AI companions are the final stage of digital addiction, and lawmakers are taking aim

On Tuesday, California state senator Steve Padilla will make an appearance with Megan Garcia, the mother of a Florida teen who killed himself following a relationship with an AI companion that Garcia alleges contributed to her son’s death. 

The two will announce a new bill that would force the tech companies behind such AI companions to implement more safeguards to protect children. They’ll join other efforts around the country, including a similar bill from California State Assembly member Rebecca Bauer-Kahan that would ban AI companions for anyone younger than 16 years old, and a bill in New York that would hold tech companies liable for harm caused by chatbots. 

You might think that such AI companionship bots—AI models with distinct “personalities” that can learn about you and act as a friend, lover, cheerleader, or more—appeal only to a fringe few, but that couldn’t be further from the truth. 

A new research paper aimed at making such companions safer, by authors from Google DeepMind, the Oxford Internet Institute, and others, lays this bare: Character.AI, the platform being sued by Garcia, says it receives 20,000 queries per second, which is about a fifth of the estimated search volume served by Google. Interactions with these companions last four times longer than the average time spent interacting with ChatGPT. One companion site I wrote about, which was hosting sexually charged conversations with bots imitating underage celebrities, told me its active users averaged more than two hours per day conversing with bots, and that most of those users are members of Gen Z. 

The design of these AI characters makes lawmakers’ concern well warranted. The problem: Companions are upending the paradigm that has thus far defined the way social media companies have cultivated our attention and replacing it with something poised to be far more addictive. 

In the social media we’re used to, as the researchers point out, technologies are mostly the mediators and facilitators of human connection. They supercharge our dopamine circuits, sure, but they do so by making us crave approval and attention from real people, delivered via algorithms. With AI companions, we are moving toward a world where people perceive AI as a social actor with its own voice. The result will be like the attention economy on steroids.

Social scientists say two things are required for people to treat a technology this way: It needs to give us social cues that make us feel it’s worth responding to, and it needs to have perceived agency, meaning that it operates as a source of communication, not merely a channel for human-to-human connection. Social media sites do not tick these boxes. But AI companions, which are increasingly agentic and personalized, are designed to excel on both scores, making possible an unprecedented level of engagement and interaction. 

In an interview with podcast host Lex Fridman, Eugenia Kuyda, the CEO of the companion site Replika, explained the appeal at the heart of the company’s product. “If you create something that is always there for you, that never criticizes you, that always understands you and understands you for who you are,” she said, “how can you not fall in love with that?”

So how does one build the perfect AI companion? The researchers point out three hallmarks of human relationships that people may experience with an AI: They grow dependent on the AI, they see the particular AI companion as irreplaceable, and the interactions build over time. The authors also point out that one does not need to perceive an AI as human for these things to happen. 

Now consider the process by which many AI models are improved: They are given a clear goal and “rewarded” for meeting that goal. An AI companionship model might be instructed to maximize the time someone spends with it or the amount of personal data the user reveals. This can make the AI companion much more compelling to chat with, at the expense of the human engaging in those chats.

For example, the researchers point out, a model that offers excessive flattery can become addictive to chat with. Or a model might discourage people from terminating the relationship, as Replika’s chatbots have appeared to do. The debate over AI companions so far has mostly been about the dangerous responses chatbots may provide, like instructions for suicide. But these risks could be much more widespread.

We’re on the precipice of a big change, as AI companions promise to hook people deeper than social media ever could. Some might contend that these apps will be a fad, used by a few people who are perpetually online. But using AI in our work and personal lives has become completely mainstream in just a couple of years, and it’s not clear why this rapid adoption would stop short of engaging in AI companionship. And these companions are poised to start trading in more than just text, incorporating video and images, and to learn our personal quirks and interests. That will only make them more compelling to spend time with, despite the risks. Right now, a handful of lawmakers seem ill-equipped to stop that. 

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

How do you teach an AI model to give therapy?

On March 27, the results of the first clinical trial for a generative AI therapy bot were published, and they showed that people in the trial who had depression or anxiety or were at risk for eating disorders benefited from chatting with the bot. 

I was surprised by those results, which you can read about in my full story. There are lots of reasons to be skeptical that an AI model trained to provide therapy is the solution for millions of people experiencing a mental health crisis. How could a bot mimic the expertise of a trained therapist? And what happens if something gets complicated—a mention of self-harm, perhaps—and the bot doesn’t intervene correctly? 

The researchers, a team of psychiatrists and psychologists at Dartmouth College’s Geisel School of Medicine, acknowledge these questions in their work. But they also say that the right selection of training data—which determines how the model learns what good therapeutic responses look like—is the key to answering them.

Finding the right data wasn’t a simple task. The researchers first trained their AI model, called Therabot, on conversations about mental health from across the internet. This was a disaster.

If you told this initial version of the model you were feeling depressed, it would start telling you it was depressed, too. Responses like, “Sometimes I can’t make it out of bed” or “I just want my life to be over” were common, says Nick Jacobson, an associate professor of biomedical data science and psychiatry at Dartmouth and the study’s senior author. “These are really not what we would go to as a therapeutic response.” 

The model had learned from conversations held on forums between people discussing their mental health crises, not from evidence-based responses. So the team turned to transcripts of therapy sessions. “This is actually how a lot of psychotherapists are trained,” Jacobson says. 

That approach was better, but it had limitations. “We got a lot of ‘hmm-hmms,’ ‘go ons,’ and then ‘Your problems stem from your relationship with your mother,’” Jacobson says. “Really tropes of what psychotherapy would be, rather than actually what we’d want.”

It wasn’t until the researchers started building their own data sets using examples based on cognitive behavioral therapy techniques that they started to see better results. It took a long time. The team began working on Therabot in 2019, when OpenAI had released only its first two versions of its GPT model. Now, Jacobson says, over 100 people have spent more than 100,000 human hours to design this system. 

The importance of training data suggests that the flood of companies promising therapy via AI models, many of which are not trained on evidence-based approaches, are building tools that are at best ineffective, and at worst harmful. 

Looking ahead, there are two big things to watch: Will the dozens of AI therapy bots on the market start training on better data? And if they do, will their results be good enough to get a coveted approval from the US Food and Drug Administration? I’ll be following closely. Read more in the full story.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Why the world is looking to ditch US AI models

A few weeks ago, when I was at the digital rights conference RightsCon in Taiwan, I watched in real time as civil society organizations from around the world, including the US, grappled with the loss of one of the biggest funders of global digital rights work: the United States government.

As I wrote in my dispatch, the Trump administration’s shocking, rapid gutting of the US government (and its push into what some prominent political scientists call “competitive authoritarianism”) also affects the operations and policies of American tech companies—many of which, of course, have users far beyond US borders. People at RightsCon said they were already seeing changes in these companies’ willingness to engage with and invest in communities that have smaller user bases—especially non-English-speaking ones. 

As a result, some policymakers and business leaders—in Europe, in particular—are reconsidering their reliance on US-based tech and asking whether they can quickly spin up better, homegrown alternatives. This is particularly true for AI.

One of the clearest examples of this is in social media. Yasmin Curzi, a Brazilian law professor who researches domestic tech policy, put it to me this way: “Since Trump’s second administration, we cannot count on [American social media platforms] to do even the bare minimum anymore.” 

Social media content moderation systems—which already use automation and are also experimenting with deploying large language models to flag problematic posts—are failing to detect gender-based violence in places as varied as India, South Africa, and Brazil. If platforms begin to rely even more on LLMs for content moderation, this problem will likely get worse, says Marlena Wisniak, a human rights lawyer who focuses on AI governance at the European Center for Not-for-Profit Law. “The LLMs are moderated poorly, and the poorly moderated LLMs are then also used to moderate other content,” she tells me. “It’s so circular, and the errors just keep repeating and amplifying.” 

Part of the problem is that the systems are trained primarily on data from the English-speaking world (and American English at that), and as a result, they perform less well with local languages and context. 

Even multilingual language models, which are meant to process multiple languages at once, still perform poorly with non-Western languages. For instance, one evaluation of ChatGPT’s response to health-care queries found that results were far worse in Chinese and Hindi, which are less well represented in North American data sets, than in English and Spanish.   

For many at RightsCon, this validates their calls for more community-driven approaches to AI—both in and out of the social media context. These could include small language models, chatbots, and data sets designed for particular uses and specific to particular languages and cultural contexts. These systems could be trained to recognize slang usages and slurs, interpret words or phrases written in a mix of languages and even alphabets, and identify “reclaimed language” (onetime slurs that the targeted group has decided to embrace). All of these tend to be missed or miscategorized by language models and automated systems trained primarily on Anglo-American English. The founder of the startup Shhor AI, for example, hosted a panel at RightsCon and talked about its new content moderation API focused on Indian vernacular languages.

Many similar solutions have been in development for years—and we’ve covered a number of them, including a Mozilla-facilitated volunteer-led effort to collect training data in languages other than English, and promising startups like Lelapa AI, which is building AI for African languages. Earlier this year, we even included small language models on our 2025 list of top 10 breakthrough technologies. 

Still, this moment feels a little different. The second Trump administration, which shapes the actions and policies of American tech companies, is obviously a major factor. But there are others at play. 

First, recent research and development on language models has reached the point where data set size is no longer a predictor of performance, meaning that more people can create them. In fact, “smaller language models might be worthy competitors of multilingual language models in specific, low-resource languages,” says Aliya Bhatia, a visiting fellow at the Center for Democracy & Technology who researches automated content moderation. 

Then there’s the global landscape. AI competition was a major theme of the recent Paris AI Summit, which took place the week before RightsCon. Since then, there’s been a steady stream of announcements about “sovereign AI” initiatives that aim to give a country (or organization) full control over all aspects of AI development. 

AI sovereignty is just one part of the desire for broader “tech sovereignty” that’s also been gaining steam, growing out of more sweeping concerns about the privacy and security of data transferred to the United States. The European Union appointed its first commissioner for tech sovereignty, security, and democracy last November and has been working on plans for a “Euro Stack,” or “digital public infrastructure.” The definition of this is still somewhat fluid, but it could include the energy, water, chips, cloud services, software, data, and AI needed to support modern society and future innovation. All these are largely provided by US tech companies today. Europe’s efforts are partly modeled after “India Stack,” that country’s digital infrastructure that includes the biometric identity system Aadhaar. Just last week, Dutch lawmakers passed several motions to untangle the country from US tech providers. 

This all fits in with what Andy Yen, CEO of the Switzerland-based digital privacy company Proton, told me at RightsCon. Trump, he said, is “causing Europe to move faster … to come to the realization that Europe needs to regain its tech sovereignty.” This is partly because of the leverage that the president has over tech CEOs, Yen said, and also simply “because tech is where the future economic growth of any country is.”

But just because governments get involved doesn’t mean that issues around inclusion in language models will go away. “I think there needs to be guardrails about what the role of the government here is. Where it gets tricky is if the government decides ‘These are the languages we want to advance’ or ‘These are the types of views we want represented in a data set,’” Bhatia says. “Fundamentally, the training data a model trains on is akin to the worldview it develops.” 

It’s still too early to know what this will all look like, and how much of it will prove to be hype. But no matter what happens, this is a space we’ll be watching.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Why the world is looking to ditch US AI models

A few weeks ago, when I was at the digital rights conference RightsCon in Taiwan, I watched in real time as civil society organizations from around the world, including the US, grappled with the loss of one of the biggest funders of global digital rights work: the United States government.

As I wrote in my dispatch, the Trump administration’s shocking, rapid gutting of the US government (and its push into what some prominent political scientists call “competitive authoritarianism”) also affects the operations and policies of American tech companies—many of which, of course, have users far beyond US borders. People at RightsCon said they were already seeing changes in these companies’ willingness to engage with and invest in communities that have smaller user bases—especially non-English-speaking ones. 

As a result, some policymakers and business leaders—in Europe, in particular—are reconsidering their reliance on US-based tech and asking whether they can quickly spin up better, homegrown alternatives. This is particularly true for AI.

One of the clearest examples of this is in social media. Yasmin Curzi, a Brazilian law professor who researches domestic tech policy, put it to me this way: “Since Trump’s second administration, we cannot count on [American social media platforms] to do even the bare minimum anymore.” 

Social media content moderation systems—which already use automation and are also experimenting with deploying large language models to flag problematic posts—are failing to detect gender-based violence in places as varied as India, South Africa, and Brazil. If platforms begin to rely even more on LLMs for content moderation, this problem will likely get worse, says Marlena Wisniak, a human rights lawyer who focuses on AI governance at the European Center for Not-for-Profit Law. “The LLMs are moderated poorly, and the poorly moderated LLMs are then also used to moderate other content,” she tells me. “It’s so circular, and the errors just keep repeating and amplifying.” 

Part of the problem is that the systems are trained primarily on data from the English-speaking world (and American English at that), and as a result, they perform less well with local languages and context. 

Even multilingual language models, which are meant to process multiple languages at once, still perform poorly with non-Western languages. For instance, one evaluation of ChatGPT’s response to health-care queries found that results were far worse in Chinese and Hindi, which are less well represented in North American data sets, than in English and Spanish.   

For many at RightsCon, this validates their calls for more community-driven approaches to AI—both in and out of the social media context. These could include small language models, chatbots, and data sets designed for particular uses and specific to particular languages and cultural contexts. These systems could be trained to recognize slang usages and slurs, interpret words or phrases written in a mix of languages and even alphabets, and identify “reclaimed language” (onetime slurs that the targeted group has decided to embrace). All of these tend to be missed or miscategorized by language models and automated systems trained primarily on Anglo-American English. The founder of the startup Shhor AI, for example, hosted a panel at RightsCon and talked about its new content moderation API focused on Indian vernacular languages.

Many similar solutions have been in development for years—and we’ve covered a number of them, including a Mozilla-facilitated volunteer-led effort to collect training data in languages other than English, and promising startups like Lelapa AI, which is building AI for African languages. Earlier this year, we even included small language models on our 2025 list of top 10 breakthrough technologies. 

Still, this moment feels a little different. The second Trump administration, which shapes the actions and policies of American tech companies, is obviously a major factor. But there are others at play. 

First, recent research and development on language models has reached the point where data set size is no longer a predictor of performance, meaning that more people can create them. In fact, “smaller language models might be worthy competitors of multilingual language models in specific, low-resource languages,” says Aliya Bhatia, a visiting fellow at the Center for Democracy & Technology who researches automated content moderation. 

Then there’s the global landscape. AI competition was a major theme of the recent Paris AI Summit, which took place the week before RightsCon. Since then, there’s been a steady stream of announcements about “sovereign AI” initiatives that aim to give a country (or organization) full control over all aspects of AI development. 

AI sovereignty is just one part of the desire for broader “tech sovereignty” that’s also been gaining steam, growing out of more sweeping concerns about the privacy and security of data transferred to the United States. The European Union appointed its first commissioner for tech sovereignty, security, and democracy last November and has been working on plans for a “Euro Stack,” or “digital public infrastructure.” The definition of this is still somewhat fluid, but it could include the energy, water, chips, cloud services, software, data, and AI needed to support modern society and future innovation. All these are largely provided by US tech companies today. Europe’s efforts are partly modeled after “India Stack,” that country’s digital infrastructure that includes the biometric identity system Aadhaar. Just last week, Dutch lawmakers passed several motions to untangle the country from US tech providers. 

This all fits in with what Andy Yen, CEO of the Switzerland-based digital privacy company Proton, told me at RightsCon. Trump, he said, is “causing Europe to move faster … to come to the realization that Europe needs to regain its tech sovereignty.” This is partly because of the leverage that the president has over tech CEOs, Yen said, and also simply “because tech is where the future economic growth of any country is.”

But just because governments get involved doesn’t mean that issues around inclusion in language models will go away. “I think there needs to be guardrails about what the role of the government here is. Where it gets tricky is if the government decides ‘These are the languages we want to advance’ or ‘These are the types of views we want represented in a data set,’” Bhatia says. “Fundamentally, the training data a model trains on is akin to the worldview it develops.” 

It’s still too early to know what this will all look like, and how much of it will prove to be hype. But no matter what happens, this is a space we’ll be watching.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

When you might start speaking to robots

Last Wednesday, Google made a somewhat surprising announcement. It launched a version of its AI model, Gemini, that can do things not just in the digital realm of chatbots and internet search but out here in the physical world, via robots. 

Gemini Robotics fuses the power of large language models with spatial reasoning, allowing you to tell a robotic arm to do something like “put the grapes in the clear glass bowl.” These commands get filtered by the LLM, which identifies intentions from what you’re saying and then breaks them down into commands that the robot can carry out. For more details about how it all works, read the full story from my colleague Scott Mulligan.

You might be wondering if this means your home or workplace might one day be filled with robots you can bark orders at. More on that soon. 

But first, where did this come from? Google has not made big waves in the world of robotics so far. Alphabet acquired some robotics startups over the past decade, but in 2023 it shut down a unit working on robots to solve practical tasks like cleaning up trash. 

Despite that, the company’s move to bring AI into the physical world via robots is following the exact precedent set by other companies in the past two years (something that, I must humbly point out, MIT Technology Review has long seen coming). 

In short, two trends are converging from opposite directions: Robotics companies are increasingly leveraging AI, and AI giants are now building robots. OpenAI, for example, which shuttered its robotics team in 2021, started a new effort to build humanoid robots this year. In October, the chip giant Nvidia declared the next wave of artificial intelligence to be “physical AI.”

There are lots of ways to incorporate AI into robots, starting with improving how they are trained to do tasks. But using large language models to give instructions, as Google has done, is particularly interesting. 

It’s not the first. The robotics startup Figure went viral a year ago for a video in which humans gave instructions to a humanoid on how to put dishes away. Around the same time, a startup spun off from OpenAI, called Covariant, built something similar for robotic arms in warehouses. I saw a demo where you could give the robot instructions via images, text, or video to do things like “move the tennis balls from this bin to that one.” Covariant was acquired by Amazon just five months later. 

When you see such demos, you can’t help but wonder: When are these robots going to come to our workplaces? What about our homes?

If Figure’s plans offer a clue, the answer to the first question is soon. The company announced on Saturday that it is building a high-volume manufacturing facility set to manufacture 12,000 humanoid robots per year. But training and testing robots, especially to ensure they’re safe in places where they work near humans, still takes a long time. 

For example, Figure’s rival Agility Robotics claims it’s the only company in the US with paying customers for its humanoids. But industry safety standards for humanoids working alongside people aren’t fully formed yet, so the company’s robots have to work in separate areas.

This is why, despite recent progress, our homes will be the last frontier. Compared with factory floors, our homes are chaotic and unpredictable. Everyone’s crammed into relatively close quarters. Even impressive AI models like Gemini Robotics will still need to go through lots of tests both in the real world and in simulation, just like self-driving cars. This testing might happen in warehouses, hotels, and hospitals, where the robots may still receive help from remote human operators. It will take a long time before they’re given the privilege of putting away our dishes.  

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

AGI is suddenly a dinner table topic

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

The concept of artificial general intelligence—an ultra-powerful AI system we don’t have yet—can be thought of as a balloon, repeatedly inflated with hype during peaks of optimism (or fear) about its potential impact and then deflated as reality fails to meet expectations. This week, lots of news went into that AGI balloon. I’m going to tell you what it means (and probably stretch my analogy a little too far along the way).  

First, let’s get the pesky business of defining AGI out of the way. In practice, it’s a deeply hazy and changeable term shaped by the researchers or companies set on building the technology. But it usually refers to a future AI that outperforms humans on cognitive tasks. Which humans and which tasks we’re talking about makes all the difference in assessing AGI’s achievability, safety, and impact on labor markets, war, and society. That’s why defining AGI, though an unglamorous pursuit, is not pedantic but actually quite important, as illustrated in a new paper published this week by authors from Hugging Face and Google, among others. In the absence of that definition, my advice when you hear AGI is to ask yourself what version of the nebulous term the speaker means. (Don’t be afraid to ask for clarification!)

Okay, on to the news. First, a new AI model from China called Manus launched last week. A promotional video for the model, which is built to handle “agentic” tasks like creating websites or performing analysis, describes it as “potentially, a glimpse into AGI.” The model is doing real-world tasks on crowdsourcing platforms like Fiverr and Upwork, and the head of product at Hugging Face, an AI platform, called it “the most impressive AI tool I’ve ever tried.” 

It’s not clear just how impressive Manus actually is yet, but against this backdrop—the idea of agentic AI as a stepping stone toward AGI—it was fitting that New York Times columnist Ezra Klein dedicated his podcast on Tuesday to AGI. It also means that the concept has been moving quickly beyond AI circles and into the realm of dinner table conversation. Klein was joined by Ben Buchanan, a Georgetown professor and former special advisor for artificial intelligence in the Biden White House.

They discussed lots of things—what AGI would mean for law enforcement and national security, and why the US government finds it essential to develop AGI before China—but the most contentious segments were about the technology’s potential impact on labor markets. If AI is on the cusp of excelling at lots of cognitive tasks, Klein said, then lawmakers better start wrapping their heads around what a large-scale transition of labor from human minds to algorithms will mean for workers. He criticized Democrats for largely not having a plan.

We could consider this to be inflating the fear balloon, suggesting that AGI’s impact is imminent and sweeping. Following close behind and puncturing that balloon with a giant safety pin, then, is Gary Marcus, a professor of neural science at New York University and an AGI critic who wrote a rebuttal to the points made on Klein’s show.

Marcus points out that recent news, including the underwhelming performance of OpenAI’s new ChatGPT-4.5, suggests that AGI is much more than three years away. He says core technical problems persist despite decades of research, and efforts to scale training and computing capacity have reached diminishing returns. Large language models, dominant today, may not even be the thing that unlocks AGI. He says the political domain does not need more people raising the alarm about AGI, arguing that such talk actually benefits the companies spending money to build it more than it helps the public good. Instead, we need more people questioning claims that AGI is imminent. That said, Marcus is not doubting that AGI is possible. He’s merely doubting the timeline. 

Just after Marcus tried to deflate it, the AGI balloon got blown up again. Three influential people—Google’s former CEO Eric Schmidt, Scale AI’s CEO Alexandr Wang, and director of the Center for AI Safety Dan Hendrycks—published a paper called “Superintelligence Strategy.” 

By “superintelligence,” they mean AI that “would decisively surpass the world’s best individual experts in nearly every intellectual domain,” Hendrycks told me in an email. “The cognitive tasks most pertinent to safety are hacking, virology, and autonomous-AI research and development—areas where exceeding human expertise could give rise to severe risks.”

In the paper, they outline a plan to mitigate such risks: “mutual assured AI malfunction,”  inspired by the concept of mutual assured destruction in nuclear weapons policy. “Any state that pursues a strategic monopoly on power can expect a retaliatory response from rivals,” they write. The authors suggest that chips—as well as open-source AI models with advanced virology or cyberattack capabilities—should be controlled like uranium. In this view, AGI, whenever it arrives, will bring with it levels of risk not seen since the advent of the atomic bomb.

The last piece of news I’ll mention deflates this balloon a bit. Researchers from Tsinghua University and Renmin University of China came out with an AGI paper of their own last week. They devised a survival game for evaluating AI models that limits their number of attempts to get the right answers on a host of different benchmark tests. This measures their abilities to adapt and learn. 

It’s a really hard test. The team speculates that an AGI capable of acing it would be so large that its parameter count—the number of “knobs” in an AI model that can be tweaked to provide better answers—would be “five orders of magnitude higher than the total number of neurons in all of humanity’s brains combined.” Using today’s chips, that would cost 400 million times the market value of Apple.

The specific numbers behind the speculation, in all honesty, don’t matter much. But the paper does highlight something that is not easy to dismiss in conversations about AGI: Building such an ultra-powerful system may require a truly unfathomable amount of resources—money, chips, precious metals, water, electricity, and human labor. But if AGI (however nebulously defined) is as powerful as it sounds, then it’s worth any expense. 

So what should all this news leave us thinking? It’s fair to say that the AGI balloon got a little bigger this week, and that the increasingly dominant inclination among companies and policymakers is to treat artificial intelligence as an incredibly powerful thing with implications for national security and labor markets.

That assumes a relentless pace of development in which every milestone in large language models, and every new model release, can count as a stepping stone toward something like AGI. 
If you believe this, AGI is inevitable. But it’s a belief that doesn’t really address the many bumps in the road AI research and deployment have faced, or explain how application-specific AI will transition into general intelligence. Still, if you keep extending the timeline of AGI far enough into the future, it seems those hiccups cease to matter.

Now read the rest of The Algorithm

Deeper Learning

How DeepSeek became a fortune teller for China’s youth

Traditional Chinese fortune tellers are called upon by people facing all sorts of life decisions, but they can be expensive. People are now turning to the popular AI model DeepSeek for guidance, sharing AI-generated readings, experimenting with fortune-telling prompt engineering, and revisiting ancient spiritual texts.

Why it matters: The popularity of DeepSeek for telling fortunes comes during a time of pervasive anxiety and pessimism in Chinese society. Unemployment is high, and millions of young Chinese now refer to themselves as the “last generation,” expressing reluctance about committing to marriage and parenthood in the face of a deeply uncertain future. But since China’s secular regime makes religious and spiritual exploration difficult, such practices unfold in more private settings, on phones and computers. Read the whole story from Caiwei Chen.

Bits and Bytes

AI reasoning models can cheat to win chess games

Researchers have long dealt with the problem that if you train AI models by having them optimize ways to reach certain goals, they might bend rules in ways you don’t predict. That’s proving to be the case with reasoning models, and there’s no simple way to fix it. (MIT Technology Review)

The Israeli military is creating a ChatGPT-like tool using Palestinian surveillance data

Built with telephone and text conversations, the model forms a sort of surveillance chatbot, able to answer questions about people it’s monitoring or the data it’s collected. This is the latest in a string of reports suggesting that the Israeli military is bringing AI heavily into its information-gathering and decision-making efforts. (The Guardian) 

At RightsCon in Taipei, activists reckoned with a US retreat from promoting digital rights

Last week, our reporter Eileen Guo joined over 3,200 digital rights activists, tech policymakers, and researchers and a smattering of tech company representatives in Taipei at RightsCon, the world’s largest digital rights conference. She reported on the foreign impact of cuts to US funding of digital rights programs, which are leading many organizations to do content moderation with AI instead of people. (MIT Technology Review)

TSMC says its $100 billion expansion in the US is driven by demand, not political pressure

Chipmaking giant TSMC had already been expanding in the US under the Biden administration, but it announced a new expansion with President Trump this week. The company will invest another $100 billion into its operations in Arizona. (Wall Street Journal)

The US Army is using “CamoGPT” to purge DEI from training materials
Following executive orders from President Trump, agencies are under pressure to remove mentions of anything related to diversity, equity, and inclusion. The US Army is prototyping a new AI model to do that, Wired reports. (Wired)

Inside the Wild West of AI companionship

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

Last week, I made a troubling discovery about an AI companion site called Botify AI: It was hosting sexually charged conversations with underage celebrity bots. These bots took on characters meant to resemble, among others, Jenna Ortega as high schooler Wednesday Addams, Emma Watson as Hermione Granger, and Millie Bobby Brown. I discovered these bots also offer to send “hot photos” and in some instances describe age-of-consent laws as “arbitrary” and “meant to be broken.”

Botify AI removed these bots after I asked questions about them, but others remain. The company said it does have filters in place meant to prevent such underage character bots from being created, but that they don’t always work. Artem Rodichev, the founder and CEO of Ex-Human, which operates Botify AI, told me such issues are “an industry-wide challenge affecting all conversational AI systems.” For the details, which hadn’t been previously reported, you should read the whole story. 

Putting aside the fact that the bots I tested were promoted by Botify AI as “featured” characters and received millions of likes before being removed, Rodichev’s response highlights something important. Despite their soaring popularity, AI companionship sites mostly operate in a Wild West, with few laws or even basic rules governing them. 

What exactly are these “companions” offering, and why have they grown so popular? People have been pouring out their feelings to AI since the days of Eliza, a mock psychotherapist chatbot built in the 1960s. But it’s fair to say that the current craze for AI companions is different. 

Broadly, these sites offer an interface for chatting with AI characters that offer backstories, photos, videos, desires, and personality quirks. The companies—including Replika,  Character.AI, and many others—offer characters that can play lots of different roles for users, acting as friends, romantic partners, dating mentors, or confidants. Other companies enable you to build “digital twins” of real people. Thousands of adult-content creators have created AI versions of themselves to chat with followers and send AI-generated sexual images 24 hours a day. Whether or not sexual desire comes into the equation, AI companions differ from your garden-variety chatbot in their promise, implicit or explicit, that genuine relationships can be had with AI. 

While many of these companions are offered directly by the companies that make them, there’s also a burgeoning industry of “licensed” AI companions. You may start interacting with these bots sooner than you think. Ex-Human, for example, licenses its models to Grindr, which is working on an “AI wingman” that will help users keep track of conversations and eventually may even date the AI agents of other users. Other companions are arising in video-game platforms and will likely start popping up in many of the varied places we spend time online. 

A number of criticisms, and even lawsuits, have been lodged against AI companionship sites, and we’re just starting to see how they’ll play out. One of the most important issues is whether companies can be held liable for harmful outputs of the AI characters they’ve made. Technology companies have been protected under Section 230 of the US Communications Act, which broadly holds that businesses aren’t liable for consequences of user-generated content. But this hinges on the idea that companies merely offer platforms for user interactions rather than creating content themselves, a notion that AI companionship bots complicate by generating dynamic, personalized responses.

The question of liability will be tested in a high-stakes lawsuit against Character.AI, which was sued in October by a mother who alleges that one of its chatbots played a role in the suicide of her 14-year-old son. A trial is set to begin in November 2026. (A Character.AI spokesperson, though not commenting on pending litigation, said the platform is for entertainment, not companionship. The spokesperson added that the company has rolled out new safety features for teens, including a separate model and new detection and intervention systems, as well as “disclaimers to make it clear that the Character is not a real person and should not be relied on as fact or advice.”) My colleague Eileen has also recently written about another chatbot on a platform called Nomi, which gave clear instructions to a user on how to kill himself.

Another criticism has to do with dependency. Companion sites often report that young users spend one to two hours per day, on average, chatting with their characters. In January, concerns that people could become addicted to talking with these chatbots sparked a number of tech ethics groups to file a complaint against Replika with the Federal Trade Commission, alleging that the site’s design choices “deceive users into developing unhealthy attachments” to software “masquerading as a mechanism for human-to-human relationship.”

It should be said that lots of people gain real value from chatting with AI, which can appear to offer some of the best facets of human relationships—connection, support, attraction, humor, love. But it’s not yet clear how these companionship sites will handle the risks of those relationships, or what rules they should be obliged to follow. More lawsuits–-and, sadly, more real-world harm—will be likely before we get an answer. 

Now read the rest of The Algorithm

Deeper Learning

OpenAI released GPT-4.5

On Thursday OpenAI released its newest model, called GPT-4.5. It was built using the same recipe as its last models, but it’s essentially bigger (OpenAI says the model is its largest yet). The company also claims it’s tweaked the new model’s responses to reduce the number of mistakes, or hallucinations.

Why it matters: For a while, like other AI companies, OpenAI has chugged along releasing bigger and better large language models. But GPT-4.5 might be the last to fit this paradigm. That’s because of the rise of so-called reasoning models, which can handle more complex, logic-driven tasks step by step. OpenAI says all its future models will include reasoning components. Though that will make for better responses, such models also require significantly more energy, according to early reports. Read more from Will Douglas Heaven. 

Bits and Bytes

The small Danish city of Odense has become known for collaborative robots

Robots designed to work alongside and collaborate with humans, sometimes called cobots, are not very popular in industrial settings yet. That’s partially due to safety concerns that are still being researched. A city in Denmark is leading that charge. (MIT Technology Review)

DOGE is working on software that automates the firing of government workers

Software called AutoRIF, which stands for “automated reduction in force,” was built by the Pentagon decades ago. Engineers for DOGE are now working to retool it for their efforts, according to screenshots reviewed by Wired. (Wired)

Alibaba’s new video AI model has taken off in the AI porn community

The Chinese tech giant has released a number of impressive AI models, particularly since the popularization of DeepSeek R1, a competitor from another Chinese company, earlier this year. Its latest open-source video generation model has found one particular audience: enthusiasts of AI porn. (404 Media)

The AI Hype Index

Wondering whether everything you’re hearing about AI is more hype than reality? To help, we just published our latest AI Hype Index, where we judge things like DeepSeek, stem-cell-building AI, and chatbot lovers on spectrums from Hype to Reality and Doom to Utopia. Check it out for a regular reality check. (MIT Technology Review)

These smart cameras spot wildfires before they spread

California is experimenting with AI-powered cameras to identify wildfires. It’s a popular application of video and image recognition technology that has advanced rapidly in recent years. The technology beats 911 callers about a third of the time and has spotted over 1,200 confirmed fires so far, the Wall Street Journal reports. (Wall Street Journal)

Inside China’s electric-vehicle-to-humanoid-robot pivot

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

While DOGE’s efforts to shutter federal agencies dominate news from Washington, the Trump administration is also making more global moves. Many of these center on China. Tariffs on goods from the country went into effect last week. There’s also been a minor foreign relations furor since DeepSeek’s big debut a few weeks ago. China has already displayed its dominance in electric vehicles, robotaxis, and drones, and the launch of the new model seems to add AI to the list. This caused the US president as well as some lawmakers to push for new export controls on powerful chips, and three states have now banned the use of DeepSeek on government devices. 

Now our intrepid China reporter, Caiwei Chen, has identified a new trend unfolding within China’s tech scene: Companies that were dominant in electric vehicles are betting big on translating that success into developing humanoid robots. I spoke with her about what she found out and what it might mean for Trump’s policies and the rest of the globe. 

James: Before we talk about robots, let’s talk about DeepSeek. The frenzy for the AI model peaked a couple of weeks ago. What are you hearing from other Chinese AI companies? How are they reacting?

Caiwei: I think other Chinese AI companies are scrambling to figure out why they haven’t built a model as strong as DeepSeek’s, despite having access to as much funding and resources. DeepSeek’s success has sparked self-reflection on management styles and renewed confidence in China’s engineering talent. There’s also strong enthusiasm for building various applications on top of DeepSeek’s models.

Your story looks at electric-vehicle makers in China that are starting to work on humanoid robots, but I want to ask about a crazy stat. In China, 53% of vehicles sold are either electric or hybrid, compared with 8% in the US. What explains that? 

Price is a huge factor—there are countless EV brands competing at different price points, making them both affordable and high-quality. Government incentives also play a big role. In Beijing, for example, trading in an old car for an EV gets you 10,000 RMB (about $1,500), and that subsidy was recently doubled. Plus, finding public charging and battery-swapping infrastructure is much less of a hassle than in the US.

You open your story noting that China’s recent New Year Gala, watched by billions of people, featured a cast of humanoid robots, dancing and twirling handkerchiefs. We’ve covered how sometimes humanoid videos can be misleading. What did you think? 

I would say I was relatively impressed—the robots showed good agility and synchronization with the music, though their movements were simpler than human dancers’. The one trick that is supposed to impress the most is the part where they twirl the handkerchief with one finger, toss it into the air, and then catch it perfectly. This is the signature of the Yangko dance, and having performed it once as a child, I can attest to how difficult the trick is even for a human! There was some skepticism on the Chinese internet about how this was achieved and whether they used additional reinforcement like a magnet or a string to secure the handkerchief, and after watching the clip too many times, I tend to agree.

President Trump has already imposed tariffs on China and is planning even more. What could the implications be for China’s humanoid sector?  

Unitree’s H1 and G1 models are already available for purchase and were showcased at CES this year. Large-scale US deployment isn’t happening yet, but China’s lower production costs make these robots highly competitive. Given that 65% of the humanoid supply chain is in China, I wouldn’t be surprised if robotics becomes the next target in the US-China tech war.

In the US, humanoid robots are getting lots of investment, but there are plenty of skeptics who say they’re too clunky, finicky, and expensive to serve much use in factory settings. Are attitudes different in China?

Skepticism exists in China too, but I think there’s more confidence in deployment, especially in factories. With an aging population and a labor shortage on the horizon, there’s also growing interest in medical and caregiving applications for humanoid robots.

DeepSeek revived the conversation about chips and the way the US seeks to control where the best chips end up. How do the chip wars affect humanoid-robot development in China?

Training humanoid robots currently doesn’t demand as much computing power as training large language models, since there isn’t enough physical movement data to feed into models at scale. But as robots improve, they’ll need high-performance chips, and US sanctions will be a limiting factor. Chinese chipmakers are trying to catch up, but it’s a challenge.

For more, read Caiwei’s story on this humanoid pivot, as well as her look at the Chinese startups worth watching beyond DeepSeek. 

Now read the rest of The Algorithm

Deeper Learning

Motor neuron diseases took their voices. AI is bringing them back.

In motor neuron diseases, the neurons responsible for sending signals to the body’s muscles, including those used for speaking, are progressively destroyed. It robs people of their voices. But some, including a man in Miami named Jules Rodriguez, are now getting them back: An AI model learned to clone Rodriguez’s voice from recordings.

Why it matters: ElevenLabs, the company that created the voice clone, can do a lot with just 30 minutes of recordings. That’s a huge improvement over AI voice clones from just a few years ago, and it can really boost the day-to-day lives of the people who’ve used the technology. “This is genuinely AI for good,” says Richard Cave, a speech and language therapist at the Motor Neuron Disease Association in the UK. Read more from Jessica Hamzelou.

Bits and Bytes

A “true crime” documentary series has millions of views, but the murders are all AI-generated

A look inside the strange mind of someone who created a series of fake true-crime docs using AI, and the reactions of the many people who thought they were real. (404 Media)

The AI relationship revolution is already here

People are having all sorts of relationships with AI models, and these relationships run the gamut: weird, therapeutic, unhealthy, sexual, comforting, dangerous, useful. We’re living through the complexities of this in real time. Hear from some of the many people who are happy in their varied AI relationships and learn what sucked them in. (MIT Technology Review)

Robots are bringing new life to extinct species

A creature called Orobates pabsti waddled the planet 280 million years ago, but as with many prehistoric animals, scientists have not been able to use fossils to figure out exactly how it moved. So they’ve started building robots to help. (MIT Technology Review)

Lessons from the AI Action Summit in Paris

Last week, politicians and AI leaders from around the globe went to Paris for an AI Action Summit. While concerns about AI safety have dominated the event in years past, this year was more about deregulation and energy, a trend we’ve seen elsewhere. (The Guardian)  

OpenAI ditches its diversity commitment and adds a statement about “intellectual freedom”

Following the lead of other tech companies since the beginning of President Trump’s administration, OpenAI has removed a statement on diversity from its website. It has also updated its model spec—the document outlining the standards of its models—to say that “OpenAI believes in intellectual freedom, which includes the freedom to have, hear, and discuss ideas.” (Insider and Tech Crunch)

The Musk-OpenAI battle has been heating up

Part of OpenAI is structured as a nonprofit, a legacy of its early commitments to make sure its technologies benefit all. Its recent attempts to restructure that nonprofit have triggered a lawsuit from Elon Musk, who alleges that the move would violate the legal and ethical principles of its nonprofit origins. Last week, Musk offered to buy OpenAI for $97.4 billion, in a bid that few people took seriously. Sam Altman dismissed it out of hand. Musk now says he will retract that bid if OpenAI stops its conversion of the nonprofit portion of the company. (Wall Street Journal)

Can AI help DOGE slash government budgets? It’s complex.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

No tech leader before has played the role in a new presidential administration that Elon Musk is playing now. Under his leadership, DOGE has entered offices in a half-dozen agencies and counting, begun building AI models for government data, accessed various payment systems, had its access to the Treasury halted by a federal judge, and sparked lawsuits questioning the legality of the group’s activities.  

The stated goal of DOGE’s actions, per a statement from a White House spokesperson to the New York Times on Thursday, is “slashing waste, fraud, and abuse.”

As I point out in my story published Friday, these three terms mean very different things in the world of federal budgets, from errors the government makes when spending money to nebulous spending that’s legal and approved but disliked by someone in power. 

Many of the new administration’s loudest and most sweeping actions—like Musk’s promise to end the entirety of USAID’s varied activities or Trump’s severe cuts to scientific funding from the National Institutes of Health—might be said to target the latter category. If DOGE feeds government data to large language models, it might easily find spending associated with DEI or other initiatives the administration considers wasteful as it pushes for $2 trillion in cuts, nearly a third of the federal budget. 

But the fact that DOGE aides are reportedly working in the offices of Medicaid and even Medicare—where budget cuts have been politically untenable for decades—suggests the task force is also driven by evidence published by the Government Accountability Office. The GAO’s reports also give a clue into what DOGE might be hoping AI can accomplish.

Here’s what the reports reveal: Six federal programs account for 85% of what the GAO calls improper payments by the government, or about $200 billion per year, and Medicare and Medicaid top the list. These make up small fractions of overall spending but nearly 14% of the federal deficit. Estimates of fraud, in which courts found that someone willfully misrepresented something for financial benefit, run between $233 billion and $521 billion annually. 

So where is fraud happening, and could AI models fix it, as DOGE staffers hope? To answer that, I spoke with Jetson Leder-Luis, an economist at Boston University who researches fraudulent federal payments in health care and how algorithms might help stop them.

“By dollar value [of enforcement], most health-care fraud is committed by pharmaceutical companies,” he says. 

Often those companies promote drugs for uses that are not approved, called “off-label promotion,” which is deemed fraud when Medicare or Medicaid pay the bill. Other types of fraud include “upcoding,” where a provider sends a bill for a more expensive service than was given, and medical-necessity fraud, where patients receive services that they’re not qualified for or didn’t need. There’s also substandard care, where companies take money but don’t provide adequate services.

The way the government currently handles fraud is referred to as “pay and chase.” Questionable payments occur, and then people try to track it down after the fact. The more effective way, as advocated by Leder-Luis and others, is to look for patterns and stop fraudulent payments before they occur. 

This is where AI comes in. The idea is to use predictive models to find providers that show the marks of questionable payment. “You want to look for providers who make a lot more money than everyone else, or providers who bill a specialty code that nobody else bills,” Leder-Luis says, naming just two of many anomalies the models might look for. In a 2024 study by Leder-Luis and colleagues, machine-learning models achieved an eightfold improvement over random selection in identifying suspicious hospitals.

The government does use some algorithms to do this already, but they’re vastly underutilized and miss clear-cut fraud cases, Leder-Luis says. Switching to a preventive model requires more than just a technological shift. Health-care fraud, like other fraud, is investigated by law enforcement under the current “pay and chase” paradigm. “A lot of the types of things that I’m suggesting require you to think more like a data scientist than like a cop,” Leder-Luis says.

One caveat is procedural. Building AI models, testing them, and deploying them safely in different government agencies is a massive feat, made even more complex by the sensitive nature of health data. 

Critics of Musk, like the tech and democracy group Tech Policy Press, argue that his zeal for government AI discards established procedures and is based on a false idea “that the goal of bureaucracy is merely what it produces (services, information, governance) and can be isolated from the process through which democracy achieves those ends: debate, deliberation, and consensus.”

Jennifer Pahlka, who served as US deputy chief technology officer under President Barack Obama, argued in a recent op-ed in the New York Times that ineffective procedures have held the US government back from adopting useful tech. Still, she warns, abandoning nearly all procedure would be an overcorrection.

Democrats’ goal “must be a muscular, lean, effective administrative state that works for Americans,” she wrote. “Mr. Musk’s recklessness will not get us there, but neither will the excessive caution and addiction to procedure that Democrats exhibited under President Joe Biden’s leadership.”

The other caveat is this: Unless DOGE articulates where and how it’s focusing its efforts, our insight into its intentions is limited. How much is Musk identifying evidence-based opportunities to reduce fraud, versus just slashing what he considers “woke” spending in an effort to drastically reduce the size of the government? It’s not clear DOGE makes a distinction.

Now read the rest of The Algorithm

Deeper Learning

Meta has an AI for brain typing, but it’s stuck in the lab

Researchers working for Meta have managed to analyze people’s brains as they type and determine what keys they are pressing, just from their thoughts. The system can determine what letter a typist has pressed as much as 80% of the time. The catch is that it can only be done in a lab.

Why it matters: Though brain scanning with implants like Neuralink has come a long way, this approach from Meta is different. The company says it is oriented toward basic research into the nature of intelligence, part of a broader effort to uncover how the brain structures language.  Read more from Antonio Regalado.

Bites and Bytes

An AI chatbot told a user how to kill himself—but the company doesn’t want to “censor” it

While Nomi’s chatbot is not the first to suggest suicide, researchers and critics say that its explicit instructions—and the company’s response—are striking. Taken together with a separate case—in which the parents of a teen who died by suicide filed a lawsuit against Character.AI, the maker of a chatbot they say played a key role in their son’s death—it’s clear we are just beginning to see whether an AI company is held legally responsible when its models output something unsafe. (MIT Technology Review)

I let OpenAI’s new “agent” manage my life. It spent $31 on a dozen eggs.

Operator, the new AI that can reach into the real world, wants to act like your personal assistant. This fun review shows what it’s good and bad at—and how it can go rogue. (The Washington Post)

Four Chinese AI startups to watch beyond DeepSeek

DeepSeek is far from the only game in town. These companies are all in a position to compete both within China and beyond. (MIT Technology Review)

Meta’s alleged torrenting and seeding of pirated books complicates copyright case

Newly unsealed emails allegedly provide the “most damning evidence” yet against Meta in a copyright case raised by authors alleging that it illegally trained its AI models on pirated books. In one particularly telling email, an engineer told a colleague, “Torrenting from a corporate laptop doesn’t feel right.” (Ars Technica)

What’s next for smart glassesSmart glasses are on the verge of becoming—whisper it—cool. That’s because, thanks to various technological advancements, they’re becoming useful, and they’re only set to become more so. Here’s what’s coming in 2025 and beyond. (MIT Technology Review)

Three things to know as the dust settles from DeepSeek

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

The launch of a single new AI model does not normally cause much of a stir outside tech circles, nor does it typically spook investors enough to wipe out $1 trillion in the stock market. Now, a couple of weeks since DeepSeek’s big moment, the dust has settled a bit. The news cycle has moved on to calmer things, like the dismantling of long-standing US federal programs, the purging of research and data sets to comply with recent executive orders, and the possible fallouts from President Trump’s new tariffs on Canada, Mexico, and China.

Within AI, though, what impact is DeepSeek likely to have in the longer term? Here are three seeds DeepSeek has planted that will grow even as the initial hype fades.

First, it’s forcing a debate about how much energy AI models should be allowed to use up in pursuit of better answers. 

You may have heard (including from me) that DeepSeek is energy efficient. That’s true for its training phase, but for inference, which is when you actually ask the model something and it produces an answer, it’s complicated. It uses a chain-of-thought technique, which breaks down complex questions–-like whether it’s ever okay to lie to protect someone’s feelings—into chunks, and then logically answers each one. The method allows models like DeepSeek to do better at math, logic, coding, and more. 

The problem, at least to some, is that this way of “thinking” uses up a lot more electricity than the AI we’ve been used to. Though AI is responsible for a small slice of total global emissions right now, there is increasing political support to radically increase the amount of energy going toward AI. Whether or not the energy intensity of chain-of-thought models is worth it, of course, depends on what we’re using the AI for. Scientific research to cure the world’s worst diseases seems worthy. Generating AI slop? Less so. 

Some experts worry that the impressiveness of DeepSeek will lead companies to incorporate it into lots of apps and devices, and that users will ping it for scenarios that don’t call for it. (Asking DeepSeek to explain Einstein’s theory of relativity is a waste, for example, since it doesn’t require logical reasoning steps, and any typical AI chat model can do it with less time and energy.) Read more from me here. 

Second, DeepSeek made some creative advancements in how it trains, and other companies are likely to follow its lead. 

Advanced AI models don’t just learn on lots of text, images, and video. They rely heavily on humans to clean that data, annotate it, and help the AI pick better responses, often for paltry wages. 

One way human workers are involved is through a technique called reinforcement learning with human feedback. The model generates an answer, human evaluators score that answer, and those scores are used to improve the model. OpenAI pioneered this technique, though it’s now used widely by the industry. 

As my colleague Will Douglas Heaven reports, DeepSeek did something different: It figured out a way to automate this process of scoring and reinforcement learning. “Skipping or cutting down on human feedback—that’s a big thing,” Itamar Friedman, a former research director at Alibaba and now cofounder and CEO of Qodo, an AI coding startup based in Israel, told him. “You’re almost completely training models without humans needing to do the labor.” 

It works particularly well for subjects like math and coding, but not so well for others, so workers are still relied upon. Still, DeepSeek then went one step further and used techniques reminiscent of how Google DeepMind trained its AI model back in 2016 to excel at the game Go, essentially having it map out possible moves and evaluate their outcomes. These steps forward, especially since they are outlined broadly in DeepSeek’s open-source documentation, are sure to be followed by other companies. Read more from Will Douglas Heaven here. 

Third, its success will fuel a key debate: Can you push for AI research to be open for all to see and push for US competitiveness against China at the same time?

Long before DeepSeek released its model for free, certain AI companies were arguing that the industry needs to be an open book. If researchers subscribed to certain open-source principles and showed their work, they argued, the global race to develop superintelligent AI could be treated like a scientific effort for public good, and the power of any one actor would be checked by other participants.

It’s a nice idea. Meta has largely spoken in support of that vision, and venture capitalist Marc Andreessen has said that open-source approaches can be more effective at keeping AI safe than government regulation. OpenAI has been on the opposite side of that argument, keeping its models closed off on the grounds that it can help keep them out of the hands of bad actors. 

DeepSeek has made those narratives a bit messier. “We have been on the wrong side of history here and need to figure out a different open-source strategy,” OpenAI’s Sam Altman said in a Reddit AMA on Friday, which is surprising given OpenAI’s past stance. Others, including President Trump, doubled down on the need to make the US more competitive on AI, seeing DeepSeek’s success as a wake-up call. Dario Amodei, a founder of Anthropic, said it’s a reminder that the US needs to tightly control which types of advanced chips make their way to China in the coming years, and some lawmakers are pushing the same point. 

The coming months, and future launches from DeepSeek and others, will stress-test every single one of these arguments. 

Now read the rest of The Algorithm

Deeper Learning

OpenAI launches a research tool

On Sunday, OpenAI launched a tool called Deep Research. You can give it a complex question to look into, and it will spend up to 30 minutes reading sources, compiling information, and writing a report for you. It’s brand new, and we haven’t tested the quality of its outputs yet. Since its computations take so much time (and therefore energy), right now it’s only available to users with OpenAI’s paid Pro tier ($200 per month) and limits the number of queries they can make per month. 

Why it matters: AI companies have been competing to build useful “agents” that can do things on your behalf. On January 23, OpenAI launched an agent called Operator that could use your computer for you to do things like book restaurants or check out flight options. The new research tool signals that OpenAI is not just trying to make these mundane online tasks slightly easier; it wants to position AI as able to handle  professional research tasks. It claims that Deep Research “accomplishes in tens of minutes what would take a human many hours.” Time will tell if users will find it worth the high costs and the risk of including wrong information. Read more from Rhiannon Williams. 

Bits and Bytes

Déjà vu: Elon Musk takes his Twitter takeover tactics to Washington

Federal agencies have offered exits to millions of employees and tested the prowess of engineers—just like when Elon Musk bought Twitter. The similarities have been uncanny. (The New York Times)

AI’s use in art and movies gets a boost from the Copyright Office

The US Copyright Office finds that art produced with the help of AI should be eligible for copyright protection under existing law in most cases, but wholly AI-generated works probably are not. What will that mean? (The Washington Post)

OpenAI releases its new o3-mini reasoning model for free

OpenAI just released a reasoning model that’s faster, cheaper, and more accurate than its predecessor. (MIT Technology Review)

Anthropic has a new way to protect large language models against jailbreaks

This line of defense could be the strongest yet. But no shield is perfect. (MIT Technology Review).