the hype of automation, agents are the industry’s It Girl, with CIOs earmarking 30% of their budget to agentic AI alone. Around 96% of CIOs reported that they are currently using or plan to use agentic AI in the next two years.

As agentic AI comes to the fore, it’s caused a shift in the kinds of work that actual humans are doing.

• 94% of CIOs say AI agent adoption will require them to expand their skillsets, focusing on soft skills instead of technical duties.
• Around 61% reported a desire to improve leadership skills, 57% reported improving storytelling skills, and 55% reported improving change management and communication.

But despite shifting priorities and increased momentum, there are still bumps in the road: While 81% acknowledge that this kind of adoption requires collaboration between teams, less than half are actually doing this.

And as investments in these deployments increase and companies become eager for returns, getting employees to trust, use and continuously integrate this tech is crucial: 93% reported that successful deployments hinge on integrations into “everyday work.”

“A huge part of my job has become about enabling employees and having them understand that AI can complement the work they do, and isn’t here to just replace them,” said Daniel Schmitt, CIO of Salesforce, in a call with reporters.

The other major concern is data. While organizations have come a long way in cleaning up their data, trust still remains a major bottleneck in deployment. Only 23% of CIOs reported that they are confident they are investing in AI with built-in data governance.

“Having a clean dataset is absolutely key,” said Schmitt. “You have to be able to trust the data that AI is using. AI is not magic, it uses the same data that humans use.”

AI firms continue planning astronomical AI infrastructure. But can the US power supply hack it?

Anthropic has joined the slew of AI firms investing billions in massive data centers throughout the US. On Wednesday, the company announced that it would invest $50 billion in American AI infrastructure, starting with data centers in Texas and New York, in partnership with Fluidstack.

Anthropic joins OpenAI, Nvidia, Oracle, Softbank and more in the race to develop these sites and evolve its AI models. But the power demands of these data centers may exceed the power grid’s capacity.

In a note published earlier this week, Morgan Stanley analysts warned that AI demand could leave the US with a “power shortfall totaling as much as 20%” for data centers through 2028, reaching a deficit of up to 13 gigawatts. 

Though tech leaders claim that the need for compute is the biggest problem facing the evolution of AI, energy supply and grid reliability present an even greater risk. The problem is that the building and deploying of these colossal server farms is far, far outpacing utility companies’ ability to upgrade the grid, Sebastian Lombardi, chair of the energy and utilities practice at law firm Day Pitney, told The Deep View.

While the problem is currently deepest felt in “pockets” of the US that have high concentrations of data centers, it’s only a matter of time before the stress on the grid and energy demand are felt all over the country, he said, possibly resulting in issues with reliability and affordability for utility payers. The rapid pace and magnitude of these buildouts are leaving utility companies and regulators scrambling to play catch-up.

“The AI data center story has complicated things. It’s created some questions about how we are going to maintain reliability,” said Lombardi. “The amount of energy that is expected to be used to power that infrastructure is quite significant.”

Despite the issue at hand, tech firms show no signs of tempering their all-out sprint. The solution might be what Lombardi calls an “all of the above” strategy for increasing the energy supply. This means a sharper focus on renewable energy, as well as more stakeholders buying into the “renaissance” that nuclear power is having. Even moonshot ideas, such as Google’s space-based Suncatcher data centers, could be worth exploring. “We all want the lights to stay on,” said Lombardi. “We may have to get away from picking winners and losers if we're going to meet the pace and the magnitude of demand.”

However, how much help these tech companies will be in solving the problem that they're effectively causing, especially as many struggle with their net-zero goals, is still unclear.

Today’s most popular AI models are great with words.

But when given tasks beyond letters and numbers, these models often fail to grasp the world around them. Conventional AI models tend to flounder when faced with real-world tasks, struggling to understand things like physics and causality. It’s why self-driving cars still struggle with edge cases, resulting in safety hazards and traffic law violations. It’s why industrial robots still need tons of training before they can be trusted to not break the things – or people – around them.

The problem is that these models can’t reconcile what they see with what’s actually real.

And from Abu Dhabi to Silicon Valley, a group of researchers from the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) is working to fix that. These researchers have their sights set on world models, or those that make decisions and act on the world around them.

“Our world model is designed to let AI understand and imagine how the world works — not just by seeing what’s happening, but by predicting what could happen next,” Hector Liu, Director at the Institute of Foundation Models (IFM), Silicon Valley Lab told The Deep View.

As it stands, tech firms are intent on using language to control AI – whether that be via chatbots, video and image generation, or agents. But conventional large language models lack what Stanford University researcher Dr. Fei-Fei Li calls “spatial intelligence,” or the ability to visualize in the way that humans do. These models are only good at predicting what to say or create based on their training data, and are unable to ground what they generate into reality.

This is the main divide between a world model and a video generation model, Liu said: One renders appearance, while the other simulates reality. 

Video generation tools like OpenAI’s Sora, Google’s Veo and xAI’s Grok Imagine can produce visually realistic scenes, but world models are designed to understand and simulate the world at large.

While a video generator creates a scene with no sense of state, a world model maintains an internal understanding of the world around it, and how that world evolves, said Liu.

“It predicts how scenes unfold over time and how they respond to actions or interventions, rather than just what they look like,” Liu said. Rather than just generating a scene, these models are interactive and reactive. If a tree falls in the world model, its virtual stump cracks, and the digital grass is flattened in its wake.

There are several companies currently in the running to create models that understand the world around them. Both Google DeepMind and Nvidia released new versions of their world models in August, for example. 

But MBZUAI’s PAN world model has several advantages over its competitors, said Liu.

• Rather than working only in narrow domains, MBZUAI’s PAN is trained for generality, said Liu, designed to transfer its knowledge across domains. It does so by combining language, vision and action data into one unified space, enabling broad simulation.
• The structure of PAN separates “reasoning from perception,” meaning seeing is distinct from thinking, said Liu. That separation provides the technical advantage of observability, preventing PAN from drifting away from real-world physics.

To measure how well PAN understands the world, MBZUAI researchers measure two main factors: long-horizon performance, or the ability to simulate a coherent world over time, and agentic usability. If something is wrong within a world model, the agent that’s working within it goes haywire.

The next step in the development of PAN is to make the model’s “imagination space,” or inner visualization capabilities, more rich and precise. This will allow the model to understand and render worlds in even finer detail. MBZUAI is also expanding beyond just vision understanding, researching modalities such as sound and motion signals, as well as using an agent to test and learn from different scenarios.

“That’s how we move from a model that only imagines the world to one that can actually think and act within it,” said Liu.

Though several developers want to build models that see the world for what it is, these systems are still in very early stages. 

Progress has been made on visual understanding, but humans have more than one sense. For a world model to be truly complete, developing a system with a strong understanding of audio, touch and physical interaction is crucial. The ideal world model not only understands all those modalities but can also create simulations in any of them. “If a modality is missing, the simulation will always be incomplete,” said Liu.

Creating an AI that can understand all of those modalities is to create a model that senses and understands almost entirely like a human does. But doing so comes with significant technical barriers, including access to substantial amounts of complex training data and potentially the need for entirely new model architecture.

But surpassing those barriers could have far-reaching implications, said Liu.

In robotics, these models can prevent the need for intense monitoring and training, limiting “real-world trial and error,” Liu said. Instead, the models that operate robots could be trained in simulations, perfecting actions and discovering mistakes before they even get onto factory floors or into homes. In self-driving cars, meanwhile, a world model could allow an autonomous driving model to rehearse thousands of traffic scenarios before the rubber hits the road.

And the possibilities extend beyond the self-piloted machines we have available today, with research being done in domains as sports strategy to simulate player outcomes, animation and digital art to design and create worlds, said Liu. More discoveries could emerge once these models are actually in the hands of the people.

“In the end, it’s about creating AI that doesn’t just react to the world but can think ahead.”

The AI industry’s budding relationship with the music sector has hit a snag.

A Munich court ruled that OpenAI violated German copyright laws after ChatGPT reproduced lyrics from popular German songs. The case, filed by German music rights group GEMA, claimed OpenAI trained ChatGPT on nine unlicensed tracks, including Herbert Grönemeyer’s “Männer” and “Bochum.”

As part of the ruling, OpenAI must pay an undisclosed fine. The company disputes the verdict, arguing ChatGPT’s lyrical reproductions stem from training on vast datasets—not individual songs—and that users are ultimately responsible for what’s generated through prompting.

Still, the decision underscores how European courts interpret AI’s production of lyrical outputs as copyright violations, cementing the EU’s strict stance on data privacy and IP protections.

The ruling diverges from the music industry’s recent embrace of AI amid ongoing debates over training data. Just last week, University Music Group, a major record label, forged a partnership with Stability AI to build AI-tools for music creation that will support the “creative and commercial success” of artists shortly after UMG’s settled a copyright lawsuit with AI music platform Udio.

That same week, performance rights organizations BMI, ASCAP and SOCAN revealed they’re accepting registrations for music that blends AI-generated music with human authorship. 

Music listeners seem increasingly open to songs touched by AI. In September, Xania Monet became the first AI-generated artist to land a multimillion-dollar record deal, with R&B hits like “Let Go, Let God” charting on Billboard airplay. A recent study even found that over half of listeners couldn’t distinguish AI-generated songs from human-made ones.

The question now isn’t whether AI will shape music, but whether it can withstand pushback from the very industry it hopes to transform.

Sam Altman is addressing the elephant in the room.

The OpenAI CEO on Thursday said in a post on X that the company is on track to earn more than $20 billion in annualized revenue run rate this year, with a path to bring in “hundreds of billion by 2030.”

The figure is a stark jump from the $13 billion in revenue that the company’s CFO Sarah Friar claimed in September. But it’s still a far cry from the $1.4 trillion in infrastructure deals with massive tech firms, which have raised concerns that the company won’t be able to cover its commitments.

Altman said that while “each doubling” of revenue is hard earned, the company is “feeling good about our prospects there,” noting that enterprise offerings, consumer devices and robotics could become promising revenue categories. Altman also alluded to directly selling compute capacity to companies and consumers.

As for the rapid pace and magnitude of the infrastructure buildout itself, Altman said the risks of not aggressively building out AI data centers are greater than having too much compute power.

“We are trying to build the infrastructure for a future economy powered by AI, and given everything we see on the horizon in our research program, this is the time to invest to be really scaling up our technology,” Altman said.

Addressing criticism that OpenAI had eyed a federal backstop for its infrastructure investments, Altman said in his post that “taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market.” 

Though Altman has full confidence that OpenAI will be a “wildly successful company,” achieving the kind of exponential growth that it’s banking on hinges not just on widespread adoption, but a large pool of users – both enterprise and consumer – willing to pay for OpenAI’s services.

And despite the fact that OpenAI now has deep financial ties with practically every major tech power player, “if we get it wrong, that’s on us,” Altman said. “If one company fails, other companies will do good work.”