Ice Lounge Media

Ice Lounge Media

The AI community is abuzz over DeepSeek R1, a new open-source reasoning model. 

The model was developed by the Chinese AI startup DeepSeek, which claims that R1 matches or even surpasses OpenAI’s ChatGPT o1 on multiple key benchmarks but operates at a fraction of the cost. 

“This could be a truly equalizing breakthrough that is great for researchers and developers with limited resources, especially those from the Global South,” says Hancheng Cao, an assistant professor in information systems at Emory University.

DeepSeek’s success is even more remarkable given the constraints facing Chinese AI companies in the form of increasing US export controls on cutting-edge chips. But early evidence shows that these measures are not working as intended. Rather than weakening China’s AI capabilities, the sanctions appear to be driving startups like DeepSeek to innovate in ways that prioritize efficiency, resource-pooling, and collaboration.

To create R1, DeepSeek had to rework its training process to reduce the strain on its GPUs, a variety released by Nvidia for the Chinese market that have their performance capped at half the speed of its top products, according to Zihan Wang, a former DeepSeek employee and current PhD student in computer science at Northwestern University. 

DeepSeek R1 has been praised by researchers for its ability to tackle complex reasoning tasks, particularly in mathematics and coding. The model employs a “chain of thought” approach similar to that used by ChatGPT o1, which lets it solve problems by processing queries step by step.

Dimitris Papailiopoulos, principal researcher at Microsoft’s AI Frontiers research lab, says what surprised him the most about R1 is its engineering simplicity. “DeepSeek aimed for accurate answers rather than detailing every logical step, significantly reducing computing time while maintaining a high level of effectiveness,” he says.

DeepSeek has also released six smaller versions of R1 that are small enough to  run locally on laptops. It claims that one of them even outperforms OpenAI’s o1-mini on certain benchmarks.“DeepSeek has largely replicated o1-mini and has open sourced it,” tweeted Perplexity CEO Aravind Srinivas. DeepSeek did not reply to MIT Technology Review’s request for comments.

Despite the buzz around R1, DeepSeek remains relatively unknown. Based in Hangzhou, China, it was founded in July 2023 by Liang Wenfeng, an alumnus of Zhejiang University with a background in information and electronic engineering. It was incubated by High-Flyer, a hedge fund that Liang founded in 2015. Like Sam Altman of OpenAI, Liang aims to build artificial general intelligence (AGI), a form of AI that can match or even beat humans on a range of tasks.

Training large language models (LLMs) requires a team of highly trained researchers and substantial computing power. In a recent interview with the Chinese media outlet LatePost, Kai-Fu Lee, a veteran entrepreneur and former head of Google China, said that only “front-row players” typically engage in building foundation models such as ChatGPT, as it’s so resource-intensive. The situation is further complicated by the US export controls on advanced semiconductors. High-Flyer’s decision to venture into AI is directly related to these constraints, however. Long before the anticipated sanctions, Liang acquired a substantial stockpile of Nvidia A100 chips, a type now banned from export to China. The Chinese media outlet 36Kr estimates that the company has over 10,000 units in stock, but Dylan Patel, founder of the AI research consultancy SemiAnalysis, estimates that it has at least 50,000. Recognizing the potential of this stockpile for AI training is what led Liang to establish DeepSeek, which was able to use them in combination with the lower-power chips to develop its models. 

Tech giants like Alibaba and ByteDance, as well as a handful of startups with deep-pocketed investors, dominate the Chinese AI space, making it challenging for small or medium-sized enterprises to compete. A company like DeepSeek, which has no plans to raise funds, is rare. 

Zihan Wang, the former DeepSeek employee, told MIT Technology Review that he had access to abundant computing resources and was given freedom to experiment when working at DeepSeek, “a luxury that few fresh graduates would get at any company.” 

In an interview with the Chinese media outlet 36Kr in July 2024 Liang said that an additional challenge Chinese companies face on top of chip sanctions, is that their AI engineering techniques tend to be less efficient. “We [most Chinese companies] have to consume twice the computing power to achieve the same results. Combined with data efficiency gaps, this could mean needing up to four times more computing power. Our goal is to continuously close these gaps,” he said.  

But DeepSeek found ways to reduce memory usage and speed up calculation without significantly sacrificing accuracy. “The team loves turning a hardware challenge into an opportunity for innovation,” says Wang.

Liang himself remains deeply involved in DeepSeek’s research process, running experiments alongside his team. “The whole team shares a collaborative culture and dedication to hardcore research,” Wang says.

As well as prioritizing efficiency, Chinese companies are increasingly embracing open-source principles. Alibaba Cloud has released over 100 new open-source AI models, supporting 29 languages and catering to various applications, including coding and mathematics. Similarly, startups like Minimax and 01.AI have open-sourced their models. 

According to a white paper released last year by the China Academy of Information and Communications Technology, a state-affiliated research institute, the number of AI large language models worldwide has reached 1,328, with 36% originating in China. This positions China as the second-largest contributor to AI, behind the United States. 

“This generation of young Chinese researchers identify strongly with open-source culture because they benefit so much from it,” says Thomas Qitong Cao, an assistant professor of technology policy at Tufts University.

“The US export control has essentially backed Chinese companies into a corner where they have to be far more efficient with their limited computing resources,” says Matt Sheehan, an AI researcher at the Carnegie Endowment for International Peace. “We are probably going to see a lot of consolidation in the future related to the lack of compute.”

That might already have started to happen. Two weeks ago, Alibaba Cloud announced that it has partnered with the Beijing-based startup 01.AI, founded by Kai-Fu Lee, to merge research teams and establish an “industrial large model laboratory.”

“It is energy-efficient and natural for some kind of division of labor to emerge in the AI industry,” says Cao, the Tufts professor. “The rapid evolution of AI demands agility from Chinese firms to survive.”

Read more

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.

OpenAI launches Operator—an agent that can use a computer for you

What’s new: After weeks of buzz, OpenAI has released Operator, its first AI agent. Operator is a web app that can carry out simple online tasks in a browser, such as booking concert tickets or filling an online grocery order. The app is powered by a new model called Computer-Using Agent—CUA for short—built on top of OpenAI’s multimodal large language model GPT-4o.

Why it matters: OpenAI claims that Operator outperforms similar rival tools, including Anthropic’s Computer Use and Google DeepMind’s Mariner. The fact that three of the world’s top AI firms have converged on the same vision of what agent-based models could be makes one thing clear. The battle for AI supremacy has a new frontier—and it’s our computer screens. Read the full story.

—Will Douglas Heaven

+ If you’re interested in reading more about AI agents, check out this piece explaining why they’re AI’s next big thing.

What’s next for robots

—James O’Donnell

In the many conversations I’ve had about robots, I’ve also found that most people tend to fall into three camps. 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 this year. Read the full story.

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

The must-reads

I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.

1 Facebook and Instagram blocked and hid abortion pill posts
But Meta denies it’s anything to do with its recent hate speech restriction U-turn. (NYT $)
+ The company’s widespread changes are making advertisers nervous. (Insider $)
+ A contraceptive drug could act as an abortion pill substitute. (The Atlantic $)

2 Donald Trump’s staff are furious with Elon Musk 
His decision to trash talk the President’s new AI deal is ruffling aides’ feathers. (Politico)
+ For once, Trump doesn’t seem to want to wade in. (CNN)
+ Stargate’s newest data center will be built in the small Texan city of Abilene. (Bloomberg $)

3 Watch the Trump administration delete agency pages in real time
An agency GitHub records the documents, handbooks and bots as they’re deleted or amended. (404 Media)

4 Central Europe’s power grid is vulnerable to attack
Its facilities’ unencrypted radio signals leave it wide open to malicious interference. (Ars Technica)
+ The race to replace the powerful greenhouse gas that underpins the power grid. (MIT Technology Review)

5 OpenAI’s conversion to becoming a for-profit is under investigation
California’s attorney general wants to know more about its asset transfer plans. (The Markup)
+ One major obstacle is determining how much equity Microsoft would hold. (FT $)

6 WeRide has its sights set on becoming a driverless power player
The Chinese company has ambitious plans to expand all over the world. (WSJ $)
+ Meanwhile, Tesla is issuing a safety update to 1.2 million cars in China. (Bloomberg $)
+ How Wayve’s driverless cars will meet one of their biggest challenges yet. (MIT Technology Review)

7 How fungi spores can help save endangered plants
But it’s a delicate balancing act. (Knowable Magazine)
+ Africa fights rising hunger by looking to foods of the past. (MIT Technology Review)

8 The fight over our tech-addled attention span
It’s not that we can’t focus—it’s what we’re focusing on. (New Yorker $) 

9 TikTok is still MIA from US app stores
Opportunists are flogging iPhones with the pre-installed app for eye-watering prices. (Insider $)

10 How random is Spotify’s shuffle, really?
And can algorithms be depended on to deal in true randomness? (FT $)

Quote of the day

“I can’t imagine that I personally can make any difference in their wealth, power or influence. But I can’t be a part of offering them my life and my joy to then turn it back around and make money off of me.”

—Michael Raine, a 50-year old Facebook and Instagram user, explains to the Washington Post why he doesn’t want to contribute to the sprawling wealth of Meta boss Mark Zuckerberg any more.

The big story

How to stop a state from sinking

April 2024

In a 10-month span between 2020 and 2021, southwest Louisiana saw five climate-related disasters, including two destructive hurricanes. As if that wasn’t bad enough, more storms are coming, and many areas are not prepared.

But some government officials and state engineers are hoping there is an alternative: elevation. The $6.8 billion Southwest Coastal Louisiana Project is betting that raising residences by a few feet will keep Louisianans in their communities.

Ultimately, it’s something of a last-ditch effort to preserve this slice of coastline, even as some locals pick up and move inland and as formal plans for managed retreat become more popular in climate-­vulnerable areas across the country and the rest of the world. Read the full story.

—Xander Peters

We can still have nice things

A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.)

+ How two enterprising actors staged a daring performance of Hamlet inside Grand Theft Auto 💀
+ Warning: these movies are dangerous!
+ Madonna released Material Girl 40 years ago this week—and changed the face of pop forever.
+ And finally, what everyone has been dying to know—do dogs really watch TV?

Read more

This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

On January 20, his first day in office, US president Donald Trump signed an executive order to withdraw the US from the World Health Organization. “Ooh, that’s a big one,” he said as he was handed the document.

The US is the biggest donor to the WHO, and the loss of this income is likely to have a significant impact on the organization, which develops international health guidelines, investigates disease outbreaks, and acts as an information-sharing hub for member states.

But the US will also lose out. “It’s a very tragic and sad event that could only hurt the United States in the long run,” says William Moss, an epidemiologist at Johns Hopkins Bloomberg School of Public Health in Baltimore.

Trump appears to take issue with the amount the US donates to the WHO. He points out that it makes a much bigger contribution than China, a country with a population four times that of the US. “It seems a little unfair to me,” he said as he prepared to sign the executive order.

It is true that the US is far and away the biggest financial supporter of the WHO. The US contributed $1.28 billion over the two-year period covering 2022 and 2023. By comparison, the second-largest donor, Germany, contributed $856 million in the same period. The US currently contributes 14.5% of the WHO’s total budget.

But it’s not as though the WHO sends a billion-dollar bill to the US. All member states are required to pay membership dues, which are calculated as a percentage of a country’s gross domestic product. For the US, this figure comes to $130 million. China pays $87.6 million. But the vast majority of the US’s contributions to the WHO are made on a voluntary basis—in recent years, the donations have been part of multibillion-dollar spending on global health by the US government. (Separately, the Bill and Melinda Gates Foundation contributed $830 million over 2022 and 2023.)

There’s a possibility that other member nations will increase their donations to help cover the shortfall left by the US’s withdrawal. But it is not clear who will step up—or what implications it will have to change the structure of donations.

Martin McKee, a professor of European public health at the London School of Hygiene and Tropical Medicine, thinks it is unlikely that European members will increase their contributions by much. China, India, Brazil, South Africa, and the Gulf states, on the other hand, may be more likely to pay more. But again, it isn’t clear how this will pan out, or whether any of these countries will expect greater influence over global health policy decisions as a result of increasing their donations.

WHO funds are spent on a range of global health projects—programs to eradicate polio, rapidly respond to health emergencies, improve access to vaccines and medicines, develop pandemic prevention strategies, and more. The loss of US funding is likely to have a significant impact on at least some of these programs.

“Diseases don’t stick to national boundaries, hence this decision is not only concerning for the US, but in fact for every country in the world,” says Pauline Scheelbeek at the London School of Hygiene and Tropical Medicine.“With the US no longer reporting to the WHO nor funding part of this process, the evidence on which public health interventions and solutions should be based is incomplete.”

“It’s going to hurt global health,” adds Moss. “It’s going to come back to bite us.”

There’s more on how the withdrawal could affect health programs, vaccine coverage, and pandemic preparedness in this week’s coverage.


Now read the rest of The Checkup

Read more from MIT Technology Review‘s archive

This isn’t the first time Donald Trump has signaled his desire for the US to leave the WHO. He proposed a withdrawal during his last term, in 2020. While the WHO is not perfect, it needs more power and funding, not less, Charles Kenny, director of technology and development at the Center for Global Development, argued at the time.

The move drew condemnation from those working in public health then, too. The editor in chief of the medical journal The Lancet called it “a crime against humanity,” as Charlotte Jee reported.

In 1974, the WHO launched an ambitious program to get lifesaving vaccines to all children around the world. Fifty years on, vaccines are thought to have averted 154 million deaths—including 146 million in children under the age of five. 

The WHO has also seen huge success in its efforts to eradicate polio. Today, wild forms of the virus have been eradicated in all but two countries. But vaccine-derived forms of the virus can still crop up around the world.

At the end of a round of discussions in September among WHO member states working on a pandemic agreement, director-general Tedros Adhanom Ghebreyesus remarked, “The next pandemic will not wait for us, whether from a flu virus like H5N1, another coronavirus, or another family of viruses we don’t yet know about.” The H5N1 virus has been circulating on US dairy farms for months now, and the US is preparing for potential human outbreaks.

From around the web

People with cancer paid $45,000 for an experimental blood-filtering treatment, delivered at a clinic in Antigua, after being misled about its effectiveness. Six of them have died since their treatments. (The New York Times)

The Trump administration has instructed federal health agencies to pause all external communications, such as health advisories, weekly scientific reports, updates to websites, and social media posts. (The Washington Post)

A new “virtual retina,” modeled on human retinas, has been developed to study the impact of retinal implants. The three-dimensional model simulates over 10,000 neurons. (Brain Stimulation)

Trump has signed an executive order stating that “it is the policy of the United States to recognize two sexes, male and female.” The document “defies decades of research into how human bodies grow and develop,” STAT reports, and represents “a dramatic failure to understand biology,” according to a neuroscientist who studies the development of sex. (STAT)

Attention, summer holiday planners: Biting sandflies in the Mediterranean region are transmitting Toscana virus at an increasing rate. The virus is a major cause of central nervous system disorders in the region. Italy saw a 2.6-fold increase in the number of reported infections between the 2016–21 period and 2022–23. (Eurosurveillance)

Read more

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.

Read more
1 93 94 95 96 97 2,624