A tiny new open-source AI model performs as well as powerful big ones

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. 

Want AI that flags hateful content? Build it.

Humane Intelligence, an organization focused on evaluating AI systems, is launching a competition that challenges developers to create a computer vision model that can track hateful image-based propaganda online. Organized in partnership with the Nordic counterterrorism group Revontulet, the bounty program opens September 26. It is open to anyone, 18 or older, who wants to compete and promises $10,000 in prizes for the winners.

This is the second of a planned series of 10 “algorithmic bias bounty” programs from Humane Intelligence, a nonprofit that investigates the societal impact of AI and was launched by the prominent AI researcher Rumman Chowdhury in 2022. The series is supported by Google.org, Google’s philanthropic arm.

“The goal of our bounty programs is to, number one, teach people how to do algorithmic assessments,” says Chowdhury, “but also, number two, to actually solve a pressing problem in the field.” 

Its first challenge asked participants to evaluate gaps in sample data sets that may be used to train models—gaps that may specifically produce output that is factually inaccurate, biased, or misleading. 

The second challenge deals with tracking hateful imagery online—an incredibly complex problem. Generative AI has enabled an explosion in this type of content, and AI is also deployed to manipulate content so that it won’t be removed from social media. For example, extremist groups may use AI to slightly alter an image that a platform has already banned, quickly creating hundreds of different copies that can’t easily be flagged by automated detection systems. Extremist networks can also use AI to embed a pattern into an image that is undetectable to the human eye but will confuse and evade detection systems. It has essentially created a cat-and-mouse game between extremist groups and online platforms. 

A core challenge of the project is that hate-based propaganda can be very dependent on its context. And someone who doesn’t have a deep understanding of certain symbols or signifiers may not be able to tell what even qualifies as propaganda for a white nationalist group. 

“If [the model] never sees an example of a hateful image from a part of the world, then it’s not going to be any good at detecting it,” says Jimmy Lin, a professor of computer science at the University of Waterloo, who is not associated with the bounty program.

This effect is amplified around the world, since many models don’t have a vast knowledge of cultural contexts. That’s why Humane Intelligence decided to partner with a non-US organization for this particular challenge. “Most of these models are often fine-tuned to US examples, which is why it’s important that we’re working with a Nordic counterterrorism group,” says Chowdhury.

Lin, though, warns that solving these problems may require more than algorithmic changes. “We have models that generate fake content. Well, can we develop other models that can detect fake generated content? Yes, that is certainly one approach to it,” he says. “But I think overall, in the long run, training, literacy, and education efforts are actually going to be more beneficial and have a longer-lasting impact. Because you’re not going to be subjected to this cat-and-mouse game.”

The challenge will run till November 7, 2024. Two winners will be selected, one for the intermediate challenge and another for the advanced; they will receive $4,000 and $6,000, respectively. Participants will also have their models reviewed by Revontulet, which may decide to add them to its current suite of tools to combat extremism. 

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

Every day across Hollywood, scores of film school graduates and production assistants work as script readers. Their job is to find the diamonds in the rough from the 50,000 or so screenplays pitched each year and flag any worth pursuing further. Each script runs anywhere from 100 to 150 pages, and it can take half a day to read one and write up a “coverage,” or summary of the strengths and weaknesses. With only about 50 of these scripts selling in a given year, readers are trained to be ruthless. 

Now the film-focused tech company Cinelytic, which works with major studios like Warner Bros. and Sony Pictures to analyze film budgets and box office potential, aims to offer script feedback with generative AI. 

Today it launched a new tool called Callaia, which amateur writers and professional script readers alike can use to analyze scripts at $79 each. 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. It also makes a recommendation on whether or not the film should be financed, giving it a rating of “pass,” “consider,” “recommend,” or “strongly recommend.” Though the foundation of the tool is built with ChatGPT’s API, the team had to coach the model on script-specific tasks like evaluating genres and writing a movie’s logline, which summarize the story in a sentence. 

“It helps people understand the script very quickly,” says Tobias Queisser, Cinelytic’s cofounder and CEO, who also had a career as a film producer. “You can look at more stories and more scripts, and not eliminate them based on factors that are detrimental to the business of finding great content.”

The idea is that Callaia will give studios a more analytical way to predict how a script may perform on the screen before spending on marketing or production. But, the company says, it’s also meant to ease the bottleneck that script readers create in the filmmaking process. With such a deluge to sort through, many scripts can make it to decision-makers only if they have a recognizable name attached. An AI-driven tool would democratize the script selection process and allow better scripts and writers to be discovered, Queisser says.

The tool’s introduction may further fuel the ongoing Hollywood debate about whether AI will help or harm its creatives. Since the public launch of ChatGPT in late 2022, the technology has drawn concern everywhere from writers’ rooms to special effects departments, where people worry that it will cheapen, augment, or replace human talent.  

In this case, Callaia’s success will depend on whether it can provide critical feedback as well as a human script reader can. 

That’s a challenge because of what GPT and other AI models are built to do, according to Tuhin Chakrabarty, a researcher who studied how well AI can analyze creative works during his PhD in computer science at Columbia University. In one of his studies, Chakrabarty and his coauthors had various AI models and a group of human experts—including professors of creative writing and a screenwriter—analyze the quality of 48 stories, 12 that appeared in the New Yorker and the rest of which were AI-generated. His team found that the two groups virtually never agreed on the quality of the works. 

“Whenever you ask an AI model about the creativity of your work, it is never going to say bad things,” Chakrabarty says. “It is always going to say good things, because it’s trained to be a helpful, polite assistant.”

Cinelytic CTO Dev Sen says this trait did present a hurdle in the design of Callaia, and that the initial output of the model was overly positive. That improved with time and tweaking. “We don’t necessarily want to be overly critical, but aim for a more balanced analysis that points out both strengths and weaknesses in the script,” he says. 

Vir Srinivas, an independent filmmaker whose film Orders from Above won Best Historical Film at Cannes in 2021, agreed to look at an example of Callaia’s output to see how well the AI model can analyze a script. I showed him what the model made of a 100-page script about a jazz trumpeter on a journey of self-discovery in San Francisco, which Cinelytic provided. Srinivas says that the coverage generated by the model didn’t go deep enough to present genuinely helpful feedback to a screenwriter.

“It’s approaching the script in too literal a sense and not a metaphorical one—something which human audiences do intuitively and unconsciously,” he says. “It’s as if it’s being forced to be diplomatic and not make any waves.”

There were other flaws, too. For example, Callaia predicted that the film would need a budget of just $5 to $10 million but also suggested that expensive A-listers like Paul Rudd would have been well suited for the lead role.

Cinelytic says it’s currently at work improving the actor recommendation component, and though the company did not provide data on how well its model analyzes a given script, Sen says feedback from 100 script readers who beta-tested the model was overwhelmingly positive. “Most of them were pretty much blown away, because they said that the coverages were on the order of, if not better than, the coverages they’re used to,” he says. 

Overall, Cinelytic is pitching Callaia as a tool meant to quickly provide feedback on lots of scripts, not to replace human script readers, who will still read and adjust the tool’s findings. Queisser, who is cognizant that whether AI can effectively write or edit creatively is hotly contested in Hollywood, is hopeful the tool will allow script readers to more quickly identify standout scripts while also providing an efficient source of feedback for writers.

“Writers that embrace our tool will have something that can help them refine their scripts and find more opportunities,” he says. “It’s positive for both sides.”

OpenAI released its advanced voice mode to more people. Here’s how to get it.

OpenAI is broadening access to Advanced Voice Mode, a feature of ChatGPT that allows you to speak more naturally with the AI model. It allows you to interrupt its responses midsentence, and it can sense and interpret your emotions from your tone of voice and adjust its responses accordingly. 

These features were teased back in May when OpenAI unveiled GPT-4o, but they were not released until July—and then just to an invite-only group. (At least initially, there seem to have been some safety issues with the model; OpenAI gave several Wired reporters access to the voice mode back in May, but the magazine reported that the company “pulled it the next morning, citing safety concerns.”) Users who’ve been able to try it have largely described the model as an impressively fast, dynamic, and realistic voice assistant—which has made its limited availability particularly frustrating to some other OpenAI users. 

Today is the first time OpenAI has promised to bring the new voice mode to a wide range of users. Here’s what you need to know.

What can it do? 

Though ChatGPT currently offers a standard voice mode to paid users, its interactions can be clunky. In the mobile app, for example, you can’t interrupt the model’s often long-winded responses with your voice, only with a tap on the screen. The new version fixes that, and also promises to modify its responses on the basis of the emotion it’s sensing from your voice. As with other versions of ChatGPT, users can personalize the voice mode by asking the model to remember facts about themselves. The new mode also has improved its pronunciation of words in non-English languages.

AI investor Allie Miller posted a demo of the tool in August, which highlighted a lot of the same strengths of OpenAI’s own release videos: The model is fast and adept at changing its accent, tone, and content to match your needs.

The update also adds new voices. Shortly after the launch of GPT-4o, OpenAI was criticized for the similarity between the female voice in its demo videos, named Sky, and that of Scarlett Johansson, who played an AI love interest in the movie Her. OpenAI then removed the voice. Now it has launched five new voices, named Arbor, Maple, Sol, Spruce, and Vale, which will be available in both the standard and advanced voice modes. MIT Technology Review has not heard them yet, but OpenAI says they were made using professional voice actors from around the world. “We interviewed dozens of actors to find those with the qualities of voices we feel people will enjoy talking to for hours—warm, approachable, inquisitive, with some rich texture and tone,” a company spokesperson says. 

Who can access it and when?

For now, OpenAI is rolling out access to Advanced Voice Mode to Plus users, who pay $20 per month for a premium version, and Team users, who pay $30 per month and have higher message limits. The next group to receive access will be those in the Enterprise and Edu tiers. The exact timing, though, is vague; an OpenAI spokesperson says the company will “gradually roll out access to all Plus and Team users and will roll out to Enterprise and Edu tiers starting next week.” The company hasn’t committed to a firm deadline for when all users in these categories will have access. A message in the ChatGPT app indicates that all Plus users will have access by “the end of fall.”

There are geographic limitations. The new feature is not yet available in the EU, the UK, Switzerland, Iceland, Norway, or Liechtenstein.

There is no immediate plan to release Advanced Voice Mode to free users. (The standard mode remains available to all paid users.)

What steps have been taken to make sure it’s safe?

As the company noted upon the initial release in July and again emphasized this week, Advanced Voice Mode has been safety-tested by external experts “who collectively speak a total of 45 different languages, and represent 29 different geographies.” The GPT-4o system card details how the underlying model handles issues like generating violent or erotic speech, imitating voices without their consent, or generating copyrighted content. 

Still, OpenAI’s models are not open-source. Compared with such models, which are more transparent about their training data and the “model weights” that govern how the AI produces responses, OpenAI’s closed-source models are harder for independent researchers to evaluate from the perspective of safety, bias, and harm.

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. 

The team, consisting of researchers from New York University, Meta, and the robotics company Hello Robot, hopes its findings will make it quicker and easier to teach robots new skills while helping them function within previously unseen domains. The approach could make it easier and cheaper to deploy robots in our homes.

“In the past, people have focused a lot on the problem of ‘How do we get robots to do everything?’ but not really asking ‘How do we get robots to do the things that they do know how to do—everywhere?’” says Mahi Shafiullah, a PhD student at New York University who worked on the project. “We looked at ‘How do you teach a robot to, say, open any door, anywhere?’”

Teaching robots new skills generally requires a lot of data, which is pretty hard to come by. Because robotic training data needs to be collected physically—a time-consuming and expensive undertaking—it’s much harder to build and scale training databases for robots than it is for types of AI like large language models, which are trained on information scraped from the internet.

To make it faster to gather the data essential for teaching a robot a new skill, the researchers developed a new version of a tool it had used in previous research: an iPhone attached to a cheap reacher-grabber stick, the kind typically used to pick up trash. 

The team used the setup to record around 1,000 demonstrations in 40 different environments, including homes in New York City and Jersey City, for each of the five tasks—some of which had been gathered as part of previous research. Then they trained learning algorithms on the five data sets to create the five RUM models.

These models were deployed on Stretch, a robot consisting of a wheeled unit, a tall pole, and a retractable arm holding an iPhone, to test how successfully they were able to execute the tasks in new environments without additional tweaking. Although they achieved a completion rate of 74.4%, the researchers were able to increase this to a 90% success rate when they took images from the iPhone and the robot’s head-mounted camera,  gave them to OpenAI’s recent GPT-4o LLM model, and asked it if the task had been completed successfully. If GPT-4o said no, they simply reset the robot and tried again.

A significant challenge facing roboticists is that training and testing their models in lab environments isn’t representative of what could happen in the real world, meaning research that helps machines to behave more reliably in new settings is much welcomed, says Mohit Shridhar, a research scientist specializing in robotic manipulation who wasn’t involved in the work. 

“It’s nice to see that it’s being evaluated in all these diverse homes and kitchens, because if you can get a robot to work in the wild in a random house, that’s the true goal of robotics,” he says.

The project could serve as a general recipe to build other utility robotics models for other tasks, helping to teach robots new skills with minimal extra work and making it easier for people who aren’t trained roboticists to deploy future robots in their homes, says Shafiullah.

“The dream that we’re going for is that I could train something, put it on the internet, and you should be able to download and run it on a robot in your home,” he says.

AI-generated content doesn’t seem to have swayed recent European elections 

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. 

Google’s new tool lets large language models fact-check their responses

As long as chatbots have been around, they have made things up. Such “hallucinations” are an inherent part of how AI models work. However, they’re a big problem for companies betting big on AI, like Google, because they make the responses it generates unreliable. 

Google is releasing a tool today to address the issue. Called DataGemma, it uses two methods to help large language models fact-check their responses against reliable data and cite their sources more transparently to users. 

The first of the two methods is called Retrieval-Interleaved Generation (RIG), which acts as a sort of fact-checker. If a user prompts the model with a question—like “Has the use of renewable energy sources increased in the world?”—the model will come up with a “first draft” answer. Then RIG identifies what portions of the draft answer could be checked against Google’s Data Commons, a massive repository of data and statistics from reliable sources like the United Nations or the Centers for Disease Control and Prevention. Next, it runs those checks and replaces any incorrect original guesses with correct facts. It also cites its sources to the user.

The second method, which is commonly used in other large language models, is called Retrieval-Augmented Generation (RAG). Consider a prompt like “What progress has Pakistan made against global health goals?” In response, the model examines which data in the Data Commons could help it answer the question, such as information about access to safe drinking water, hepatitis B immunizations, and life expectancies. With those figures in hand, the model then builds its answer on top of the data and cites its sources.

“Our goal here was to use Data Commons to enhance the reasoning of LLMs by grounding them in real-world statistical data that you could source back to where you got it from,” says Prem Ramaswami, head of Data Commons at Google. Doing so, he says, will “create more trustable, reliable AI.”

It is only available to researchers for now, but Ramaswami says access could widen further after more testing. If it works as hoped, it could be a real boon for Google’s plan to embed AI deeper into its search engine.  

However, it comes with a host of caveats. First, the usefulness of the methods is limited by whether the relevant data is in the Data Commons, which is more of a data repository than an encyclopedia. It can tell you the GDP of Iran, but it’s unable to confirm the date of the First Battle of Fallujah or when Taylor Swift released her most recent single. In fact, Google’s researchers found that with about 75% of the test questions, the RIG method was unable to obtain any usable data from the Data Commons. And even if helpful data is indeed housed in the Data Commons, the model doesn’t always formulate the right questions to find it. 

Second, there is the question of accuracy. When testing the RAG method, researchers found that the model gave incorrect answers 6% to 20% of the time. Meanwhile, the RIG method pulled the correct stat from Data Commons only about 58% of the time (though that’s a big improvement over the 5% to 17% accuracy rate of Google’s large language models when they’re not pinging Data Commons). 

Ramaswami says DataGemma’s accuracy will improve as it gets trained on more and more data. The initial version has been trained on only about 700 questions, and fine-tuning the model required his team to manually check each individual fact it generated. To further improve the model, the team plans to increase that data set from hundreds of questions to millions.