ChatGPT can now search the web for up-to-date answers to a user’s queries, OpenAI announced today. 

Until now, ChatGPT was mostly restricted to generating answers from its training data, which is current up to October 2023 for GPT-4o, and had limited web search capabilities. Searches about generalized topics will still draw on this information from the model itself, but now ChatGPT will automatically search the web in response to queries about recent information such as sports, stocks, or news of the day, and can deliver rich multi-media results. Users can also manually trigger a web search, but for the most part, the chatbot will make its own decision about when an answer would benefit from information taken from the web, says Adam Fry, OpenAI’s product lead for search.

“Our goal is to make ChatGPT the smartest assistant, and now we’re really enhancing its capabilities in terms of what it has access to from the web,” Fry tells MIT Technology Review. The feature is available today for the chatbot’s paying users. 

While ChatGPT search, as it is known, is initially available to paying customers, OpenAI intends to make it available for free later, even when people are logged out. The company also plans to combine search with its voice features and Canvas, its interactive platform for coding and writing, although these capabilities will not be available in today’s initial launch.

The company unveiled a standalone prototype of web search in July. Those capabilities are now built directly into the chatbot. OpenAI says it has “brought the best of the SearchGPT experience into ChatGPT.” 

OpenAI is the latest tech company to debut an AI-powered search assistant, challenging similar tools from competitors such as Google, Microsoft, and startup Perplexity. Meta, too, is reportedly developing its own AI search engine. As with Perplexity’s interface, users of ChatGPT search can interact with the chatbot in natural language, and it will offer an AI-generated answer with sources and links to further reading. In contrast, Google’s AI Overviews offer a short AI-generated summary at the top of the website, as well as a traditional list of indexed links. 

These new tools could eventually challenge Google’s 90% market share in online search. AI search is a very important way to draw more users, says Chirag Shah, a professor at the University of Washington, who specializes in online search. But he says it is unlikely to chip away at Google’s search dominance. Microsoft’s high-profile attempt with Bing barely made a dent in the market, Shah says. 

Instead, OpenAI is trying to create a new market for more powerful and interactive AI agents, which can take complex actions in the real world, Shah says. 

The new search function in ChatGPT is a step toward these agents. 

It can also deliver highly contextualized responses that take advantage of chat histories, allowing users to go deeper in a search. Currently, ChatGPT search is able to recall conversation histories and continue the conversation with questions on the same topic. 

ChatGPT itself can also remember things about users that it can use later —sometimes it does this automatically, or you can ask it to remember something. Those “long-term” memories affect how it responds to chats. Search doesn’t have this yet—a new web search starts from scratch— but it should get this capability in the “next couple of quarters,” says Fry. When it does, OpenAI says it will allow it to deliver far more personalized results based on what it knows.

“Those might be persistent memories, like ‘I’m a vegetarian,’ or it might be contextual, like ‘I’m going to New York in the next few days,’” says Fry. “If you say ‘I’m going to New York in four days,’ it can remember that fact and the nuance of that point,” he adds. 

To help develop ChatGPT’s web search, OpenAI says it leveraged its partnerships with news organizations such as Reuters, the Atlantic, Le Monde, the Financial Times, Axel Springer, Condé Nast, and Time. However, its results include information not only from these publishers, but any other source online that does not actively block its search crawler.   

It’s a positive development that ChatGPT will now be able to retrieve information from these reputable online sources and generate answers based on them, says Suzan Verberne, a professor of natural-language processing at Leiden University, who has studied information retrieval. It also allows users to ask follow-up questions.

But despite the enhanced ability to search the web and cross-check sources, the tool is not immune from the persistent tendency of AI language models to make things up or get it wrong. When MIT Technology Review tested the new search function and asked it for vacation destination ideas, ChatGPT suggested “luxury European destinations” such as Japan, Dubai, the Caribbean islands, Bali, the Seychelles, and Thailand. It offered as a source an article from the Times, a British newspaper, which listed these locations as well as those in Europe as luxury holiday options.

“Especially when you ask about untrue facts or events that never happened, the engine might still try to formulate a plausible response that is not necessarily correct,” says Verberne. There is also a risk that misinformation might seep into ChatGPT’s answers from the internet if the company has not filtered its sources well enough, she adds. 

Another risk is that the current push to access the web through AI search will disrupt the internet’s digital economy, argues Benjamin Brooks, a fellow at Harvard University’s Berkman Klein Center, who previously led public policy for Stability AI, in an op-ed published by MIT Technology Review today.

“By shielding the web behind an all-knowing chatbot, AI search could deprive creators of the visits and ‘eyeballs’ they need to survive,” Brooks writes.

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

War is a catalyst for change, an expert in AI and warfare told me in 2022. At the time, the war in Ukraine had just started, and the military AI business was booming. Two years later, things have only ramped up as geopolitical tensions continue to rise.

Silicon Valley players are poised to benefit. One of them is Palmer Luckey, the founder of the virtual-reality headset company Oculus, which he sold to Facebook for $2 billion. After Luckey’s highly public ousting from Meta, he founded Anduril, which focuses on drones, cruise missiles, and other AI-enhanced technologies for the US Department of Defense. The company is now valued at $14 billion. My colleague James O’Donnell interviewed Luckey about his new pet project: headsets for the military. 

Luckey is increasingly convinced that the military, not consumers, will see the value of mixed-reality hardware first: “You’re going to see an AR headset on every soldier, long before you see it on every civilian,” he says. In the consumer world, any headset company is competing with the ubiquity and ease of the smartphone, but he sees entirely different trade-offs in defense. Read the interview here. 

The use of AI for military purposes is controversial. Back in 2018, Google pulled out of the Pentagon’s Project Maven, an attempt to build image recognition systems to improve drone strikes, following staff walkouts over the ethics of the technology. (Google has since returned to offering services for the defense sector.) There has been a long-standing campaign to ban autonomous weapons, also known as “killer robots,” which powerful militaries such as the US have refused to agree to.  

But the voices that boom even louder belong to an influential faction in Silicon Valley, such as Google’s former CEO Eric Schmidt, who has called for the military to adopt and invest more in AI to get an edge over adversaries. Militaries all over the world have been very receptive to this message.

That’s good news for the tech sector. Military contracts are long and lucrative, for a start. Most recently, the Pentagon purchased services from Microsoft and OpenAI to do search, natural-language processing, machine learning, and data processing, reports The Intercept. In the interview with James, Palmer Luckey says the military is a perfect testing ground for new technologies. Soldiers do as they are told and aren’t as picky as consumers, he explains. They’re also less price-sensitive: Militaries don’t mind spending a premium to get the latest version of a technology.

But there are serious dangers in adopting powerful technologies prematurely in such high-risk areas. Foundation models pose serious national security and privacy threats by, for example, leaking sensitive information, argue researchers at the AI Now Institute and Meredith Whittaker, president of the communication privacy organization Signal, in a new paper. Whittaker, who was a core organizer of the Project Maven protests, has said that the push to militarize AI is really more about enriching tech companies than improving military operations. 

Despite calls for stricter rules around transparency, we are unlikely to see governments restrict their defense sectors in any meaningful way beyond voluntary ethical commitments. We are in the age of AI experimentation, and militaries are playing with the highest stakes of all. And because of the military’s secretive nature, tech companies can experiment with the technology without the need for transparency or even much accountability. That suits Silicon Valley just fine. 

---

Now read the rest of The Algorithm

Deeper Learning

How Wayve’s driverless cars will meet one of their biggest challenges yet

The UK driverless-car startup Wayve is headed west. The firm’s cars learned to drive on the streets of London. But Wayve has announced that it will begin testing its tech in and around San Francisco as well. And that brings a new challenge: Its AI will need to switch from driving on the left to driving on the right.

Full speed ahead: As visitors to or from the UK will know, making that switch is harder than it sounds. Your view of the road, how the vehicle turns—it’s all different. The move to the US will be a test of Wayve’s technology, which the company claims is more general-purpose than what many of its rivals are offering. Across the Atlantic, the company will now go head to head with the heavyweights of the growing autonomous-car industry, including Cruise, Waymo, and Tesla. Join Will Douglas Heaven on a ride in one of its cars to find out more. 

Bits and Bytes

Kids are learning how to make their own little language models
Little Language Models is a new application from two PhD researchers at MIT’s Media Lab that helps children understand how AI models work—by getting to build small-scale versions themselves. (MIT Technology Review) 

Google DeepMind is making its AI text watermark open source
Google DeepMind has developed a tool for identifying AI-generated text called SynthID, which is part of a larger family of watermarking tools for generative AI outputs. The company is applying the watermark to text generated by its Gemini models and making it available for others to use too. (MIT Technology Review) 

Anthropic debuts an AI model that can “use” a computer
The tool enables the company’s Claude AI model to interact with computer interfaces and take actions such as moving a cursor, clicking on things, and typing text. It’s a very cumbersome and error-prone version of what some have said AI agents will be able to do one day. (Anthropic) 

Can an AI chatbot be blamed for a teen’s suicide?
A 14-year-old boy committed suicide, and his mother says it was because he was obsessed with an AI chatbot created by Character.AI. She is suing the company. Chatbots have been touted as cures for loneliness, but critics say they actually worse isolation.  (The New York Times) 

Google, Microsoft, and Perplexity are promoting scientific racism in search results
The internet’s biggest AI-powered search engines are featuring the widely debunked idea that white people are genetically superior to other races. (Wired) 

Google DeepMind has developed a tool for identifying AI-generated text and is making it available open source. 

The tool, called SynthID, is part of a larger family of watermarking tools for generative AI outputs. The company unveiled a watermark for images last year, and it has since rolled out one for AI-generated video. In May, Google announced it was applying SynthID in its Gemini app and online chatbots and made it freely available on Hugging Face, an open repository of AI data sets and models. Watermarks have emerged as an important tool to help people determine when something is AI generated, which could help counter harms such as misinformation. 

“Now, other [generative] AI developers will be able to use this technology to help them detect whether text outputs have come from their own [large language models], making it easier for more developers to build AI responsibly,” says Pushmeet Kohli, the vice president of research at Google DeepMind. 

SynthID works by adding an invisible watermark directly into the text when it is generated by an AI model. 

Large language models work by breaking down language into “tokens” and then predicting which token is most likely to follow the other. Tokens can be a single character, word, or part of a phrase, and each one gets a percentage score for how likely it is to be the appropriate next word in a sentence. The higher the percentage, the more likely the model is going to use it. 

SynthID introduces additional information at the point of generation by changing the probability that tokens will be generated, explains Kohli. 

To detect the watermark and determine whether text has been generated by an AI tool, SynthID compares the expected probability scores for words in watermarked and unwatermarked text. 

Google DeepMind found that using the SynthID watermark did not compromise the quality, accuracy, creativity, or speed of generated text. That conclusion was drawn from a massive live experiment of SynthID’s performance after the watermark was deployed in its Gemini products and used by millions of people. Gemini allows users to rank the quality of the AI model’s responses with a thumbs-up or a thumbs-down. 

Kohli and his team analyzed the scores for around 20 million watermarked and unwatermarked chatbot responses. They found that users did not notice a difference in quality and usefulness between the two. The results of this experiment are detailed in a paper published in Nature today. Currently SynthID for text only works on content generated by Google’s models, but the hope is that open-sourcing it will expand the range of tools it’s compatible with. 

SynthID does have other limitations. The watermark was resistant to some tampering, such as cropping text and light editing or rewriting, but it was less reliable when AI-generated text had been rewritten or translated from one language into another. It is also less reliable in responses to prompts asking for factual information, such as the capital city of France. This is because there are fewer opportunities to adjust the likelihood of the next possible word in a sentence without changing facts. 

“Achieving reliable and imperceptible watermarking of AI-generated text is fundamentally challenging, especially in scenarios where LLM outputs are near deterministic, such as factual questions or code generation tasks,” says Soheil Feizi, an associate professor at the University of Maryland, who has studied the vulnerabilities of AI watermarking.  

Feizi says Google DeepMind’s decision to open-source its watermarking method is a positive step for the AI community. “It allows the community to test these detectors and evaluate their robustness in different settings, helping to better understand the limitations of these techniques,” he adds. 

There is another benefit too, says João Gante, a machine-learning engineer at Hugging Face. Open-sourcing the tool means anyone can grab the code and incorporate watermarking into their model with no strings attached, Gante says. This will improve the watermark’s privacy, as only the owner will know its cryptographic secrets. 

“With better accessibility and the ability to confirm its capabilities, I want to believe that watermarking will become the standard, which should help us detect malicious use of language models,” Gante says. 

But watermarks are not an all-purpose solution, says Irene Solaiman, Hugging Face’s head of global policy. 

“Watermarking is one aspect of safer models in an ecosystem that needs many complementing safeguards. As a parallel, even for human-generated content, fact-checking has varying effectiveness,” she says. 

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

I’ve recently been feeling heartbroken. A very close friend recently cut off contact with me. I don’t really understand why, and my attempts at fixing the situation have backfired. Situations like this are hurtful and confusing. So it’s no wonder that people are increasingly turning to AI chatbots to help solve them. And there’s good news: AI might actually be able to help. 

Researchers from Google DeepMind recently trained a system of large language models to help people come to agreement over complex but important social or political issues. The AI model was trained to identify and present areas where people’s ideas overlapped. With the help of this AI mediator, small groups of study participants became less divided in their positions on various issues. You can read more from Rhiannon Williams here.   

One of the best uses for AI chatbots is for brainstorming. I’ve had success in the past using them to draft more assertive or persuasive emails for awkward situations, such as complaining about services or negotiating bills. This latest research suggests they could help us to see things from other people’s perspectives too. So why not use AI to patch things up with my friend? 

I described the conflict, as I see it, to ChatGPT and asked for advice about what I should do. The response was very validating, because the AI chatbot supported the way I had approached the problem. The advice it gave was along the lines of what I had thought about doing anyway. I found it helpful to chat with the bot and get more ideas about how to deal with my specific situation. But ultimately, I was left dissatisfied, because the advice was still pretty generic and vague (“Set your boundary calmly” and “Communicate your feelings”) and didn’t really offer the kind of insight a therapist might. 

And there’s another problem: Every argument has two sides. I started a new chat, and described the problem as I believe my friend sees it. The chatbot supported and validated my friend’s decisions, just as it did for me. On one hand, this exercise helped me see things from her perspective. I had, after all, tried to empathize with the other person, not just win an argument. But on the other hand, I can totally see a situation where relying too much on the advice of a chatbot that tells us what we want to hear could cause us to double down, preventing us from seeing things from the other person’s perspective. 

This served as a good reminder: An AI chatbot is not a therapist or a friend. While it can parrot the vast reams of internet text it’s been trained on, it doesn’t understand what it’s like to feel sadness, confusion, or joy. That’s why I would tread with caution when using AI chatbots for things that really matter to you, and not take what they say at face value. 

An AI chatbot can never replace a real conversation, where both sides are willing to truly listen and take the other’s point of view into account. So I decided to ditch the AI-assisted therapy talk and reached out to my friend one more time. Wish me luck! 

---

Now read the rest of The Algorithm

Deeper Learning

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 1,000 responses on average, and as many as one in 100 responses in the worst case.

Why this matters: 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. But the rise of chatbots, which enable individuals to interact with models directly, brings a new spin to the problem. Read more from Will Douglas Heaven. 

Bits and Bytes

Intro to AI: a beginner’s guide to artificial intelligence from MIT Technology Review
There is an overwhelming amount of AI news, and it is a lot to keep up with. Do 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. 

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: a lack of data. (MIT Technology Review) 

Cracks are starting to appear in Microsoft’s “bromance” with OpenAI 
As part of OpenAI’s transition from a research lab to a for-profit company, it has tried to renegotiate its deal with Microsoft to secure more computing power and funding. Meanwhile, Microsoft has started to invest in other AI projects, such as DeepMind cofounder Mustafa Suleyman’s Inflection AI, to reduce its reliance on OpenAI—much to Sam Altman’s chagrin. 
(The New York Times) 

Millions of people are using abusive AI “nudify” bots on Telegram 
The messaging app is a hotbed for popular AI bots that “remove clothes” from photos of people to create nonconsensual deepfake images. (Wired) 

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. 

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

David Baker is sleep-deprived but happy. He’s just won the Nobel prize, after all. 

The call from the Royal Swedish Academy of Sciences woke him in the middle of the night. Or rather, his wife did. She answered the phone at their home in Washington, D.C. and screamed that he’d won the Nobel Prize for Chemistry. The prize is the ultimate recognition of his work as a biochemist at the University of Washington.

“I woke up at two [a.m.] and basically didn’t sleep through the whole day, which was all parties and stuff,” he told me the day after the announcement. “I’m looking forward to getting back to normal a little bit today.”

Last week was a major milestone for AI, with two Nobel prizes awarded for AI-related discoveries. 

Baker wasn’t alone in winning the Nobel Prize for Chemistry. The Royal Swedish Academy of Sciences awarded it to Demis Hassabis, the cofounder and CEO of Google DeepMind, and John M. Jumper, a director at the same company, too. Google DeepMind was awarded for its research on AlphaFold, a tool which can predict how proteins are structured, while Baker was recognized for his work using AI to design new proteins. Read more about it here. 

Meanwhile, the physics prize went to Geoffrey Hinton, a computer scientist whose pioneering work on deep learning in the 1980s and ’90s underpins all of the most powerful AI models in the world today, and fellow computer scientist John Hopfield, who invented a type of pattern-matching neural network that can store and reconstruct data. Read more about it here.

Speaking to reporters after the prize was announced, Hassabis said he believes that it will herald more AI tools being used for significant scientific discoveries. 

But there is one problem. AI needs masses of high-quality data to be useful for science, and databases containing that sort of data are rare, says Baker. 

The prize is a recognition for the whole community of people working as protein designers. It will help move protein design from the “lunatic fringe of stuff that no one ever thought would be useful for anything to being at the center stage,” he says.  

AI has been a gamechanger for biochemists like Baker. Seeing what DeepMind was able to do with AlphaFold made it clear that deep learning was going to be a powerful tool for their work. 

“There’s just all these problems that were really hard before that we are now having much more success with thanks to generative AI methods. We can do much more complicated things,” Baker says. 

Baker is already busy at work. He says his team is focusing on designing enzymes, which carry out all the chemical reactions that living things rely upon to exist. His team is also working on medicines that only act at the right time and place in the body. 

But Baker is hesitant in calling this a watershed moment for AI in science. 

In AI there’s a saying: Garbage in, garbage out. If the data that is fed into AI models is not good, the outcomes won’t be dazzling either. 

The power of the Chemistry Nobel Prize-winning AI tools lies in the Protein Data Bank (PDB), a rare treasure trove of high-quality, curated and standardized data. This is exactly the kind of data that AI needs to do anything useful. But the current trend in AI development is training ever-larger models on the entire content of the internet, which is increasingly full of AI-generated slop. This slop in turn gets sucked into datasets and pollutes the outcomes, leading to bias and errors. That’s just not good enough for rigorous scientific discovery.

“If there were many databases as good as the PDB, I would say, yes, this [prize] probably is just the first of many, but it is kind of a unique database in biology,” Baker says. “It’s not just the methods, it’s the data. And there aren’t so many places where we have that kind of data.”

---

Now read the rest of The Algorithm

Deeper Learning

Adobe wants to make it easier for artists to blacklist their work from AI scraping

Adobe has announced a new tool to help creators watermark their work and opt out of having it used to train generative AI models. The web app, called Adobe Content Authenticity, also gives artists the opportunity to add “content credentials,” including their verified identity, social media handles, or other online domains, to their work.

A digital signature: Content credentials are based on C2PA, an internet protocol that uses cryptography to securely label images, video, and audio with information clarifying where they came from—the 21st-century equivalent of an artist’s signature. Creators can apply them to their content regardless of whether it was created using Adobe tools. The company is launching a public beta in early 2025. Read more from Rhiannon Williams here.

Bits and Bytes

Why artificial intelligence and clean energy need each other
A geopolitical battle is raging over the future of AI. The key to winning it is a clean-energy revolution, argue Michael Kearney and Lisa Hansmann, from Engine Ventures, a firm that invests in startups commercializing breakthrough science and engineering. They believe that AI’s huge power demands represent a chance to scale the next generation of clean energy technologies. (MIT Technology Review)

The state of AI in 2025
AI investor Nathan Benaich and Air Street Capital have released their annual analysis of the state of AI. Their predictions for the next year? Big, proprietary models will start to lose their edge, and labs will focus more on planning and reasoning. Perhaps unsurprisingly, the investor also bets that a handful of AI companies will begin to generate serious revenue. 

Silicon Valley, the new lobbying monster
Big Tech’s tentacles reach everywhere in Washington DC. This is a fascinating look at how tech companies lobby politicians to influence how AI is regulated in the United States.  (The New Yorker) 

In a second Nobel win for AI, the Royal Swedish Academy of Sciences has awarded half the 2024 prize in chemistry to Demis Hassabis, the cofounder and CEO of Google DeepMind, and John M. Jumper, a director at the same company, for their work on using artificial intelligence to predict the structures of proteins. The other half goes to David Baker, a professor of biochemistry at the University of Washington, for his work on computational protein design. The winners will share a prize pot of 11 million Swedish kronor ($1 million). 

The potential impact of this research is enormous. Proteins are fundamental to life, but understanding what they do involves figuring out their structure—a very hard puzzle that once took months or years to crack for each type of protein. By cutting down the time it takes to predict a protein’s structure, computational tools such as those developed by this year’s award winners are helping scientists gain a greater understanding of how proteins work and opening up new avenues of research and drug development. The technology could unlock more efficient vaccines, speed up research on cures for cancer, or lead to completely new materials.

Hassabis and Jumper created AlphaFold, which in 2020 solved a problem scientists have been wrestling with for decades: predicting the three-dimensional structure of a protein from a sequence of amino acids. The AI tool has since been used to predict the shapes of all proteins known to science.

Their latest model, AlphaFold 3, can predict the structures of DNA, RNA, and molecules like ligands, which are essential to drug discovery. DeepMind has also released the source code and database of its results to scientists for free. 

“I’ve dedicated my career to advancing AI because of its unparalleled potential to improve the lives of billions of people,” said Demis Hassabis. “AlphaFold has already been used by more than two million researchers to advance critical work, from enzyme design to drug discovery. I hope we’ll look back on AlphaFold as the first proof point of AI’s incredible potential to accelerate scientific discovery,” he added.

Baker has created several AI tools for designing and predicting the structure of proteins, such as a family of programs called Rosetta. In 2022, his lab created an open-source AI tool called ProteinMPNN that could help researchers discover previously unknown proteins and design entirely new ones. It helps researchers who have an exact protein structure in mind find amino acid sequences that fold into that shape.

Most recently, in late September, Baker’s lab announced it had developed custom molecules that allow scientists to precisely target and eliminate proteins associated with diseases in living cells. 

“[Proteins] evolved over the course of evolution to solve the problems that organisms faced during evolution. But we face new problems today, like covid. If we could design proteins that were as good at solving new problems as the ones that evolved during evolution are at solving old problems, it would be really, really powerful,” Baker told MIT Technology Review in 2022.  

This article has been updated with a quote from Demis Hassabis.

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

Chatting with an AI chatbot is so 2022. The latest hot AI toys take advantage of multimodal models, which can handle several things at the same time, such as images, audio, and text. 

Exhibit A: Google’s NotebookLM. NotebookLM is a research tool the company launched with little fanfare a year ago. A few weeks ago, Google added an AI podcasting tool called Audio Overview to NotebookLM, which allows users to create podcasts about anything. Add a link to, for example, your LinkedIn profile, and the AI podcast hosts will boost your ego for nine minutes. The feature has become a surprise viral hit. I wrote about all the weird and amazing ways people are using it here. 

To give you a taste, I created a podcast of our 125th-anniversary magazine issue. The AI does a great job of picking some highlights from the magazine and giving you the gist of what they are about. Have a listen below. 

Multimodal generative content has also become markedly better in a very short time. In September 2022, I covered Meta’s first text-to-video model, Make-A-Video. Next to today’s technology, those videos look clunky and silly. Meta just announced its competitor to OpenAI’s Sora, called Movie Gen. The tool allows users to use text prompts to create custom videos and sounds, edit existing videos, and make images into videos.

The way we interact with AI systems is also changing, becoming less reliant on text. OpenAI’s new Canvas interface allows users to collaborate on projects with ChatGPT. Instead of relying on a traditional chat window, which requires users to do several rounds of prompting and regenerating text to get the desired result, Canvas allows people to select bits of text or code to edit. 

Even search is getting a multimodal upgrade. In addition to inserting ads into AI overviews, Google has rolled out a new feature where users can upload a video and use their voice to search for things. In a demo at Google I/O, the company showed how you can open the Google Lens app, take a video of fish swimming in an aquarium, and ask a question about them. Google’s Gemini model will then search the web and offer you an answer in the form of Google’s AI summary. 

What unites these features is a more interactive, customizable interface and the ability to apply AI tools to lots of different types of source material. NotebookLM was the first AI product in a while that brought me wonder and delight, partly because of how different, realistic, and unexpected the AI voices were. But the fact that NotebookLM’s Audio Overviews became a hit despite being a side feature hidden inside a bigger product just goes to show that AI developers don’t really know what they are doing. Hard to believe now, but ChatGPT itself was an unexpected hit for OpenAI.

We are a couple of years into the multibillion-dollar generative AI boom. The huge investment in AI has contributed to rapid improvement in the quality of the resulting content. But we’ve yet to see a killer app, and these new multimodal applications are a result of the immense pressure AI companies are under to make money and deliver. Tech companies are throwing different AI tools at people and seeing what sticks. 

---

Now read the rest of The Algorithm

Deeper Learning

AI-generated images can teach robots how to act

Image-generating AI models have been used to  create training data for robots. The new system, called Genima,  fine-tunes the image-generating AI model Stable Diffusion to draw robots’ movements, helping guide them both in simulations and in the real world. 

What’s the big deal: Genima 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. Read more from Rhiannon Williams here. 

Bits and Bytes

This startup uses AI to detect wildfires 
Our 2024 list of Climate Tech Companies to Watch is here! One company on the list is Pano AI, which uses computer vision and ultra-high-definition cameras to alert firefighters to new blazes. (MIT Technology Review) 

How Sam Altman concentrated power to his own hands
And then there was one. With OpenAI now valued at $157 billion, Bloomberg details how the company lost most of its top executives and shifted to an Altman-led profit-making monster.  (Bloomberg) 

Eight scientists, a billion dollars, and the moonshot agency trying to make Britain great again
A nice profile on the UK’s new Advanced Research and Invention Agency, or ARIA. The agency is the UK’s answer to DARPA in the US. It is funding projects such as Turing Award winner Yoshua Bengio’s project to prevent AI catastrophes. (Wired) 

Why women in tech are sounding an alarm
Tech’s AI mania is encouraging the field to backtrack on years of diversity and inclusion efforts, at the expense of women. (The Information) 

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

In AI research, everyone seems to think that bigger is better. The idea is that more data, more computing power, and more parameters will lead to models that are more powerful. This thinking started with a landmark paper from 2017, in which Google researchers introduced the transformer architecture underpinning today’s language model boom and helped embed the “scale is all you need” mindset into the AI community. Today, big tech companies seem to be competing over scale above everything else.

“It’s like, how big is your model, bro?” says Sasha Luccioni, the AI and climate lead at the AI startup Hugging Face. Tech companies just add billions more parameters, which means an average person couldn’t download the models and tinker with them, even if they were open-source (which they mostly aren’t). The AI models of today are just “way too big,” she says.  

With scale come a slew of problems, such as invasive data-gathering practices and child sexual abuse material in data sets, as Luccioni and coauthors detail in a new paper. To top it off, bigger models also have a far bigger carbon footprint, because they require more energy to run. 

Another problem that scale brings is the extreme concentration of power, says Luccioni. Scaling up costs tons of money, and only elite researchers working in Big Tech have the resources to build and operate models like that. 

“There’s this bottleneck that’s created by a very small number of rich and powerful companies who use AI as part of their core product,” she says. 

It doesn’t have to be like this. I just published a story on a new multimodal large language model that is small but mighty. Researchers at the Allen Institute for Artificial Intelligence (Ai2) built an open-source family of models called Molmo, which achieve impressive performance with a fraction of the resources used to build state-of-the-art models. 

The organization claims that its biggest Molmo model, which has 72 billion parameters, outperforms OpenAI’s GPT-4o, which is estimated to have over a trillion parameters, in tests that measure things like understanding images, charts, and documents.  

Meanwhile, Ai2 says a smaller Molmo model, with 7 billion parameters, comes close to OpenAI’s state-of-the-art model in performance, an achievement it ascribes to vastly more efficient data collection and training methods. Read more about it from me here. Molmo shows we don’t need massive data sets and massive models that take tons of money and energy to train. 

Breaking out of the “scale is all you need” mindset was one of the biggest challenges for the researchers who built Molmo, says Ani Kembhavi, a senior director of research at Ai2. 

“When we started this project, we were like, we have to think completely out of the box, because there has to be a better way to train models,” he says. The team wanted to prove that open models can be as powerful as closed, proprietary ones, and that required them to build models that were accessible and didn’t cost millions of dollars to train. 

Molmo shows that “less is more, small is big, open [is as good as] closed,” Kembhavi says. 

There’s another good case for scaling down. Bigger models tend to be able to do a wider range of things than end users actually need, says Luccioni. 

“Most of the time, you don’t need a model that does everything. You need a model that does a specific task that you want it to do. And for that, bigger models are not necessarily better,” she says.

Instead, we need to change the ways we measure AI performance to focus on things that actually matter, says Luccioni. For example, in a cancer detection algorithm, instead of using a model that can do all sorts of things and is trained on the internet, perhaps we should be prioritizing factors such as accuracy, privacy, or whether the model is trained on data that you can trust, she says. 

But that would require a higher level of transparency than is currently the norm in AI. Researchers don’t really know how or why their models do what they do, and don’t even really have a grasp of what goes into their data sets. Scaling is a popular technique because researchers have found that throwing more stuff at models seems to make them perform better. The research community and companies need to shift the incentives so that tech companies will be required to be more mindful and transparent about what goes into their models, and help us do more with less. 

“You don’t need to assume [AI models] are a magic box and going to solve all your issues,” she says. 

---

Now read the rest of The Algorithm

Deeper Learning

An AI script editor could help decide what films get made in Hollywood

Every day across Hollywood, scores of people read through scripts on behalf of studios, trying to find the diamonds in the rough among the many thousands sent in every year. Each script runs up to 150 pages, and it can take half a day to read one and write up a summary. With only about 50 of these scripts selling in a given year, readers are trained to be ruthless. 

Lights, camera, AI: Now the tech company Cinelytic, which works with major studios like Warner Bros. and Sony Pictures, aims to offer script feedback with generative AI. It launched a new tool called Callaia that analyzes scripts. Using AI, it takes Callaia less than a minute to write its own “coverage,” which includes a synopsis, a list of comparable films, grades for areas like dialogue and originality, and actor recommendations. Read more from James O’Donnell here.

Bits and Bytes

California’s governor has vetoed the state’s sweeping AI legislation
Governor Gavin Newsom vetoed SB 1047, a bill that required pre-deployment safety testing of large AI systems, and gave the state’s attorney general the right to sue AI companies for serious harm. He said he thought the bill focused too much on the largest models without considering broader harms and risks. Critics of AI’s rapid growth have expressed dismay at the decision. (The New York Times) 

Sorry, AI won’t “fix” climate change
OpenAI’s CEO Sam Altman claims AI will deliver an “Intelligence Age,” unleashing “unimaginable” prosperity and “astounding triumphs” like “fixing the climate.” But tech breakthroughs alone can’t solve global warming. In fact, as it stands, AI is making the problem much worse. (MIT Technology Review) 

How turning OpenAI into a real business is tearing it apart
In yet another organizational shakeup, the startup lost its CTO Mira Murati and other senior leaders. OpenAI is riddled with chaos that stems from its CEO’s push to transform it from a nonprofit research lab into a for-profit organization. Insiders say this shift has “corrupted” the company’s culture. (The Wall Street Journal)

Why Microsoft made a deal to help restart Three Mile Island
A once-shuttered nuclear plant could soon be used to power Microsoft’s massive investment in AI development. (MIT Technology Review) 

OpenAI released its advanced voice mode to more people. Here’s how to get it.
The company says the updated version responds to your emotions and tone of voice, and allows you to interrupt it midsentence. (MIT Technology Review) 

The FTC is cracking down on AI scams
The agency launched “Operation AI Comply” and says it will investigate AI-infused frauds and other types of deception, such as chatbots giving “legal advice,” AI tools that let people create fake online reviews, and false claims of huge earnings from AI-powered business opportunities.
(The FTC) 

Want AI that flags hateful content? Build it.
A new competition promises $10,000 in prizes to anyone who can track hateful images online. (MIT Technology Review) 

The Allen Institute for Artificial Intelligence (Ai2), a research nonprofit, is releasing a family of open-source multimodal language models, called Molmo, that it says perform as well as top proprietary models from OpenAI, Google, and Anthropic. 

The organization claims that its biggest Molmo model, which has 72 billion parameters, outperforms OpenAI’s GPT-4o, which is estimated to have over a trillion parameters, in tests that measure things like understanding images, charts, and documents.  

Meanwhile, Ai2 says a smaller Molmo model, with 7 billion parameters, comes close to OpenAI’s state-of-the-art model in performance, an achievement it ascribes to vastly more efficient data collection and training methods. 

What Molmo shows is that open-source AI development is now on par with closed, proprietary models, says Ali Farhadi, the CEO of Ai2. And open-source models have a significant advantage, as their open nature means other people can build applications on top of them. The Molmo demo is available here, and it will be available for developers to tinker with on the Hugging Face website. (Certain elements of the most powerful Molmo model are still shielded from view.) 

Other large multimodal language models are trained on vast data sets containing billions of images and text samples that have been hoovered from the internet, and they can include several trillion parameters. This process introduces a lot of noise to the training data and, with it, hallucinations, says Ani Kembhavi, a senior director of research at Ai2. In contrast, Ai2’s Molmo models have been trained on a significantly smaller and more curated data set containing only 600,000 images, and they have between 1 billion and 72 billion parameters. This focus on high-quality data, versus indiscriminately scraped data, has led to good performance with far fewer resources, Kembhavi says.

Ai2 achieved this by getting human annotators to describe the images in the model’s training data set in excruciating detail over multiple pages of text. They asked the annotators to talk about what they saw instead of typing it. Then they used AI techniques to convert their speech into data, which made the training process much quicker while reducing the computing power required. 

These techniques could prove really useful if we want to meaningfully govern the data that we use for AI development, says Yacine Jernite, who is the machine learning and society lead at Hugging Face, and was not involved in the research. 

“It makes sense that in general, training on higher-quality data can lower the compute costs,” says Percy Liang, the director of the Stanford Center for Research on Foundation Models, who also did not participate in the research. 

Another impressive capability is that the model can “point” at things, meaning it can analyze elements of an image by identifying the pixels that answer queries.

In a demo shared with MIT Technology Review, Ai2 researchers took a photo outside their office of the local Seattle marina and asked the model to identify various elements of the image, such as deck chairs. The model successfully described what the image contained, counted the deck chairs, and accurately pinpointed to other things in the image as the researchers asked. It was not perfect, however. It could not locate a specific parking lot, for example. 

Other advanced AI models are good at describing scenes and images, says Farhadi. But that’s not enough when you want to build more sophisticated web agents that can interact with the world and can, for example, book a flight. Pointing allows people to interact with user interfaces, he says. 

Jernite says Ai2 is operating with a greater degree of openness than we’ve seen from other AI companies. And while Molmo is a good start, he says, its real significance will lie in the applications developers build on top of it, and the ways people improve it.

Farhadi agrees. AI companies have drawn massive, multitrillion-dollar investments over the past few years. But in the past few months, investors have expressed skepticism about whether that investment will bring returns. Big, expensive proprietary models won’t do that, he argues, but open-source ones can. He says the work shows that open-source AI can also be built in a way that makes efficient use of money and time. 

“We’re excited about enabling others and seeing what others would build with this,” Farhadi says. 

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

If it’s not broken, don’t fix it. That’s the approach bad state actors seem to have taken when it comes to how they mess with elections around the world.

When the generative-AI boom first kicked off, one of the biggest concerns among pundits and experts was that hyperrealistic AI deepfakes could be used to influence elections. But new research from the Alan Turing Institute in the UK shows that those fears might have been overblown. AI-generated falsehoods and deepfakes seem to have had no effect on election results in the UK, France, and the European Parliament, as well as other elections around the world so far this year.

Instead of using generative AI to interfere in elections, state actors such as Russia are relying on well-established techniques—such as social bots that flood comment sections—to sow division and create confusion, says Sam Stockwell, the researcher who conducted the study. Read more about it from me here.

But one of the most consequential elections of the year is still ahead of us. In just over a month, Americans will head to the polls to choose Donald Trump or Kamala Harris as their next president. Are the Russians saving their GPUs for the US elections? 

So far, that does not seem to be the case, says Stockwell, who has been monitoring viral AI disinformation around the US elections too. Bad actors are “still relying on these well-established methods that have been used for years, if not decades, around things such as social bot accounts that try to create the impression that pro-Russian policies are gaining traction among the US public,” he says. 

And when they do try to use generative-AI tools, they don’t seem to pay off, he adds. For example, one information campaign with strong ties to Russia, called Copy Cop, has been trying to use chatbots to rewrite genuine news stories on Russia’s war in Ukraine to reflect pro-Russian narratives. 

The problem? They’re forgetting to remove the prompts from the articles they publish. 

In the short term, there are a few things that the US can do to counter more immediate harms, says Stockwell. For example, some states, such as Arizona and Colorado, are already conducting red-teaming workshops with election polling officials and law enforcement to simulate worst-case scenarios involving AI threats on Election Day. There also needs to be heightened collaboration between social media platforms, their online safety teams, fact-checking organizations, disinformation researchers, and law enforcement to ensure that viral influencing efforts can be exposed, debunked, and taken down, says Stockwell. 

But while state actors aren’t using deepfakes, that hasn’t stopped the candidates themselves. Most recently Donald Trump has used AI-generated images implying that Taylor Swift had endorsed him. (Soon after, the pop star offered her endorsement to Harris.) 

Earlier this year I wrote a piece exploring the brave new world of hyperrealistic deepfakes and what the technology is doing to our information landscape. As I wrote then, there is a real risk of creating so much skepticism and distrust in our information landscape that bad actors, or opportunistic politicians, can take advantage of this trust vacuum and lie about the authenticity of real content. This is called the “liar’s dividend.” 

There is an urgent need for guidelines on how politicians use AI. We currently lack accountability or clear red lines as to how political candidates can use AI in an ethical manner within the election context, says Stockwell. The more we see political candidates carry out practices like sharing AI-generated adverts without labels or or making accusations that other candidates’ activities are AI-generated, the more it becomes normalized, he adds. And everything we’ve seen so far suggests that these elections are only the beginning. 

---

Now read the rest of The Algorithm

Deeper Learning

AI models let robots carry out tasks in unfamiliar environments

It’s tricky to get robots to do things in environments they’ve never seen before. Typically, researchers need to train them on new data for every new place they encounter, which can become very time-consuming and expensive.

Now researchers have developed a series of AI models that teach robots to complete basic tasks in new surroundings without further training or fine-tuning. The five AI models, called robot utility models (RUMs), allow machines to complete five separate tasks—opening doors and drawers, and picking up tissues, bags, and cylindrical objects—in unfamiliar environments with a 90% success rate. This approach could make it easier and cheaper to deploy robots in our homes. Read more from Rhiannon Williams here.

Bits and Bytes

There are more than 120 AI bills in Congress right now
US policymakers have an “everything everywhere all at once” approach to regulating artificial intelligence, with bills that are as varied as the definitions of AI itself.
(MIT Technology Review)

Google is funding an AI-powered satellite constellation to spot wildfires faster
The full FireSat system should be able to detect tiny fires anywhere in the world—and provide updated images every 20 minutes. (MIT Technology Review) 

A project analyzing human language usage shut down because “generative AI has polluted the data”
Wordfreq, an open-source project that scraped the internet to analyze how humans use language, found that post-2021, there is too much AI-generated text online to make any reliable analyses. (404 Media) 

Data center emissions are probably 662% higher than Big Tech claims
AI models take a lot of energy to run and train, and tech companies have emphasized their efforts to counter their emissions. There is, however, a lot of “creative accounting” happening when it comes to calculating carbon footprints, and new analysis shows that data center emissions from these companies is likely 7.62 times higher than officially reported.
(The Guardian) 

AI-generated falsehoods and deepfakes seem to have had no effect on election results in the UK, France, and the European Parliament this year, according to new research. 

Since the beginning of the generative-AI boom, there has been widespread fear that AI tools could boost bad actors’ ability to spread fake content with the potential to interfere with elections or even sway the results. Such worries were particularly heightened this year, when billions of people were expected to vote in over 70 countries. 

Those fears seem to have been unwarranted, says Sam Stockwell, the researcher at the Alan Turing Institute who conducted the study. He focused on three elections over a four-month period from May to August 2024, collecting data on public reports and news articles on AI misuse. Stockwell identified 16 cases of AI-enabled falsehoods or deepfakes that went viral during the UK general election and only 11 cases in the EU and French elections combined, none of which appeared to definitively sway the results. The fake AI content was created by both domestic actors and groups linked to hostile countries such as Russia. 

These findings are in line with recent warnings from experts that the focus on election interference is distracting us from deeper and longer-lasting threats to democracy.   

AI-generated content seems to have been ineffective as a disinformation tool in most European elections this year so far. This, Stockwell says, is because most of the people who were exposed to the disinformation already believed its underlying message (for example, that levels of immigration to their country are too high). Stockwell’s analysis showed that people who were actively engaging with these deepfake messages by resharing and amplifying them had some affiliation or previously expressed views that aligned with the content. So the material was more likely to strengthen preexisting views than to influence undecided voters. 

Tried-and-tested election interference tactics, such as flooding comment sections with bots and exploiting influencers to spread falsehoods, remained far more effective. Bad actors mostly used generative AI to rewrite news articles with their own spin or to create more online content for disinformation purposes. 

“AI is not really providing much of an advantage for now, as existing, simpler methods of creating false or misleading information continue to be prevalent,” says Felix Simon, a researcher at the Reuters Institute for Journalism, who was not involved in the research. 

However, it’s hard to draw firm conclusions about AI’s impact upon elections at this stage, says Samuel Woolley, a disinformation expert at the University of Pittsburgh. That’s in part because we don’t have enough data.

“There are less obvious, less trackable, downstream impacts related to uses of these tools that alter civic engagement,” he adds.

Stockwell agrees: Early evidence from these elections suggests that AI-generated content could be more effective for harassing politicians and sowing confusion than changing people’s opinions on a large scale. 

Politicians in the UK, such as former prime minister Rishi Sunak, were targeted by AI deepfakes that, for example, showed them promoting scams or admitting to financial corruption. Female candidates were also targeted with nonconsensual sexual deepfake content, intended to disparage and intimidate them. 

“There is, of course, a risk that in the long run, the more that political candidates are on the receiving end of online harassment, death threats, deepfake pornographic smears—that can have a real chilling effect on their willingness to, say, participate in future elections, but also obviously harm their well-being,” says Stockwell. 

Perhaps more worrying, Stockwell says, his research indicates that people are increasingly unable to discern the difference between authentic and AI-generated content in the election context. Politicians are also taking advantage of that. For example, political candidates in the European Parliament elections in France have shared AI-generated content amplifying anti-immigration narratives without disclosing that they’d been made with AI. 

“This covert engagement, combined with a lack of transparency, presents in my view a potentially greater risk to the integrity of political processes than the use of AI by the general population or so-called ‘bad actors,’” says Simon.