Concrete is the most widely used building material on Earth, and producing it is one of the largest single sources of carbon emissions. One promising way to reduce its environmental footprint is to 3D-print concrete, laying it down bead by bead like a giant icing-piping robot. This process eliminates the labor-intensive formwork of pouring it into molds, and places the material only where a structure needs it.

But many of the most efficient designs created by computers are impossible for today’s printers to build. Engineers use a technique called topology optimization to find the strongest structure that uses the least amount of material. But those mathematically ideal designs, with their intricate, spider-web shapes, don’t account for the physical limitations of large-scale concrete printers with their thick nozzles, limited turning, and need to print in one continuous motion.

Now a team of MIT researchers has developed a way to close that gap. Their framework, described in a new article in Additive Manufacturing, bakes a printer’s real fabrication limits directly into the optimization, so the design that comes out is one a machine can build and print with little or no manual redesign. They demonstrated it by designing, printing, and load-testing a 2.3-meter concrete bridge and found that today’s printing hardware, not the concrete itself, limits how light a structure can be.

“We were finding a lot of cracks you can fall through when it comes to translating these super-optimal designs into manufacturable designs,” says co-first author Hajin Kim-Tackowiak PhD ’26, a postdoc in MIT’s Department of Civil and Environmental Engineering (CEE). “Those cracks were like chasms.”

Designing for what can be built

To pin down the constraints, the team worked with the people who run the large-scale printing machines at Autodesk’s facility in Boston.

“They pointed at some of our sharp angles, and they went, 'I don't feel safe printing something like that,'” Kim-Tackowiak recalls.

Those conversations surfaced three key limitations: how thick each printed bead must be, how sharply the nozzle can turn, and the need to print in a single continuous line. The researchers translated each constraint directly into the mathematical rules of their framework.

Existing 3D-printed structures are typically produced with older methods that optimize the shape first, and then require “a massive amount of post-processing,” taking days to run, Kim-Tackowiak explains. By contrast, the team’s framework generated fully printable designs in about two minutes on a laptop. When the team needed to slightly reduce the bridge’s size on the day of printing, they simply reran the optimization and had an updated design five to 10 minutes later. 

“Reaching that speed at all is recent,” says co-first author Zane Schemmer, a PhD student in CEE. The math the method relies on, mixed-integer optimization, was long considered too hard to use. “You go back five, 10 years ago, the solver we used, even three years ago, could not solve these problems,” he says. “This field has been avoided, because everyone thinks that’s not an avenue we can go down. But with new algorithms and resources, it’s becoming a way we can start to frame problems.”

A bridge reveals the real limitation

To validate the framework, the researchers went back to Autodesk’s facility to print a 2.3-meter-long concrete bridge.

“The bridge took about 30 minutes to make and was built from off-the-shelf mortar,” says senior author Josephine Carstensen, the Gilbert W. Winslow (1937) Career Development Professor in Civil Engineering.

In testing, the roughly 900-pound structure held more than 2,000 pounds spread across it with virtually no measurable bending, closely matching the team’s simulations.

But the test also revealed the study’s biggest surprise. “What we found was our result was super over-engineered,” Kim-Tackowiak says. “From zero to 200,000 pounds, your design is entirely driven by these 'can I build it or not' constraints. And then, after 200,000 pounds, you can start to think about the physics.” In other words, the limits of current printing technology, not the strength of concrete, were dictating how efficient the structure could be.

A roadmap for better printers

Because the framework finds the mathematically best possible design, the researchers could measure exactly how much each hardware limitation costs in material.

“With mixed-integer optimization, we can find the global optimum, the best solution there is, as opposed to just a good solution,” Carstensen says. “Because we know we’re finding the best solution out there, we can also quantify: If we had a machine that could do other things, what would that mean for how much material we’re using?”

The single biggest lever was the width of the printed bead. The bridge used a 4 centimeter bead. The analysis showed a machine that was able to lay a 1cm bead could cut material use by as much as 76 percent while staying “well within safety margins,” Carstensen says. The result surprised her. “I thought the continuous path would be the problem, the one that had the highest effect,” she says. “But it wasn’t. It was the bead width.”

The result is a roadmap for printer-makers showing that modest hardware improvements could unlock large gains in efficiency and cut concrete’s carbon footprint.

Part of what made the bridge possible is that every piece is in compression. “With concrete, it’s really good when you push on it, really bad when you pull on it,” Schemmer says. “We're able to guarantee that every piece of concrete that you see is in compression, there’s no part that’s being pulled on.”

The savings come not only from using less material, but from skipping molds entirely, an advantage that grows for one-off shapes. Carstensen sees early promise in disaster relief, “You can quickly put up new infrastructure without needing to make formwork.”

The bridge’s compression-only nature showed itself dramatically after testing. It had held more than 2,000 pounds without budging, but when a worker lifted one corner a few inches to sweep beneath it, it broke. The failure wasn’t a design flaw so much as a demonstration of the principle behind it: Concrete is weak when pulled, and the lift put parts of the bridge in tension they were never meant to carry. “It’s optimal in one way, but it’s definitely not optimal in every way,” Kim-Tackowiak says.

That points to the team’s next step of reinforced concrete. “We know a pure concrete structure is not necessarily going to be the most optimal thing, so we’re moving it more into the world we live in today, which is reinforced concrete,” Kim-Tackowiak says, “though working out how to feed rebar into a printed concrete structure,” she adds, “is proving its own challenge.”

The work was funded by the National Science Foundation and supported by the MIT Center for Advanced Production Technologies. Joining Kim-Tackowiak, Schemmer, and Carstensen on the paper are co-authors Pittipat Wongsittikan, a PhD student in the MIT Building Technology Architecture program, and Jackson Jewett MEng ’18, PhD ’25, a former MIT postdoc.

It’s a fact of life that the electrical activity of neurons will vary during repetitions of the same task, even when the ultimate outcome is the same. A new study shows that a lot of ongoing fluctuations in the brain’s activity could be explained by the influence local electric fields exerted on the neurons, a phenomenon called “ephaptic coupling.” The finding, published in Cerebral Cortex, adds to evidence that the brain’s electric fields act as important control signals for underlying brain function.

“The brain is a rollicking sea of electrical influences,” says study co-author Earl K. Miller, Picower Professor of Neuroscience in The Picower Institute for Learning and Memory and MIT’s Department of Brain and Cognitive Sciences. “But the traditional view of brain function focuses only on the spiking and synaptic connections among individual neurons. Now, there is growing evidence for electric field effects. For instance, in this study we show that neural variability is explained by how ephaptic effects are influencing neural activity.”

In 2022 and 2023, Miller and fellow author Dimitris Pinotsis, associate professor at City St George’s, University of London, published several studies showing that local electric fields in the brain’s cortex not only reflected the information neurons were processing better than any individual neuron did, but also that the fields actively helped to organize the underlying neural spiking that executes that processing. Like an orchestra conductor, the electric waves can conduct crowds of neurons so that they are “playing the same tune.” They further theorize that fields physically exert influence on the structure of the brain via cytoelectric coupling, in which the fields alter the cytoskeleton of neurons, optimizing them to oscillate in synchrony.

Because electric fields can be manipulated, Miller and Pinotsis argue in the new study that understanding how they influence momentary brain function could open the door to therapeutic interventions designed to improve it when it is faltering in disease. It would be difficult to adjust every neural connection, but ephaptic coupling suggests that intervening at the level of electric fields could accomplish that therapeutic end, the researchers say.

“Properly devised electric field manipulations could help patients rewire faulty circuits,” Pinotsis and Miller wrote.

In the duo’s prior studies, they analyzed signals averaged over time, documenting that in general, even though local (or “mesoscale”) electric fields in the cortex arise from the electrical activity of individual neurons, the field ultimately represents and coordinates their function. Think of it this way: Neurons are like individual citizens, and the electric fields are their government. Once the citizens establish a government with their individual votes, they are then subject to and unified by the laws the government creates and enforces. 

In the new study, the team asked whether mesoscale electric fields not only provide this ephaptic influence overall during working memory tasks, but also trial by trial. After all, that’s closer to the timescale of actual brain operations that matter both for healthy function and in disease. 

So the scientists looked anew at the data they recorded as animals played a simple video game. The animals were shown a dot in one of six positions around a screen. After the dot disappeared, the animals had to hold its former position in memory because to succeed in the game and earn a reward, they had to glance when cued to indicate the direction where the dot had appeared. Meanwhile, the scientists recorded neural electrical spiking and more collective local field potentials. Using that information, they calculated the local prevailing electric field at each moment.

In their statistical analysis of the data, they made several findings. One, as expected, was that neural activity varied sometimes quite widely trial by trial during the task. Another, using a mathematical technique called Granger Causality, showed that the direction of influence between the electric field and the neural activity was strongly in favor of the field. In other words, in the coupling between the two, the fields were dominant.

“We found that electric fields that emerge from neural activity, captured with LFPs [local field potentials], turn around and influence this activity in a top-down fashion (ephaptic coupling),” the researchers wrote.

Moreover, the team’s modeling and calculations showed that the strength of the ephaptic coupling between the field and the neural activity was proportional to the variations in the LFP power — another sign that the fields influenced the neural activity.

“The larger the variability, the more evident the top-down organizing effects,” the researchers wrote. “The emerging picture is that electric fields serve as control parameters.”

The U.K. Medical Council, the U.S. Army Research Office, the U.S. Office of Naval Research, the Freedom Together Foundation, and the Picower Institute funded the study.

A high-fat, low-carbohydrate diet, also called a ketogenic diet, can help some people lose weight by forcing their bodies to burn fat for fuel instead of sugar. 

In recent years, scientists have been exploring how this type of diet might affect other aspects of health and disease, including cancer. While some research has shown that the diet may protect against the development of colon cancer, a new study by MIT researchers suggests that in the small intestine, a ketogenic diet may increase the risk of cancer.

“Ketogenic diets have distinct effects on different tissues even within the gastrointestinal tract. I think the message here is that we need to be very careful in generalizing the effects that these diets can have, because what might be beneficial for one tissue may be detrimental for another tissue,” says Omer Yilmaz, director of the MIT Stem Cell Initiative, an associate professor of biology at MIT, and a member of MIT’s Koch Institute for Integrative Cancer Research.

Yilmaz is the senior author of the study, which appears today in Nature. MIT postdocs Jessica Shay and Fangtao Chi are the lead authors of the paper. Researchers from the labs of Alex K. Shalek, director of MIT’s Institute for Medical Engineering and Science, and Matthew Vander Heiden, director of the Koch Institute, also contributed to the study.

Diet and cancer

Ketogenic diets, originally developed in the 1920s as a way to treat epilepsy, have been adapted in the past few decades as a strategy to lose weight or increase lifespan. The diet comprises a high percentage of fat, low percentage of carbohydrates, and normal or reduced amounts of protein.

This type of diet forces the body to burn fatty acids for energy in place of carbohydrates such as glucose. Burning these lipids produces ketone bodies — primarily β-hydroxybutyrate (BHB) and acetoacetate — as byproducts of fatty acid metabolism. These ketone bodies are also generated when people fast or follow very low-calorie diets, which force the body to burn its own fatty stores.

A 2022 Nature study suggested that ketogenic diets have a protective effect against colon cancer and that BHB — the most abundant ketone body — is responsible for this effect. In the new Nature study, the MIT team wanted to explore whether ketogenic diets might have a similar protective effect in the small intestine.

The researchers fed mice who were genetically predisposed to developing intestinal cancer either a ketogenic diet, a control diet, or a high fat/high calorie diet. They found that mice on a ketogenic diet were more likely to develop tumors of the small intestine than those on a control diet. While they did not become obese, mice on the ketogenic diet developed tumors at rates similar to or even higher than those of mice on an obesogenic high fat/high calorie diet. 

Additional studies revealed that ketone bodies did not play a role in tumor development. Instead, tumor growth was driven by how intestinal cells burn dietary fat for energy — a metabolic pathway called fatty acid oxidation. This pathway activates a family of proteins called PPARs, which signal stem cells to multiply more rapidly, increasing the chance that some become cancerous. 

This stem cell proliferation can be beneficial in certain situations, such as when the intestinal lining needs to be repaired after illness or injury. However, too much proliferation can tip cells toward becoming cancerous. 

“Having more stem cells means that when you injure the small intestine, it can repair itself better, but the downside is that having more active stem cells can lead to tumor formation,” Yilmaz says.

Opposite effects

Surprisingly, the same ketogenic diet that promoted tumors in the small intestine had the opposite effect in the colon. The researchers found, similar to the earlier Nature study back in 2022, that a ketogenic diet suppressed the development of colon tumors. However, the new findings suggest that ketone bodies are not responsible for this protective effect.

“Given how much attention has been paid to ketone bodies like BHB, both as a commercial health trend and in recent high-profile studies suggesting BHB suppresses colon cancer, we fully expected them to be the direct drivers. Instead, our experiments in genetically engineered mice revealed that these molecules are essentially metabolic bystanders. The real surprise is that tumor acceleration is driven entirely by how stem cells process and burn the heavy influx of dietary fat itself,” Yilmaz says.

The researchers now hope to further study why ketogenic diets have such different effects in the colon and the small intestine. As ketogenic diets continue to gain popularity, understanding these tissue-specific effects will be critical for guiding their use, the researchers say.

“The deeper question is why the same diet has opposite consequences in two adjacent parts of the gut. That is what we are working to understand next,” Chi says.

The findings carry practical implications. Because the diet’s effects — both the tumor acceleration in the small intestine and the protection in the colon — are driven entirely by fat metabolism rather than the ketones themselves, commercial ketone supplements or drinks would not be expected to mimic either the risks or the benefits discovered in this study. This may be especially relevant given that small intestinal tumors have been rising in incidence in recent decades, with the greatest impact on patients with inherited conditions that predispose them to intestinal cancer, such as familial adenomatous polyposis.

The research was funded, in part, by the National Institutes of Health, a Pew-Stewart Trust scholar award, the Kathy and Curt Marble cancer research award, a Koch Institute-Dana Farber/Harvard Cancer Center Bridge project grant, the American Federation for Aging Research, the MIT Stem Cell Initiative, a Damon Runyon Postdoctoral Research Fellowship, and the Koch Institute Support (core) grant from the National Cancer Institute.

Systems using artificial intelligence to enhance forecasting, planning, and decision-making in businesses have been proliferating in recent years, but in many cases, they lack the detailed, specific information about the organization itself, limiting the usefulness of those tools. 

Devavrat Shah, a principal investigator at MIT’s Laboratory for Information and Decision Systems (LIDS), faculty member with the department of Electrical Engineering and Computer Science (EECS), and member of the Institute for Data, Systems, and Society (IDSS), has been focused on how to design methods that can handle second-by-second decision-making using limited computational resources. 

“In a sense, with a small amount of resource, you have to do a lot of heavy lifting,” he says. As a researcher, “my interest is in the ability to develop methods that can extract information from data at scale in as effective a manner as possible.”

The Andrew (1956) and Erna Viterbi Professor has been teaching at MIT since 2005. 

In 2019, he also co-founded a spinoff company called Ikigai Labs. Ikigai built a foundation model for tabular, time series data based on years of research in Shah’s lab, which was patented and licensed by MIT to the company. This model can take input from enterprise data from varied sources, continuously and at scale, so that it learns as it goes along by testing its predictions against real outcomes.

Shah explains that the system is an extension of the kind of graphical models that are used, for example, by GPS devices to convert a sparse amount of data received from satellites into an accurate model of a position on the Earth’s surface, or by communication system like that in a digital watch that communicates at high speed in an energy-efficient manner. 

“My interest was: How does one design such graphical models for generic, tabular data?” he says.

While most AI models have been taught using text and images, this system takes tabular data as its input — structured data such as the familiar kind of row-and-column format used in spreadsheets. And then it provides the kind of real-time planning, on a vastly larger scale. 

The idea for Ikigai was to provide forecasting and decision-making technology for large businesses, such as consumer goods manufacturers and pharmaceutical companies.

Shah gives the example of how a consumer electronics company might use this system. 

“Let’s say you’re making headphones and all sorts of different things. And each of the products that you manufacture has lots of small pieces that come from different parts of the world. And once the device is sold, it needs to be supported and maintained. And you have to come up with new versions of the product, you have to market them, you have to price them … So the questions you would typically ask would be: If I were to sell these next quarter or next year, how many will be sold in different places, and what would happen to demand if I change the price, or if I introduce promotion?”

He adds that all of these processes are interdependent, and at every stage of the processes decisions have to be made that have implications over time. “At some level,” he says, “digitizing these processes and being able to do predictions and constantly optimize is what leads to ultimately better business operations.”

Ikigai was recently acquired by the international firm Celonis, where Shah is now chief scientist in addition to his roles at MIT. Ultimately, he hopes the model he developed for Ikigai will help Celonis deliver tools that can integrate with a company’s own data and business processes in order to provide real-world analyses that can help make forecasts, plans, and decisions.

Shah adds that Celonis has specialized in digitizing and automating operations for more than 1,400 large companies around the world. Now that these systems are fully digitized, they provide a platform for Ikigai’s software to take the next step, reading the data from these digitized systems in order to provide detailed models to allow simulation of different options, predict optimum strategies, and forecast the results of a given set of decisions. 

“Once the digital layer of these processes exists and this information layer exists,” Shah says, “now, on top of it, we can put the Ikigai stack to enable decision-making at a much larger scale than otherwise.”

While so many companies are working on various aspects of AI, “we are very much focused on part of the domain that the rest of the world is not paying attention to,” which is the area of structured or time-domain data. By starting from such data, he says, it provides a very cost-effective version of AI. 

“A narrower focus comes with sharper technology,” he says, “but it’s broad enough that it’s very valuable.”

Shah adds, “The recent buzzword that’s become pertinent in the modern AI popular press is a ‘world model.’ In a sense, this is trying to build the enterprise process world model, so to speak.”

In an increasingly crowded, for-profit landscape for scholarly research, the health of a publishing program is often measured by the influence of its publications. This year, three MIT Press journals demonstrated their stature by earning the highest impact factors in their disciplines.

Computational Linguistics ranked first in the Linguistics category, International Security led the International Relations category, and The Review of Economics and Statistics topped the Social Sciences, Mathematical Models category in Clarivate’s 2026 journal impact factor rankings.

For the MIT Press, this achievement highlights the distinctive strength of its journals program. Although relatively small compared to other commercial and university press publishers, MIT Press journals consistently publish widely cited scholarship across a broad range of disciplines, from social science and the humanities to neuroscience and artificial intelligence. 

Clarivate’s impact factors capture the previous year’s scholarly citation activity, but the influence of MIT Press journals often extends well beyond academia. In recent months, International Security articles have been cited by Foreign Policy, Foreign Affairs, The Conversation, CBC, and Brookings. The journal has also published research with significant real-world policy relevance, including a widely discussed article by MIT political scientist Caitlin Talmadge that anticipated how a limited strike on Iran could escalate into attempts to disrupt shipping through the Strait of Hormuz, triggering a broader military and economic crisis. 

“I am proud and humbled that International Security has had the number one impact factor in International Relations for two years running,” says Jacqueline Hazelton, editor of International Security. “Thanks are due to our generous reviewers, our brilliant authors, our talented editors who handle the often-thankless work of copy editing and production, and, of course, our readers. We plan to continue leading the field in IR/security studies with rigorous scholarship that challenges the conventional wisdom, identifies new threats and opportunities, engages with policy and theory, and illuminates history.”

The MIT Press journals team is small, with under 10 people working across production, sales, and marketing; but that small team collaborates with the editorial staff of 50 disparate journals to publish around 2,500 articles annually. “Some of the joy I take in editing International Security stems from working with the people at MIT Press,” Hazelton adds. “They are helpful and patient. They know what to do, and they do it.”

“The journals division at MIT Press has undergone significant change over the past decade — from business model upheaval and rapid technological advances to the ongoing challenge of competing with commercial publishers many times our size,” says Nick Lindsay, director of journals and institutional partnerships at the MIT Press. “Through it all, the journals group has adapted and evolved to meet those challenges and remains a home for experimentation and fair and equitable publishing.”

The MIT Press’ reputation for influential publishing has attracted many prestigious partners to its journals program, including Harvard University, the American Academy of Arts and Sciences, and the University of California at Berkeley. Amid this growth and development, the program continues to launch and support new journals in emerging and interdisciplinary fields while upholding the high editorial and publishing standards that have made it what it is today.

“Computational Linguistics has long stood for depth and rigor, and in a field that moves remarkably fast, our aspiration is for it to remain a home for work that lasts — scholarship the community can keep building on for years to come,” says Wei Lu, editor of Computational Linguistics. “We are very proud of this result, which reflects both the strength of the work our authors publish and the care our reviewers and editors bring to the journal. We are grateful to MIT Press for being such a steadfast partner.”

This strong performance extended well beyond the press’ three top-ranked publications. Transactions of the Association for Computational Linguistics was ranked 2nd in the Linguistics Category out of 312 journals; Global Environmental Politics was 2nd in the International Relations category out of 173 journals; and The Review of Economics and Statistics was 17th in the Economics category among 626 journals. Other highlights include Harvard Data Science Review ranking 7th in Statistics and Probability; European Societies ranking No. 13 in Sociology; and Neurobiology of Language ranking No. 13 in Psychology, Experimental.

Overall, 13 MIT Press journals earned impact factors that place them in the top quartile of their area of publishing, including: 

• Computational Linguistics 
• European Societies
• Evolutionary Computation 
• Global Environmental Politics
• Harvard Data Science Review
• International Security
• Journal of Cold War Studies
• The Journal of Interdisciplinary History
• Linguistic Inquiry
• Neurobiology of Language
• Quantitative Science Studies
• The Review of Economics and Statistics
• Transactions of the Association for Computational Linguistics

Together, these rankings point to the strong reputation that the MIT Press has built for its journals portfolio, a relatively small program that shapes conversations across the humanities, social sciences, and STEM fields.

The wiring and rewiring of the brain never ends. Neural pathways are constantly being reshaped as we interact with the world and learn new things. At MIT’s McGovern Institute for Brain Research and York University in Toronto, Ontario, scientists are combining detailed analysis of brain activity with computational modeling to better understand that change.

McGovern Institute postdoc Lynn Sörensen, McGovern investigator and MIT Professor James DiCarlo, and York University Assistant Professor Kohitij Kar, worked together to compare what happened when animals and an artificial neural network with brain-like architecture were trained to visually identify the same objects. As the model’s performance improved, it reorganized itself in ways that closely paralleled changes the team detected in the animal brains.

Their open-access work, reported July 8 in the journal Nature Communications, shows how changes in visual processing support animals’ ability to learn to discriminate new kinds of objects. By modeling these changes, the researchers hope to better predict how training reshapes perception, which could one day inform educational strategies for a wide range of learners.

Subtle changes

Learning about a new object calls on many parts of the brain. Visual-processing areas work together to make sense of information taken in through the eyes, then communicate with other brain areas to give the visual information meaning and guide behavior. Multiple parts of this system likely change during learning, and the research team wanted a clearer understanding of how that change is distributed.

Neuroscientists have debated how much change occurs in the brain’s visual-processing areas when an animal learns to recognize new objects. Some suspected that visual-processing pathways remain largely unchanged during learning to avoid broadly disrupting visual perception, but others have reported changes in activity within dedicated visual-processing areas with this kind of learning in humans and other primates.  

To take a closer look, the team focused on neural activity in a key component of the brain’s visual object-processing network, the inferior temporal (IT) cortex. By the time visual information reaches the IT cortex, key object features are clearly represented — so much so that it’s possible to “decode” what object the subject is seeing and even predict what errors it’s likely to make in identifying it, simply by analyzing patterns of neural activity there.

The team recorded neural activity in the IT cortex from animals as they looked at and identified images of objects. Some of the animals were untrained, so the images they saw had little meaning to them. Others had already learned to identify similar objects, so they could usually discriminate between elephants, chairs, and other select objects, even when those objects were presented at different sizes, from different angles, or against different backgrounds than the ones they had seen before.

The broad pattern of activity in the IT cortex was largely similar in trained and untrained animals, suggesting that learning had not dramatically rewritten this high-level visual representation. Still, the group found subtle but reliable differences in the way neurons in the IT cortex responded to images in animals that had learned to recognize the kinds of objects they were shown, compared to the untrained animals.

Modeling learning

The group turned to computational models to investigate how those modest changes might contribute to learning. Sörensen trained a suite of artificial neural networks whose internal components had been mapped to the IT cortex to identify the same categories of objects the animals had seen. The models were designed to learn using gradient descent, meaning they continually improved their accuracy by adjusting their parameters in response to errors.

Only some of the animal models showed learning behavior that matched that of the subjects. In those that did, the IT-like stage changed in ways that resembled the learning-related changes the researchers had observed in the IT cortex of trained animals.

While gradient descent is commonly used to train artificial intelligence, it is generally considered biologically implausible as a direct model of how the brain learns. The researchers say the strong match in learning effects between the animals and their model demonstrates that these kinds of artificial neural networks can offer insights into biological learning at a useful level of abstraction, even if the brain does not learn in the same way.

“This shows that you can actually build in silico versions of future experiments,” Sörensen says. “I think that gives us this playground of asking ‘what if’ questions — and potentially predicting new things that go beyond the experimenter’s intuition.”

Most of the changes that allowed for learning in the model occurred outside of the IT cortex. “This tells us that there is a lot between the area we recorded from and the final behavioral readout that needs to change during this process,” Kar says. He adds that the team’s model will be useful as researchers look more deeply into how downstream brain areas contribute to learning.

The researchers stress that their study allowed more granular measurements of brain activity than would be possible in humans, and because the animal brains are organized similarly to our own, their experiments have direct relevance to human learning. They say understanding the impact of plasticity in the subjects’ IT cortex could help researchers design new learning strategies for humans.

“Our prior conceptual working model of you learning new objects was that your brain makes changes to synaptic connections that are largely downstream of your visual system, so you don’t destroy your visual system,” says DiCarlo, who is also the Peter de Florez Professor of Brain and Cognitive Sciences and director of the MIT Siegel Family Quest for Intelligence. “You wouldn’t want your whole visual system to become an elephant detector [just because you’ve learned to identify an elephant]. But this study went beyond that to say actually, when you learn ‘elephant,’ your IT does change a little bit to make it a little more relevant to elephants.”

That likely has consequences for recognizing other visual features, too. Subtle changes in the IT cortex that support elephant recognition might also make you better at identifying things other than elephants, DiCarlo says. Likewise, the same changes might make it a little harder to identify something else.

These kinds of consequences may be difficult to predict intuitively, but become obvious with computational modeling. For instance, the team’s models revealed that after learning to recognize new objects, the IT cortex contained more information about objects’ locations. By providing insights like these, models could aid the design of more effective training strategies for visual tasks, including for people with altered sensory processing, who may learn from visual information in atypical ways.

Artificial intelligence has rapidly transformed software engineering. Generative AI and large language models (LLMs) can create huge volumes of code and documentation; machine-learning algorithms can monitor performance and detect security vulnerabilities. But when the task is to conceive, design, and make a complex physical system such as a jet engine, are those AI tools equally transformative?

This past semester, the JARVIS Challenge (Jet-engine AI Research and Validation Intensive Sprint) set out to explore whether AI can compress the design-build-test cycle, asking MIT undergraduates to discover whether AI can help them to build faster and better. 

“The JARVIS challenge showed that AI can substantially accelerate safety-critical hardware engineering, but engineering judgment remains the decisive differentiator. An AI-native engineer is not defined by using AI, but by leading it — knowing when to trust it, when to challenge it, and how to translate AI outputs into working hardware. Manufacturing — not engineering design or analysis — remained the fundamental rate-limiting step,” says Professor Zolti Spakovszky, director of the MIT Gas Turbine Laboratory.

The teams, the tools, the task

The challenge gave undergraduates four weeks to design, fabricate, assemble, and test a small gas turbine aero engine, using AI as their primary engineering partner. The objective: build a “JARVIS-class” single-spool jet engine producing 50–100 pounds of thrust, running on Jet-A, and completing five 60-second runs. Teams had total freedom over design, materials, and fabrication. 

Representing nearly every department in the School of Engineering, 31 students organized into seven teams, ranging from all first-years to senior-heavy groups. Many of the competitors initially had little experience in turbomachinery, compressible flows, or, in the case of the younger students, even thermodynamics. Many had never seen the inside of a gas turbine before signing up to build one.  

At their disposal: MIT’s machine shops and manufacturing vendors; commercial software including Concepts NREC, SolidWorks, and ABAQUS; and various test rigs for characterizing and assembling individual components.

The teams also had access to MIT Parley, a newly launched platform that aggregates frontier large language models through a single interface. Through Parley, JARVIS leads could see directly how the students were using the AI tools, including their prompts, the cost per prompt, the specific LLMs being used, and other critical information. The JARVIS leads secured early access to Parley for all participants, and with financial support from MIT Lincoln Laboratory, the Department of Mechanical Engineering, and corporate sponsors Safran, Voyager Technologies, and Beehive Industries, students had access to essentially unlimited use of AI.

The sponsors were drawn by recruiting interest and genuine curiosity about how AI might reshape engineering workflows. 

“We see this as the future of engineering,” Ryan (Hal) Hefron of Voyager Technologies told the students. “You’re honing skills that are not just nice to have — they’re going to be the future baseline in the engineering workforce.”

Vincent Garnier, managing director of Safran Tech, watched the competition unfold with excitement. “JARVIS was a genuine experiment, a learning endeavor. We frankly didn’t know what to expect, from the students or from the AI models. What struck me coming from the students was: first, the enthusiasm to explore; then, as the project developed, they all came to the cool-headed realization of what AI could or could not help them with, and then almost instantly adapted for that,” he says. “It makes me confident that this generation of leading engineers will probably not fall prey to easy and shortsighted use of AI, and will do so by keeping ever more in contact with experiments — physical or thought experiments.”

The faculty leadership — professors Zachary Cordero, Zolti Spakovszky, Masha Folk, and Andreea Bobu of the Department of Aeronautics and Astronautics, along with Lincoln Laboratory engineers and a team of teaching assistants — were there to ensure safety. In weekly progress reviews, they would critically evaluate the student progress and assess how the students were using AI.

Spakovszky developed a careful technique for guiding teams in the right direction without giving away answers or providing help. After a team’s presentation, he might ask: “Do you know what a rabbet fit is? Take in the comment.”

Where AI helps and hurts

By the end of week 1, one team withdrew from the competition; the others had, with varying degrees of success, developed an initial design for their gas turbines. Different teams used AI to summarize textbooks, teach them to use design software, source vendors, create Excel sheets, answer specific questions, find references, and create comparative analysis between design decisions. One team created an agent in Parley and tasked it with serving as their project manager. 

By week 2, teams had to start working on detailed CAD designs, ordering parts, and prototyping their combustors. This is where the teams started to hit limitations in their use of AI. While Claude and ChatGPT were good at offering design alternatives and filling knowledge gaps, teams found that the hallucinations, sycophancy, and lack of physical understanding that have become notorious features of generative AI were undermining their confidence and slowing them down. 

“AI is a helpful tool, great at finding information, helping organize things, and can write well, but it can’t do design,” says Elizabeth Tupaj, a member of team 811 Crew. “The moment the engineer doesn’t know what is going on and the AI is in charge is the moment the design becomes unreliable, at least with AI at its present capabilities.”

Teaching assistant John Zhang notes, “seeing this firsthand with the students reminded me how much first impressions matter. If the students couldn’t get answers from the AI early on, they quickly grew frustrated and formed a lasting opinion that precluded them from using it later.” 

In the final weeks, the finalists hit another obstacle no AI could solve: working with vendors. “AI searches found vendors we had no rapport with, who had no interest in our tight timeline,” students reported. “The vendors who came through were the ones our team had personal relationships with.”

Of the three finalists, only Fast and Fractured achieved first-attempt ignition of their mini-combustor. The team had used AI heavily for trade studies and architecture comparisons, arriving at a viable design despite none of them having prior gas turbine experience.

“The JARVIS Challenge showed what’s possible when you combine AI-enabled design with motivated students and a culture of rapid experimentation,” says Masha Folk, the Charles Stark Draper Career Development Professor of Aeronautics and Astronautics. “The moment that stood out most was when the first student-designed combustor was installed on the test stand. It ignited flawlessly, ramped to full power, transitioned to dual-fuel operation, and then sustained stable combustion on 100 percent Jet-A fuel. This was proof that we can dramatically accelerate the cycle of design, build, and test while giving students hands-on experience with a real engineering challenge.”

At the vanguard of AI-native engineering

By the end of May, the two more senior teams – Fast and Fractured and 811 Crew – had completed full engine tests. Fast and Fractured, with their AI-assisted design, were delayed by vendor headaches week after week, but finally made it to test. Unfortunately, their hot fire was cut short when the rotor rubbed and seized against the stationary housing. Team 811 Crew, however, who had more exposure to turbomachinery and propulsion concepts going into the competition, emerged victorious. Their engine started, successfully transitioned to Jet-A, and generated net thrust. 

“As we stood there with the air-starter, hearing their engines spool up and watching them spit fire, it felt like my heart was racing out of my chest. There were so many ways it could go wrong! What these students accomplished in such a short time span is nothing short of amazing,” says PhD student Joe Chiapperi. 

The 811 team had been resistant to using AI throughout the competition, trusting instead to their fundamentals and teamwork. “We had people who were at least somewhat familiar with the design software, mechanical engineers who knew how to build anything, and aerospace engineers who had taken classes on the design of gas turbine engines specifically,” says Tupaj. 

From the start of the JARVIS Challenge, younger students used Parley more frequently and cleverly, while the juniors and seniors leveraged deeper experience. 

“JARVIS taught me that getting value from AI takes two things: enough expertise to judge what it tells you and catch it when it’s wrong, and enough curiosity to actually lean on it where it could help,” says Professor Andreea Bobu. “The team that moved fastest in the sprint was experienced and leaned heavily on AI to get there. The team that eventually won was more resistant to AI; they had the expertise, but that skepticism made them slower. The sweet spot seems to be knowing enough to stay in charge of the tool, and being eager enough to pick it up in the first place. To me, that’s the real opportunity ahead: training the next generation of engineers who have the judgment to direct these AI tools and the instinct to reach for them.”

The competition’s clearest finding: engineering experience is a multiplier, and the human factor remains a vital element. Mastering the first principles and fundamental concepts breeds good engineering judgment and the ability to navigate strings of tough decisions in the face of incomplete information. And when it comes to building safety-critical physical systems, nothing can replace human hands and human accountability. 

“JARVIS has shown that AI copilots can have a multiplicative effect on engineering productivity, with judgment and first-principles thinking serving as the key differentiators among teams,” adds teaching assistant Kyle Woody. 

But the implications of AI in aerospace are significant. If small teams using well-managed AI copilots can compress design-build-test cycles from years to weeks, the consequences for workforce structure, R&D timelines, and competitive dynamics could be substantial. The students who tackled the JARVIS Challenge are among the first engineers to grapple with those stakes not as a thought experiment, but in a machine shop, with a jet engine on the test stand.

“JARVIS highlighted the power of AI in the design of physical systems,” says Cordero, associate director of the MIT Gas Turbine Laboratory. “But it also showed that the key to unlocking that power is education, through coursework, internships, and hands-on extracurriculars like MIT Motorsports and Rocket Team. Performance in JARVIS correlated strongly with year in school. My main takeaway is that in the AI era, education is more valuable than ever.”

Tissue engineers are finding ways to grow living organs and tissues from cells, with the aim of replacing diseased and damaged counterparts in the body. Scientists have successfully grown artificial muscles, livers, kidneys, skin, and other tissues. But there’s been no reliable way to engineer precisely patterned networks of blood vessels, some of which can be finer than a human hair. 

Without a vascular network to deliver nutrients, any artificial tissues, no matter how life-like, can’t function. 

Now MIT engineers have found they can engineer and control the growth of blood vessels by mechanically stretching them. 

The team has built a human “blood vessel on a chip,” composed of a central artery made from human endothelial cells, that is embedded in a gel that also contains a small magnet. The researchers studied how the main artery responded as they jostled the gel back and forth using an external magnet to move the magnet embedded within the gel.

They found that the simple mechanical action of repeatedly jostling the artery stimulated the artery to sprout other, smaller capillaries. By changing the direction in which the artery is jostled or stretched, the researchers could redirect the growing new vessels. And stretching the artery by various degrees influenced how many more new vessels sprouted. 

Their results, reported in the Proceedings of the National Academy of Sciences, offer scientists a new way to engineer artificial blood vessels and program the patterns in which they grow. 

“Healthy tissues depend on organized blood vessel networks, but state-of-the-art protocols don't enable fabricating such networks within engineered tissues,” says Ritu Raman, associate professor of mechanical engineering at MIT and the study’s co-lead author. “The ability to program blood vessel growth with physical cues may enable reproducible and scalable fabrication of engineered tissues that can be implanted in the body to restore function after debilitating disease or injury.”

The study’s MIT co-authors include Sina Kheiri, Jessica Shah, Shashaank Venkatesh, and Roger Kamm, along with Peiyuan Chai and Ryan Flynn at Harvard University. 

“Moving is good”

Blood vessels are tricky to grow and control using conventional fabrication techniques. While 3D printers can produce vessels at the scale of major arteries and veins, the technology is not precise enough to print intricate networks of much finer, thread-like capillaries. Scientists have had some success with growing blood vessels from individual cells, by cultivating them in Petri dishes filled with nutrients and growth factors. But controlling how and where they grow remains a challenge. 

“You can try to pattern chemical cues, like growth factors, to direct where vessels grow, but you can’t do this very precisely,” Raman says. “We thus need other types of patternable cues that can help us build tissues with organized vessels.”

She and her students wondered whether they could grow and control new blood vessels using a protocol they previously developed to grow artificial muscles and nerves. In their previous works, the team engineered a small chip filled with a gel that they infused with nutrients and growth factors. They embedded a small magnet within the gel, and then carpeted the surface of the gel with live muscle or neuron cells. They then manipulated an external magnet to pull the embedded magnet, and the cell-covered gel, back and forth. This work revealed that mechanical “exercise,” pulling the cells back and forth, directly influenced how the cells grew. 

In their new work, the team used a similar setup to see if they could grow and control new blood vessels. 

The researchers built a “blood-vessel-on-a-chip,” smaller than a postage stamp, and filled it with a similar nutrient-rich gel containing a small magnet. They poked a thin tube lengthwise through the gel to create a hollow channel, and coated the channel with live endothelial cells, which naturally grow and fuse to form blood vessels in the body. Once the cells took on the channel’s shape, they started sprouting new, capillary-like vessels in the gel. 

Placing the device under a motorized stage fitted with small, suspended magnets, the researchers moved the magnets back and forth in different directions, and by various degrees, and observed whether and how blood vessels sprouted from the central artery in response. 

“The main takeaway is: Stretching the blood vessel back and forth seems to enhance the number of new capillaries that grow,” Raman says. 

If the main artery were simply left alone in the gel, it would grow some new vessels in random locations along its length. But when the artery was jostled, significantly more vessels sprouted. When the team used the magnets to stretch the gel back and forth, by 5 percent of the gel’s total width, many new vessels grew out from the main artery. When they stretched by 15 percent, fewer vessels sprouted, but those that did grew longer. And when the team changed the direction of stretching, the new vessels followed in response, taking turns and following the pattern of the team’s mechanical stimulation.

“We’re finding that moving is good, which is always the takeaway of everything we do in our lab,” Raman says. “Mechanical forces play an important role in our bodies. That means that if you want to grow more or less vessels, or shorter or longer vessels, or vessels in certain directions, we now know how to do that.”

A gatekeeping gene

The researchers went a step further to investigate why blood vessels grow in response to mechanical forces. To do so, they looked to gene editing, and the role of one particular gene: Piezo1. 

Raman had recently attended a talk by molecular biologist Ardem Patapoutian. In 2021, Patapoutian received the Nobel Prize in Physiology or Medicine for his discovery of ion channels in cell membranes that open and close in response to mechanical pressure. These channels, named PIEZO1 and PIEZO2, act as a cell’s gatekeepers, controlling what goes in and what comes out of a cell. Both types of channels, Patapoutian found, are regulated by their respective genes, also named PIEZO1 and PIEZO2. 

After his talk, Raman showed Patapoutian her group’s experimental results, which showed a connection between blood vessel growth and mechanical stimulation. Patapoutian in turn proposed that the explanation could be the PIEZO1 channel; by mechanically exercising the central artery, Raman may have been stimulating ion channels in the artery’s cells to open, triggering new blood vessels to grow. 

To test this hypothesis, Raman looked to knock down the PIEZO1 gene. If this gene were less active, and fewer blood vessels grew as a result, then it would mean that blood vessels do indeed grow in response to mechanical stimulation, and specifically, through the activation of PIEZO1 ion channels. 

The team repeated their experiments, this time with endothelial cells that were genetically edited to suppress the PIEZO1 gene. Sure enough, they observed that significantly fewer new blood vessels sprouted, even as they mechanically exercised the central artery.

Now that the team has found a way to grow and control blood vessel growth, they plan to apply the protocol to grow organized networks of vessels to supply artificial organs and tissues. “We are now investigating how precisely patterning blood vessel growth can help improve muscle function,” says co-author Jessica Shah.

This work was supported, in part, by the U.S. Department of War Army Research Office Early Career Program and PECASE Grant, and a Department of War DURIP Program Grant.

Professor Arthur Bahr has been named head of the MIT Literature Section, effective July 1. 

“Arthur is an exceptional scholar and a proven leader. I am confident that he will guide the unit with judgment, insight, and a deep commitment to its continued success,” says Agustín Rayo, the Kenan Sahin Dean of the School of Humanities, Arts, and Social Sciences. “I very much look forward to having him join the school’s leadership team.”

Bahr’s work blends formalist and materialist approaches to find literary resonance in the physical particularities of medieval manuscripts. He joined the MIT faculty in 2007 and helped lead the Ancient and Medieval Studies program in 2009-18 and 2022-23, working with colleagues from across the Institute to strengthen and expand the program. He has also been curriculum chair and undergraduate officer of the Literature Section.

“Lit@MIT has some of the world’s most innovative literary scholars and some of the Institute’s most dedicated teachers,” Bahr says. “It has also been my home for nearly 20 years, and I feel both humbled and energized by the opportunity to help shape its future. 

“Literature creates opportunities to slow down and reflect on what really matters, and in a fast-paced, increasingly automated world, those skills are more vital than ever,” he continues.

Bahr succeeds Associate Professor Sandy Alexandre, who served as head of the unit since July 2025.

Bahr is the author of “Chasing the Pearl-Manuscript: Speculation, Shapes, Delight” (University of Chicago Press, 2025); “Fragments and Assemblages: Forming Compilations of Medieval London” (University of Chicago Press, 2013); and co-editor of “Medieval English Manuscripts: Form, Aesthetics, and the Literary Text,” a special volume of The Chaucer Review (47.4, April 2013). His essays have appeared in ELH, Studies in the Age of Chaucer, Studies in Philology, and The Chaucer Review, among others. 

Bahr has been named a SHASS Faculty Fellow for the spring 2027 semester. His next project combines his interest in manuscripts with his training as a figure skating judge to explore analogies between sheets of parchment and sheets of ice, as sites of performance, inscription, and erasure.  

Bahr was named a MacVicar Faculty Fellow in 2015. He received the James A. (’48) and Ruth Levitan Award for Excellence in Teaching in 2012.

Bahr has served MIT as chair of the Committee on the Undergraduate Program from 2019 to 2021, and served on the pandemic-era Academic Policies and Regulations Team. He was also a subcommittee chair of the Education Group of Task Force 2021 and Beyond, and member of the subsequent Refinement and Implementation Committee on the Undergraduate Program. 

Bahr earned his undergraduate degree from Amherst College and his PhD in English Language and Literature from the University of California at Berkeley.

In May 2019, the government of Baltimore, Maryland, fell into chaos. Cybercriminals had locked the city out of many of its critical files and demanded payment to decrypt them. The city refused to pay ransom. The attack incapacitated a swath of services, including real estate transactions and bill payment, and recovery costs soared into the millions.

The syllabus of class 11.074/11.274 (Cybersecurity Clinic), a course in the MIT Department of Urban Studies and Planning (DUSP), includes a case study on Baltimore’s situation as an example of increasingly common ransomware attacks on municipal governments and other public agencies. To counter such threats, Lecturer Jungwoo Chun and Ford Professor of Urban and Environmental Planning Lawrence Susskind launched the MIT Cybersecurity Clinic in 2019. They have offered the course nearly every semester since.

Much like a legal or medical clinic, the course doubles as hands-on training for students and a pro-bono service to at-risk communities. After completing instructional modules and passing a certification exam, students are assigned in teams to a client. By the end of the semester, each team creates a report assessing the client’s vulnerabilities to cyberattack and recommending steps to improve protection. So far, the clinic has provided more than 40 assessments, confidential and free of charge, primarily for New England municipalities and health-care organizations.

In 2025, the FBI’s Internet Crime Complaint Center documented an average of 2,765 cyberattacks targeting Americans every day. When these attacks strike cities and towns, the fallout goes beyond finances, says Chun: “There’s a terrifying, cascading effect on every dimension of our lives.” 

In recent years, cyberattacks targeting the kinds of client communities served by MIT’s clinic have imperiled water supplies, impeded 911 and police services, and exposed citizens’ personal data.

Despite being gateways to essential infrastructure, many small municipalities and hospitals lack in-house staff trained in cybersecurity. Demand for such experts far exceeds supply in today’s labor market, and public sector budgets rarely can match the high salaries private companies offer qualified candidates.

According to Comparitech, from 2018 to 2024, there have been 525 ransomware attacks on U.S. government entities, approximately one every five days, leading to an estimated $1.09 billion in downtime costs.  

“Underfunded public and not-for-profit bodies need to follow a self-help pathway,” Susskind says. “There are many low-cost moves that these organizations can implement with a little coaching from a free-service clinic.”

Defensive social engineering

Some might be surprised to find a university cybersecurity program housed outside the computer science department. Chun is an applied social scientist with expertise in public policy and planning, and Susskind is a leading scholar of conflict resolution and consensus building. They call the approach they’ve developed for the clinic “defensive social engineering” to emphasize that cybersecurity isn’t solely a technical challenge.

Chun acknowledges that the rapid development of artificial intelligence has created alarming new tools for criminals — “now AI can not only identify the vulnerability, but do the attack itself, which is really scary” — and an ever-evolving menu of software claims to guard against these attacks. Accordingly, the course spends considerable time on the technical aspects of cybersecurity. “But at the end of the day,” Chun says, “the biggest attack vector is still through humans.”

The term “social engineering” commonly refers to ways cybercrime victims are manipulated into compromising security (for example, by sending money to a scammer, downloading malicious code, or disclosing sensitive information). Susskind and Chun’s concept of defensive social engineering is similarly grounded in human psychology. The approach emphasizes that cybersecurity must be part of everyone’s job, technical or otherwise.

“It’s about people knowing what to do, people making the right choices,” says Chun. “It’s helping them use the resources and budget they have now on things that can be long-lasting, rather than just spending on the latest antivirus software.”

“Students with computer science backgrounds are surprised by the importance we attach to helping clients build organizational capacity,” says Susskind. “Students need to understand the leadership dynamics in their client communities. The IT director can’t just do what she or he wants. They depend on the local government for their budget. They need approval to hire new staff.”

On the other hand, Susskind says, students from planning or social science backgrounds often study smart city innovations without learning much about the technologies needed to manage the associated risks. And there are aspects of AI and advanced system design — along with cyber law and other topics critical to cybersecurity — that engineering students may not learn in their other courses. The Cybersecurity Clinic aims to round out the knowledge of students from every discipline. The course aims to broaden those students’ knowledge, too, by inviting at least half a dozen guest speakers each semester from industry, other universities and MIT academic departments, industry, and/or relevant public agencies.

This past spring, for example, the lineup of lecturers included Dan Ricci, the founder of Industrial Data Works, on the modeling of risk in energy systems within budget-constrained environments; Gus Serino, president of I&C Secure Inc., on operational-technology cybersecurity for industrial control systems; and representatives from the MassCyberCenter and the Cybersecurity Infrastructure Security Agency providing overviews of their respective state- and federal-level organizations’ programs and initiatives.

“There are highly specialized things to learn, especially about the ways AI is changing cybersecurity, that we need help teaching,” Susskind says. “The rate at which the field of cybersecurity is changing means that most academics will have a very hard time keeping up.”

A roadmap for improvement

Clinic students spend the first four weeks of the semester preparing for field assignments. A series of online modules, supplemented by class discussion, outline the scope and nature of cyberattacks against critical urban infrastructure; review the 23 risk areas most relevant to their type of clients; and provide guidance for each step of the assessment process. This includes simulations of tricky client interactions. What if clients don’t take students seriously, or fail to provide the necessary information? What if they argue to receive a more positive assessment than the facts warrant?

“I’ve never ever had a class that prepared us for such realistic scenarios before,” says Diego Contreras, a rising senior majoring in computer science and engineering who completed the course this spring.

The modules culminate in an exam students must pass on their first try to receive a field assignment. For the remainder of the semester, they’ll receive continued support via weekly class meetings and get faculty input on their drafted reports, but the onus is on students to coordinate their team’s activities and build client trust.

“You represent MIT, and that is quite the responsibility,” Contreras says. “This course has given me people skills I wouldn’t have developed in any other context.”

“The most delicate aspect of the project was balancing our assessment findings,” says Zev Moore ’26, who took the class last fall as a senior studying mathematical economics and finance. “Our approach was to provide important feedback while simultaneously validating the positive security measures our client already had in place, which ensured our report felt like a collaborative roadmap for improvement.”

Certain key recommendations show up in the majority of reports. For example, clients are advised to inventory all hardware and software tied into their network and track who has access; patch software and back up data regularly; require multi-factor authentication and frequent password updates; train employees not to open attachments from unknown parties; prepare an attack response plan that clarifies lines of authority and includes the organization’s stance on paying ransoms; and only use vendors with good cybersecurity hygiene.

“None of these items is costly,” Susskind says. “Together, they will probably avoid 80 percent or more of the possible cost and danger of cyberattacks.”

Spreading the model

To date, more than 120 students have completed the full course at MIT. The online modules that prepare students for certification are freely available to the public as a massive open online course on MITx called Cybersecurity for Critical Urban Infrastructure, which has attracted tens of thousands of learners. The modules are also used by universities with their own cybersecurity clinics — a growing cohort, thanks in part to a consortium (with 61 member institutions and counting) co-founded by MIT in 2021 with the University of California at Berkeley, Indiana University, and the University of Alabama.

Most student teams wrap up client work after finalizing their recommendations; a few have volunteered to stay on after semester’s end to advise on implementation. In either case, Susskind and Chun check in periodically with clients for at least two years following each engagement.

“We often hear of the vulnerability assessment report serving as the organization's blueprint for their short-term, mid-term, and long-term agenda to be more prepared for future attacks,” says Chun. “We primarily work with IT directors or chief technology officers, and many of them have been telling us post-engagement that they shared the MIT report with the city or town leadership and were able to convince them they needed extra budget or a specific line item. They were using the student report as leverage to say, ‘it’s not just me saying it. We have a credible team who dedicated their time and these are the findings.’

“It's really a humbling experience,” Chun adds, “when some of our past clients reach out to us again after some time to say: ‘Now we have different people, we just purchased new equipment. Can we do this all over again?’”

Robots walking down the street, surrounded by astounded onlookers, is an increasingly common sight. But these machines aren’t yet the do-it-all assistants you’d want working in a kitchen or factory, and a major bottleneck is data. Much like humans, robots learn best by experience. The challenge is that it’s labor-intensive and time-consuming to physically teach these machines so many actions across different settings. 

“One natural idea is to use simulation as a training ground. While there has been significant progress over the last few years in the physics engines that power robotics simulators, one of the remaining challenges has been creating sufficiently rich and diverse simulation content to capture the complexity of the real world,” says Russ Tedrake, the Toyota Professor of Electrical Engineering and Computer Science (EECS), Aeronautics and Astronautics, and Mechanical Engineering at MIT, and a principal investigator at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).

It turns out that AI agents, or semi-autonomous programs that “think” and complete well-defined tasks, could help produce the lifelike virtual settings that robots need. The new “SceneSmith” system developed by researchers at MIT CSAIL and Toyota Research Institute uses three agents to piece together the objects, walls, and overall look of a 3D scene. Its recreations of indoor spaces such as restaurants, bedrooms, and hotels are more realistic and detailed than prior systems, helping robots practice skills and try out different ways of doing tasks before they’re powered on. In turn, engineers save time on real-world testing.

The agents have a sense of how everyday places are supposed to look because they each call on a multi-modal system called a vision-language model (VLM), specifically the state-of-the-art VLM GPT-5.2. It’s trained on lots of text and images from the internet to handle more visual prompts. This advanced model gives each agent a sort of spatial knowledge: First, a “designer” agent generates the elements of a scene, then a “critic” advises whether it looks realistic, and finally, an “orchestrator” manages their back-and-forth, deciding when the design is done. Once the three VLMs wrap up their creative collaboration, the scene is ready to load directly into physics simulation software.

“We’ve found that the system can construct 3D scenes the way a human designer would,” says MIT EECS PhD student Nicholas Pfaff, a CSAIL researcher and a lead author on a paper with Tedrake presenting the work. “We made over 1,300 scenes using a leading VLM that has internet-scale priors, and it made insanely creative and diverse arrangements. I hadn’t taught the system to do that in the prompts; it just improvised.”

Talk to my agent

Thanks to VLM agents, you can ask SceneSmith to do things like “generate a garage with a car, a workbench, tires stacked in the corner, and a ladder against the wall,” and get a virtual playground rich with objects a robot can tinker with. These rooms are decorated with up to six times more items per scene than prior methods, making them great for helping robots learn skills such as putting a cup in the sink, placing fruit on plates, and moving a soda can from a shelf to a table.

With so many rich virtual environments handy, you can evaluate whether your robot is ready for deployment without so much trial and error in the physical world. The researchers tested out different action plans (also called “policies”) in SceneSmith’s digital worlds, generating 100 unique spaces in the process. A VLM agent evaluated each attempt, and it found the robot’s plans were faulty, with the machine often failing at its chores. Humans agreed with the model’s verdicts over 99 percent of the time, which could help roboticists weed out flawed approaches in simulation before a robot moves in the real world.

But how realistic are these virtual worlds, really? It can be difficult to prove outright, so the researchers approached the question from several angles. The most telling test: they dropped a pretrained robot policy — an AI controller trained largely on real-world data, which had never seen a SceneSmith scene — into the generated environments. In one test, users told the system to “take the apple from the bowl and place it onto the cutting board,” and the simulated robot did exactly that. If the scenes didn’t closely resemble the real settings the policy had learned from, it simply wouldn’t have worked. 

The team also teleoperated robots through the virtual spaces, guiding them to open cabinets, put away bottles, and navigate between rooms. Their experiments revealed that the environments hold up under sustained physical interaction, expanding beyond visual inspection.

Behind the scenes

The agents that SceneSmith uses each have a well-defined role in the generative process, fleshing out scenes in stages. They essentially create a floor plan and bring it to life. 

Let’s say you wanted to create a scene similar to the first floor of a house. The “designer” VLM would start with a general layout, which the “critic” reviews, and then the “orchestrator” signs off. The agents repeat this approach for each step: adding furniture, placing objects on walls and then ceilings, and finally, dropping in objects that robots can manipulate. For example, the VLMs can add cabinets that the robots can open and close — an articulated item, which prior baselines didn’t often have.

At each stage, the second VLM ensures the scene is practical, advising that a bathtub is removed from a living room, for example. The third VLM ensures a high-quality scene is generated, even taking the design process a few turns back if the visuals aren’t up to par. Once the three VLMs wrap up their creative collaboration, the mechanics of the physical world are added via simulation software.

With a sound understanding of how rooms should look, where objects should be placed, and real-world physics, SceneSmith has a noticeable edge over prior methods. Compared to scene-generation baselines such as “HSM” and “Holodeck,” SceneSmith made environments with more objects, including a private office, a pottery store, and even a Minecraft-themed gaming room.

SceneSmith was also a favorite among over 200 users. They found the system’s visuals to be more realistic over 90 percent of the time. They also observed that, generally speaking, it followed prompts more closely than other approaches did. In other words, it was the best at generating the virtual playgrounds users actually wanted to see.

A system of many talents

Realism, diversity, and richness are all strong suits for SceneSmith, even when it comes to generating individual 3D objects. You can prompt it to create a rolling serving cart, and it’ll make a 2D image that it then turns into a detailed model with physical properties like mass, friction, and inertia.

Such a detailed process does come with a speed trade-off, though. It can take multiple hours to produce a single scene because the agents are creating and closely scrutinizing each object. With more computing power, the system could see dramatic increases in efficiency. CSAIL engineers are also hoping to expand to deformable objects (like sponges), should extensive 3D libraries become available.

“SceneSmith represents a significant advance in this regard by providing an agentic framework for generating simulation-ready indoor environments just from a simple text prompt,” says Jeremy Binagia, an applied scientist at Amazon Robotics who wasn’t involved in the research. “It advances the state of the art in several ways, including pushing the limits of the density of objects in the simulated environment, ensuring that all of the objects are physically accurate (versus just being visually realistic), and creating assets that are not constrained to a fixed library, since they can be generated via text-to-3D.”

Pfaff and Tedrake wrote the paper with Thomas Cohn SM ’24, an MIT PhD student and CSAIL researcher; and Toyota Research Institute roboticists Sergey Zakharov and Rick Cory SM ’08, PhD ’10. Their work was supported, in part, by Amazon, the U.S. Office of Naval Research, the Toyota Research Institute, and the U.S. National Science Foundation.

The team presented their findings as a spotlight at last week’s International Conference on Machine Learning. 

With the exploding popularity of generative artificial intelligence, many open-source models are now available online for anyone to adapt for their task, such as generating product renderings in a certain artistic style. 

But these models also find their way into the hands of nefarious actors who may optimize them to produce illegal content, like hate speech or child sexual abuse material (CSAM). This is a growing problem — the National Center for Missing and Exploited Children received more than 1.5 million reports of AI-generated CSAM in 2025, an increase from 67,000 in 2024.

Engineers usually test AI for harmful capabilities by prompting the model and inspecting its outputs, but this is impossible for CSAM, since it is illegal in the U.S to generate such content, regardless of intent.

To avoid this dilemma and improve AI safety, a team of MIT scientists, led by graduate student Vinith Suriyakumar and associate professors Ashia Wilson and Marzyeh Ghassemi, joined forces with researchers from Thorn to develop a new auditing approach that determines whether a model can produce CSAM, without prompting it. Thorn is a child safety nonprofit whose mission is to transform how children are protected from sexual abuse and exploitation in the digital age.

Their technique examines how the inner workings of a model have been adapted, but it never generates an output. By examining hidden representations, it can reliably infer whether a model has been specialized to produce harmful imagery.

When tested, the auditing procedure identified model variations that had been specialized to generate CSAM with 100 percent accuracy. A hosting platform could use this technique to flag unsafe models and quickly remove them or prevent them from being uploaded in the first place.

“This unlocks a new avenue for platforms that host open-source models and for law enforcement to actually test whether a model is capable of generating CSAM. Before, we had no way of measuring this. It was a huge blind spot that some people were taking advantage of. Now, we can address an AI safety problem that is having severe negative impacts,” says Vinith Suriyakumar, an MIT electrical engineering and computer science (EECS) graduate student and lead author of a paper on this technique.

Suriyakamur and Wilson, the Lister Borthers Career Develop Professor in EECS and a principal investigator in the Laboratory for Information and Decision Systems (LIDS), are joined on the paper by Lena Stempfle, an MIT postdoc; Ghassemi, an associate professor in EECS and a member of the Institute of Medical Engineering Sciences (IMES) and LIDS; and others at Boston University and Thorn. The paper was be presented as a spotlight at the “Trustworthy AI for Good” workshop at the International Conference on Machine Learning.

Auditing adaptations

Recent techniques have made it easier for users to specialize a generative AI model for their task through a process known as fine-tuning. 

Rather than retraining the entire model on a task-specific dataset, individuals can utilize an algorithm called low-rank adaptation (LoRA) to specialize the model in a more efficient manner.

This has led to a wave of new generative AI model variants for a variety of purposes, like producing watercolor images that mimic an artistic movement. But it has also enabled malicious actors to create models that can generate high-quality CSAM and other harmful imagery.

To audit a model, engineers typically prompt it for harmful content and check its outputs, but this manual auditing procedure is not scalable. In addition, repeatedly generating heinous images can have negative psychological impacts on human evaluators. 

This evaluation method quickly falls apart when testing CSAM, which is illegal to generate for any purpose in the U.S. and many other international jurisdictions.

“We are in this very difficult situation where, based on the law itself, we cannot use the de facto means of evaluation. We had to throw out the entire toolkit and take a different approach,” Suriyakumar says.

After learning about this conundrum, the researchers joined forces with Thorn, to address this issue.

A nongenerative solution

Instead of focusing on outputs, the researchers targeted the modifications a LoRA algorithm makes during fine-tuning. 

Their technique probes these modifications, called LoRA adaptors, to determine whether a model has been specialized for a harmful capability, without generating an output.

Using a technique called Gaussian probing, the researchers feed the model a set of random data points and analyze how it manipulates those data within its multilayer internal structure. 

“We never run the model all the way to the end or prompt the model, so we never generate images,” Suriyakumar explains.

The researchers capture those modifications at multiple time points within the model’s inner structure and average them to summarize how the LoRA adaptor changed the model’s computation. They found these responses to be a strong signal of how a model had been specialized.

They tested their method on variations of three types of models, comparing the results to ground-truth data from LoRA adaptors known for generating CSAM, other harmful images, and safe content. 

Their method was 100 percent accurate in identifying models that had been adapted to generate CSAM. 

“There is a huge bucket of child safety concerns with AI, and these are real concerns that need to be addressed. A lot of children are being harmed by AI deepfakes. We’ve shown that Gaussian probing can be a very useful tool, and we hope the research community really pours more attention into this problem,” Wilson says.

Importantly, their technique is scalable and would be relatively inexpensive to implement. Since thousands of model variations are published online every month, scalability is key to help auditors remove harmful adaptations before they are widely distributed.

Gaussian probing is also more robust than some other auditing techniques, since a nefarious actor would need to carefully alter the inner workings of the base model to avoid detection.

In the future, the researchers want to evaluate their technique on a larger set of model variations and explore whether Gaussian probing can detect harmful capabilities in base models before they are adapted.

“Now we have a technological approach to partially address this concern. So much effort was poured into this collaboration, which enabled us to tackle a really hard problem that is harming so many children, nationally and around the world. Hopefully, we can have a transformative impact in this area,” Ghassemi says.

This work was supported, in part, by the Bridgewater AIA Labs Research Fellowship.

Infrared cameras can be used to spot useful information that our eyes can’t see, such as gases escaping from a pipeline, chemicals in the atmosphere, or heat leaking from a building. But sensing infrared light in sophisticated ways still requires expensive and bulky systems.

Now MIT researchers have created a chip-based optical device that can dynamically control incoming infrared light, to act as a tunable lens that gathers additional information for infrared cameras. Each microscopic pixel of the device’s lens can control infrared light independently, allowing it to change its focus and help cameras detect different signals without moving parts.

The system is described in a paper published in Nature Communications. The researchers also explain how they built a lab-scale demonstration using mostly conventional manufacturing processes in a semiconductor chip factory, suggesting the approach could be implemented at industrial scales.

The technology could enable compact, tunable infrared cameras for more dynamic thermal imaging, chemical sensing, pollution monitoring, and even new kinds of optical computing.

“This could give us more information as we study space, or help with environmental protections where you want to monitor for specific compounds in the atmosphere,” explains first author Cosmin-Constantin Popescu PhD ’25. “Thermal imaging is another application, and you can think of military applications where night vision goggles are currently being used. Basically, a lot of organic molecules absorb in the mid-infrared wavelength, and you could use this system to detect them.”

Joining Popescu on the paper are MIT PhD students Maarten Robbert Anton Peters and Khoi Phuong Dao; Dynasil company scientists Oleg Maksimov and Harish Bhandari; University of Central Florida PhD candidate Kathleen Richardson and scientist Rashi Sharma; University of Washington Professor Arka Majumdar; Korea Advanced Institute of Science and Technology Associate Professor Hyun Jung Kim; MIT postdoc Rui Chen; Luigi Ranno PhD ’25; Brian Mills ’20, PhD ’26; Draper Laboratory scientist Dennis Calahan; MIT principal investigator Tian Gu; and Juejun Hu, MIT’s John F. Elliott Professor of Materials Science and Engineering.

Chip-based lenses

In recent years, researchers have developed ways to dynamically control light by etching tiny, precise patterns on transparent materials known as “metasurfaces,” which could enable more compact, programmable cameras and other advanced optical devices.

Hu’s research group at MIT has experimented with a class of metasurfaces that shift from solid to liquid after heat is applied. The phase changes can be harnessed to control how the materials interact with light. In 2021, Hu and collaborators created a miniature lens that could adjust its focus to different depths through such phase changes.

The device worked reliably, but it could only adjust focus all at once across the entire material, which is how most metasurfaces work. For their new study the researchers wanted to build on that approach to control light independently at each microscopic pixel of the material.

“Most active metasurfaces trying to do single-pixel tuning need wires going to every pixel, and how you route the wires becomes a big issue,” Hu explains. “The best approach so far has been one-dimensional pixel control with a bunch of wires.”

The researchers also wanted to create a system that worked with the mid-infrared wavelength of light, which the human eye can’t see but is useful for detecting heat signatures and molecules including methane and propane. Mid-infrared detection devices are already used to detect gas leaks and study Earth’s atmosphere, and for a number of defense and aerospace applications.

To build their system, the researchers adapted an approach commonly used in displays in which two layers of neatly packed copper wires are placed on top of each other perpendicularly. Below the wires is a layer of doped silicon that generates heat at the cross points of the wires and sits on top of the phase-change material. The silicon’s heat is used to switch each pixel of the material between crystalline and amorphous structures, which changes how the material interacts with the infrared light coming in. The silicon also includes a diode selector, which helps prevent unintended currents from leaking through neighboring pixels.

“We did some calculations showing this architecture allows us to scale to potentially millions of pixels without having any issues with the [unintended] currents,” Hu says. “The key innovation is this crossbar architecture, which creates a scalable way to increase the pixel-level switching of metasurfaces. We didn’t invent this architecture — it’s used in displays — but it’s the first time anyone’s used it for active phase-change metasurfaces to show you can get pixel-level control. People have been working toward two-dimensional pixel-level control for a long time, and it’s the first time anyone’s implemented it.”

The researchers worked with equipment in MIT.nano and with a factory that manufactures semiconductor chips, ultimately creating a two-dimensional system that featured a 6-by-6 metasurface pixel array. They tested their system and found it could switch on and off reliably.

“We found this mesh architecture to be very resilient,” Popescu says. “You don’t want these materials to switch once and not work anymore. You want it to switch a large number of times: maybe tens of thousands of times or more.”

Scaling up

The researchers say integrating part of their system’s design into existing semiconductor manufacturing should help it move beyond a research prototype.

“As you want to scale up, you need something that’s part of a consistent process, and that’s why chip foundry manufacturing becomes so important,” Hu says. “Working with a semiconductor foundry with well-defined process control is very powerful. It also allows you to implement each of the components into a single efficient process.”

The researchers are working to add more pixels to their array and develop more robust versions of their system so that it can start capturing more infrared information.

“In lots of cases when you’re taking images, you have prior knowledge of what you’re looking for,” Hu says. “You might be looking for a human in a dark room, or some specific features in an image, like a tree, and that prior information can be useful because now you can configure this system to specifically highlight those features.”

Hu also notes that researchers have used metasurfaces to emulate computational neural networks that power AI systems, though he notes that applications could be farther away from taking hold.

“This could enable more effective optical computing, where metasurfaces are used to encode network weights in neural networks,” Hu explains. “When light passes through the material, it interacts with the metasurface, and that information gets encoded in such a way that you can infer computational results. Researchers have already used this approach to emulate very complex neural networks.”

The work was supported, in part, by the U.S. Air Force, the U.S. National Science Foundation, the National Research Foundation of Korea, and the Draper Scholar Program.