MIT Technology Review
AI’s next era is moving far beyond chatbots, into agent teams, scientific discovery, robotics, warfare and systems that model the physical world.
10 things that matter in AI right now
Author: Will Douglas Heaven, Caiwei Chen, Grace Huckins, Eileen Guo, James O’Donnell, Rhiannon Williams, Michelle Kim
What is really worth your attention in the busy, buzzy world of AI? Our reporters and editors have spent years thinking about this question, charting AI’s progress and mapping out what’s next. Now, for the first time, we’ve distilled our answers into a single list.
From AI lab assistants and models that aim to shape the future of robotics to superscammers and algorithmically advised warfare, here are the big ideas, trends, and advances that are driving progress, shifting power dynamics, and shaping what’s possible tomorrow.
When OpenAI launched ChatGPT as an experimental prototype in late 2022, it caught fire and became an everyday everything app for hundreds of millions of people. The entire tech industry was consumed by the inferno, with companies racing to spin up rival products.
The ashes still haven’t settled, but that hasn’t stopped people from asking what’s next. Spoiler alert: It’s more LLMs. But better. Let’s call them LLMs+.
The challenge at hand is to get LLMs to work through complex and multipart problems that would take humans days or weeks to solve—and to do it by themselves for long periods of time. To get there, a few things need to happen.
First, LLMs must become more efficient and cheaper to run. One approach, called mixture-of-experts, splits an LLM up into smaller parts and gives each one expertise in a different type of task, so only some parts need to run at a given time. Another tactic could be to ditch transformers—the type of neural network underpinning almost all LLMs—in favor of diffusion models, which are more typically used for image and video generation. There are also more experimental approaches: Last year, the Chinese AI firm DeepSeek showed off a way to encode text in images, which cuts down on computation costs.
Another crucial area of progress has to do with what’s known as an LLM’s context window. This is the amount of text (or video) that a model can take in at once, equivalent to its working memory. A couple of years ago, LLMs could process several thousand tokens (words or parts of words) in one go, or a few dozen pages of text. The latest models now have context windows up to a million tokens long, which is like a whole stack of books.
But the bigger the context window and the longer the task, the more likely models are to go off the rails. There are breakthroughs happening there, too. One recent paper from MIT CSAIL introduced what the researchers call recursive LLMs. A recursive model breaks its input up into chunks and sends each chunk to a copy of itself, which in turn might break those chunks up again and send the results to even more copies. Multiple LLMs processing smaller pieces of information seem to be far more reliable for longer, harder tasks. The result is still LLMs, but not exactly as we’ ve come to know them.
—Will Douglas Heaven
Agent orchestration
The assembly line upended manufacturing last century. Could teams of agents do the same for white-collar work?
When people say AI will speed up drug development or fear that it will bring about mass layoffs, they’re thinking about agents—AI that can do stuff.
Now, after much hype, the first bona fide multi-agent tools are starting to show their colors.
OpenClaw, a personal AI assistant that you could talk to from your phone, got everyone’s attention. But beneath the buzz, OpenClaw had a limited set of tricks—and a saboteur’s approach to security. Still, it felt like the future. And so companies from Nvidia to Tencent built their own safer, more reliable bots on top of OpenClaw’s open-source code.
But the real power of agents comes when they can work as a team rather than as lone wolves carrying out singular tasks, such as using a browser to make a restaurant reservation. New tools can yoke together multiple agents, give each of them a job, and orchestrate their behaviors to complete tasks more complex than one agent could do alone.
For example, Claude Code, released by Anthropic last year, lets you launch and coordinate several coding agents working on different parts of the code base at the same time. Agents can also be given specific roles: One writes code, another tests it, a third fixes bugs, and so on. Such tools promise to turn coders into project managers, letting them delegate and oversee many more tasks.
But coding was just the start. The latest multi-agent tools are aimed at people who don’t need or want to develop software. Desktop apps such as Anthropic’s Claude Cowork (which the firm claims to have built using Claude Code in just 10 days), OpenAI’s Codex, and Perplexity’s Computer are all pitched as productivity tools for the white-collar set. Their agents can coordinate bespoke workflows across a wide range of tasks, from managing inboxes to handling customer complaints. Similarly, multi-agent tools like Google DeepMind’s Co-Scientist let researchers use teams of AI agents to coordinate literature searches, generate and test hypotheses, design experiments, and more.
In theory, networks of AI agents could do for white-collar knowledge work what Henry Ford–style assembly lines did for manufacturing in the early 20th century.
That’s the vision, at least. Because this technology also comes with huge risks. It’s no secret that LLMs can be unpredictable. That’s an annoyance when chatbots are stuck inside their screen, but when they start interacting more with our ubiquitous digital infrastructure—from health care to missile launchers—it could be disastrous.—WDH
China’s open-source bet
The country’s top AI labs are undercutting US competitors and winning over developers by making their best models free.
Silicon Valley AI companies have a playbook: Keep the secret sauce behind an API, and charge for every drop. China’s leading labs are playing a different game. They ship models as downloadable “open-weight” packages, so developers can adapt them without paying a US gatekeeper.
This strategy went mainstream after DeepSeek open-sourced its R1 model in January 2025. On raw capability, the gap between US and Chinese labs seemed to have suddenly narrowed. But China also won something stickier: goodwill.
China rode that momentum hard. A year after DeepSeek R1 debuted, other Chinese open-source giants are following the same blueprint and racing to release more capable models. Players include Z.ai (formerly Zhipu), Moonshot, Alibaba’s Qwen, and MiniMax. A study by MIT and Hugging Face found that Chinese open-weight models made up about 17% of global AI model downloads over the year ending in August 2025—surpassing the US share.
But the open-source ideal runs into hard realities. Chinese models avoid outputs that conflict with government policy. And Anthropic has accused several labs of using fraudulent means to extract capabilities from Claude through distillation, a process where outputs of one model train another.
Despite pushback from the West, much of the Global South is embracing Chinese models. Singapore’s government-backed program chose Qwen as the basis for its latest regional model, and Malaysia has announced that its sovereign AI ecosystem will run on DeepSeek.
While US CEOs believe models should stay proprietary, that doesn’t mean Chinese labs are purely idealistic. Open source is free advertising and also a shrewd workaround. Without access to cutting-edge chips restricted by US export controls, releasing models openly is a way to accelerate the cycle of external contributions and compensate for constrained compute. That naturally translates into API usage and revenue.
— Caiwei Chen
Artificial scientists
AI has already proved itself a valuable scientific tool. Could it take on a more central role in the research process?
AI companies frequently invoke the possibility of scientific discovery as a justification for their existence: If the technology eventually cures cancer and solves climate change, then all the emissions and slop will have been well worth it.
Already, LLMs can assist scientists in all sorts of ways. They can point people to relevant studies in the literature, draft journal articles, and, of course, write code. But AI companies and researchers alike have a much more ambitious vision for AI co-scientists. They want to develop systems that can act as a full member of a scientific team or, even more ambitiously, initiate and carry out projects with limited human guidance.
Google DeepMind has invested heavily in scientific AI for years, and it paid off in 2024 when Demis Hassabis and John Jumper, the company’s CEO and director, won the Nobel Prize in chemistry for AlphaFold, a specialized system that can predict the three-dimensional structure of a protein.
Now its competitors are working to catch up. In October 2025, OpenAI launched a team devoted to AI for science, and Anthropic announced several Claude features geared toward the biological sciences around the same time. OpenAI in particular has called building an autonomous researcher its “North Star”; Google released its own AI co-scientist tool in February 2025.
Under the hood, many of these AI-for-science systems are in fact multiple specialized AI agents working in concert. Google’s co-scientist system—which individual researchers can now apply to use—includes a supervisor agent, a generation agent, and a ranking agent, among several others; that coordination allows it to generate potential hypotheses and research plans in response to a goal provided by a human scientist. Similarly, researchers at Stanford’s AI for Science Lab devised a “virtual lab” made up of agents that took on the roles of specialists in different scientific fields. They found that their system could design new antibody fragments that bind to SARS-CoV-2, the virus that causes covid-19.
Unlike human scientists, however, those teams of agents can’t yet go out and test their ideas in the lab. To overcome that limitation, some researchers are plugging LLMs into experiment-running robots. In February, OpenAI announced that it had connected GPT-5 directly with automated biological laboratories built by the company Ginkgo Bioworks so that the AI system could iteratively propose experiments and interpret the results with limited human involvement. This approach allowed the system to run a gargantuan number of experiments and create a recipe that reduced the cost of synthesizing a particular protein by 40%.
AI-powered science seems like a win for frontier labs and for society at large. But research suggests it could have unintended consequences. A recent Nature study found that while individual scientists see professional advantages from adopting AI, science on the whole may suffer, because AI reduces the scope of what the community investigates. That might be because AI is especially good at analyzing preexisting data and literature, so scientists who use it gravitate toward established topics. Maintaining the vibrance and diversity of science in the AI era may require concerted effort from the scientific community. — Grace Huckins
Weaponized deepfakes
AI-generated imagery of people doing things they haven’t done in real life is increasingly being deployed in malicious ways.
For years, experts have warned that deepfakes— AI-generated videos, images, or audio of people doing or saying things they never did— could be used maliciously.
These dangers are now here. The widespread availability of easy-to-use and cheap (or free) generative models has made it simpler to fake reality in a way that’s difficult to spot.
Weaponized deepfakes may look startlingly real. And they are already inciting violence, trying to change minds, and generally sowing mistrust.
Experts worry that deepfakes will further crater critical thinking skills, as well as our trust in institutions and each other. And the impacts will weigh disproportionately on women and marginalized groups. A 2023 study found that 98% of deepfakes were pornographic in nature, and 99% depicted women. And one report estimated that 81% of the millions of sexualized images posted with Grok’s “edit image” function depicted women. xAI has since blocked the nudity feature in jurisdictions where it is illegal.
There’s also been an explosion of political deepfakes. The Trump administration, for example, has made and shared AI-generated images and videos, many seemingly to sway public opinion.
Suggested solutions include instituting new technical safeguards and detection methods, encouraging users to take more protective actions, and crafting new legislation or applying existing regulatory frameworks, like copyright law.
But these all have limits. Technical solutions can be bypassed. Getting people to change how they behave is unrealistic. Regulations require enforcement— and while President Trump has signed legislation that criminalizes deepfake porn, his administration continues to post other harmful deepfakes.
The problem could get much worse. There are midterm elections in the US in November, and the federal agencies that traditionally addressed election-related information integrity have been weakened. So have many research groups that fight election-related disinformation.—Eileen Guo
The new war room
Militaries already use AI to detect what humans might miss. Now they also want an advice engine for commanders to consult in battle.
To call the conflict in Iran the first “AI war” would be, in many ways, incorrect. Algorithms that scour surveillance footage for, say, trucks with mounted machine guns go back to the war in Afghanistan.
Here’s what is new: conversational AI systems that commanders turn to for both analysis and advice. These engines are built on large language models, and they’re reshaping how militaries work with Big Tech and make life-and-death decisions.
A decade ago, AI tools started to automate the work a junior intelligence analyst might do—for example, picking out a signal on a satellite feed. Systems like the US military’s Maven, built mainly on technology from the surveillance giant Palantir, fed that sort of analysis into tools commanders use to select targets.
Now LLMs are making these systems capable of more. One US defense official told MIT Technology Review that today’s personnel might give chatbots a list of potential targets to help decide which to strike first. And it’s not just the US; China is commissioning similar tools, according to Georgetown University’s Center for Security and Emerging Technology.
One problem with this will be obvious to anyone who has used generative AI: It can produce different outputs from the same prompt, and its recommendations are not always useful, precise, or correct.
Military experts also warn that officers who use systems that compress the world into a neat battlefield dashboard might trust the system too much, or that such systems might give tech firms undue influence over what information gets seen—all without meaningful public oversight.
The OpenAIs of the world, though, see an opportunity to win lucrative defense contracts. And the Pentagon even plans to allow companies to train new models on classified data. That would mean intelligence could go into the models themselves, presenting new security risks and bringing Silicon Valley closer to the Pentagon than ever before. — James O’Donnell
Humanoid data
Robotics companies want tremendous amounts of data on how we move our hands and limbs, and their tactics are getting strange.
An app that pays people to film themselves microwaving food. A website that recommends a game in which players remotely control a robotic arm in Shenzhen.
What on earth is happening? Well, just as our words became training data for large language models, robotics companies are betting that data about the way we move will help them build more capable humanoid robots. Though humanoids are trickier to train than robotic arms, companies see humanoids more easily slotting into the places where people work today—and someday replacing them entirely.
This new notion for how to train humanoids arguably began with the launch of ChatGPT in 2022. LLMs were able to generate text through exposure to massive amounts of training data. Roboticists wanted to apply these scaling laws to robotics but lacked an internet-size collection of data describing how we move.
Put off by how difficult this would be to amass, companies used workarounds, like teaching robots to move in simulations. However, simulations never perfectly model how things like friction or elasticity work in the real world, so the robots trained in them tended to (literally) stumble.
Now companies building humanoid robots have decided that collecting real-world data, as cumbersome as it is, could yield a massive payoff.
Early efforts were quaint and academic. Labs collected hours of data from people doing household tasks, like making waffles or cleaning their desks, while wearing cameras or handheld grippers. The data was shared openly. But as venture capital money poured into robotics—$6.1 billion in 2025 for humanoids alone—the race to create this training data has gotten more competitive, and more elaborate.
There are now training centers in China where people wear exoskeletons and virtual-reality hardware while they do the same repetitive task, like wiping a table, hundreds of times per day. Gig workers in Nigeria, Argentina, and India are filming themselves doing chores at home.
All this points to a future of work in which physical laborers increasingly become data collectors. But training robots on movement is complicated. It’s not clear that it’s even possible to do it at the scale potentially needed to yield breakthroughs, let alone build a profitable business.
How many thousands of clips of someone opening a microwave would it take to teach a robot to cook dinner? Perhaps we’re about to find out. — JO
World models
Today’s AI is still unreliable. Some researchers think solving that problem requires teaching AI systems to understand the world around them.
AI systems already have impressive mastery over the digital world, but the physical world is still humanity’s domain. To get to an AI system that can fold laundry or navigate a city, you need something called a world model.
World models are not a new idea, but recent innovations from Google DeepMind and Stanford professor Fei-Fei Li’s World Labs— among other developments— have brought them to the forefront. Proponents argue that world models will allow researchers to overcome the well-known limitations of LLMs and better realize AI’s promise for robotics.
Definitions of the term “world model” vary, but they all center on the ways in which systems represent the external world. Research suggests that LLMs might not have reliable and robust world models: One study found that models trained to provide driving directions between different points in Manhattan will fail miserably if forced to take occasional detours. An AI system that could draw on a detailed, map-like representation of the city streets would probably do much better.
Many researchers think that world models will prove essential to robotics. They could, for instance, facilitate the development of robots that explore the deep sea and assist health-care providers. But for now, the applications are more modest; for instance, billions of Pokémon Go images are being used to build the first pieces of a model that, developers hope, could help guide delivery robots.
Google DeepMind and World Labs are currently focusing their efforts on building models that can generate interactive, 3D virtual environments from text, images, and (in the case of World Labs) video prompts. Such tools could be used to streamline the design of video games and VR experiences. The real breakthroughs are likely to come from integrating such systems into flexible, intelligent agents that can represent their environments, predict the consequences of their actions, and then decide what to do.—GH
Supercharged scams
AI tools are making it easier than ever for online criminals to trick people and steal money and valuable confidential data.
ChatGPT opened people’s eyes to how easily generative AI could churn out vast amounts of human-seeming text. This quickly caught the attention of criminals, who began using LLMs to produce malicious emails—both spam and targeted attacks designed to steal money and sensitive information.
Since then, cybercriminals have adopted AI tools to supercharge their operations. They’ve done everything from composing phishing emails and creating deepfake clips to making malicious software (a.k.a. malware) harder to detect. They can also use AI to automate the search for vulnerabilities in networks and computer systems, quickly generate ransom notes, and analyze vast swathes of stolen data to pinpoint what’s most valuable.
AI is lowering the barriers for would-be attackers, providing them with an arsenal of capabilities, and making it faster, cheaper, and easier for them to try to infiltrate targets. For example, scam centers across Southeast Asia are embracing inexpensive AI tools to target greater numbers of potential victims and switch to new locations, Interpol has warned. Similarly, the United Arab Emirates recently claimed it foiled a series of attacks on its vital sectors. And because these scattergun attacks can be pumped out at a colossal scale, they don’t need to be very sophisticated.
Many organizations are struggling to cope with the sheer volume. The problem is likely to get much worse as more criminals try their luck, and as the capabilities of publicly available generative AI systems improve. This spring, the AI company Anthropic claimed that its Mythos model had found thousands of critical vulnerabilities, including some in every major operating system and web browser. Anthropic delayed the model’s initial release and set up a consortium of tech companies called Project Glasswing, aiming to put these capabilities to work for defensive purposes.
Right now, cybersecurity researchers are optimistic that sloppier attacks can be thwarted through basic defenses, highlighting just how important it is to keep on top of software updates and stick to network security protocols. How well positioned we’ll be to ward off more sophisticated attacks in the future, however, is much less clear.
The good news is that AI is also being used to defend. Just one example: Each day, Microsoft processes more than 100 trillion signals its AI systems flag as potentially malicious or suspicious. The company says that between April 2024 and April 2025, it blocked $4 billion in scams and fraudulent transactions, many of which may have been aided by AI content.—Rhiannon Williams
Resistance
A populist backlash is building against AI.
Turns out not everyone wants to live in the future that AI companies are building. People from all walks of life are speaking out against rising electricity bills from data centers, disappearing jobs, chatbots’ impact on teen mental health, the military’s use of AI, and copyright infringement—among other concerns.
This anti-AI movement is taking shape around the world. In February, hundreds of people marched past the London headquarters of OpenAI, Google DeepMind, and Meta in one of the largest protests against AI to date. And in the US in March, an unlikely coalition of MAGA Republicans, democratic socialists, labor activists, and church leaders signed a Pro-Human AI Declaration, articulating the principle that AI should serve humanity, not replace it.
In March, the biggest flash point was the US military’s use of the technology. In the wake of OpenAI’s deal with the Pentagon earlier this year, users uninstalled ChatGPT in droves, while protesters chalked messages such as “What are the safeguards?” around OpenAI’s headquarters in San Francisco. In April, a Texas man allegedly threw a Molotov cocktail at OpenAI CEO Sam Altman’s home in San Francisco and was found carrying an anti-AI diatribe.
The backlash reflects deep anxieties. Last year, a Pew poll found that half of Americans are concerned about the increased use of AI in daily life, with many believing it will erode people’s ability to think creatively and form meaningful relationships. Another survey found that three-quarters of Americans worry AI could pose a threat to humanity.
People have practical concerns, too. College graduates are having a harder time finding jobs. And a survey late last year indicated that even though AI is not yet generating substantial economic value, employers are preemptively laying off workers. (Some argue that AI is just a convenient excuse for cost-cutting.) Employees are protesting these kinds of layoffs while labor unions mobilize for better worker protections.
Parents are also sounding the alarm. Lawsuits alleging that chatbots drove teens to suicide or self-harm are piling up. In some cities, parents are signing petitions to demand a two-year moratorium on AI in schools.
Some of the pushback is shaping policy. In New York and California, new rules have put safeguards on AI companionship bots. Meanwhile, artists are winning small battles to protect copyright laws. In March, after fierce blowback from artists, the UK government backtracked on plans to let AI companies train their models on copyrighted content without permission.
But some of the sharpest resistance is coming from communities where data centers are built, fueled by concerns that these facilities are driving up utility bills, creating pollution, and consuming rural land.
In the US, activists stalled more than $98 billion in data-center development in the second quarter of 2025 alone. In response, President Trump secured a pledge from AI company executives in March to cover the energy costs related to their data centers by building or buying from new power plants.
Credits: TCA, LLC.