What’s next for Chinese open-source AI

MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.

The past year has marked a turning point for Chinese AI. Since DeepSeek released its R1 reasoning model in January 2025, Chinese companies have repeatedly delivered AI models that match the performance of leading Western models at a fraction of the cost. 

Just last week the Chinese firm Moonshot AI released its latest open-weight model, Kimi K2.5, which came close to top proprietary systems such as Anthropic’s Claude Opus on some early benchmarks. The difference: K2.5 is roughly one-seventh Opus’s price.

On Hugging Face, Alibaba’s Qwen family—after ranking as the most downloaded model series in 2025 and 2026—has overtaken Meta’s Llama models in cumulative downloads. And a recent MIT study found that Chinese open-source models have surpassed US models in total downloads. For developers and builders worldwide, access to near-frontier AI capabilities has never been this broad or this affordable.

But these models differ in a crucial way from most US models like ChatGPT or Claude, which you pay to access and can’t inspect. The Chinese companies publish their models’ weights—numerical values that get set when a model is trained—so anyone can download, run, study, and modify them. 

If these open-source AI models keep getting better, they will not just offer the cheapest options for people who want access to frontier AI capabilities; they will change where innovation happens and who sets the standards. 

Here’s what may come next.

China’s commitment to open source will continue

When DeepSeek launched R1, much of the initial shock centered on its origin. Suddenly, a Chinese team had released a reasoning model that could stand alongside the best systems from US labs. But the long tail of DeepSeek’s impact had less to do with nationality than with distribution. R1 was released as an open-weight model under a permissive MIT license, allowing anyone to download, inspect, and deploy it. On top of that, DeepSeek also published a paper detailing its training process and techniques. For developers who access models via an API, DeepSeek also undercut competitors on price, offering access at a fraction the cost of OpenAI’s o1, the leading proprietary reasoning model at the time.

Within days of its release, DeepSeek replaced ChatGPT as the most downloaded free app in the US App Store. The moment spilled beyond developer circles into financial markets, triggering a sharp sell-off in US tech stocks that briefly erased roughly $1 trillion in market value. Almost overnight, DeepSeek went from a little-known spin-off team backed by a quantitative hedge fund to the most visible symbol of China’s push for open-source AI.

China’s decision to lean in to open source isn’t surprising. It has the world’s second-largest concentration of AI talent after the US. plus a vast, well-resourced tech industry. After ChatGPT broke into the mainstream, China’s AI sector went through a reckoning—and emerged determined to catch up. Pursuing an open-source strategy was seen as the fastest way to close the gap by rallying developers, spreading adoption, and setting standards.

DeepSeek’s success injected confidence into an industry long used to following global standards rather than setting them. “Thirty years ago, no Chinese person would believe they could be at the center of global innovation,” says Alex Chenglin Wu, CEO and founder of Atoms, an AI agent company and prominent contributor to China’s open-source ecosystem. “DeepSeek shows that with solid technical talent, a supportive environment, and the right organizational culture, it’s possible to do truly world-class work.”

DeepSeek’s breakout moment wasn’t China’s first open-source success. Alibaba’s Qwen Lab had been releasing open-weight models for years. By September 2024,  well before DeepSeek’s V3 launch, Alibaba was saying that global downloads had exceeded 600 million. On Hugging Face, Qwen accounted for more than 30% of all model downloads in 2024. Other institutions, including the Beijing Academy of Artificial Intelligence and the AI firm Baichuan, were also releasing open models as early as 2023. 

But since the success of DeepSeek, the field has widened rapidly. Companies such as Z.ai (formerly Zhipu), MiniMax, Tencent, and a growing number of smaller labs have released models that are competitive on reasoning, coding, and agent-style tasks. The growing number of capable models has sped up progress. Capabilities that once took months to make it to the open-source world now emerge within weeks, even days.

“Chinese AI firms have seen real gains from the open-source playbook,” says Liu Zhiyuan, a professor of computer science at Tsinghua University and chief scientist at the AI startup ModelBest. “By releasing strong research, they build reputation and gain free publicity.”

Beyond commercial incentives, Liu says, open source has taken on cultural and strategic weight. “In the Chinese programmer community, open source has become politically correct,” he says, framing it as a response to US dominance in proprietary AI systems.

That shift is also reflected at the institutional level. Universities including Tsinghua have begun encouraging AI development and open-source contributions, while policymakers have moved to formalize those incentives. In August, China’s State Council released a draft policy encouraging universities to reward open-source work, proposing that students’ contributions on platforms such as GitHub or Gitee could eventually be counted toward academic credit.

With growing momentum and a reinforcing feedback loop, China’s push for open-source models is likely to continue in the near term, though its long-term sustainability still hinges on financial results, says Tiezhen Wang, who helps lead work on global AI at Hugging Face. In January, the model labs Z.ai and MiniMax went public in Hong Kong. “Right now, the focus is on making the cake bigger,” says Wang. “The next challenge is figuring out how each company secures its share.”

The next wave of models will be narrower—and better

Chinese open-source models are leading not just in download volume but also in variety. Alibaba’s Qwen has become one of the most diversified open model families in circulation, offering a wide range of variants optimized for different uses. The lineup ranges from lightweight models that can run on a single laptop to large, multi-hundred-billion-parameter systems designed for data-center deployment. Qwen features many task-optimized variants created by the community: the “instruct” models are good at following orders, and “code” variants specialize in coding.

Although this strategy isn’t unique to Chinese labs, Qwen was the first open model family to roll out so many high-quality options that it started to feel like a full product line—one that’s free to use.

The open-weight nature of these releases also makes it easy for others to adapt them through techniques like fine-tuning and distillation, which means training a smaller model to mimic a larger one.  According to ATOM (American Truly Open Models), a project by the AI researcher Nathan Lambert, by August 4, 2025, model variations derived from Qwen were “more than 40%” of new Hugging Face language-model derivatives, while Llama had fallen to about 15%. This means that Qwen has become the default base model for all the “remixes.”

This pattern has made the case for smaller, more specialized models. “Compute and energy are real constraints for any deployment,” Liu says. He told MIT Technology Review that the rise of small models is about making AI cheaper to run and easier for more people to use. His company, ModelBest, focuses on small language models designed to run locally on devices such as phones, cars, and other consumer hardware.

While an average user might interact with AI only through the web or an app for simple conversations, power users of AI models with some technical background are experimenting with giving AI more autonomy to solve large-scale problems. OpenClaw, an open-source AI agent that recently went viral within the AI hacker world, allows AI to take over your computer—it can run 24-7, going through your emails and work tasks without supervision. 

OpenClaw, like many other open-source tools, allows users to connect to different AI models via an application programming interface, or API. Within days of OpenClaw’s release, the team revealed that Kimi’s K2.5 had surpassed Claude Opus and became the most used AI model—by token count, meaning it was handling more total text processed across user prompts and model responses.

Cost has been a major reason Chinese models have gained traction, but it would be a mistake to treat them as mere “dupes” of Western frontier systems, Wang suggests. Like any product, a model only needs to be good enough for the job at hand. 

The landscape of open-source models in China is also getting more specialized. Research groups such as Shanghai AI Laboratory have released models geared toward scientific and technical tasks; several projects from Tencent have focused specifically on music generation. Ubiquant, a quantitative finance firm like DeepSeek’s parent High-Flyer, has released an open model aimed at medical reasoning.

In the meantime, innovative architectural ideas from Chinese labs are being picked up more broadly. DeepSeek has published work exploring model efficiency and memory; techniques that compress the model’s attention “cache,” reducing memory and inference costs while mostly preserving performance, have drawn significant attention in the research community. 

“The impact of these research breakthroughs is amplified because they’re open-sourced and can be picked up quickly across the field,” says Wang.

Chinese open models will become infrastructure for global AI builders

The adoption of Chinese models is picking up in Silicon Valley, too. Martin Casado, a general partner at Andreessen Horowitz, has put a number on it: Among startups pitching with open-source stacks, there’s about an 80% chance they’re running on Chinese open models, according to a post he made on X. Usage data tells a similar story. OpenRouter,  a middleman that tracks how people use different AI models through its API, shows Chinese open models rising from almost none in late 2024 to nearly 30% of usage in some recent weeks.

The demand is also rising globally. Z.ai limited new subscriptions to its GLM coding plan (a coding tool based on its flagship GLM models) after demand surged, citing compute constraints. What’s notable is where the demand is coming from: CNBC reports that the system’s user base is primarily concentrated in the United States and China, followed by India, Japan, Brazil, and the UK.

“The open-source ecosystems in China and the US are tightly bound together,” says Wang at Hugging Face. Many Chinese open models still rely on Nvidia and US cloud platforms to train and serve them, which keeps the business ties tangled. Talent is fluid too: Researchers move across borders and companies, and many still operate as a global community, sharing code and ideas in public.

That interdependence is part of what makes Chinese developers feel optimistic about this moment: The work travels, gets remixed, and actually shows up in products. But openness can also accelerate the competition. Dario Amodei, the CEO of Anthropic, made a version of this point after DeepSeek’s 2025 releases: He wrote that export controls are “not a way to duck the competition” between the US and China, and that AI companies in the US “must have better models” if they want to prevail. 

For the past decade, the story of Chinese tech in the West has been one of big expectations that ran into scrutiny, restrictions, and political backlash. This time the export isn’t just an app or a consumer platform. It’s the underlying model layer that other people build on. Whether that will play out differently is still an open question.

America’s coming war over AI regulation

MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.

In the final weeks of 2025, the battle over regulating artificial intelligence in the US reached a boiling point. On December 11, after Congress failed twice to pass a law banning state AI laws, President Donald Trump signed a sweeping executive order seeking to handcuff states from regulating the booming industry. Instead, he vowed to work with Congress to establish a “minimally burdensome” national AI policy, one that would position the US to win the global AI race. The move marked a qualified victory for tech titans, who have been marshaling multimillion-dollar war chests to oppose AI regulations, arguing that a patchwork of state laws would stifle innovation.

In 2026, the battleground will shift to the courts. While some states might back down from passing AI laws, others will charge ahead, buoyed by mounting public pressure to protect children from chatbots and rein in power-hungry data centers. Meanwhile, dueling super PACs bankrolled by tech moguls and AI-safety advocates will pour tens of millions into congressional and state elections to seat lawmakers who champion their competing visions for AI regulation. 

Trump’s executive order directs the Department of Justice to establish a task force that sues states whose AI laws clash with his vision for light-touch regulation. It also directs the Department of Commerce to starve states of federal broadband funding if their AI laws are “onerous.” In practice, the order may target a handful of laws in Democratic states, says James Grimmelmann, a law professor at Cornell Law School. “The executive order will be used to challenge a smaller number of provisions, mostly relating to transparency and bias in AI, which tend to be more liberal issues,” Grimmelmann says.

For now, many states aren’t flinching. On December 19, New York’s governor, Kathy Hochul, signed the Responsible AI Safety and Education (RAISE) Act, a landmark law requiring AI companies to publish the protocols used to ensure the safe development of their AI models and report critical safety incidents. On January 1, California debuted the nation’s first frontier AI safety law, SB 53—which the RAISE Act was modeled on—aimed at preventing catastrophic harms such as biological weapons or cyberattacks. While both laws were watered down from earlier iterations to survive bruising industry lobbying, they struck a rare, if fragile, compromise between tech giants and AI safety advocates.

If Trump targets these hard-won laws, Democratic states like California and New York will likely take the fight to court. Republican states like Florida with vocal champions for AI regulation might follow suit. Trump could face an uphill battle. “The Trump administration is stretching itself thin with some of its attempts to effectively preempt [legislation] via executive action,” says Margot Kaminski, a law professor at the University of Colorado Law School. “It’s on thin ice.”

But Republican states that are anxious to stay off Trump’s radar or can’t afford to lose federal broadband funding for their sprawling rural communities might retreat from passing or enforcing AI laws. Win or lose in court, the chaos and uncertainty could chill state lawmaking. Paradoxically, the Democratic states that Trump wants to rein in—armed with big budgets and emboldened by the optics of battling the administration—may be the least likely to budge.

In lieu of state laws, Trump promises to create a federal AI policy with Congress. But the gridlocked and polarized body won’t be delivering a bill this year. In July, the Senate killed a moratorium on state AI laws that had been inserted into a tax bill, and in November, the House scrapped an encore attempt in a defense bill. In fact, Trump’s bid to strong-arm Congress with an executive order may sour any appetite for a bipartisan deal. 

The executive order “has made it harder to pass responsible AI policy by hardening a lot of positions, making it a much more partisan issue,” says Brad Carson, a former Democratic congressman from Oklahoma who is building a network of super PACs backing candidates who support AI regulation. “It hardened Democrats and created incredible fault lines among Republicans,” he says. 

While AI accelerationists in Trump’s orbit—AI and crypto czar David Sacks among them—champion deregulation, populist MAGA firebrands like Steve Bannon warn of rogue superintelligence and mass unemployment. In response to Trump’s executive order, Republican state attorneys general signed a bipartisan letter urging the FCC not to supersede state AI laws.

With Americans increasingly anxious about how AI could harm mental health, jobs, and the environment, public demand for regulation is growing. If Congress stays paralyzed, states will be the only ones acting to keep the AI industry in check. In 2025, state legislators introduced more than 1,000 AI bills, and nearly 40 states enacted over 100 laws, according to the National Conference of State Legislatures.

Efforts to protect children from chatbots may inspire rare consensus. On January 7, Google and Character Technologies, a startup behind the companion chatbot Character.AI, settled several lawsuits with families of teenagers who killed themselves after interacting with the bot. Just a day later, the Kentucky attorney general sued Character Technologies, alleging that the chatbots drove children to suicide and other forms of self-harm. OpenAI and Meta face a barrage of similar suits. Expect more to pile up this year. Without AI laws on the books, it remains to be seen how product liability laws and free speech doctrines apply to these novel dangers. “It’s an open question what the courts will do,” says Grimmelmann. 

While litigation brews, states will move to pass child safety laws, which are exempt from Trump’s proposed ban on state AI laws. On January 9, OpenAI inked a deal with a former foe, the child-safety advocacy group Common Sense Media, to back a ballot initiative in California called the Parents & Kids Safe AI Act, setting guardrails around how chatbots interact with children. The measure proposes requiring AI companies to verify users’ age, offer parental controls, and undergo independent child-safety audits. If passed, it could be a blueprint for states across the country seeking to crack down on chatbots. 

Fueled by widespread backlash against data centers, states will also try to regulate the resources needed to run AI. That means bills requiring data centers to report on their power and water use and foot their own electricity bills. If AI starts to displace jobs at scale, labor groups might float AI bans in specific professions. A few states concerned about the catastrophic risks posed by AI may pass safety bills mirroring SB 53 and the RAISE Act. 

Meanwhile, tech titans will continue to use their deep pockets to crush AI regulations. Leading the Future, a super PAC backed by OpenAI president Greg Brockman and the venture capital firm Andreessen Horowitz, will try to elect candidates who endorse unfettered AI development to Congress and state legislatures. They’ll follow the crypto industry’s playbook for electing allies and writing the rules. To counter this, super PACs funded by Public First, an organization run by Carson and former Republican congressman Chris Stewart of Utah, will back candidates advocating for AI regulation. We might even see a handful of candidates running on anti-AI populist platforms.

In 2026, the slow, messy process of American democracy will grind on. And the rules written in state capitals could decide how the most disruptive technology of our generation develops far beyond America’s borders, for years to come.

What’s next for AI in 2026

MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.

In an industry in constant flux, sticking your neck out to predict what’s coming next may seem reckless. (AI bubble? What AI bubble?) But for the last few years we’ve done just that—and we’re doing it again. 

How did we do last time? We picked five hot AI trends to look out for in 2025, including what we called generative virtual playgrounds, a.k.a world models (check: From Google DeepMind’s Genie 3 to World Labs’s Marble, tech that can generate realistic virtual environments on the fly keeps getting better and better); so-called reasoning models (check: Need we say more? Reasoning models have fast become the new paradigm for best-in-class problem solving); a boom in AI for science (check: OpenAI is now following Google DeepMind by setting up a dedicated team to focus on just that); AI companies that are cozier with national security (check: OpenAI reversed position on the use of its technology for warfare to sign a deal with the defense-tech startup Anduril to help it take down battlefield drones); and legitimate competition for Nvidia (check, kind of: China is going all in on developing advanced AI chips, but Nvidia’s dominance still looks unassailable—for now at least). 

So what’s coming in 2026? Here are our big bets for the next 12 months. 

More Silicon Valley products will be built on Chinese LLMs

The last year shaped up as a big one for Chinese open-source models. In January, DeepSeek released R1, its open-source reasoning model, and shocked the world with what a relatively small firm in China could do with limited resources. By the end of the year, “DeepSeek moment” had become a phrase frequently tossed around by AI entrepreneurs, observers, and builders—an aspirational benchmark of sorts. 

It was the first time many people realized they could get a taste of top-tier AI performance without going through OpenAI, Anthropic, or Google.

Open-weight models like R1 allow anyone to download a model and run it on their own hardware. They are also more customizable, letting teams tweak models through techniques like distillation and pruning. This stands in stark contrast to the “closed” models released by major American firms, where core capabilities remain proprietary and access is often expensive.

As a result, Chinese models have become an easy choice. Reports by CNBC and Bloomberg suggest that startups in the US have increasingly recognized and embraced what they can offer.

One popular group of models is Qwen, created by Alibaba, the company behind China’s largest e-commerce platform, Taobao. Qwen2.5-1.5B-Instruct alone has 8.85 million downloads, making it one of the most widely used pretrained LLMs. The Qwen family spans a wide range of model sizes alongside specialized versions tuned for math, coding, vision, and instruction-following, a breadth that has helped it become an open-source powerhouse.

Other Chinese AI firms that were previously unsure about committing to open source are following DeepSeek’s playbook. Standouts include Zhipu’s GLM and Moonshot’s Kimi. The competition has also pushed American firms to open up, at least in part. In August, OpenAI released its first open-source model. In November, the Allen Institute for AI, a Seattle-based nonprofit, released its latest open-source model, Olmo 3. 

Even amid growing US-China antagonism, Chinese AI firms’ near-unanimous embrace of open source has earned them goodwill in the global AI community and a long-term trust advantage. In 2026, expect more Silicon Valley apps to quietly ship on top of Chinese open models, and look for the lag between Chinese releases and the Western frontier to keep shrinking—from months to weeks, and sometimes less.

—Caiwei Chen

The US will face another year of regulatory tug-of-war

T​​he battle over regulating artificial intelligence is heading for a showdown. On December 11, President Donald Trump signed an executive order aiming to neuter state AI laws, a move meant to handcuff states from keeping the growing industry in check. In 2026, expect more political warfare. The White House and states will spar over who gets to govern the booming technology, while AI companies wage a fierce lobbying campaign to crush regulations, armed with the narrative that a patchwork of state laws will smother innovation and hobble the US in the AI arms race against China.

Under Trump’s executive order, states may fear being sued or starved federal funding if they clash with his vision for light-touch regulation. Big Democratic states like California—which just enacted the nation’s first frontier AI law requiring companies to publish safety testing for their AI models—will take the fight to court, arguing that only Congress can override state laws. But states that can’t afford to lose federal funding, or fear getting in Trump’s crosshairs, might fold. Still, expect to see more state lawmaking on hot-button issues, especially where Trump’s order gives states a green light to legislate. With chatbots accused of triggering teen suicides and data centers sucking up more and more energy, states will face mounting public pressure to push for guardrails. 

In place of state laws, Trump promises to work with Congress to establish a federal AI law. Don’t count on it. Congress failed to pass a moratorium on state legislation twice in 2025, and we aren’t holding out hope that it will deliver its own bill this year. 

AI companies like OpenAI and Meta will continue to deploy powerful super-PACs to support political candidates who back their agenda and target those who stand in their way. On the other side, super-PACs supporting AI regulation will build their own war chests to counter. Watch them duke it out at next year’s midterm elections.

The further AI advances, the more people will fight to steer its course, and 2026 will be another year of regulatory tug-of-war—with no end in sight.

—Michelle Kim

Chatbots will change the way we shop

Imagine a world in which you have a personal shopper at your disposal 24-7—an expert who can instantly recommend a gift for even the trickiest-to-buy-for friend or relative, or trawl the web to draw up a list of the best bookcases available within your tight budget. Better yet, they can analyze a kitchen appliance’s strengths and weaknesses, compare it with its seemingly identical competition, and find you the best deal. Then once you’re happy with their suggestion, they’ll take care of the purchasing and delivery details too.

But this ultra-knowledgeable shopper isn’t a clued-up human at all—it’s a chatbot. This is no distant prediction, either. Salesforce recently said it anticipates that AI will drive $263 billion in online purchases this holiday season. That’s some 21% of all orders. And experts are betting on AI-enhanced shopping becoming even bigger business within the next few years. By 2030, between $3 trillion and $5 trillion annually will be made from agentic commerce, according to research from the consulting firm McKinsey. 

Unsurprisingly, AI companies are already heavily invested in making purchasing through their platforms as frictionless as possible. Google’s Gemini app can now tap into the company’s powerful Shopping Graph data set of products and sellers, and can even use its agentic technology to call stores on your behalf. Meanwhile, back in November, OpenAI announced a ChatGPT shopping feature capable of rapidly compiling buyer’s guides, and the company has struck deals with Walmart, Target, and Etsy to allow shoppers to buy products directly within chatbot interactions. 

Expect plenty more of these kinds of deals to be struck within the next year as consumer time spent chatting with AI keeps on rising, and web traffic from search engines and social media continues to plummet. 

—Rhiannon Williams

An LLM will make an important new discovery

I’m going to hedge here, right out of the gate. It’s no secret that large language models spit out a lot of nonsense. Unless it’s with monkeys-and-typewriters luck, LLMs won’t discover anything by themselves. But LLMs do still have the potential to extend the bounds of human knowledge.

We got a glimpse of how this could work in May, when Google DeepMind revealed AlphaEvolve, a system that used the firm’s Gemini LLM to come up with new algorithms for solving unsolved problems. The breakthrough was to combine Gemini with an evolutionary algorithm that checked its suggestions, picked the best ones, and fed them back into the LLM to make them even better.

Google DeepMind used AlphaEvolve to come up with more efficient ways to manage power consumption by data centers and Google’s TPU chips. Those discoveries are significant but not game-changing. Yet. Researchers at Google DeepMind are now pushing their approach to see how far it will go.

And others have been quick to follow their lead. A week after AlphaEvolve came out, Asankhaya Sharma, an AI engineer in Singapore, shared OpenEvolve, an open-source version of Google DeepMind’s tool. In September, the Japanese firm Sakana AI released a version of the software called SinkaEvolve. And in November, a team of US and Chinese researchers revealed AlphaResearch, which they claim improves on one of AlphaEvolve’s already better-than-human math solutions.

There are alternative approaches too. For example, researchers at the University of Colorado Denver are trying to make LLMs more inventive by tweaking the way so-called reasoning models work. They have drawn on what cognitive scientists know about creative thinking in humans to push reasoning models toward solutions that are more outside the box than their typical safe-bet suggestions.

Hundreds of companies are spending billions of dollars looking for ways to get AI to crack unsolved math problems, speed up computers, and come up with new drugs and materials. Now that AlphaEvolve has shown what’s possible with LLMs, expect activity on this front to ramp up fast.    

—Will Douglas Heaven

Legal fights heat up

For a while, lawsuits against AI companies were pretty predictable: Rights holders like authors or musicians would sue companies that trained AI models on their work, and the courts generally found in favor of the tech giants. AI’s upcoming legal battles will be far messier.

The fights center on thorny, unresolved questions: Can AI companies be held liable for what their chatbots encourage people to do, as when they help teens plan suicides? If a chatbot spreads patently false information about you, can its creator be sued for defamation? If companies lose these cases, will insurers shun AI companies as clients?

In 2026, we’ll start to see the answers to these questions, in part because some notable cases will go to trial (the family of a teen who died by suicide will bring OpenAI to court in November).

At the same time, the legal landscape will be further complicated by President Trump’s executive order from December—see Michelle’s item above for more details on the brewing regulatory storm.

No matter what, we’ll see a dizzying array of lawsuits in all directions (not to mention some judges even turning to AI amid the deluge).

—James O’Donnell

What’s next for AI and math

MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.

The way DARPA tells it, math is stuck in the past. In April, the US Defense Advanced Research Projects Agency kicked off a new initiative called expMath—short for Exponentiating Mathematics—that it hopes will speed up the rate of progress in a field of research that underpins a wide range of crucial real-world applications, from computer science to medicine to national security.

“Math is the source of huge impact, but it’s done more or less as it’s been done for centuries—by people standing at chalkboards,” DARPA program manager Patrick Shafto said in a video introducing the initiative. 

The modern world is built on mathematics. Math lets us model complex systems such as the way air flows around an aircraft, the way financial markets fluctuate, and the way blood flows through the heart. And breakthroughs in advanced mathematics can unlock new technologies such as cryptography, which is essential for private messaging and online banking, and data compression, which lets us shoot images and video across the internet.

But advances in math can be years in the making. DARPA wants to speed things up. The goal for expMath is to encourage mathematicians and artificial-intelligence researchers to develop what DARPA calls an AI coauthor, a tool that might break large, complex math problems into smaller, simpler ones that are easier to grasp and—so the thinking goes—quicker to solve.

Mathematicians have used computers for decades, to speed up calculations or check whether certain mathematical statements are true. The new vision is that AI might help them crack problems that were previously uncrackable.  

But there’s a huge difference between AI that can solve the kinds of problems set in high school—math that the latest generation of models has already mastered—and AI that could (in theory) solve the kinds of problems that professional mathematicians spend careers chipping away at.

On one side are tools that might be able to automate certain tasks that math grads are employed to do; on the other are tools that might be able to push human knowledge beyond its existing limits.

Here are three ways to think about that gulf.

1/ AI needs more than just clever tricks

Large language models are not known to be good at math. They make things up and can be persuaded that 2 + 2 = 5. But newer versions of this tech, especially so-called large reasoning models (LRMs) like OpenAI’s o3 and Anthropic’s Claude 4 Thinking, are far more capable—and that’s got mathematicians excited.

This year, a number of LRMs, which try to solve a problem step by step rather than spit out the first result that comes to them, have achieved high scores on the American Invitational Mathematics Examination (AIME), a test given to the top 5% of US high school math students.

At the same time, a handful of new hybrid models that combine LLMs with some kind of fact-checking system have also made breakthroughs. Emily de Oliveira Santos, a mathematician at the University of São Paulo, Brazil, points to Google DeepMind’s AlphaProof, a system that combines an LLM with DeepMind’s game-playing model AlphaZero, as one key milestone. Last year AlphaProof became the first computer program to match the performance of a silver medallist at the International Math Olympiad, one of the most prestigious mathematics competitions in the world.

And in May, a Google DeepMind model called AlphaEvolve discovered better results than anything humans had yet come up with for more than 50 unsolved mathematics puzzles and several real-world computer science problems.

The uptick in progress is clear. “GPT-4 couldn’t do math much beyond undergraduate level,” says de Oliveira Santos. “I remember testing it at the time of its release with a problem in topology, and it just couldn’t write more than a few lines without getting completely lost.” But when she gave the same problem to OpenAI’s o1, an LRM released in January, it nailed it.

Does this mean such models are all set to become the kind of coauthor DARPA hopes for? Not necessarily, she says: “Math Olympiad problems often involve being able to carry out clever tricks, whereas research problems are much more explorative and often have many, many more moving pieces.” Success at one type of problem-solving may not carry over to another.

Others agree. Martin Bridson, a mathematician at the University of Oxford, thinks the Math Olympiad result is a great achievement. “On the other hand, I don’t find it mind-blowing,” he says. “It’s not a change of paradigm in the sense that ‘Wow, I thought machines would never be able to do that.’ I expected machines to be able to do that.”

That’s because even though the problems in the Math Olympiad—and similar high school or undergraduate tests like AIME—are hard, there’s a pattern to a lot of them. “We have training camps to train high school kids to do them,” says Bridson. “And if you can train a large number of people to do those problems, why shouldn’t you be able to train a machine to do them?”

Sergei Gukov, a mathematician at the California Institute of Technology who coaches Math Olympiad teams, points out that the style of question does not change too much between competitions. New problems are set each year, but they can be solved with the same old tricks.

“Sure, the specific problems didn’t appear before,” says Gukov. “But they’re very close—just a step away from zillions of things you have already seen. You immediately realize, ‘Oh my gosh, there are so many similarities—I’m going to apply the same tactic.’” As hard as competition-level math is, kids and machines alike can be taught how to beat it.

That’s not true for most unsolved math problems. Bridson is president of the Clay Mathematics Institute, a nonprofit US-based research organization best known for setting up the Millenium Prize Problems in 2000—seven of the most important unsolved problems in mathematics, with a $1 million prize to be awarded to the first person to solve each of them. (One problem, the Poincaré conjecture, was solved in 2010; the others, which include P versus NP and the Riemann hypothesis, remain open). “We’re very far away from AI being able to say anything serious about any of those problems,” says Bridson.

And yet it’s hard to know exactly how far away, because many of the existing benchmarks used to evaluate progress are maxed out. The best new models already outperform most humans on tests like AIME.

To get a better idea of what existing systems can and cannot do, a startup called Epoch AI has created a new test called FrontierMath, released in December. Instead of co-opting math tests developed for humans, Epoch AI worked with more than 60 mathematicians around the world to come up with a set of math problems from scratch.

FrontierMath is designed to probe the limits of what today’s AI can do. None of the problems have been seen before and the majority are being kept secret to avoid contaminating training data. Each problem demands hours of work from expert mathematicians to solve—if they can solve it at all: some of the problems require specialist knowledge to tackle.

FrontierMath is set to become an industry standard. It’s not yet as popular as AIME, says de Oliveira Santos, who helped develop some of the problems: “But I expect this to not hold for much longer, since existing benchmarks are very close to being saturated.”

On AIME, the best large language models (Anthropic’s Claude 4, OpenAI’s o3 and o4-mini, Google DeepMind’s Gemini 2.5 Pro, X-AI’s Grok 3) now score around 90%. On FrontierMath, 04-mini scores 19% and Gemini 2.5 Pro scores 13%. That’s still remarkable, but there’s clear room for improvement.     

FrontierMath should give the best sense yet just how fast AI is progressing at math. But there are some problems that are still too hard for computers to take on.

2/ AI needs to manage really vast sequences of steps

Squint hard enough and in some ways math problems start to look the same: to solve them you need to take a sequence of steps from start to finish. The problem is finding those steps. 

“Pretty much every math problem can be formulated as path-finding,” says Gukov. What makes some problems far harder than others is the number of steps on that path. “The difference between the Riemann hypothesis and high school math is that with high school math the paths that we’re looking for are short—10 steps, 20 steps, maybe 40 in the longest case.” The steps are also repeated between problems.

“But to solve the Riemann hypothesis, we don’t have the steps, and what we’re looking for is a path that is extremely long”—maybe a million lines of computer proof, says Gukov.

Finding very long sequences of steps can be thought of as a kind of complex game. It’s what DeepMind’s AlphaZero learned to do when it mastered Go and chess. A game of Go might only involve a few hundred moves. But to win, an AI must find a winning sequence of moves among a vast number of possible sequences. Imagine a number with 100 zeros at the end, says Gukov.

But that’s still tiny compared with the number of possible sequences that could be involved in proving or disproving a very hard math problem: “A proof path with a thousand or a million moves involves a number with a thousand or a million zeros,” says Gukov. 

No AI system can sift through that many possibilities. To address this, Gukov and his colleagues developed a system that shortens the length of a path by combining multiple moves into single supermoves. It’s like having boots that let you take giant strides: instead of taking 2,000 steps to walk a mile, you can now walk it in 20.

The challenge was figuring out which moves to replace with supermoves. In a series of experiments, the researchers came up with a system in which one reinforcement-learning model suggests new moves and a second model checks to see if those moves help.

They used this approach to make a breakthrough in a math problem called the Andrews-Curtis conjecture, a puzzle that has been unsolved for 60 years. It’s a problem that every professional mathematician will know, says Gukov.

(An aside for math stans only: The AC conjecture states that a particular way of describing a type of set called a trivial group can be translated into a different but equivalent description with a certain sequence of steps. Most mathematicians think the AC conjecture is false, but nobody knows how to prove that. Gukov admits himself that it is an intellectual curiosity rather than a practical problem, but an important problem for mathematicians nonetheless.)

Gukov and his colleagues didn’t solve the AC conjecture, but they found that a counterexample (suggesting that the conjecture is false) proposed 40 years ago was itself false. “It’s been a major direction of attack for 40 years,” says Gukov. With the help of AI, they showed that this direction was in fact a dead end.   

“Ruling out possible counterexamples is a worthwhile thing,” says Bridson. “It can close off blind alleys, something you might spend a year of your life exploring.” 

True, Gukov checked off just one piece of one esoteric puzzle. But he thinks the approach will work in any scenario where you need to find a long sequence of unknown moves, and he now plans to try it out on other problems.

“Maybe it will lead to something that will help AI in general,” he says. “Because it’s teaching reinforcement learning models to go beyond their training. To me it’s basically about thinking outside of the box—miles away, megaparsecs away.”  

3/ Can AI ever provide real insight?

Thinking outside the box is exactly what mathematicians need to solve hard problems. Math is often thought to involve robotic, step-by-step procedures. But advanced math is an experimental pursuit, involving trial and error and flashes of insight.

That’s where tools like AlphaEvolve come in. Google DeepMind’s latest model asks an LLM to generate code to solve a particular math problem. A second model then evaluates the proposed solutions, picks the best, and sends them back to the LLM to be improved. After hundreds of rounds of trial and error, AlphaEvolve was able to come up with solutions to a wide range of math problems that were better than anything people had yet come up with. But it can also work as a collaborative tool: at any step, humans can share their own insight with the LLM, prompting it with specific instructions.

This kind of exploration is key to advanced mathematics. “I’m often looking for interesting phenomena and pushing myself in a certain direction,” says Geordie Williamson, a mathematician at the University of Sydney in Australia. “Like: ‘Let me look down this little alley. Oh, I found something!’”

Williamson worked with Meta on an AI tool called PatternBoost, designed to support this kind of exploration. PatternBoost can take a mathematical idea or statement and generate similar ones. “It’s like: ‘Here’s a bunch of interesting things. I don’t know what’s going on, but can you produce more interesting things like that?’” he says.

Such brainstorming is essential work in math. It’s how new ideas get conjured. Take the icosahedron, says Williamson: “It’s a beautiful example of this, which I kind of keep coming back to in my own work.” The icosahedron is a 20-sided 3D object where all the faces are triangles (think of a 20-sided die). The icosahedron is the largest of a family of exactly five such objects: there’s the tetrahedron (four sides), cube (six sides), octahedron (eight sides), and dodecahedron (12 sides).

Remarkably, the fact that there are exactly five of these objects was proved by mathematicians in ancient Greece. “At the time that this theorem was proved, the icosahedron didn’t exist,” says Williamson. “You can’t go to a quarry and find it—someone found it in their mind. And the icosahedron goes on to have a profound effect on mathematics. It’s still influencing us today in very, very profound ways.”

For Williamson, the exciting potential of tools like PatternBoost is that they might help people discover future mathematical objects like the icosahedron that go on to shape the way math is done. But we’re not there yet. “AI can contribute in a meaningful way to research-level problems,” he says. “But we’re certainly not getting inundated with new theorems at this stage.”

Ultimately, it comes down to the fact that machines still lack what you might call intuition or creative thinking. Williamson sums it up like this: We now have AI that can beat humans when it knows the rules of the game. “But it’s one thing for a computer to play Go at a superhuman level and another thing for the computer to invent the game of Go.”

“I think that applies to advanced mathematics,” he says. “Breakthroughs come from a new way of thinking about something, which is akin to finding completely new moves in a game. And I don’t really think we understand where those really brilliant moves in deep mathematics come from.”

Perhaps AI tools like AlphaEvolve and PatternBoost are best thought of as advance scouts for human intuition. They can discover new directions and point out dead ends, saving mathematicians months or years of work. But the true breakthroughs will still come from the minds of people, as has been the case for thousands of years.

For now, at least. “There’s plenty of tech companies that tell us that won’t last long,” says Williamson. “But you know—we’ll see.” 

What’s next for robots

MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.

Jan Liphardt teaches bioengineering at Stanford, but to many strangers in Los Altos, California, he is a peculiar man they see walking a four-legged robotic dog down the street. 

Liphardt has been experimenting with building and modifying robots for years, and when he brings his “dog” out in public, he generally gets one of three reactions. Young children want to have one, their parents are creeped out, and baby boomers try to ignore it. “They’ll quickly walk by,” he says, “like, ‘What kind of dumb new stuff is going on here?’” 

In the many conversations I’ve had about robots, I’ve also found that most people tend to fall into these three camps, though I don’t see such a neat age division. Some are upbeat and vocally hopeful that a future is just around the corner in which machines can expertly handle much of what is currently done by humans, from cooking to surgery. Others are scared: of job losses, injuries, and whatever problems may come up as we try to live side by side. 

The final camp, which I think is the largest, is just unimpressed. We’ve been sold lots of promises that robots will transform society ever since the first robotic arm was installed on an assembly line at a General Motors plant in New Jersey in 1961. Few of those promises have panned out so far. 

But this year, there’s reason to think that even those staunchly in the “bored” camp will be intrigued by what’s happening in the robot races. Here’s a glimpse at what to keep an eye on. 

Humanoids are put to the test

The race to build humanoid robots is motivated by the idea that the world is set up for the human form, and that automating that form could mean a seismic shift for robotics. It is led by some particularly outspoken and optimistic entrepreneurs, including Brett Adcock, the founder of Figure AI, a company making such robots that’s valued at more than $2.6 billion (it’s begun testing its robots with BMW). Adcock recently told Time, “Eventually, physical labor will be optional.” Elon Musk, whose company Tesla is building a version called Optimus, has said humanoid robots will create “a future where there is no poverty.” A robotics company called Eliza Wakes Up is taking preorders for a $420,000 humanoid called, yes, Eliza.

In June 2024, Agility Robotics sent a fleet of its Digit humanoid robots to GXO Logistics, which moves products for companies ranging from Nike to Nestlé. The humanoids can handle most tasks that involve picking things up and moving them somewhere else, like unloading pallets or putting boxes on a conveyor. 

There have been hiccups: Highly polished concrete floors can cause robots to slip at first, and buildings need good Wi-Fi coverage for the robots to keep functioning. But charging is a bigger issue. Agility’s current version of Digit, with a 39-pound battery, can run for two to four hours before it needs to charge for one hour, so swapping out the robots for fresh ones is a common task on each shift. If there are a small number of charging docks installed, the robots can theoretically charge by shuffling among the docks themselves overnight when some facilities aren’t running, but moving around on their own can set off a building’s security system. “It’s a problem,” says CTO Melonee Wise.

Wise is cautious about whether humanoids will be widely adopted in workplaces. “I’ve always been a pessimist,” she says. That’s because getting robots to work well in a lab is one thing, but integrating them into a bustling warehouse full of people and forklifts moving goods on tight deadlines is another task entirely.

If 2024 was the year of unsettling humanoid product launch videos, this year we will see those humanoids put to the test, and we’ll find out whether they’ll be as productive for paying customers as promised. Now that Agility’s robots have been deployed in fast-paced customer facilities, it’s clear that small problems can really add up. 

Then there are issues with how robots and humans share spaces. In the GXO facility the two work in completely separate areas, Wise says, but there are cases where, for example, a human worker might accidentally leave something obstructing a charging station. That means Agility’s robots can’t return to the dock to charge, so they need to alert a human employee to move the obstruction out of the way, slowing operations down.  

It’s often said that robots don’t call out sick or need health care. But this year, as fleets of humanoids arrive on the job, we’ll begin to find out the limitations they do have.

Learning from imagination

The way we teach robots how to do things is changing rapidly. It used to be necessary to break their tasks down into steps with specifically coded instructions, but now, thanks to AI, those instructions can be gleaned from observation. Just as ChatGPT was taught to write through exposure to trillions of sentences rather than by explicitly learning the rules of grammar, robots are learning through videos and demonstrations. 

That poses a big question: Where do you get all these videos and demonstrations for robots to learn from?

Nvidia, the world’s most valuable company, has long aimed to meet that need with simulated worlds, drawing on its roots in the video-game industry. It creates worlds in which roboticists can expose digital replicas of their robots to new environments to learn. A self-driving car can drive millions of virtual miles, or a factory robot can learn how to navigate in different lighting conditions.

In December, the company went a step further, releasing what it’s calling a “world foundation model.” Called Cosmos, the model has learned from 20 million hours of video—the equivalent of watching YouTube nonstop since Rome was at war with Carthage—that can be used to generate synthetic training data.

Here’s an example of how this model could help in practice. Imagine you run a robotics company that wants to build a humanoid that cleans up hospitals. You can start building this robot’s “brain” with a model from Nvidia, which will give it a basic understanding of physics and how the world works, but then you need to help it figure out the specifics of how hospitals work. You could go out and take videos and images of the insides of hospitals, or pay people to wear sensors and cameras while they go about their work there.

“But those are expensive to create and time consuming, so you can only do a limited number of them,” says Rev Lebaredian, vice president of simulation technologies at Nvidia. Cosmos can instead take a handful of those examples and create a three-dimensional simulation of a hospital. It will then start making changes—different floor colors, different sizes of hospital beds—and create slightly different environments. “You’ll multiply that data that you captured in the real world millions of times,” Lebaredian says. In the process, the model will be fine-tuned to work well in that specific hospital setting. 

It’s sort of like learning both from your experiences in the real world and from your own imagination (stipulating that your imagination is still bound by the rules of physics). 

Teaching robots through AI and simulations isn’t new, but it’s going to become much cheaper and more powerful in the years to come. 

A smarter brain gets a smarter body

Plenty of progress in robotics has to do with improving the way a robot senses and plans what to do—its “brain,” in other words. Those advancements can often happen faster than those that improve a robot’s “body,” which determine how well a robot can move through the physical world, especially in environments that are more chaotic and unpredictable than controlled assembly lines. 

The military has always been keen on changing that and expanding the boundaries of what’s physically possible. The US Navy has been testing machines from a company called Gecko Robotics that can navigate up vertical walls (using magnets) to do things like infrastructure inspections, checking for cracks, flaws, and bad welding on aircraft carriers. 

There are also investments being made for the battlefield. While nimble and affordable drones have reshaped rural battlefields in Ukraine, new efforts are underway to bring those drone capabilities indoors. The defense manufacturer Xtend received an $8.8 million contract from the Pentagon in December 2024 for its drones, which can navigate in confined indoor spaces and urban environments. These so-called “loitering munitions” are one-way attack drones carrying explosives that detonate on impact.

“These systems are designed to overcome challenges like confined spaces, unpredictable layouts, and GPS-denied zones,” says Rubi Liani, cofounder and CTO at Xtend. Deliveries to the Pentagon should begin in the first few months of this year. 

Another initiative—sparked in part by the Replicator project, the Pentagon’s plan to spend more than $1 billion on small unmanned vehicles—aims to develop more autonomously controlled submarines and surface vehicles. This is particularly of interest as the Department of Defense focuses increasingly on the possibility of a future conflict in the Pacific between China and Taiwan. In such a conflict, the drones that have dominated the war in Ukraine would serve little use because battles would be waged almost entirely at sea, where small aerial drones would be limited by their range. Instead, undersea drones would play a larger role.

All these changes, taken together, point toward a future where robots are more flexible in how they learn, where they work, and how they move. 

Jan Liphardt from Stanford thinks the next frontier of this transformation will hinge on the ability to instruct robots through speech. Large language models’ ability to understand and generate text has already made them a sort of translator between Liphardt and his robot.

“We can take one of our quadrupeds and we can tell it, ‘Hey, you’re a dog,’ and the thing wants to sniff you and tries to bark,” he says. “Then we do one word change—‘You’re a cat.’ Then the thing meows and, you know, runs away from dogs. And we haven’t changed a single line of code.”

Correction: A previous version of this story incorrectly stated that the robotics company Eliza Wakes Up has ties to a16z.