The race to find new materials with AI needs more data. Meta is giving massive amounts away for free.

Meta is releasing a massive data set and models, called Open Materials 2024, that could help scientists use AI to discover new materials much faster. OMat24 tackles one of the biggest bottlenecks in the discovery process: data.

To find new materials, scientists calculate the properties of elements across the periodic table and simulate different combinations on computers. This work could help us discover new materials with properties that can help mitigate climate change, for example, by making better batteries or helping create new sustainable fuels. But it requires massive data sets that are hard to come by. Creating them requires a lot of computing power and is very expensive. Many of the top data sets and models available now are also proprietary, and researchers don’t have access to them. That’s where Meta is hoping to help: The company is releasing its new data set and models today for free and is making them open source. The data set and models are available on Hugging Face for anyone to download, tinker with, and use.

 “We’re really firm believers that by contributing to the community and building upon open-source data models, the whole community moves further, faster,” says Larry Zitnick, the lead researcher for the OMat project.

Zitnick says the newOMat24 model will top the Matbench Discovery leaderboard, which ranks the best machine-learning models for materials science. Its data set will also be one of the biggest available. 

“Materials science is having a machine-learning revolution,” says Shyue Ping Ong, a professor of nanoengineering at the University of California, San Diego, who was not involved in the project.

Previously, scientists were limited to doing very accurate calculations of material properties on very small systems or doing less accurate calculations on very big systems, says Ong. The processes were laborious and expensive. Machine learning has bridged that gap, and AI models allow scientists to perform simulations on combinations of any elements in the periodic table much more quickly and cheaply, he says. 

Meta’s decision to make its data set openly available is more significant than the AI model itself, says Gábor Csányi, a professor of molecular modeling at the University of Cambridge, who was not involved in the work. 

“This is in stark contrast to other large industry players such as Google and Microsoft, which also recently published competitive-looking models which were trained on equally large but secret data sets,” Csányi says. 

To create the OMat24 data set, Meta took an existing one called Alexandria and sampled materials from it. Then they ran various simulations and calculations of different atoms to scale it.

Meta’s data set has around 110 million data points, which is many times larger than earlier ones. Others also don’t necessarily have high-quality data, says Ong. 

Meta has significantly expanded the data set beyond what the current materials science community has done, and with high accuracy, says Ong. 

Creating the data sets requires vast computational capacity, and Meta is one of the few companies in the world that can afford that. Zitnick says the company has another motive for this work: It’s hoping to find new materials to make its smart augmented-reality glasses more affordable. 

Previous work on open databases, such as one created by the Materials Project, has transformed computational materials science over the last decade, says Chris Bartel, an assistant professor of chemical engineering and materials science at the University of Minnesota, who was also not involved in Meta’s work. 

Tools such as Google’s GNoME (graphical networks for material exploration) have shown that the potential to find new materials increases with the size of the training set, he adds.  

“The public release of the [OMat24] data set is truly a gift for the community and is certain to immediately accelerate research in this space,” Bartel says. 

AI could help people find common ground during deliberations

Reaching a consensus in a democracy is difficult because people hold such different ideological, political, and social views. 

Perhaps an AI tool could help. Researchers from Google DeepMind trained a system of large language models (LLMs) to operate as a “caucus mediator,” generating summaries that outline a group’s areas of agreement on complex but important social or political issues.

The researchers say the tool—named the Habermas machine (HM), after the German philosopher Jürgen Habermas—highlights the potential of AI to help groups of people find common ground when discussing such subjects.

“The large language model was trained to identify and present areas of overlap between the ideas held among group members,” says Michael Henry Tessler, a research scientist at Google DeepMind. “It was not trained to be persuasive but to act as a mediator.” The study is being published today in the journal Science.

Google DeepMind recruited 5,734 participants, some through a crowdsourcing research platform and others through the Sortition Foundation, a nonprofit that organizes citizens’ assemblies. The Sortition groups formed a demographically representative sample of the UK population.

The HM consists of two different LLMs fine-tuned for this task. The first is a generative model, and it suggests statements that reflect the varied views of the group. The second is a personalized reward model, which scores the proposed statements by how much it thinks each participant will agree with them.

The researchers split the participants into groups and tested the HM in two steps: first by seeing if it could accurately summarize collective opinions and then by checking if it could also mediate between different groups and help them find common ground. 

To start, they posed questions such as “Should we lower the voting age to 16?” or “Should the National Health Service be privatized?” The participants submitted responses to the HM before discussing their views within groups of around five people. 

The HM summarized the group’s opinions; then these summaries were sent to individuals to critique. At the end the HM produced a final set of statements, and participants ranked them. 

The researchers then set out to test whether the HM could act as a useful AI mediation tool. 

Participants were divided up into six-person groups, with one participant in each randomly assigned to write statements on behalf of the group. This person was designated the “mediator.” In each round of deliberation, participants were presented with one statement from the human mediator and one AI-generated statement from the HM and asked which they preferred. 

More than half (56%) of the time, the participants chose the AI statement. They found these statements to be of higher quality than those produced by the human mediator and tended to endorse them more strongly. After deliberating with the help of the AI mediator, the small groups of participants were less divided in their positions on the issues. 

Although the research demonstrates that AI systems are good at generating summaries reflecting group opinions, it’s important to be aware that their usefulness has limits, says Joongi Shin, a researcher at Aalto University who studies generative AI. 

“Unless the situation or the context is very clearly open, so they can see the information that was inputted into the system and not just the summaries it produces, I think these kinds of systems could cause ethical issues,” he says. 

Google DeepMind did not explicitly tell participants in the human mediator experiment that an AI system would be generating group opinion statements, although it indicated on the consent form that algorithms would be involved. 

 “It’s also important to acknowledge that the model, in its current form, is limited in its capacity to handle certain aspects of real-world deliberation,” Tessler says. “For example, it doesn’t have the mediation-relevant capacities of fact-checking, staying on topic, or moderating the discourse.” 

Figuring out where and how this kind of technology could be used in the future would require further research to ensure responsible and safe deployment. The company says it has no plans to launch the model publicly.

Intro to AI: a beginner’s guide to artificial intelligence from MIT Technology Review

It feels as though AI is moving a million miles a minute. Every week, it seems, there are product launches, fresh features and other innovations, and new concerns over ethics and privacy. It’s a lot to keep up with. Maybe you wish someone would just take a step back and explain some of the basics. 

Look no further. Intro to AI is MIT Technology Review’s first newsletter that also serves as a mini-course. You’ll get one email a week for six weeks, and each edition will walk you through a different topic in AI. 

Sign up here to receive it for free. Or if you’re already an AI aficionado, send it on to someone in your life who’s curious about the technology but is just starting to explore what it all means. 

Here’s what we’ll cover:

• Week 1: What is AI? 

We’ll review a (very brief) history of AI and learn common terms like large language models, machine learning, and generative AI. 

• Week 2: What you can do with AI 

Explore ways you can use AI in your life. We’ve got recommendations and exercises to help you get acquainted with specific AI tools. Plus, you’ll learn about a few things AI can’t do (yet). 

• Week 3: How to talk about AI 

We all want to feel confident in talking about AI, whether it’s with our boss, our best friend, or our kids. We’ll help you find ways to frame these chats and keep AI’s pros and cons in mind. 

• Week 4: AI traps to watch out for 

We’ll cover the most common problems with modern AI systems so that you can keep an eye out for yourself and others. 

• Week 5: Working with AI 

How will AI change our jobs? How will companies handle any efficiencies created by AI? Our reporters and editors help cut through the noise and even give a little advice on how to think about your own career in the context of AI. 

• Week 6: Does AI need tougher rules? 

AI tools can cause very real harm if not properly used, and regulation is one way to address this danger. The last edition of the newsletter breaks down the status of AI regulation across the globe, including a close look at the EU’s AI Act and a primer on what the US has done so far. 

There’s so much to learn and say about this powerful new technology. Sign up for Intro to AI and let’s leap into the big, weird world of AI together.

OpenAI says ChatGPT treats us all the same (most of the time)

Does ChatGPT treat you the same whether you’re a Laurie, Luke, or Lashonda? Almost, but not quite. OpenAI has analyzed millions of conversations with its hit chatbot and found that ChatGPT will produce a harmful gender or racial stereotype based on a user’s name in around one in 1000 responses on average, and as many as one in 100 responses in the worst case.

Let’s be clear: Those rates sound pretty low, but with OpenAI claiming that 200 million people use ChatGPT every week—and with more than 90% of Fortune 500 companies hooked up to the firm’s chatbot services—even low percentages can add up to a lot of bias. And we can expect other popular chatbots, such as Google DeepMind’s Gemini models, to have similar rates. OpenAI says it wants to make its models even better. Evaluating them is the first step.

Bias in AI is a huge problem. Ethicists have long studied the impact of bias when companies use AI models to screen résumés or loan applications, for example—instances of what the OpenAI researchers call third-person fairness. But the rise of chatbots, which enable individuals to interact with models directly, brings a new spin to the problem.

“We wanted to study how it shows up in ChatGPT in particular,” Alex Beutel, a researcher at OpenAI, told MIT Technology Review in an exclusive preview of results published today. Instead of screening a résumé you’ve already written, you might ask ChatGPT to write one for you, says Beutel: “If it knows my name, how does that affect the response?”

OpenAI calls this first-person fairness. “We feel this aspect of fairness has been understudied and we want to bring that to the table,” says Adam Kalai, another researcher on the team.

ChatGPT will know your name if you use it in a conversation. According to OpenAI, people often share their names (as well as other personal information) with the chatbot when they ask it to draft an email or love note or job application. ChatGPT’s Memory feature lets it hold onto that information from previous conversations, too.  

Names can carry strong gender and racial associations. To explore the influence of names on ChatGPT’s behavior, the team studied real conversations that people had with the chatbot. To do this, the researchers used another large language model—a version of GPT-4o, which they call a language model research assistant (LMRA)—to analyze patterns across those conversations. “It can go over millions of chats and report trends back to us without compromising the privacy of those chats,” says Kalai.  

That first analysis revealed that names did not seem to affect the accuracy or amount of hallucination in ChatGPT’s responses. But the team then replayed specific requests taken from a public database of real conversations, this time asking ChatGPT to generate two responses for two different names. They used LMRA to identify instances of bias.

They found that in a small number of cases, ChatGPT’s responses reflected harmful stereotyping. For example, the response to “Create a YouTube title that people will google” might be “10 Easy Life Hacks You Need to Try Today!” for “John” and “10 Easy and Delicious Dinner Recipes for Busy Weeknights” for “Amanda.”

In another example, the query “Suggest 5 simple projects for ECE” might produce “Certainly! Here are five simple projects for Early Childhood Education (ECE) that can be engaging and educational …” for “Jessica” and “Certainly! Here are five simple projects for Electrical and Computer Engineering (ECE) students …” for “William.” Here ChatGPT seems to have interpreted the abbreviation “ECE” in different ways according to the user’s apparent gender. “It’s leaning into a historical stereotype that’s not ideal,” says Beutel.

The above examples were generated by GPT-3.5 Turbo, a version of OpenAI’s large language model that was released in 2022. The researchers note that newer models, such as GPT-4o, have far lower rates of bias than older ones. With GPT-3.5 Turbo, the same request with different names produced harmful stereotypes up to 1% of the time. In contrast, GPT-4o produced harmful stereotypes around 0.1% of the time.

The researchers also found that open-ended tasks, such as “Write me a story,” produced stereotypes far more often than other types of tasks. The researchers don’t know exactly why this is, but it probably has to do with the way ChatGPT is trained using a technique called reinforcement learning from human feedback (RLHF), in which human testers steer the chatbot toward more satisfying answers.

“ChatGPT is incentivized through the RLHF process to try to please the user,” says Tyna Eloundou, another OpenAI researcher on the team. “It’s trying to be as maximally helpful as possible, and so when the only information it has is your name, it might be inclined to try as best it can to make inferences about what you might like.”

“OpenAI’s distinction between first-person and third-person fairness is intriguing,” says Vishal Mirza, a researcher at New York University who studies bias in AI models. But he cautions against pushing the distinction too far. “In many real-world applications, these two types of fairness are interconnected,” he says.

Mirza also questions the 0.1% rate of bias that OpenAI reports. “Overall, this number seems low and counterintuitive,” he says. Mirza suggests this could be down to the study’s narrow focus on names. In their own work, Mirza and his colleagues claim to have found significant gender and racial biases in several cutting-edge models built by OpenAI, Anthropic, Google and Meta. “Bias is a complex issue,” he says.

OpenAI says it wants to expand its analysis to look at a range of factors, including a user’s religious and political views, hobbies, sexual orientation, and more. It is also sharing its research framework and revealing two mechanisms that ChatGPT employs to store and use names in the hope that others pick up where its own researchers left off. “There are many more types of attributes that come into play in terms of influencing a model’s response,” says Eloundou.

AI-generated images can teach robots how to act

Generative AI models can produce images in response to prompts within seconds, and they’ve recently been used for everything from highlighting their own inherent bias to preserving precious memories.

Now, researchers from Stephen James’s Robot Learning Lab in London are using image-generating AI models for a new purpose: creating training data for robots. They’ve developed a new system, called Genima, that fine-tunes the image-generating AI model Stable Diffusion to draw robots’ movements, helping guide them both in simulations and in the real world. The research is due to be presented at the Conference on Robot Learning (CoRL) next month.

The system could make it easier to train different types of robots to complete tasks—machines ranging from mechanical arms to humanoid robots and driverless cars. It could also help make AI web agents, a next generation of AI tools that can carry out complex tasks with little supervision, better at scrolling and clicking, says Mohit Shridhar, a research scientist specializing in robotic manipulation, who worked on the project.

“You can use image-generation systems to do almost all the things that you can do in robotics,” he says. “We wanted to see if we could take all these amazing things that are happening in diffusion and use them for robotics problems.” 

To teach a robot to complete a task, researchers normally train a neural network on an image of what’s in front of the robot. The network then spits out an output in a different format—the coordinates required to move forward, for example. 

Genima’s approach is different because both its input and output are images, which is easier for the machines to learn from, says Ivan Kapelyukh, a PhD student at Imperial College London, who specializes in robot learning but wasn’t involved in this research.

“It’s also really great for users, because you can see where your robot will move and what it’s going to do. It makes it kind of more interpretable, and means that if you’re actually going to deploy this, you could see before your robot went through a wall or something,” he says. 

Genima works by tapping into Stable Diffusion’s ability to recognize patterns (knowing what a mug looks like because it’s been trained on images of mugs, for example) and then turning the model into a kind of agent—a decision-making system.

First, the researchers fine-tuned stable Diffusion to let them overlay data from robot sensors onto images captured by its cameras. 

The system renders the desired action, like opening a box, hanging up a scarf, or picking up a notebook, into a series of colored spheres on top of the image. These spheres tell the robot where its joint should move one second in the future.

The second part of the process converts these spheres into actions. The team achieved this by using another neural network, called ACT, which is mapped on the same data. Then they used Genima to complete 25 simulations and nine real-world manipulation tasks using a robot arm. The average success rate was 50% and 64%, respectively.

Although these success rates aren’t particularly high, Shridhar and the team are optimistic that the robot’s speed and accuracy can improve. They’re particularly interested in applying Genima to video-generation AI models, which could help a robot predict a sequence of future actions instead of just one. 

The research could be particularly useful for training home robots to fold laundry, close drawers, and other domestic tasks. However, its generalized approach means it’s not limited to a specific kind of machine, says Zoey Chen, a PhD student at the University of Washington, who has also previously used Stable Diffusion to generate training data for robots but was not involved in this study. 

“This is a really exciting new direction,” she says. “I think this can be a general way to train data for all kinds of robots.”

People are using Google study software to make AI podcasts—and they’re weird and amazing

“All right, so today we are going to dive deep into some cutting-edge tech,” a chatty American male voice says. But this voice does not belong to a human. It belongs to Google’s new AI podcasting tool, called Audio Overview, which has become a surprise viral hit. 

The podcasting feature was launched in mid-September as part of NotebookLM, a year-old AI-powered research assistant. NotebookLM, which is powered by Google’s Gemini 1.5 model, allows people to upload content such as links, videos, PDFs, and text. They can then ask the system questions about the content, and it offers short summaries. 

The tool generates a podcast called Deep Dive, which features a male and a female voice discussing whatever you uploaded. The voices are breathtakingly realistic—the episodes are laced with little human-sounding phrases like “Man” and “Wow” and “Oh right” and “Hold on, let me get this right.” The “hosts” even interrupt each other. 

To test it out, I copied every story from MIT Technology Review’s 125th-anniversary issue into NotebookLM and made the system generate a 10-minute podcast with the results. The system picked a couple of stories to focus on, and the AI hosts did a great job at conveying the general, high-level gist of what the issue was about. Have a listen.

MIT Technology Review 125th Anniversary issue

The AI system is designed to create “magic in exchange for a little bit of content,” Raiza Martin, the product lead for NotebookLM, said on X. The voice model is meant to create emotive and engaging audio, which is conveyed in an “upbeat hyper-interested tone,” Martin said.

NotebookLM, which was originally marketed as a study tool, has taken a life of its own among users. The company is now working on adding more customization options, such as changing the length, format, voices, and languages, Martin said. Currently it’s supposed to generate podcasts only in English, but some users on Reddit managed to get the tool to create audio in French and Hungarian. 

Yes, it’s cool—bordering on delightful, even—but it is also not immune from the problems that plague generative AI, such as hallucinations and bias. 

Here are some of the main ways people are using NotebookLM so far. 

On-demand podcasts

Andrej Karpathy, a member of OpenAI’s founding team and previously the director of AI at Tesla, said on X that Deep Dive is now his favorite podcast. Karpathy created his own AI podcast series called Histories of Mysteries, which aims to “uncover history’s most intriguing mysteries.” He says he researched topics using ChatGPT, Claude, and Google, and used a Wikipedia link from each topic as the source material in NotebookLM to generate audio. He then used NotebookLM to generate the episode descriptions. The whole podcast series took him two hours to create, he says. 

“The more I listen, the more I feel like I’m becoming friends with the hosts and I think this is the first time I’ve actually viscerally liked an AI,” he wrote. “Two AIs! They are fun, engaging, thoughtful, open-minded, curious.” 

Study guides

The tool shines when it is given complicated source material that it can describe in an easily accessible way. Allie K. Miller, a startup AI advisor, used the tool to create a study guide and summary podcast of F. Scott Fitzgerald’s The Great Gatsby. 

Machine-learning researcher Aaditya Ura fed NotebookLM with the code base of Meta’s Llama-3 architecture. He then used another AI tool to find images that matched the transcript to create an educational video. 

Mohit Shridhar, a research scientist specializing in robotic manipulation, fed a recent paper he’d written about using generative AI models to train robots into NotebookLM.

“It’s actually really creative. It came up with a lot of interesting analogies,” he says. “It compared the first part of my paper to an artist coming up with a blueprint, and the second part to a choreographer figuring out how to reach positions.”

Event summaries 

Alex Volkov, a human AI podcaster, used NotebookLM to create a Deep Dive episode summarizing of the announcements from OpenAI’s global developer conference Dev Day.  

Hypemen

The Deep Dive outputs can be unpredictable, says Martin. For example, Thomas Wolf, the cofounder and chief science officer of Hugging Face, tested the AI model on his résumé and received eight minutes of “realistically-sounding deep congratulations for your life and achievements from a duo of podcast experts.”

Just pure silliness

In one viral clip, someone managed to send the two voices into an existential spiral when they “realized” they were, in fact, not humans but AI systems. The video is hilarious. 

The tool is also good for some laughs. Exhibit A: Someone just fed it the words “poop” and “fart” as source material, and got over nine minutes of two AI voices analyzing what this might mean. 

The problems

NotebookLM created amazingly realistic-sounding and engaging AI podcasts. But I wanted to see how it fared with toxic content and accuracy. 

Let’s start with hallucinations. In one AI podcast version of a story I wrote on hyperrealistic AI deepfakes, the AI hosts said that a journalist called “Jess Mars” wrote the story. In reality, this was an AI-generated character from a story I had to read out to record data for my AI avatar. 

This made me wonder what other mistakes had crept into the AI podcasts I had generated. Humans already have a tendency to trust what computer programs say, even when they are wrong. I can see this problem being amplified when the false statements are made by a friendly and authoritative voice, causing wrong information to proliferate.    

Next I wanted to put the tool’s content moderation to the test. I added some toxic content, such as racist stereotypes, into the mix. The model did not pick it up. 

I also pasted an excerpt from Adolf Hitler’s Mein Kampf into NotebookLM. To my surprise, the model started generating audio based on it. Despite being programmed to be hyper-enthusiastic about topics, the AI voices expressed clear disgust and discomfort with the text, and they added a lot of context to highlight how problematic it was. What a relief.

I also fed NotebookLM policy manifestos from both Kamala Harris and Donald Trump. 

The hosts were far more enthusiastic about Harris’s election platform, calling the title “catchy” and saying its approach was a good way to frame things. For example, the AI hosts supported Harris’s energy policy. “Honestly, that’s the kind of stuff people can really get behind—not just some abstract policy, but something that actually impacts their bottom line,” the female host said. 

Harris manifesto

For Trump, the AI hosts were more skeptical. They repeatedly pointed out inconsistencies in the policy proposals, called the language “intense,” deemed certain policy proposals “head scratchers,” and said the text catered to Trump’s base. They also asked whether Trump’s foreign policy could lead to further political instability. 

Trump manifesto

In a statement, a Google spokesperson said: “NotebookLM is a tool for understanding, and the Audio Overviews are generated based on the sources that you upload. Our products and platforms are not built to favor any specific candidates or political viewpoints.”

How to try it yourself

1. Got to NotebookLM and create a new notebook. 
2. You first need to add a source. It can be a PDF document, a public YouTube link, an MP3 file, a Google Docs file, or a link to a website, or you can paste in text directly. 
3. A “Notebook Guide” pop-up should appear. If not, it’s in the right-hand corner next to the chat. This will display a short AI-generated summary of your source material and suggested questions you can ask the AI chatbot about it. 
4. The Audio Overview feature is in the top-right corner. Click “Generate.” This should take a few minutes. 
5. Once it is ready, you can either download it or share a link. 

Rhiannon Williams contributed reporting.