Researchers from MIT and Dana-Farber Cancer Institute have discovered that a class of peptides expressed in pancreatic cancer cells could be a promising target for T-cell therapies and other approaches that attack pancreatic tumors.

Known as cryptic peptides, these molecules are produced from sequences in the genome that were not thought to encode proteins. Such peptides can also be found in some healthy cells, but in this study, the researchers identified about 500 that appear to be found only in pancreatic tumors.

The researchers also showed they could generate T cells targeting those peptides. Those T cells were able to attack pancreatic tumor organoids derived from patient cells, and they significantly slowed down tumor growth in a study of mice.

“Pancreas cancer is one of the most challenging cancers to treat. This study identifies an unexpected vulnerability in pancreas cancer cells that we may be able to exploit therapeutically,” says Tyler Jacks, the David H. Koch Professor of Biology at MIT and a member of the Koch Institute for Integrative Cancer Research.

Jacks and William Freed-Pastor, a physician-scientist in the Hale Family Center for Pancreatic Cancer Research at Dana-Farber Cancer Institute and an assistant professor at Harvard Medical School, are the senior authors of the study, which appears today in Science. Zackery Ely PhD ’22 and Zachary Kulstad, a former research technician at Dana-Farber Cancer Institute and the Koch Institute, are the lead authors of the paper.

Cryptic peptides

Pancreatic cancer has one of the lowest survival rates of any cancer — about 10 percent of patients survive for five years after their diagnosis.

Most pancreatic cancer patients receive a combination of surgery, radiation treatment, and chemotherapy. Immunotherapy treatments such as checkpoint blockade inhibitors, which are designed to help stimulate the body’s own T cells to attack tumor cells, are usually not effective against pancreatic tumors. However, therapies that deploy T cells engineered to attack tumors have shown promise in clinical trials.

These therapies involve programming the T-cell receptor (TCR) of T cells to recognize a specific peptide, or antigen, found on tumor cells. There are many efforts underway to identify the most effective targets, and researchers have found some promising antigens that consist of mutated proteins that often show up when pancreatic cancer genomes are sequenced.

In the new study, the MIT and Dana-Farber team wanted to extend that search into tissue samples from patients with pancreatic cancer, using immunopeptidomics — a strategy that involves extracting the peptides presented on a cell surface and then identifying the peptides using mass spectrometry.

Using tumor samples from about a dozen patients, the researchers created organoids — three-dimensional growths that partially replicate the structure of the pancreas. The immunopeptidomics analysis, which was led by Jennifer Abelin and Steven Carr at the Broad Institute, found that the majority of novel antigens found in the tumor organoids were cryptic antigens. Cryptic peptides have been seen in other types of tumors, but this is the first time they have been found in pancreatic tumors.

Each tumor expressed an average of about 250 cryptic peptides, and in total, the researchers identified about 1,700 cryptic peptides.

“Once we started getting the data back, it just became clear that this was by far the most abundant novel class of antigens, and so that’s what we wound up focusing on,” Ely says.

The researchers then performed an analysis of healthy tissues to see if any of these cryptic peptides were found in normal cells. They found that about two-thirds of them were also found in at least one type of healthy tissue, leaving about 500 that appeared to be restricted to pancreatic cancer cells.

“Those are the ones that we think could be very good targets for future immunotherapies,” Freed-Pastor says.

Programmed T cells

To test whether these antigens might hold potential as targets for T-cell-based treatments, the researchers exposed about 30 of the cancer-specific antigens to immature T cells and found that 12 of them could generate large populations of T cells targeting those antigens.

The researchers then engineered a new population of T cells to express those T-cell receptors. These engineered T cells were able to destroy organoids grown from patient-derived pancreatic tumor cells. Additionally, when the researchers implanted the organoids into mice and then treated them with the engineered T cells, tumor growth was significantly slowed.

This is the first time that anyone has demonstrated the use of T cells targeting cryptic peptides to kill pancreatic tumor cells. Even though the tumors were not completely eradicated, the results are promising, and it is possible that the T-cells’ killing power could be strengthened in future work, the researchers say.

Freed-Pastor’s lab is also beginning to work on a vaccine targeting some of the cryptic antigens, which could help stimulate patients’ T cells to attack tumors expressing those antigens. Such a vaccine could include a collection of the antigens identified in this study, including those frequently found in multiple patients.

This study could also help researchers in designing other types of therapy, such as T cell engagers — antibodies that bind an antigen on one side and T cells on the other, which allows them to redirect any T cell to kill tumor cells.

Any potential vaccine or T cell therapy is likely a few years away from being tested in patients, the researchers say.

The research was funded in part by the Hale Family Center for Pancreatic Cancer Research, the Lustgarten Foundation, Stand Up To Cancer, the Pancreatic Cancer Action Network, the Burroughs Wellcome Fund, a Conquer Cancer Young Investigator Award, the National Institutes of Health, and the National Cancer Institute.

It’s been a scientific truth so universally acknowledged that it’s taught in classrooms and repeated in pop-science videos: An egg is strongest when dropped vertically, on its ends. But when MIT engineers actually put this assumption to the test, they cracked open a surprising revelation. 

Their experiments revealed that eggs dropped on their sides — not their tips — are far more resilient, thanks to a clever physics trick: Sideways eggs bend like shock absorbers, trading stiffness for superior energy absorption. Their open-access findings, published today in Communications Physics, don’t just rewrite the rules of the classic egg drop challenge — they’re a lesson in intellectual humility and curiosity. Even “settled” science can yield surprises when approached with rigor and an open mind.

At first glance, an eggshell may seem fragile, but its strength is a marvel of physics. Crack an egg on its side for your morning omelet and it breaks easily. Intuitively, we believe eggs are harder to break when positioned vertically. This notion has long been a cornerstone of the classic “egg drop challenge,” a popular science activity in STEM classrooms across the country that introduces students to physics concepts of impact, force, kinetic energy, and engineering design.

The annual egg drop competition is a highlight of first-year orientation in the MIT Department of Civil and Environmental Engineering. “Every year we follow the scientific literature and talk to the students about how to position the egg to avoid breakage on impact,” says Tal Cohen, associate professor of civil and environmental engineering and mechanical engineering. “But about three years ago, we started to question whether vertical really is stronger.” 

That curiosity sparked an initial experiment by Cohen’s research group, which leads the department’s egg drop event. They decided to put their remaining box of eggs to the test in the lab. “We expected to confirm the vertical side was tougher based on what we had read online,” says Cohen. “But when we looked at the data — it was really unclear.”

What began as casual inquiry evolved into a research project. To rigorously investigate the strength of both egg orientations, the researchers conducted two types of experiments: static compression tests, which applied gradually increasing force to measure stiffness and toughness; and dynamic drop tests, to quantify the likelihood of breaking on impact.

“In the static testing, we wanted to keep an egg at a standstill and push on it until it cracked,” explains Avishai Jeselsohn, an undergraduate researcher and an author in the study. “We used thin paper supports to precisely orient the eggs vertically and horizontally.”

What the researchers found was it required the same amount of force to initiate a crack in both orientations. “However, we noticed a key difference in how much the egg compressed before it broke, says Joseph Bonavia, PhD candidate who contributed to the work. “The horizontal egg compressed more under the same amount of force, meaning it was more compliant.”

Using mechanical modeling and numerical simulations to validate results of their experiments, the researchers concluded that even though the force to crack the egg was consistent, the horizontal eggs absorbed more energy due to their compliance. “This suggested that in situations where energy absorption is important, like in a drop, the horizontal orientation might be more resilient. We then performed the dynamic drop tests to see if this held true in practice,” says Jeselsohn.

The researchers designed a drop setup using solenoids and 3D-printed supports, ensuring simultaneous release and consistent egg orientation. Eggs were dropped from various heights to observe breakage patterns. The result: Horizontal eggs cracked less frequently when dropped from the same height.

“This confirmed what we saw in the static tests,” says Jeselsohn. “Even though both orientations experienced similar peak forces, the horizontal eggs absorbed energy better and were more resistant to breaking.”

Challenging common notions

The study reveals a misconception in popular science regarding the strength of an egg when subjected to impact. Even seasoned researchers in fracture mechanics initially assumed that vertical oriented eggs would be stronger. “It’s a widespread, accepted belief, referenced in many online sources,” notes Jeselsohn.

Everyday experience may reinforce that misconception. After all, we often crack eggs on their sides when cooking. “But that’s not the same as resisting impact,” explains Brendan Unikewicz, a PhD candidate and author on the paper. “Cracking an egg for cooking involves applying locally focused force for a clean break to retrieve the yolk, while its resistance to breaking from a drop involves distributing and absorbing energy across the shell.”

The difference is subtle but significant. A vertically oriented egg, while stiffer, is more brittle under sudden force. A horizontal egg, being more compliant, bends and absorbs energy over a greater distance — similar to how bending your knees during a fall softens the blow.

“In a way, our legs are ‘weaker’ when bent, but they’re actually tougher in absorbing impact,” Bonavia adds. “It’s the same with the egg. Toughness isn’t just about resisting force — it’s about how that force is dissipated.”

The research findings offer more than insight into egg behavior — they underscore a broader scientific principle: that widely accepted “truths” are worth re-examining.

Which came first?

“It’s great to see an example of ‘received wisdom’ being tested scientifically and shown to be incorrect. There are many such examples in the scientific literature, and it’s a real problem in some fields because it can be difficult to secure funding to challenge an existing, ‘well-known’ theory,” says David Taylor, emeritus professor in the Department of Mechanical, Manufacturing and Biomedical Engineering at Trinity College Dublin, who was not affiliated with the study.

The authors hope their findings encourage young people to remain curious and recognize just how much remains to be discovered in the physical world.

“Our paper is a reminder of the value in challenging common notions and relying on empirical evidence, rather than intuition,” says Cohen. “We hope our work inspires students to stay curious, question even the most familiar assumptions, and continue thinking critically about the physical world around them. That’s what we strive to do in our group — constantly challenge what we’re taught through thoughtful inquiry.”

In addition to Cohen, who serves as senior author on the paper, co-authors include lead authors Antony Sutanto MEng ’24 and Suhib Abu-Qbeitah, a postdoc at Tel Aviv University, as well as the following MIT affiliates: Avishai Jeselsohn, an undergraduate in mechanical engineering; Brendan Unikewicz, a PhD candidate in mechanical engineering; Joseph Bonavia, a PhD candidate in mechanical engineering; Stephen Rudolph, a lab instructor in civil and environmental engineering; Hudson Borja da Rocha, an MIT postdoc in civil and environmental engineering; and Kiana Naghibzadeh, Engineering Excellence Postdoctoral Fellow in civil and environmental engineering. The research was funded by U.S. Office of Naval Research with support from the U.S. National Science Foundation. 

MIT engineers are getting in on the robotic ping pong game with a powerful, lightweight design that returns shots with high-speed precision.

The new table tennis bot comprises a multijointed robotic arm that is fixed to one end of a ping pong table and wields a standard ping pong paddle. Aided by several high-speed cameras and a high-bandwidth predictive control system, the robot quickly estimates the speed and trajectory of an incoming ball and executes one of several swing types — loop, drive, or chop — to precisely hit the ball to a desired location on the table with various types of spin.

In tests, the engineers threw 150 balls at the robot, one after the other, from across the ping pong table. The bot successfully returned the balls with a hit rate of about 88 percent across all three swing types. The robot’s strike speed approaches the top return speeds of human players and is faster than that of other robotic table tennis designs.

Now, the team is looking to increase the robot’s playing radius so that it can return a wider variety of shots. Then, they envision the setup could be a viable competitor in the growing field of smart robotic training systems.

Beyond the game, the team says the table tennis tech could be adapted to improve the speed and responsiveness of humanoid robots, particularly for search-and-rescue scenarios, and situations in a which a robot would need to quickly react or anticipate.

“The problems that we’re solving, specifically related to intercepting objects really quickly and precisely, could potentially be useful in scenarios where a robot has to carry out dynamic maneuvers and plan where its end effector will meet an object, in real-time,” says MIT graduate student David Nguyen.

Nguyen is a co-author of the new study, along with MIT graduate student Kendrick Cancio and Sangbae Kim, associate professor of mechanical engineering and head of the MIT Biomimetics Robotics Lab. The researchers will present the results of those experiments in a paper at the IEEE International Conference on Robotics and Automation (ICRA) this month.

Precise play

Building robots to play ping pong is a challenge that researchers have taken up since the 1980s. The problem requires a unique combination of technologies, including high-speed machine vision, fast and nimble motors and actuators, precise manipulator control, and accurate, real-time prediction, as well as higher-level planning of game strategy.

“If you think of the spectrum of control problems in robotics, we have on one end manipulation, which is usually slow and very precise, such as picking up an object and making sure you’re grasping it well. On the other end, you have locomotion, which is about being dynamic and adapting to perturbations in your system,” Nguyen explains. “Ping pong sits in between those. You’re still doing manipulation, in that you have to be precise in hitting the ball, but you have to hit it within 300 milliseconds. So, it balances similar problems of dynamic locomotion and precise manipulation.”

Ping pong robots have come a long way since the 1980s, most recently with designs by Omron and Google DeepMind that employ artificial intelligence techniques to “learn” from previous ping pong data, to improve a robot’s performance against an increasing variety of strokes and shots. These designs have been shown to be fast and precise enough to rally with intermediate human players.

“These are really specialized robots designed to play ping pong,” Cancio says. “With our robot, we are exploring how the techniques used in playing ping pong could translate to a more generalized system, like a humanoid or anthropomorphic robot that can do many different, useful things.”

Game control

For their new design, the researchers modified a lightweight, high-power robotic arm that Kim’s lab developed as part of the MIT Humanoid — a bipedal, two-armed robot that is about the size of a small child. The group is using the robot to test various dynamic maneuvers, including navigating uneven and varying terrain as well as jumping, running, and doing backflips, with the aim of one day deploying such robots for search-and-rescue operations.

Each of the humanoid’s arms has four joints, or degrees of freedom, which are each controlled by an electrical motor. Cancio, Nguyen, and Kim built a similar robotic arm, which they adapted for ping pong by adding an additional degree of freedom in the wrist to allow for control of a paddle.

The team fixed the robotic arm to a table at one end of a standard ping pong table and set up high-speed motion capture cameras around the table to track balls that are bounced at the robot. They also developed optimal control algorithms that predict, based on the principles of math and physics, what speed and paddle orientation the arm should execute to hit an incoming ball with a particular type of swing: loop (or topspin), drive (straight-on), or chop (backspin).

They implemented the algorithms using three computers that simultaneously processed camera images, estimated a ball’s real-time state, and translated these estimations to commands for the robot’s motors to quickly react and take a swing.

After consecutively bouncing 150 balls at the arm, they found the robot’s hit rate, or accuracy of returning the ball, was about the same for all three types of swings: 88.4 percent for loop strikes, 89.2 percent for chops, and 87.5 percent for drives. They have since tuned the robot’s reaction time and found the arm hits balls faster than existing systems, at velocities of 20 meters per second.

In their paper, the team reports that the robot’s strike speed, or the speed at which the paddle hits the ball, is on average 11 meters per second. Advanced human players have been known to return balls at speeds of between 21 to 25 meters second. Since writing up the results of their initial experiments, the researchers have further tweaked the system, and have recorded strike speeds of up to 19 meters per second (about 42 miles per hour).

“Some of the goal of this project is to say we can reach the same level of athleticism that people have,” Nguyen says. “And in terms of strike speed, we’re getting really, really close.”

Their follow-up work has also enabled the robot to aim. The team incorporated control algorithms into the system that predict not only how but where to hit an incoming ball. With its latest iteration, the researchers can set a target location on the table, and the robot will hit a ball to that same location.

Because it is fixed to the table, the robot has limited mobility and reach, and can mostly return balls that arrive within a crescent-shaped area around the midline of the table. In the future, the engineers plan to rig the bot on a gantry or wheeled platform, enabling it to cover more of the table and return a wider variety of shots.

“A big thing about table tennis is predicting the spin and trajectory of the ball, given how your opponent hit it, which is information that an automatic ball launcher won’t give you,” Cancio says. “A robot like this could mimic the maneuvers that an opponent would do in a game environment, in a way that helps humans play and improve.”

This research is supported, in part, by the Robotics and AI Institute.

A human clearing junk out of an attic can often guess the contents of a box simply by picking it up and giving it a shake, without the need to see what’s inside. Researchers from MIT, Amazon Robotics, and the University of British Columbia have taught robots to do something similar.

They developed a technique that enables robots to use only internal sensors to learn about an object’s weight, softness, or contents by picking it up and gently shaking it. With their method, which does not require external measurement tools or cameras, the robot can accurately guess parameters like an object’s mass in a matter of seconds.

This low-cost technique could be especially useful in applications where cameras might be less effective, such as sorting objects in a dark basement or clearing rubble inside a building that partially collapsed after an earthquake.

Key to their approach is a simulation process that incorporates models of the robot and the object to rapidly identify characteristics of that object as the robot interacts with it. 

The researchers’ technique is as good at guessing an object’s mass as some more complex and expensive methods that incorporate computer vision. In addition, their data-efficient approach is robust enough to handle many types of unseen scenarios.

“This idea is general, and I believe we are just scratching the surface of what a robot can learn in this way. My dream would be to have robots go out into the world, touch things and move things in their environments, and figure out the properties of everything they interact with on their own,” says Peter Yichen Chen, an MIT postdoc and lead author of a paper on this technique.

His coauthors include fellow MIT postdoc Chao Liu; Pingchuan Ma PhD ’25; Jack Eastman MEng ’24; Dylan Randle and Yuri Ivanov of Amazon Robotics; MIT professors of electrical engineering and computer science Daniela Rus, who leads MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL); and Wojciech Matusik, who leads the Computational Design and Fabrication Group within CSAIL. The research will be presented at the International Conference on Robotics and Automation.

Sensing signals

The researchers’ method leverages proprioception, which is a human or robot’s ability to sense its movement or position in space.

For instance, a human who lifts a dumbbell at the gym can sense the weight of that dumbbell in their wrist and bicep, even though they are holding the dumbbell in their hand. In the same way, a robot can “feel” the heaviness of an object through the multiple joints in its arm.

“A human doesn’t have super-accurate measurements of the joint angles in our fingers or the precise amount of torque we are applying to an object, but a robot does. We take advantage of these abilities,” Liu says.

As the robot lifts an object, the researchers’ system gathers signals from the robot’s joint encoders, which are sensors that detect the rotational position and speed of its joints during movement. 

Most robots have joint encoders within the motors that drive their moveable parts, Liu adds. This makes their technique more cost-effective than some approaches because it doesn’t need extra components like tactile sensors or vision-tracking systems.

To estimate an object’s properties during robot-object interactions, their system relies on two models: one that simulates the robot and its motion and one that simulates the dynamics of the object.

“Having an accurate digital twin of the real-world is really important for the success of our method,” Chen adds.

Their algorithm “watches” the robot and object move during a physical interaction and uses joint encoder data to work backward and identify the properties of the object.

For instance, a heavier object will move slower than a light one if the robot applies the same amount of force.

Differentiable simulations

They utilize a technique called differentiable simulation, which allows the algorithm to predict how small changes in an object’s properties, like mass or softness, impact the robot’s ending joint position. The researchers built their simulations using NVIDIA’s Warp library, an open-source developer tool that supports differentiable simulations.

Once the differentiable simulation matches up with the robot’s real movements, the system has identified the correct property. The algorithm can do this in a matter of seconds and only needs to see one real-world trajectory of the robot in motion to perform the calculations.

“Technically, as long as you know the model of the object and how the robot can apply force to that object, you should be able to figure out the parameter you want to identify,” Liu says.

The researchers used their method to learn the mass and softness of an object, but their technique could also determine properties like moment of inertia or the viscosity of a fluid inside a container.

Plus, because their algorithm does not need an extensive dataset for training like some methods that rely on computer vision or external sensors, it would not be as susceptible to failure when faced with unseen environments or new objects.

In the future, the researchers want to try combining their method with computer vision to create a multimodal sensing technique that is even more powerful.

“This work is not trying to replace computer vision. Both methods have their pros and cons. But here we have shown that without a camera we can already figure out some of these properties,” Chen says.

They also want to explore applications with more complicated robotic systems, like soft robots, and more complex objects, including sloshing liquids or granular media like sand.

In the long run, they hope to apply this technique to improve robot learning, enabling future robots to quickly develop new manipulation skills and adapt to changes in their environments.

“Determining the physical properties of objects from data has long been a challenge in robotics, particularly when only limited or noisy measurements are available. This work is significant because it shows that robots can accurately infer properties like mass and softness using only their internal joint sensors, without relying on external cameras or specialized measurement tools,” says Miles Macklin, senior director of simulation technology at NVIDIA, who was not involved with this research.

This work is funded, in part, by Amazon and the GIST-CSAIL Research Program.

Dangers come but dangers also go, and when they do, the brain has an “all-clear” signal that teaches it to extinguish its fear. A new study in mice by MIT neuroscientists shows that the signal is the release of dopamine along a specific interregional brain circuit. The research therefore pinpoints a potentially critical mechanism of mental health, restoring calm when it works, but prolonging anxiety or even post-traumatic stress disorder when it doesn’t.

“Dopamine is essential to initiate fear extinction,” says Michele Pignatelli di Spinazzola, co-author of the new study from the lab of senior author Susumu Tonegawa, Picower Professor of biology and neuroscience at the RIKEN-MIT Laboratory for Neural Circuit Genetics within The Picower Institute for Learning and Memory at MIT, and a Howard Hughes Medical Institute (HHMI) investigator.

In 2020, Tonegawa’s lab showed that learning to be afraid, and then learning when that’s no longer necessary, result from a competition between populations of cells in the brain’s amygdala region. When a mouse learns that a place is “dangerous” (because it gets a little foot shock there), the fear memory is encoded by neurons in the anterior of the basolateral amygdala (aBLA) that express the gene Rspo2. When the mouse then learns that a place is no longer associated with danger (because they wait there and the zap doesn’t recur), neurons in the posterior basolateral amygdala (pBLA) that express the gene Ppp1r1b encode a new fear extinction memory that overcomes the original dread. Notably, those same neurons encode feelings of reward, helping to explain why it feels so good when we realize that an expected danger has dwindled.

In the new study, the lab, led by former members Xiangyu Zhang and Katelyn Flick, sought to determine what prompts these amygdala neurons to encode these memories. The rigorous set of experiments the team reports in the Proceedings of the National Academy of Sciences show that it’s dopamine sent to the different amygdala populations from distinct groups of neurons in the ventral tegmental area (VTA).

“Our study uncovers a precise mechanism by which dopamine helps the brain unlearn fear,” says Zhang, who also led the 2020 study and is now a senior associate at Orbimed, a health care investment firm. “We found that dopamine activates specific amygdala neurons tied to reward, which in turn drive fear extinction. We now see that unlearning fear isn’t just about suppressing it — it’s a positive learning process powered by the brain’s reward machinery. This opens up new avenues for understanding and potentially treating fear-related disorders, like PTSD.”

Forgetting fear

The VTA was the lab’s prime suspect to be the source of the signal because the region is well known for encoding surprising experiences and instructing the brain, with dopamine, to learn from them. The first set of experiments in the paper used multiple methods for tracing neural circuits to see whether and how cells in the VTA and the amygdala connect. They found a clear pattern: Rspo2 neurons were targeted by dopaminergic neurons in the anterior and left and right sides of the VTA. Ppp1r1b neurons received dopaminergic input from neurons in the center and posterior sections of the VTA. The density of connections was greater on the Ppp1r1b neurons than for the Rspo2 ones.

The circuit tracing showed that dopamine is available to amygdala neurons that encode fear and its extinction, but do those neurons care about dopamine? The team showed that indeed they express “D1” receptors for the neuromodulator. Commensurate with the degree of dopamine connectivity, Ppp1r1b cells had more receptors than Rspo2 neurons.

Dopamine does a lot of things, so the next question was whether its activity in the amygdala actually correlated with fear encoding and extinction. Using a method to track and visualize it in the brain, the team watched dopamine in the amygdala as mice underwent a three-day experiment. On Day One, they went to an enclosure where they experienced three mild shocks on the feet. On Day Two, they went back to the enclosure for 45 minutes, where they didn’t experience any new shocks — at first, the mice froze in anticipation of a shock, but then relaxed after about 15 minutes. On Day Three they returned again to test whether they had indeed extinguished the fear they showed at the beginning of Day Two.

The dopamine activity tracking revealed that during the shocks on Day One, Rspo2 neurons had the larger response to dopamine, but in the early moments of Day Two, when the anticipated shocks didn’t come and the mice eased up on freezing, the Ppp1r1b neurons showed the stronger dopamine activity. More strikingly, the mice that learned to extinguish their fear most strongly also showed the greatest dopamine signal at those neurons.

Causal connections

The final sets of experiments sought to show that dopamine is not just available and associated with fear encoding and extinction, but also actually causes them. In one set, they turned to optogenetics, a technology that enables scientists to activate or quiet neurons with different colors of light. Sure enough, when they quieted VTA dopaminergic inputs in the pBLA, doing so impaired fear extinction. When they activated those inputs, it accelerated fear extinction. The researchers were surprised that when they activated VTA dopaminergic inputs into the aBLA they could reinstate fear even without any new foot shocks, impairing fear extinction.

The other way they confirmed a causal role for dopamine in fear encoding and extinction was to manipulate the amygdala neurons’ dopamine receptors. In Ppp1r1b neurons, over-expressing dopamine receptors impaired fear recall and promoted extinction, whereas knocking the receptors down impaired fear extinction. Meanwhile in the Rspo2 cells, knocking down receptors reduced the freezing behavior.

“We showed that fear extinction requires VTA dopaminergic activity in the pBLA Ppp1r1b neurons by using optogenetic inhibition of VTA terminals and cell-type-specific knockdown of D1 receptors in these neurons,” the authors wrote.

The scientists are careful in the study to note that while they’ve identified the “teaching signal” for fear extinction learning, the broader phenomenon of fear extinction occurs brainwide, rather than in just this single circuit.

But the circuit seems to be a key node to consider as drug developers and psychiatrists work to combat anxiety and PTSD, Pignatelli di Spinazzola says.

“Fear learning and fear extinction provide a strong framework to study generalized anxiety and PTSD,” he says. “Our study investigates the underlying mechanisms suggesting multiple targets for a translational approach, such as pBLA and use of dopaminergic modulation.”

Marianna Rizzo is also a co-author of the study. Support for the research came from the RIKEN Center for Brain Science, the HHMI, the Freedom Together Foundation, and The Picower Institute.

Inside every human cell, 2 meters of DNA is crammed into a nucleus that is only one-hundredth of a millimeter in diameter.

To fit inside that tiny space, the genome must fold into a complex structure known as chromatin, made up of DNA and proteins. The structure of that chromatin, in turn, helps to determine which of the genes will be expressed in a given cell. Neurons, skin cells, and immune cells each express different genes depending on which of their genes are accessible to be transcribed.

Deciphering those structures experimentally is a time-consuming process, making it difficult to compare the 3D genome structures found in different cell types. MIT Professor Bin Zhang is taking a computational approach to this challenge, using computer simulations and generative artificial intelligence to determine these structures.

“Regulation of gene expression relies on the 3D genome structure, so the hope is that if we can fully understand those structures, then we could understand where this cellular diversity comes from,” says Zhang, an associate professor of chemistry.

From the farm to the lab

Zhang first became interested in chemistry when his brother, who was four years older, bought some lab equipment and started performing experiments at home.

“He would bring test tubes and some reagents home and do the experiment there. I didn’t really know what he was doing back then, but I was really fascinated with all the bright colors and the smoke and the odors that could come from the reactions. That really captivated my attention,” Zhang says.

His brother later became the first person from Zhang’s rural village to go to college. That was the first time Zhang had an inkling that it might be possible to pursue a future other than following in the footsteps of his parents, who were farmers in China’s Anhui province.

“Growing up, I would have never imagined doing science or working as a faculty member in America,” Zhang says. “When my brother went to college, that really opened up my perspective, and I realized I didn’t have to follow my parents’ path and become a farmer. That led me to think that I could go to college and study more chemistry.”

Zhang attended the University of Science and Technology in Hefei, China, where he majored in chemical physics. He enjoyed his studies and discovered computational chemistry and computational research, which became his new fascination.

“Computational chemistry combines chemistry with other subjects I love — math and physics — and brings a sense of rigor and reasoning to the otherwise more empirical rules,” he says. “I could use programming to solve interesting chemistry problems and test my own ideas very quickly.”

After graduating from college, he decided to continue his studies in the United States, which he recalled thinking was “the pinnacle of academics.” At Caltech, he worked with Thomas Miller, a professor of chemistry who used computational methods to understand molecular processes such as protein folding.

For Zhang’s PhD research, he studied a transmembrane protein that acts as a channel to allow other proteins to pass through the cell membrane. This protein, called translocon, can also open a side gate within the membrane, so that proteins that are meant to be embedded in the membrane can exit directly into the membrane.

“It’s really a remarkable protein, but it wasn’t clear how it worked,” Zhang says. “I built a computational model to understand the molecular mechanisms that dictate what are the molecular features that allow certain proteins to go into the membrane, while other proteins get secreted.”

Turning to the genome

After finishing grad school, Zhang’s research focus shifted from proteins to the genome. At Rice University, he did a postdoc with Peter Wolynes, a professor of chemistry who had made many key discoveries in the dynamics of protein folding. Around the time that Zhang joined the lab, Wolynes turned his attention to the structure of the genome, and Zhang decided to do the same.

Unlike proteins, which tend to have highly structured regions that can be studied using X-ray crystallography or cryo-EM, DNA is a very globular molecule that doesn’t lend itself to those types of analysis.

A few years earlier, in 2009, researchers at the Broad Institute, the University of Massachusetts Medical School, MIT, and Harvard University had developed a technique for studying the genome’s structure by cross-linking DNA in a cell’s nucleus. Researchers can then determine which segments are located near each other by shredding the DNA into many tiny pieces and sequencing it.

Zhang and Wolynes used data generated by this technique, known as Hi-C, to explore the question of whether DNA forms knots when it’s condensed in the nucleus, similar to how a strand of Christmas lights may become tangled when crammed into a box for storage.

“If DNA was just like a regular polymer, you would expect that it will become tangled and form knots. But that could be very detrimental for biology, because the genome is not just sitting there passively. It has to go through cell division, and also all this molecular machinery has to interact with the genome and transcribe it into RNA, and having knots will create a lot of unnecessary barriers,” Zhang says.

They found that, unlike Christmas lights, DNA does not form any knots even when packed into the cell nucleus, and they built a computational model allowing them to test hypotheses for how the genome is able to avoid those entanglements.

Since joining the MIT faculty in 2016, Zhang has continued developing models of how the genome behaves in 3D space, using molecular dynamic simulations. In one area of research, his lab is studying how differences between the genome structures of neurons and other brain cells give rise to their unique functions, and they are also exploring how misfolding of the genome may lead to diseases such as Alzheimer’s.

When it comes to connecting genome structure and function, Zhang believes that generative AI methods will also be essential. In a recent study, he and his students reported a new computational model, ChromoGen, that uses generative AI to predict the 3D structures of genomic regions, based on their DNA sequences.

“I think that in the future, we will have both components: generative AI and also theoretical chemistry-based approaches,” he says. “They nicely complement each other and allow us to both build accurate 3D structures and understand how those structures arise from the underlying physical forces.” 

As the world struggles to reduce climate-warming carbon emissions, India has pledged to do its part, and its success is critical: In 2023, India was the third-largest carbon emitter worldwide. The Indian government has committed to having net-zero carbon emissions by 2070.

To fulfill that promise, India will need to decarbonize its electric power system, and that will be a challenge: Fully 60 percent of India’s electricity comes from coal-burning power plants that are extremely inefficient. To make matters worse, the demand for electricity in India is projected to more than double in the coming decade due to population growth and increased use of air conditioning, electric cars, and so on.

Despite having set an ambitious target, the Indian government has not proposed a plan for getting there. Indeed, as in other countries, in India the government continues to permit new coal-fired power plants to be built, and aging plants to be renovated and their retirement postponed.

To help India define an effective — and realistic — plan for decarbonizing its power system, key questions must be addressed. For example, India is already rapidly developing carbon-free solar and wind power generators. What opportunities remain for further deployment of renewable generation? Are there ways to retrofit or repurpose India’s existing coal plants that can substantially and affordably reduce their greenhouse gas emissions? And do the responses to those questions differ by region?

With funding from IHI Corp. through the MIT Energy Initiative (MITEI), Yifu Ding, a postdoc at MITEI, and her colleagues set out to answer those questions by first using machine learning to determine the efficiency of each of India’s current 806 coal plants, and then investigating the impacts that different decarbonization approaches would have on the mix of power plants and the price of electricity in 2035 under increasingly stringent caps on emissions.

First step: Develop the needed dataset

An important challenge in developing a decarbonization plan for India has been the lack of a complete dataset describing the current power plants in India. While other studies have generated plans, they haven’t taken into account the wide variation in the coal-fired power plants in different regions of the country. “So, we first needed to create a dataset covering and characterizing all of the operating coal plants in India. Such a dataset was not available in the existing literature,” says Ding.

Making a cost-effective plan for expanding the capacity of a power system requires knowing the efficiencies of all the power plants operating in the system. For this study, the researchers used as their metric the “station heat rate,” a standard measurement of the overall fuel efficiency of a given power plant. The station heat rate of each plant is needed in order to calculate the fuel consumption and power output of that plant as plans for capacity expansion are being developed.

Some of the Indian coal plants’ efficiencies were recorded before 2022, so Ding and her team used machine-learning models to predict the efficiencies of all the Indian coal plants operating now. In 2024, they created and posted online the first comprehensive, open-sourced dataset for all 806 power plants in 30 regions of India. The work won the 2024 MIT Open Data Prize. This dataset includes each plant’s power capacity, efficiency, age, load factor (a measure indicating how much of the time it operates), water stress, and more.

In addition, they categorized each plant according to its boiler design. A “supercritical” plant operates at a relatively high temperature and pressure, which makes it thermodynamically efficient, so it produces a lot of electricity for each unit of heat in the fuel. A “subcritical” plant runs at a lower temperature and pressure, so it’s less thermodynamically efficient. Most of the Indian coal plants are still subcritical plants running at low efficiency.

Next step: Investigate decarbonization options

Equipped with their detailed dataset covering all the coal power plants in India, the researchers were ready to investigate options for responding to tightening limits on carbon emissions. For that analysis, they turned to GenX, a modeling platform that was developed at MITEI to help guide decision-makers as they make investments and other plans for the future of their power systems.

Ding built a GenX model based on India’s power system in 2020, including details about each power plant and transmission network across 30 regions of the country. She also entered the coal price, potential resources for wind and solar power installations, and other attributes of each region. Based on the parameters given, the GenX model would calculate the lowest-cost combination of equipment and operating conditions that can fulfill a defined future level of demand while also meeting specified policy constraints, including limits on carbon emissions. The model and all data sources were also released as open-source tools for all viewers to use.

Ding and her colleagues — Dharik Mallapragada, a former principal research scientist at MITEI who is now an assistant professor of chemical and biomolecular energy at NYU Tandon School of Engineering and a MITEI visiting scientist; and Robert J. Stoner, the founding director of the MIT Tata Center for Technology and Design and former deputy director of MITEI for science and technology — then used the model to explore options for meeting demands in 2035 under progressively tighter carbon emissions caps, taking into account region-to-region variations in the efficiencies of the coal plants, the price of coal, and other factors. They describe their methods and their findings in a paper published in the journal Energy for Sustainable Development.

In separate runs, they explored plans involving various combinations of current coal plants, possible new renewable plants, and more, to see their outcome in 2035. Specifically, they assumed the following four “grid-evolution scenarios:”

Baseline: The baseline scenario assumes limited onshore wind and solar photovoltaics development and excludes retrofitting options, representing a business-as-usual pathway.

High renewable capacity: This scenario calls for the development of onshore wind and solar power without any supply chain constraints.

Biomass co-firing: This scenario assumes the baseline limits on renewables, but here all coal plants — both subcritical and supercritical — can be retrofitted for “co-firing” with biomass, an approach in which clean-burning biomass replaces some of the coal fuel. Certain coal power plants in India already co-fire coal and biomass, so the technology is known.

Carbon capture and sequestration plus biomass co-firing: This scenario is based on the same assumptions as the biomass co-firing scenario with one addition: All of the high-efficiency supercritical plants are also retrofitted for carbon capture and sequestration (CCS), a technology that captures and removes carbon from a power plant’s exhaust stream and prepares it for permanent disposal. Thus far, CCS has not been used in India. This study specifies that 90 percent of all carbon in the power plant exhaust is captured.

Ding and her team investigated power system planning under each of those grid-evolution scenarios and four assumptions about carbon caps: no cap, which is the current situation; 1,000 million tons (Mt) of carbon dioxide (CO2) emissions, which reflects India’s announced targets for 2035; and two more-ambitious targets, namely 800 Mt and 500 Mt. For context, CO2 emissions from India’s power sector totaled about 1,100 Mt in 2021. (Note that transmission network expansion is allowed in all scenarios.)

Key findings

Assuming the adoption of carbon caps under the four scenarios generated a vast array of detailed numerical results. But taken together, the results show interesting trends in the cost-optimal mix of generating capacity and the cost of electricity under the different scenarios.

Even without any limits on carbon emissions, most new capacity additions will be wind and solar generators — the lowest-cost option for expanding India’s electricity-generation capacity. Indeed, this is observed to be the case now in India. However, the increasing demand for electricity will still require some new coal plants to be built. Model results show a 10 to 20 percent increase in coal plant capacity by 2035 relative to 2020.

Under the baseline scenario, renewables are expanded up to the maximum allowed under the assumptions, implying that more deployment would be economical. More coal capacity is built, and as the cap on emissions tightens, there is also investment in natural gas power plants, as well as batteries to help compensate for the now-large amount of intermittent solar and wind generation. When a 500 Mt cap on carbon is imposed, the cost of electricity generation is twice as high as it was with no cap.

The high renewable capacity scenario reduces the development of new coal capacity and produces the lowest electricity cost of the four scenarios. Under the most stringent cap — 500 Mt — onshore wind farms play an important role in bringing the cost down. “Otherwise, it’ll be very expensive to reach such stringent carbon constraints,” notes Ding. “Certain coal plants that remain run only a few hours per year, so are inefficient as well as financially unviable. But they still need to be there to support wind and solar.” She explains that other backup sources of electricity, such as batteries, are even more costly. 

The biomass co-firing scenario assumes the same capacity limit on renewables as in the baseline scenario, and the results are much the same, in part because the biomass replaces such a low fraction — just 20 percent — of the coal in the fuel feedstock. “This scenario would be most similar to the current situation in India,” says Ding. “It won’t bring down the cost of electricity, so we’re basically saying that adding this technology doesn’t contribute effectively to decarbonization.”

But CCS plus biomass co-firing is a different story. It also assumes the limits on renewables development, yet it is the second-best option in terms of reducing costs. Under the 500 Mt cap on CO2 emissions, retrofitting for both CCS and biomass co-firing produces a 22 percent reduction in the cost of electricity compared to the baseline scenario. In addition, as the carbon cap tightens, this option reduces the extent of deployment of natural gas plants and significantly improves overall coal plant utilization. That increased utilization “means that coal plants have switched from just meeting the peak demand to supplying part of the baseline load, which will lower the cost of coal generation,” explains Ding.

Some concerns

While those trends are enlightening, the analyses also uncovered some concerns for India to consider, in particular, with the two approaches that yielded the lowest electricity costs.

The high renewables scenario is, Ding notes, “very ideal.” It assumes that there will be little limiting the development of wind and solar capacity, so there won’t be any issues with supply chains, which is unrealistic. More importantly, the analyses showed that implementing the high renewables approach would create uneven investment in renewables across the 30 regions. Resources for onshore and offshore wind farms are mainly concentrated in a few regions in western and southern India. “So all the wind farms would be put in those regions, near where the rich cities are,” says Ding. “The poorer cities on the eastern side, where the coal power plants are, will have little renewable investment.”

So the approach that’s best in terms of cost is not best in terms of social welfare, because it tends to benefit the rich regions more than the poor ones. “It’s like [the government will] need to consider the trade-off between energy justice and cost,” says Ding. Enacting state-level renewable generation targets could encourage a more even distribution of renewable capacity installation. Also, as transmission expansion is planned, coordination among power system operators and renewable energy investors in different regions could help in achieving the best outcome.

CCS plus biomass co-firing — the second-best option for reducing prices — solves the equity problem posed by high renewables, and it assumes a more realistic level of renewable power adoption. However, CCS hasn’t been used in India, so there is no precedent in terms of costs. The researchers therefore based their cost estimates on the cost of CCS in China and then increased the required investment by 10 percent, the “first-of-a-kind” index developed by the U.S. Energy Information Administration. Based on those costs and other assumptions, the researchers conclude that coal plants with CCS could come into use by 2035 when the carbon cap for power generation is less than 1,000 Mt.

But will CCS actually be implemented in India? While there’s been discussion about using CCS in heavy industry, the Indian government has not announced any plans for implementing the technology in coal-fired power plants. Indeed, India is currently “very conservative about CCS,” says Ding. “Some researchers say CCS won’t happen because it’s so expensive, and as long as there’s no direct use for the captured carbon, the only thing you can do is put it in the ground.” She adds, "It’s really controversial to talk about whether CCS will be implemented in India in the next 10 years.”

Ding and her colleagues hope that other researchers and policymakers — especially those working in developing countries — may benefit from gaining access to their datasets and learning about their methods. Based on their findings for India, she stresses the importance of understanding the detailed geographical situation in a country in order to design plans and policies that are both realistic and equitable.

MIT Provost Cynthia Barnhart has announced that Vice Provost for the Arts Philip S. Khoury will step down from the position on Aug. 31. Khoury, the Ford International Professor of History, served in the role for 19 years. After a sabbatical, he will rejoin the faculty in the School of Humanities, Arts, and Social Sciences (SHASS).

“Since arriving at MIT in 1981, Philip has championed what he calls the Institute’s ‘artistic ecosystem,’ which sits at the intersection of technology, science, the humanities, and the arts. Thanks to Philip’s vision, this ecosystem is now a foundational element of MIT’s educational and research missions and a critical component of how we advance knowledge, understanding, and discovery in service to the world,” says Barnhart.

Khoury was appointed associate provost in 2006 by then-MIT president Susan Hockfield, with a double portfolio enhancing the Institute’s nonacademic arts programs and beginning a review of MIT’s international activities. Those programs include the List Visual Arts Center, the MIT Museum, the Center for Art, Science and Technology (CAST), and the Council for the Arts at MIT (CAMIT). After five years, the latter half of this portfolio evolved into the Office of the Vice Provost for International Activities. 

Khoury devoted most of his tenure to expanding the Institute’s arts infrastructure, promoting the visibility of its stellar arts faculty, and guiding the growth of student participation in the arts. Today, more than 50 percent of MIT undergraduates take arts classes, with more than 1,500 studying music.

“Philip has been a remarkable leader at MIT over decades. He has ensured that the arts are a prominent part of the MIT ‘mens-et-manus’ [‘mind-and-hand’] experience and that our community has the opportunity to admire, learn from, and participate in creative thinking in all realms,” says L. Rafael Reif, the Ray and Maria Stata Professor of Electrical Engineering and Computer Science and MIT president emeritus. “A historian — and a humanist at heart — Philip also played a crucial role in helping MIT develop a thoughtful international strategy in research and education."

“I will miss my colleagues first and foremost as I leave this position behind,” says Khoury. “But I have been proud to see the quality of the faculty grow and the student interest in the arts grow almost exponentially, along with an awareness of how the arts are prospering at MIT.”

Stream of creativity

During his time as vice provost, he partnered with then-School of Architecture and Planning (SAP) dean Adèle Santos and SHASS dean Deborah Fitzgerald to establish the CAST in 2012. The center encourages artistic collaborations and provides seed funds and research grants to students and faculty. 

Khoury also helped oversee a significant expansion of the Institute’s art facilities, including the unique multipurpose design of the Theater Arts Building, the new MIT Museum, and the Edward and Joyce Linde Music Building. Along with the List Visual Arts Center, which will celebrate its 40th anniversary this year, these vibrant spaces “offer an opportunity for our students to do something different from what they came to MIT to do in science and engineering,” Khoury suggests. “It gives them an outlet to do other kinds of experimentation.”

“What makes the arts so successful here is that they are very much in the stream of creativity, which science and technology are all about,” he adds.

One of Khoury’s other long-standing goals has been to elevate the recognition of the arts faculty, “to show that the quality of what we do in those areas matches the quality of what we do in engineering and science,” he says.

“I will always remember Philip Khoury’s leadership and advocacy as dean of the School of Humanities and Social Sciences for changing the definition of the ‘A’ in SHASS from ‘and’ to ‘Arts.’ That small change had large implications for professional careers for artists, enrollments, and subject options that remain a source of renewal and strength to this day,” says Institute Professor Marcus Thompson.

Most recently, Khoury and his team, in collaboration with faculty, students, and staff from across the Institute, oversaw the development and production of MIT’s new festival of the arts, known as Artfinity. Launched in February and open to the public, the Institute-sponsored, campus-wide festival featured a series of 80 performing and visual arts events.

International activities

Khoury joined the faculty as an assistant professor in 1981 and later served as dean of SHASS between 1991 and 2006. In 2002, he was appointed the inaugural Kenan Sahin Dean of SHASS.

His academic focus made him a natural choice for the first coordinator of MIT international activities, a role he served in from 2006 to 2011. During that time, he traveled widely to learn more about the ways MIT faculty were engaged abroad, and he led the production of an influential report on the state of MIT’s international activities.

“We wanted to create a strategy, but not a foreign policy,” Khoury said of the report.

Khoury’s time in the international role led him to consider ways that collaborations with other countries should be balanced so as not to diminish MIT’s offerings at home, he says. He also looked for ways to encourage more collaborations with countries in sub-Saharan Africa, South America, and parts of the Middle East.

Future plans

Khoury was instrumental in establishing the Future of the Arts at MIT Committee, which was charged by Provost Barnhart in June 2024 in collaboration with Dean Hashim Sarkis of the School of Architecture and Planning and Dean Agustín Rayo of SHASS. The committee aims to find new ways to envision the place of arts at the Institute — a task that was last undertaken in 1987, he says. The committee submitted a draft report to Provost Barnhart in April. 

“I think it will hit the real sweet spot of where arts meet science and technology, but not where art is controlled by science and technology,” Khoury says. “I think the promotion of that, and the emphasis on that, among other connections with art, are really what we should be pushing for and developing.”

After he steps down as vice provost, Khoury plans to devote more time to writing two books: a personal memoir and a book about the Middle East. And he is looking forward to seeing how the arts at MIT will flourish in the near future. “I feel elated about where we’ve landed and where we’ll continue to go,” he says.

As Barnhart noted in her letter to the community, the Future of the Arts at MIT Committee's efforts combined with Khoury staying on through the end of the summer, provides President Kornbluth, the incoming provost, and Khoury with the opportunity to reflect on the Institute’s path forward in this critical space.

What would a behind-the-scenes look at a video generated by an artificial intelligence model be like? You might think the process is similar to stop-motion animation, where many images are created and stitched together, but that’s not quite the case for “diffusion models” like OpenAl's SORA and Google's VEO 2.

Instead of producing a video frame-by-frame (or “autoregressively”), these systems process the entire sequence at once. The resulting clip is often photorealistic, but the process is slow and doesn’t allow for on-the-fly changes. 

Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Adobe Research have now developed a hybrid approach, called “CausVid,” to create videos in seconds. Much like a quick-witted student learning from a well-versed teacher, a full-sequence diffusion model trains an autoregressive system to swiftly predict the next frame while ensuring high quality and consistency. CausVid’s student model can then generate clips from a simple text prompt, turning a photo into a moving scene, extending a video, or altering its creations with new inputs mid-generation.

This dynamic tool enables fast, interactive content creation, cutting a 50-step process into just a few actions. It can craft many imaginative and artistic scenes, such as a paper airplane morphing into a swan, woolly mammoths venturing through snow, or a child jumping in a puddle. Users can also make an initial prompt, like “generate a man crossing the street,” and then make follow-up inputs to add new elements to the scene, like “he writes in his notebook when he gets to the opposite sidewalk.”

The CSAIL researchers say that the model could be used for different video editing tasks, like helping viewers understand a livestream in a different language by generating a video that syncs with an audio translation. It could also help render new content in a video game or quickly produce training simulations to teach robots new tasks.

Tianwei Yin SM ’25, PhD ’25, a recently graduated student in electrical engineering and computer science and CSAIL affiliate, attributes the model’s strength to its mixed approach.

“CausVid combines a pre-trained diffusion-based model with autoregressive architecture that’s typically found in text generation models,” says Yin, co-lead author of a new paper about the tool. “This AI-powered teacher model can envision future steps to train a frame-by-frame system to avoid making rendering errors.”

Yin’s co-lead author, Qiang Zhang, is a research scientist at xAI and a former CSAIL visiting researcher. They worked on the project with Adobe Research scientists Richard Zhang, Eli Shechtman, and Xun Huang, and two CSAIL principal investigators: MIT professors Bill Freeman and Frédo Durand.

Caus(Vid) and effect

Many autoregressive models can create a video that’s initially smooth, but the quality tends to drop off later in the sequence. A clip of a person running might seem lifelike at first, but their legs begin to flail in unnatural directions, indicating frame-to-frame inconsistencies (also called “error accumulation”).

Error-prone video generation was common in prior causal approaches, which learned to predict frames one by one on their own. CausVid instead uses a high-powered diffusion model to teach a simpler system its general video expertise, enabling it to create smooth visuals, but much faster.

CausVid displayed its video-making aptitude when researchers tested its ability to make high-resolution, 10-second-long videos. It outperformed baselines like “OpenSORA” and “MovieGen,” working up to 100 times faster than its competition while producing the most stable, high-quality clips.

Then, Yin and his colleagues tested CausVid’s ability to put out stable 30-second videos, where it also topped comparable models on quality and consistency. These results indicate that CausVid may eventually produce stable, hours-long videos, or even an indefinite duration.

A subsequent study revealed that users preferred the videos generated by CausVid’s student model over its diffusion-based teacher.

“The speed of the autoregressive model really makes a difference,” says Yin. “Its videos look just as good as the teacher’s ones, but with less time to produce, the trade-off is that its visuals are less diverse.”

CausVid also excelled when tested on over 900 prompts using a text-to-video dataset, receiving the top overall score of 84.27. It boasted the best metrics in categories like imaging quality and realistic human actions, eclipsing state-of-the-art video generation models like “Vchitect” and “Gen-3.”

While an efficient step forward in AI video generation, CausVid may soon be able to design visuals even faster — perhaps instantly — with a smaller causal architecture. Yin says that if the model is trained on domain-specific datasets, it will likely create higher-quality clips for robotics and gaming.

Experts say that this hybrid system is a promising upgrade from diffusion models, which are currently bogged down by processing speeds. “[Diffusion models] are way slower than LLMs [large language models] or generative image models,” says Carnegie Mellon University Assistant Professor Jun-Yan Zhu, who was not involved in the paper. “This new work changes that, making video generation much more efficient. That means better streaming speed, more interactive applications, and lower carbon footprints.”

The team’s work was supported, in part, by the Amazon Science Hub, the Gwangju Institute of Science and Technology, Adobe, Google, the U.S. Air Force Research Laboratory, and the U.S. Air Force Artificial Intelligence Accelerator. CausVid will be presented at the Conference on Computer Vision and Pattern Recognition in June.

For the Abdul Latif Jameel Water and Food Systems Lab (J-WAFS), 2025 marks a decade of translating groundbreaking research into tangible solutions for global challenges. Few examples illustrate that mission better than NONA Technologies. With support from a J-WAFS Solutions grant, MIT electrical engineering and biological engineering Professor Jongyoon Han and his team developed a portable desalination device that transforms seawater into clean drinking water without filters or high-pressure pumps. 

The device stands apart from traditional systems because conventional desalination technologies, like reverse osmosis, are energy-intensive, prone to fouling, and typically deployed at large, centralized plants. In contrast, the device developed in Han’s lab employs ion concentration polarization technology to remove salts and particles from seawater, producing potable water that exceeds World Health Organization standards. It is compact, solar-powered, and operable at the push of a button — making it an ideal solution for off-grid and disaster-stricken areas.

This research laid the foundation for spinning out NONA Technologies along with co-founders Junghyo Yoon PhD ’21 from Han’s lab and Bruce Crawford MBA ’22, to commercialize the technology and address pressing water-scarcity issues worldwide. “This is really the culmination of a 10-year journey that I and my group have been on,” said Han in an earlier MIT News article. “We worked for years on the physics behind individual desalination processes, but pushing all those advances into a box, building a system, and demonstrating it in the ocean ... that was a really meaningful and rewarding experience for me.” You can watch this video showcasing the device in action.

Moving breakthrough research out of the lab and into the world is a well-known challenge. While traditional “seed” grants typically support early-stage research at Technology Readiness Level (TRL) 1-2, few funding sources exist to help academic teams navigate to the next phase of technology development. The J-WAFS Solutions Program is strategically designed to address this critical gap by supporting technologies in the high-risk, early-commercialization phase that is often neglected by traditional research, corporate, and venture funding. By supporting technologies at TRLs 3-5, the program increases the likelihood that promising innovations will survive beyond the university setting, advancing sufficiently to attract follow-on funding.

Equally important, the program gives academic researchers the time, resources, and flexibility to de-risk their technology, explore customer need and potential real-world applications, and determine whether and how they want to pursue commercialization. For faculty-led teams like Han’s, the J-WAFS Solutions Program provided the critical financial runway and entrepreneurial guidance needed to refine the technology, test assumptions about market fit, and lay the foundation for a startup team. While still in the MIT innovation ecosystem, Nona secured over $200,000 in non-dilutive funding through competitions and accelerators, including the prestigious MIT delta v Educational Accelerator. These early wins laid the groundwork for further investment and technical advancement.

Since spinning out of MIT, NONA has made major strides in both technology development and business viability. What started as a device capable of producing just over half-a-liter of clean drinking water per hour has evolved into a system that now delivers 10 times that capacity, at 5 liters per hour. The company successfully raised a $3.5 million seed round to advance its portable desalination device, and entered into a collaboration with the U.S. Army Natick Soldier Systems Center, where it co-developed early prototypes and began generating revenue while validating the technology. Most recently, NONA was awarded two SBIR Phase I grants totaling $575,000, one from the National Science Foundation and another from the National Institute of Environmental Health Sciences.

Now operating out of Greentown Labs in Somerville, Massachusetts, NONA has grown to a dedicated team of five and is preparing to launch its nona5 product later this year, with a wait list of over 1,000 customers. It is also kicking off its first industrial pilot, marking a key step toward commercial scale-up. “Starting a business as a postdoc was challenging, especially with limited funding and industry knowledge,” says Yoon, who currently serves as CTO of NONA. “J-WAFS gave me the financial freedom to pursue my venture, and the mentorship pushed me to hit key milestones. Thanks to J-WAFS, I successfully transitioned from an academic researcher to an entrepreneur in the water industry.”

NONA is one of several J-WAFS-funded technologies that have moved from the lab to market, part of a growing portfolio of water and food solutions advancing through MIT’s innovation pipeline. As J-WAFS marks a decade of catalyzing innovation in water and food, NONA exemplifies what is possible when mission-driven research is paired with targeted early-stage support and mentorship.

To learn more or get involved in supporting startups through the J-WAFS Solutions Program, please contact jwafs@mit.edu.

As the saying goes, time is money. That’s certainly evident in the transportation sector, where people will pay more for direct flights, express trains, and other ways to get somewhere quickly.

Still, it is difficult to measure precisely how much people value their time. Now, a paper co-authored by an MIT economist uses ride-sharing data to reveal multiple implications of personalized pricing.

By focusing on a European ride-sharing platform that auctions its rides, the researchers found that people are more responsive to prices than to wait times. They also found that people pay more to save time during the workday, and that when people pay more to avoid waiting, it notably increases business revenues. And some segments of consumers are distinctly more willing than others to pay higher prices.

Specifically, when people can bid for rides that arrive sooner, the amount above the minimum price the platform can charge increases by 5.2 percent. Meanwhile, the gap between offered prices and the maximum that consumers are willing to pay decreases by 2.5 percent. In economics terms, this creates additional “surplus” value for firms, while lowering the “consumer surplus” in these transactions.

“One of the important quantities in transportation is the value of time,” says MIT economist Tobias Salz, co-author of a new paper detailing the study’s findings. “We came across a setting that offered a very clean way of examining this quantity, where the value of time is revealed by people’s transportation choices.”

The paper, “Personalized Pricing and the Value of Time: Evidence from Auctioned Cab Rides,” is being published in Econometrica. The authors are Nicholas Buchholz, an assistant professor of economics at Princeton University; Laura Doval, a professor at Columbia Business School; Jakub Kastl, a professor of economics at Princeton University; Filip Matejka, a professor at Charles University in Prague; and Salz, the Castle Krob Career Development Associate Professor of Economics in MIT’s Department of Economics.

It is not easy to study how much money people will spend to save time — and time alone. Transportation is one sector where it is possible to do so, though not the only one. People will also pay more for, say, an express pass to avoid long lines at an amusement park. But data for those scenarios, even when available, may contain complicating factors. (Also, the value of time shouldn’t be confused with how much people pay for services charged by the hour, from accountants to tuba lessons.)

In this case, however, the researchers were provided data from Liftago, a ride-sharing platform in Prague with a distinctive feature: It lets drivers bid on a passenger’s business, with the wait time until the auto arrives as one of the factors involved. Drivers can also indicate when they will be available. In studying how passengers compare offers with different wait times and prices, the researchers see exactly how much people are paying not to wait, other things being equal. All told, they examined 1.9 million ride requests and 5.2 million bids.

“It’s like an eBay for taxis,” Salz says. “Instead of assigning the driver to you, drivers bid for the passengers’ business. With this, we can very directly observe how people make their choices. How they value time is revealed by the wait and the prices attached to that. In many settings we don’t observe that directly, so it’s a very clean comparison that rids the data of a lot of confounds.”

The data set allows the researchers to examine many aspects of personalized pricing and the way it affects the transportation market in this setting. That produces a set of insights on its own, along with the findings on time valuation. 

Ultimately, the researchers found that the elasticity of prices — how much they change — ranged from four to 10 times as much as the elasticity of wait times, meaning people are more keen on avoiding high prices.

The team found the overall value of time in this context is $13.21 per hour for users of the ride-share platform, though the researchers note that is not a universal measure of the value of time and is dependent on this setting. The study also shows that bids increase during work hours.

Additionally, the research reveals a split among consumers: The top quartile of bids placed a value on time that is 3.5 times higher than the value of the bids in the bottom quartile.

Then there is still the question of how much personalized pricing benefits consumers, providers, or both. The numbers, again, show that the overall surplus increases — meaning business benefits — while the consumer surplus is reduced. However, the data show an even more nuanced picture. Because the top quartile of bidders are paying substantially more to avoid longer waits, they are the ones who absorb the brunt of the costs in this kind of system.

“The majority of consumers still benefit,” Salz says. “The consumers hurt by this have a very high willingness to pay. The source of welfare gains is that most consumers can be brought into the market. But the flip side is that the firm, by knowing every consumer’s choke point, can extract the surplus. Welfare goes up, the ride-sharing platform captures most of that, and drivers — interestingly — also benefit from the system, although they do not have access to the data.”

Economic theory and other transportation studies alone would not necessarily have predicted the study’s results and various nuances.

“It was not clear a priori whether consumers benefit,” Salz observes. “That is not something you would know without going to the data.”

While this study might hold particular interest for firms and others interested in transportation, mobility, and ride-sharing, it also fits into a larger body of economics research about information in markets and how its presence, or absence, influences consumer behavior, consumer welfare, and the functioning of markets.

“The [research] umbrella here is really information about where to find trading partners and what their willingness to pay is,” Salz says. “What I’m broadly interested in is these types of information frictions and how they determine market outcomes, how they might impact consumers, and be used by firms.”

The research was supported, in part, by the National Bureau of Economic Research, the U.S. Department of Transportation, and the National Science Foundation. 

Farmers today face a number of challenges, from supply chain stability to nutrient and waste management. But hanging over everything is the need to maintain profitability amid changing markets and increased uncertainty.

Fyto, founded by former MIT staff member Jason Prapas, is offering a highly automated cultivation system to address several of farmers’ biggest problems at once.

At the heart of Fyto’s system is Lemna, a genus of small aquatic plants otherwise known as duckweed. Most people have probably seen thick green mats of Lemna lying on top of ponds and swamps. But Lemna is also rich in protein and capable of doubling in biomass every two days. Fyto has built an automated cropping system that uses nitrogen-rich wastewater from dairy farms to grow Lemna in shallow pools on otherwise less productive farmland. On top of the pools, the company has built what it believes are the largest agricultural robots in the world, which monitor plant health and harvest the Lemna sustainably. The Lemna can then be used on farms as a high-protein cattle feed or fertilizer supplement.

Fyto’s systems are designed to rely on minimal land, water, and labor while creating a more sustainable, profitable food system.

“We developed from scratch a robotic system that takes the guesswork out of farming this crop,” says Prapas, who previously led the translational research program of MIT’s Tata Center. “It looks at the crop on a daily basis, takes inventory to know how many plants there are, how much should be harvested to have healthy growth the next day, can detect if the color is slightly off or there are nutrient deficiencies, and can suggest different interventions based on all that data.”

From kiddie pools to cow farms

Prapas’ first job out of college was with an MIT spinout called Green Fuel that harvested algae to make biofuel. He went back to school for a master’s and then a PhD in mechanical engineering, but he continued working with startups. Following his PhD at Colorado State University, he co-founded Factor[e] Ventures to fund and incubate startups focused on improving energy access in emerging markets.

Through that work, Prapas was introduced to MIT’s Tata Center for Technology and Design.

“We were really interested in the new technologies being developed at the MIT Tata Center, and in funding new startups taking on some of these global climate challenges in emerging markets,” Prapas recalls. “The Tata Center was interested in making sure these technologies get put into practice rather than patented and put on a shelf somewhere. It was a good synergy.”

One of the people Prapas got to know was Rob Stoner, the founding director of the Tata Center, who encouraged Prapas to get more directly involved with commercializing new technologies. In 2017, Prapas joined the Tata Center as the translational research director. During that time, Prapas worked with MIT students, faculty, and staff to test their inventions in the real world. Much of that work involved innovations in agriculture.

“Farming is a fact of life for a lot of folks around the world — both subsistence farming but also producing food for the community and beyond,” Prapas says. “That has huge implications for water usage, electricity consumption, labor. For years, I’d been thinking about how we make farming a more attractive endeavor for people: How do we make it less back-breaking, more efficient, and more economical?”

Between his work at MIT and Factor[e], Prapas visited hundreds of farms around the world, where he started to think about the lack of good choices for farming inputs like animal feed and fertilizers. The problem represented a business opportunity.

Fyto began with kiddie pools. Prapas started growing aquatic plants in his backyard, using them as a fertilizer source for vegetables. The experience taught him how difficult it would be to train people to grow and harvest Lemna at large scales on farms.

“I realized we’d have to invent both the farming method — the agronomy — and the equipment and processes to grow it at scale cost effectively,” Prapas explains.

Prapas started discussing his ideas with others around 2019.

“The MIT and Boston ecosystems are great for pitching somewhat crazy ideas to willing audiences and seeing what sticks,” Prapas says. “There’s an intangible benefit of being at MIT, where you just can’t help but think of bold ideas and try putting them into practice.”

Prapas, who left MIT to lead Fyto in 2019, partnered with Valerie Peng ’17, SM ’19, then a graduate student at MIT who became his first hire.

“Farmers work so hard, and I have so much respect for what they do,” says Peng, who serves as Fyto’s head of engineering. “People talk about the political divide, but there’s a lot of alignment around using less, doing more with what you have, and making our food systems more resilient to drought, supply chain disruptions, and everything else. There’s more in common with everyone than you’d expect.”

A new farming method

Lemna can produce much more protein per acre than soy, another common source of protein on farms, but it requires a lot of nitrogen to grow. Fortunately, many types of farmers, especially large dairy farmers, have abundant nitrogen sources in the waste streams that come from washing out cow manure.

“These waste streams are a big problem: In California it’s believed to be one of the largest source of greenhouse gas emissions in the agriculture sector despite the fact that hundreds of crops are grown in California,” Prapas says.

For the last few years, Fyto has run its systems in pilots on farms, trialing the crop as feed and fertilizer before delivering to its customers. The systems Fyto has deployed so far are about 50 feet wide, but it is actively commissioning its newest version that’s 160 feet wide. Eventually, Fyto plans to sell the systems directly to farmers.

Fyto is currently awaiting California’s approval for use in feed, but Lemna has already been approved in Europe. Fyto has also been granted a fertilizer license on its plant-based fertilizer, with promising early results in trials, and plans to sell new fertilizer products this year.

Although Fyto is focused on dairy farms for its early deployments, it has also grown Lemna using manure from chicken, and Prapas notes that even people like cheese producers have a nitrogen waste problem that Fyto could solve.

“Think of us like a polishing step you could put on the end of any system that has an organic waste stream,” Prapas says. “In that situation, we’re interested in growing our crops on it. We’ve had very few things that the plant can’t grow on. Globally, we see this as a new farming method, and that means it’s got a lot of potential applications.”

What if data could help predict a patient’s prognosis, streamline hospital operations, or optimize human resources in medicine? A book fresh off the shelves, “The Analytics Edge in Healthcare,” shows that this is already happening, and demonstrates how to scale it. 

Authored by Dimitris Bertsimas, MIT’s vice provost for open learning, along with two of Bertsimas’ former students — Agni Orfanoudaki PhD ’21, associate professor of operations management at University of Oxford’s Saïd Business School, and Holly Wiberg PhD ’22, assistant professor of public policy and operations research at Carnegie Mellon University — the book provides a practical introduction to the field of health care analytics. With an emphasis on real-world applications, the first part of the book establishes technical foundations — spanning machine learning and optimization — while the second part of the book presents integrated case studies that cover various clinical specialties and problem types using descriptive, predictive, and prescriptive analytics. 

Part of a broader series, “The Analytics Edge in Healthcare” demonstrates how to leverage data and models to make better decisions within the health care sector, while its predecessor, “The Analytics Edge,” dives into the science of using data to build models, improve decisions, and add value to institutions and individuals. 

Bertsimas, who is also the associate dean of business analytics and the Boeing Leaders for Global Operations Professor of Management at the MIT Sloan School of Management, is the innovator behind 15.071 (The Analytics Edge), a course on MIT Open Learning’s MITx that has attracted hundreds of thousands of online learners and served as the inspiration behind the book series. Bertsimas took a break from research and his work at MIT Open Learning to discuss how the field of analytics is transforming the health care system and share some surprising ways analytics are already being used in hospitals. 

Q: How is the field of analytics changing the way hospitals provide care and manage their operations?

A: As an academic, I’ve always aspired to educate, write publications, and utilize what we do in practice. Therefore, I founded Holistic Hospital Optimization (H20) with the goal of optimizing hospital operations with machine learning to improve patient care. We have developed a variety of tools at MIT and implemented them at hospitals around the world. For example, we manage patients’ length of stay and their deterioration indexes (a computerized tool that predicts a patient’s risk of clinical deterioration); we manage nurse optimization and how hospitals can allocate human resources appropriately; and we optimize blocks for surgeries. This is the beginning of a change where analytics and AI methods are now being utilized quite widely. My hope would be that this work and this book will accelerate the effect of using these tools. 

Additionally, I have taught a nine-lecture course twice with Agni and Holly at the Hartford Hospital System, where I realized that these analytics methods — which are typically not taught in medical schools — can be demonstrated for health care practitioners, including physicians, nurses, and administrators. To have an impact, you need to have appropriate methods, implement them, and apply them, but you also need to educate people on how to use them. This links well with my role at Open Learning, where our objective is to educate learners globally. In fact, Open Learning is launching this fall Universal AI, a dynamic online learning experience that provides comprehensive knowledge on artificial intelligence, preparing a global audience of learners for employment in our rapidly evolving job market. 

Q: What are some surprising ways analytics are being used in health care that most people wouldn’t expect?

A: Using analytics, we have reduced patients’ length of stay at Hartford Hospital from 5.67 days to five days. We have an algorithm that predicts patients’ probability of being released; therefore, doctors prioritize the patients with the highest probability, preparing them for discharge. This means that the hospital can treat far more patients, and the patients stay in the hospital less time.

Furthermore, when hospitals saw an increase in nurse turnover during the Covid-19 pandemic, we developed an analytics system that takes into account equity and fairness and decreases overtime costs, giving preferred slots to nurses and decreasing overall turnover substantially. These are just two examples; there are many others where an analytical perspective to health care and medicine has made a material difference. 

Q: Looking ahead, how do you see artificial intelligence shaping the future of health care?

A: In a very significant way — we use machine learning to make better predictions, but generative AI can explain them. I already see a movement in that direction. It’s really the evolution of AI that made this possible, and it is exciting. It’s also important for the world, because of its capabilities to improve care and save lives. 

For example, through our program at the Hartford Hospital System, we discovered that a patient was getting worse and predicted through analytics that they would get even worse. After our prediction, the doctors examined the patient more closely and discovered the patient had an early case of sepsis, a life-threatening condition in which the body responds improperly to an infection. If we hadn’t detected sepsis earlier, the patient might have died. This made an actual difference in saving a person’s life. 

Q: If you had to describe “The Analytics Edge in Healthcare” in one or two words, what would they be, and why? 

A: The book is a phased transition in health care because it is capable of affecting the health care sector in a way that has not been done before. The book really outlines my work in health care and its applications in the last decade.

If there’s one thing that characterizes driving in any major city, it’s the constant stop-and-go as traffic lights change and as cars and trucks merge and separate and turn and park. This constant stopping and starting is extremely inefficient, driving up the amount of pollution, including greenhouse gases, that gets emitted per mile of driving. 

One approach to counter this is known as eco-driving, which can be installed as a control system in autonomous vehicles to improve their efficiency.

How much of a difference could that make? Would the impact of such systems in reducing emissions be worth the investment in the technology? Addressing such questions is one of a broad category of optimization problems that have been difficult for researchers to address, and it has been difficult to test the solutions they come up with. These are problems that involve many different agents, such as the many different kinds of vehicles in a city, and different factors that influence their emissions, including speed, weather, road conditions, and traffic light timing.

“We got interested a few years ago in the question: Is there something that automated vehicles could do here in terms of mitigating emissions?” says Cathy Wu, the Thomas D. and Virginia W. Cabot Career Development Associate Professor in the Department of Civil and Environmental Engineering and the Institute for Data, Systems, and Society (IDSS) at MIT, and a principal investigator in the Laboratory for Information and Decision Systems. “Is it a drop in the bucket, or is it something to think about?,” she wondered.

To address such a question involving so many components, the first requirement is to gather all available data about the system, from many sources. One is the layout of the network’s topology, Wu says, in this case a map of all the intersections in each city. Then there are U.S. Geological Survey data showing the elevations, to determine the grade of the roads. There are also data on temperature and humidity, data on the mix of vehicle types and ages, and on the mix of fuel types.

Eco-driving involves making small adjustments to minimize unnecessary fuel consumption. For example, as cars approach a traffic light that has turned red, “there’s no point in me driving as fast as possible to the red light,” she says. By just coasting, “I am not burning gas or electricity in the meantime.” If one car, such as an automated vehicle, slows down at the approach to an intersection, then the conventional, non-automated cars behind it will also be forced to slow down, so the impact of such efficient driving can extend far beyond just the car that is doing it.

That’s the basic idea behind eco-driving, Wu says. But to figure out the impact of such measures, “these are challenging optimization problems” involving many different factors and parameters, “so there is a wave of interest right now in how to solve hard control problems using AI.” 

The new benchmark system that Wu and her collaborators developed based on urban eco-driving, which they call “IntersectionZoo,” is intended to help address part of that need. The benchmark was described in detail in a paper presented at the 2025 International Conference on Learning Representation in Singapore.

Looking at approaches that have been used to address such complex problems, Wu says an important category of methods is multi-agent deep reinforcement learning (DRL), but a lack of adequate standard benchmarks to evaluate the results of such methods has hampered progress in the field.

The new benchmark is intended to address an important issue that Wu and her team identified two years ago, which is that with most existing deep reinforcement learning algorithms, when trained for one specific situation (e.g., one particular intersection), the result does not remain relevant when even small modifications are made, such as adding a bike lane or changing the timing of a traffic light, even when they are allowed to train for the modified scenario.

In fact, Wu points out, this problem of non-generalizability “is not unique to traffic,” she says. “It goes back down all the way to canonical tasks that the community uses to evaluate progress in algorithm design.” But because most such canonical tasks do not involve making modifications, “it’s hard to know if your algorithm is making progress on this kind of robustness issue, if we don’t evaluate for that.”

While there are many benchmarks that are currently used to evaluate algorithmic progress in DRL, she says, “this eco-driving problem features a rich set of characteristics that are important in solving real-world problems, especially from the generalizability point of view, and that no other benchmark satisfies.” This is why the 1 million data-driven traffic scenarios in IntersectionZoo uniquely position it to advance the progress in DRL generalizability.  As a result, “this benchmark adds to the richness of ways to evaluate deep RL algorithms and progress.”

And as for the initial question about city traffic, one focus of ongoing work will be applying this newly developed benchmarking tool to address the particular case of how much impact on emissions would come from implementing eco-driving in automated vehicles in a city, depending on what percentage of such vehicles are actually deployed.

But Wu adds that “rather than making something that can deploy eco-driving at a city scale, the main goal of this study is to support the development of general-purpose deep reinforcement learning algorithms, that can be applied to this application, but also to all these other applications — autonomous driving, video games, security problems, robotics problems, warehousing, classical control problems.”

Wu adds that “the project’s goal is to provide this as a tool for researchers, that’s openly available.” IntersectionZoo, and the documentation on how to use it, are freely available at GitHub.

Wu is joined on the paper by lead authors Vindula Jayawardana, a graduate student in MIT’s Department of Electrical Engineering and Computer Science (EECS); Baptiste Freydt, a graduate student from ETH Zurich; and co-authors Ao Qu, a graduate student in transportation; Cameron Hickert, an IDSS graduate student; and Zhongxia Yan PhD ’24. 

Tuberculosis, the world’s deadliest infectious disease, is estimated to infect around 10 million people each year, and kills more than 1 million annually. Once established in the lungs, the bacteria’s thick cell wall helps it to fight off the host immune system.

Much of that cell wall is made from complex sugar molecules known as glycans, but it’s not well-understood how those glycans help to defend the bacteria. One reason for that is that there hasn’t been an easy way to label them inside cells.

MIT chemists have now overcome that obstacle, demonstrating that they can label a glycan called ManLAM using an organic molecule that reacts with specific sulfur-containing sugars. These sugars are found in only three bacterial species, the most notorious and prevalent of which is Mycobacterium tuberculosis, the microbe that causes TB.

After labeling the glycan, the researchers were able to visualize where it is located within the bacterial cell wall, and to study what happens to it throughout the first few days of tuberculosis infection of host immune cells.

The researchers now hope to use this approach to develop a diagnostic that could detect TB-associated glycans, either in culture or in a urine sample, which could offer a cheaper and faster alternative to existing diagnostics. Chest X-rays and molecular diagnostics are very accurate but are not always available in developing nations where TB rates are high. In those countries, TB is often diagnosed by culturing microbes from a sputum sample, but that test has a high false negative rate, and it can be difficult for some patients, especially children, to provide a sputum sample. This test also requires many weeks for the bacteria to grow, delaying diagnosis.

“There aren’t a lot of good diagnostic options, and there are some patient populations, including children, who have a hard time giving samples that can be analyzed. There’s a lot of impetus to develop very simple, fast tests,” says Laura Kiessling, the Novartis Professor of Chemistry at MIT and the senior author of the study.

MIT graduate student Stephanie Smelyansky is the lead author of the paper, which appears this week in the Proceedings of the National Academy of Sciences. Other authors include Chi-Wang Ma, an MIT postdoc; Victoria Marando PhD ’23; Gregory Babunovic, a postdoc at the Harvard T.H. Chan School of Public Health; So Young Lee, an MIT graduate student; and Bryan Bryson, an associate professor of biological engineering at MIT.

Labeling glycans

Glycans are found on the surfaces of most cells, where they perform critical functions such as mediating communication between cells.In bacteria, glycans help the microbes to enter host cells, and they also appear to communicate with the host immune system, in some cases blocking the immune response.

“Mycobacterium tuberculosis has a really elaborate cell envelope compared to other bacteria, and it’s a rich structure that’s composed of a lot of different glycans,” Smelyansky says. “Something that’s often underappreciated is the fact that these glycans can also interact with our host cells. When our immune cells recognize these glycans, instead of sending out a danger signal, it can send the opposite message, that there’s no danger.”

Glycans are notoriously difficult to tag with any kind of probe, because unlike proteins or DNA, they don’t have distinctive sequences or chemical reactivities that can be targeted. And unlike proteins, they are not genetically encoded, so cells can’t be genetically engineered to produce sugars labeled with fluorescent tags such as green fluorescent protein.

One of the key glycans in M. tuberculosis, known as ManLAM, contains a rare sugar known as MTX, which is unusual in that it has a thioether — a sulfur atom sandwiched between two carbon atoms. This chemical group presented an opportunity to use a small-molecule tag that had been previously developed for labeling methionine, an  amino acid that contains a similar group.

The researchers showed that they could use this tag, known as an oxaziridine, to label ManLAM in M. tuberculosis. The researchers linked the oxaziridine to a fluorescent probe and showed that in M. tuberculosis, this tag showed up in the outer layer of the cell wall. When the researchers exposed the label to Mycobacterium smegmatis, a related bacterium that does not cause disease and does not have the sugar MTX, they saw no fluorescent signal.

“This is the first approach that really selectively allows us to visualize one glycan in particular,” Smelyansky says.

Better diagnostics

The researchers also showed that after labeling ManLAM in M. tuberculosis cells, they could track the cells as they infected immune cells called macrophages. Some tuberculosis researchers had hypothesized that the bacterial cells shed ManLAM once inside a host cell, and that those free glycans then interact with the host immune system. However, the MIT team found that the glycan appears to remain in the bacterial cell walls for at least the first few days of infection.

“The bacteria still have their cell walls attached to them. So it may be that some glycan is being released, but the majority of it is retained on the bacterial cell surface, which has never been shown before,” Smelyansky says.

The researchers now plan to use this approach to study what happens to the bacteria following treatment with different antibiotics, or immune stimulation of the macrophages. It could also be used to study in more detail how the bacterial cell wall is assembled, and how ManLAM helps bacteria get into macrophages and other cells.

“Having a handle to follow the bacteria is really valuable, and it will allow you to visualize processes, both in cells and in animal models, that were previously invisible,” Kiessling says.

She also hopes to use this approach to create new diagnostics for tuberculosis. There is currently a diagnostic in development that uses antibodies to detect ManLAM in a urine sample. However, this test only works well in patients with very active cases of TB, especially people who are immunosuppressed because of HIV or other conditions.

Using their small-molecule sensor instead of antibodies, the MIT team hopes to develop a more sensitive test that could detect ManLAM in the urine even when only small quantities are present.

“This is a beautifully elegant approach to selectively label the surface of mycobacteria, enabling real-time monitoring of cell wall dynamics in this important bacterial family. Such investigations will inform the development of novel strategies to diagnose, prevent, and treat mycobacterial disease, most notably tuberculosis, which remains a global health challenge,” says Todd Lowary, a distinguished research fellow at the Institute of Biological Chemistry, Academia Sinica, Taipei Taiwan, who was not involved in the research.

The research was funded by the National Institute of Allergy and Infectious Disease, the National Institutes of Health, the National Science Foundation, and the Croucher Fellowship.

MIT physicists have captured the first images of individual atoms freely interacting in space. The pictures reveal correlations among the “free-range” particles that until now were predicted but never directly observed. Their findings, appearing today in the journal Physical Review Letters, will help scientists visualize never-before-seen quantum phenomena in real space.

The images were taken using a technique developed by the team that first allows a cloud of atoms to move and interact freely. The researchers then turn on a lattice of light that briefly freezes the atoms in their tracks, and apply finely tuned lasers to quickly illuminate the suspended atoms, creating a picture of their positions before the atoms naturally dissipate.

The physicists applied the technique to visualize clouds of different types of atoms, and snapped a number of imaging firsts. The researchers directly observed atoms known as “bosons,” which bunched up in a quantum phenomenon to form a wave. They also captured atoms known as “fermions” in the act of pairing up in free space — a key mechanism that enables superconductivity.

“We are able to see single atoms in these interesting clouds of atoms and what they are doing in relation to each other, which is beautiful,” says Martin Zwierlein, the Thomas A. Frank Professor of Physics at MIT.

In the same journal issue, two other groups report using similar imaging techniques, including a team led by Nobel laureate Wolfgang Ketterle, the John D. MacArthur Professor of Physics at MIT. Ketterle’s group visualized enhanced pair correlations among bosons, while the other group, from École Normale Supérieure in Paris, led by Tarik Yefsah, a former postdoc in Zwierlein’s lab, imaged a cloud of noninteracting fermions.

The study by Zwierlein and his colleagues is co-authored by MIT graduate students Ruixiao Yao, Sungjae Chi, and Mingxuan Wang, and MIT assistant professor of physics Richard Fletcher.

Inside the cloud

A single atom is about one-tenth of a nanometer in diameter, which is one-millionth of the thickness of a strand of human hair. Unlike hair, atoms behave and interact according to the rules of quantum mechanics; it is their quantum nature that makes atoms difficult to understand. For example, we cannot simultaneously know precisely where an atom is and how fast it is moving.

Scientists can apply various methods to image individual atoms, including absorption imaging, where laser light shines onto the atom cloud and casts its shadow onto a camera screen.

“These techniques allow you to see the overall shape and structure of a cloud of atoms, but not the individual atoms themselves,” Zwierlein notes. “It’s like seeing a cloud in the sky, but not the individual water molecules that make up the cloud.”

He and his colleagues took a very different approach in order to directly image atoms interacting in free space. Their technique, called “atom-resolved microscopy,” involves first corralling a cloud of atoms in a loose trap formed by a laser beam. This trap contains the atoms in one place where they can freely interact. The researchers then flash on a lattice of light, which freezes the atoms in their positions. Then, a second laser illuminates the suspended atoms, whose fluorescence reveals their individual positions.

“The hardest part was to gather the light from the atoms without boiling them out of the optical lattice,” Zwierlein says. “You can imagine if you took a flamethrower to these atoms, they would not like that. So, we’ve learned some tricks through the years on how to do this. And it’s the first time we do it in-situ, where we can suddenly freeze the motion of the atoms when they’re strongly interacting, and see them, one after the other. That’s what makes this technique more powerful than what was done before.”

Bunches and pairs

The team applied the imaging technique to directly observe interactions among both bosons and fermions. Photons are an example of a boson, while electrons are a type of fermion. Atoms can be bosons or fermions, depending on their total spin, which is determined by whether the total number of their protons, neutrons, and electrons is even or odd. In general, bosons attract, whereas fermions repel.

Zwierlein and his colleagues first imaged a cloud of bosons made up of sodium atoms. At low temperatures, a cloud of bosons forms what’s known as a Bose-Einstein condensate — a state of matter where all bosons share one and the same quantum state. MIT’s Ketterle was one of the first to produce a Bose-Einstein condensate, of sodium atoms, for which he shared the 2001 Nobel Prize in Physics.

Zwierlein’s group now is able to image the individual sodium atoms within the cloud, to observe their quantum interactions. It has long been predicted that bosons should “bunch” together, having an increased probability to be near each other. This bunching is a direct consequence of their ability to share one and the same quantum mechanical wave. This wave-like character was first predicted by physicist Louis de Broglie. It is the “de Broglie wave” hypothesis that in part sparked the beginning of modern quantum mechanics.

“We understand so much more about the world from this wave-like nature,” Zwierlein says. “But it’s really tough to observe these quantum, wave-like effects. However, in our new microscope, we can visualize this wave directly.”

In their imaging experiments, the MIT team were able to see, for the first time in situ, bosons bunch together as they shared one quantum, correlated de Broglie wave. The team also imaged a cloud of two types of lithium atoms. Each type of atom is a fermion, that naturally repels its own kind, but that can strongly interact with other particular fermion types. As they imaged the cloud, the researchers observed that indeed, the opposite fermion types did interact, and formed fermion pairs — a coupling that they could directly see for the first time.

“This kind of pairing is the basis of a mathematical construction people came up with to explain experiments. But when you see pictures like these, it’s showing in a photograph, an object that was discovered in the mathematical world,” says study co-author Richard Fletcher. “So it’s a very nice reminder that physics is about physical things. It’s real.”

Going forward, the team will apply their imaging technique to visualize more exotic and less understood phenomena, such as “quantum Hall physics” — situations when interacting electrons display novel correlated behaviors in the presence of a magnetic field.

“That’s where theory gets really hairy — where people start drawing pictures instead of being able to write down a full-fledged theory because they can’t fully solve it,” Zwierlein says. “Now we can verify whether these cartoons of quantum Hall states are actually real. Because they are pretty bizarre states.”

This work was supported, in part, by National Science Foundation through the MIT-Harvard Center for Ultracold Atoms, as well as by the Air Force Office of Scientific Research, the Army Research Office, the Department of Energy, the Defense Advanced Projects Research Agency, a Vannevar Bush Faculty Fellowship, and the David and Lucile Packard Foundation.

Caring well for the elderly is a familiar challenge. Some elderly people need close medical attention in facilities; others struggle with reduced capabilities while not wanting to leave their homes. For families, finding good care is hard and expensive, and already-burdened family members often pick up the slack.

The problem is expanding as birthrates drop while some segments of the population live longer, meaning that a growing portion of the population is elderly. In the U.S., there are currently three states currently where at least 20 percent of the population is 65 and older. (Yes, Florida is one.) But by 2050, demographic trends suggest, there will be 43 states with that profile.

In age terms, “America is becoming Florida,” quips MIT economist Jonathan Gruber. “And it’s not just America. The whole world is aging rapidly. The share of the population over 65 is growing rapidly everywhere, and within that, the share of the elderly that are over 85 is growing rapidly.”

In a new edited volume, Gruber and several other scholars explore the subject from a global perspective. The book, “Long-Term Care around the World,” is published this month by the University of Chicago Press. The co-editors are Gruber, the Ford Professor of Economics and chair of the Department of Economics at MIT; and Kathleen McGarry, a professor of economics at Stony Brook University.

The book looks at 10 relatively wealthy countries and how they approach the problem of long-term care. In their chapter about the U.S., Gruber and McGarry emphasize a remarkable fact: About one-third of long-term care for the elderly in the U.S. is informal, provided by family and friends, despite limited time and resources. Overall, long-term care is 2 percent of U.S. GDP.

“We have two fundamental long-term care problems in the U.S.,” Gruber says. “Too much informal care at home, and, relatedly, not enough options for elders to live with effective care in ‘congregate housing’ [or elder communities], even if they’re not sick enough for a nursing facility.”

The nature of the problem

The needs of the elderly sit in plain sight. In the U.S., about 30 percent of people 65 and over, and 60 percent of people 85 and over report limitations in basic activities. Getting dressed and taking baths are among the most common daily problems; shopping for groceries and managing money are also widely reported issues. Additionally, these limitations have mental health implications. About 10 percent of the elderly report depression, rising to 30 percent among those who struggle with three or more types of basic daily tasks.

Even so, the U.S. is not actually heavily dotted with nursing homes. In a country of about 330 million people, with 62 million being 65 and over, it’s unusual for an elderly person to be in one.

“We all think of nursing homes as where you go when you’re old, but there are only about 1.2 million people in nursing homes in America,” Gruber observes. “Which is a lot, but tiny compared to the share of people who are elderly in the U.S. and who have needs. Most people who have needs get them met at home.”

And while nursing homes can be costly, home care is too. Given an average U.S. salary of $23 per hour for a home health care aide, annual costs can reach six figures even with half-time care. As a result, many families simply help their elderly relatives as best they can.

Therefore, Gruber has found, we must account for the informal costs of elder care, too. Ultimately, Gruber says, informal help represents “an inefficient system of people taking care of their elderly parents at home, which is a stress on the family, and the elders don’t get enough care.”

To be sure, some people buy private long-term care insurance to defray these costs. But this is a tricky market, where insurers are concerned about “adverse selection,” people buying policies with a distinct need for them (beyond what insurers can detect). Rates therefore can seem high, and for limited, conditional benefits. Research by MIT economist Amy Finkelstein has shown that only 18 percent of long-term insurance policies are used.

“Private long-term care insurance is a market that just hasn’t worked well,” Gruber says. “It’s basically a fixed amount of money, should you meet certain conditions. And people are surprised by that, and it doesn’t meet their needs, and it’s expensive. We need a public solution.”

Congregate housing, a possible solution

Looking at long-term care internationally helps identify what those solutions might be. The U.S. does not neglect elder care, but could clearly broaden its affordable options.

“On the one hand, what jumped out at me is how normal the U.S. is,” Gruber says. “We’re in the middle of the pack in terms of the share of GDP we spend on long-term care.” However, some European countries that spend a similar share and also rely heavily on informal elder care, including Italy and Spain, have notably lower levels of GDP per capita.

Some other European countries with income levels closer to the U.S., including Germany and the Netherlands, do spend more on long-term elder care. The Netherlands tops the list by devoting about 4 percent of its GDP to this area.

However, in the U.S., the issue is not so much drastically changing how much it spends on long-term elder care, but how it spends. The Dutch have a relatively more extensive system of elder communities — the “congregate housing” for the elderly who are not desperately unwell, but simply find self-reliance increasingly hard.

“That’s the huge missing hole in the U.S. long-term care system, what do we do with people who aren’t sick enough for a nursing home, but probably shouldn’t be at home,” Gruber says. “Right now they stay at home, they’re lonely, they’re not getting services, their kids are super-stressed out, and they’re pulling millions of people out of the labor force, especially women. Everyone is unhappy about it, and they’re not growing GDP, so it’s hurting our economy and our well-being.”

Overall, then, Gruber thinks further investment in elder-care communities would be an example of effective government spending that can address the brewing crisis in long-term care — although it would require new federal legislation in a highly polarized political environment.

Could that happen? Could the U.S. invest more now and realize long-term financial benefits, while allowing working-age employees to spend more time at their jobs rather than acting as home caregivers? Making people more aware of the issue, Gruber thinks, is a necessary starting point.

“If anything might be bipartisan, it could be long-term care,” Gruber says. “Everybody has parents. A solution has to be bipartisan. Long-term care may be one of those areas where it’s possible.”

Support for the research was provided, in part, by the National Institute on Aging.

A team from MIT Lincoln Laboratory has built and demonstrated the wideband selective propagation radar (WiSPR), a system capable of seeing out various distances at millimeter-wave (mmWave or MMW) frequencies. Typically, these high frequencies, which range from 30 to 300 gigahertz (GHz), are employed for only short-range operations. Using transmit-and-receive electronically scanned arrays of many antenna elements each, WiSPR produces narrow beams capable of quickly scanning around an area to detect objects of interest. The narrow beams can also be manipulated into broader beams for communications.

"Building a system with sufficient sensitivity to operate over long distances at these frequencies for radar and communications functions is challenging," says Greg Lyons, a senior staff member in the Airborne Radar Systems and Techniques Group, part of Lincoln Laboratory's ISR Systems and Technology R&D area. "We have many radar experts in our group, and we all debated whether such a system was even feasible. Much innovation is happening in the commercial sector, and we leveraged those advances to develop this multifunctional system."

The high signal bandwidth available at mmWave makes these frequencies appealing. Available licensed frequencies are quickly becoming overloaded, and harnessing mmWave frequencies frees up considerable bandwidth and reduces interference between systems. A high signal bandwidth is useful in a communications system to transmit more information, and in a radar system to improve range resolution (i.e., ability of radar to distinguish between objects in the same angular direction but at different distances from the radar).

The phases for success

In 2019, the laboratory team set out to assess the feasibility of their mmWave radar concept. Using commercial off-the-shelf radio-frequency integrated circuits (RFICs), which are chips that send and receive radio waves, they built a fixed-beam system (only capable of staring in one direction, not scanning) with horn antennas. During a demonstration on a foggy day at Joint Base Cape Cod, the proof-of-concept system successfully detected calibration objects at unprecedented ranges.  

"How do you build a prototype for what will eventually be a very complicated system?" asks program manager Christopher Serino, an assistant leader of the Airborne Radar Systems and Techniques Group. "From this feasibility testing, we showed that such a system could actually work, and identified the technology challenges. We knew those challenges would require innovative solutions, so that's where we focused our initial efforts."

WiSPR is based on multiple-element antenna arrays. Whether serving a radar or communications function, the arrays are phased, which means the phase between each antenna element is adjusted. This adjustment ensures all phases add together to steer the narrow beams in the desired direction. With this configuration of multiple elements phased up, the antenna becomes more directive in sending and receiving energy toward one location. (Such phased arrays are becoming ubiquitous in technologies like 5G smartphones, base stations, and satellites.)

To enable the tiny beams to continuously scan for objects, the team custom-built RFICs using state-of-the-art semiconductor technology and added digital capabilities to the chips. By controlling the behavior of these chips with custom firmware and software, the system can search for an object and, after the object is found, keep it in "track" while the search for additional objects continues — all without physically moving antennas or relying on an operator to tell the system what to do next.

"Phasing up elements in an array to get gain in a particular direction is standard practice," explains Deputy Program Manager David Conway, a senior staff member in the Integrated RF and Photonics Group. "What isn't standard is having this many elements with the RF at millimeter wavelengths still working together, still summing up their energy in transmit and receive, and capable of quickly scanning over very wide angles."

Line 'em up and cool 'em down

For the communications function, the team devised a novel beam alignment procedure.

"To be able to combine many antenna elements to have a radar reach out beyond typical MMW operating ranges — that's new," Serino says. "To be able to electronically scan the beams around as a radar with effectively zero latency between beams at these frequencies — that's new. Broadening some of those beams so you're not constantly reacquiring and repointing during communications — that's also new."

Another innovation key to WiSPR's development is a cooling arrangement that removes the large amount of heat dissipated in a small area behind the transmit elements, each having their own power amplifier.

Last year, the team demonstrated their prototype WiSPR system at the U.S. Army Aberdeen Proving Ground in Maryland, in collaboration with the U.S. Army Rapid Capabilities and Critical Technologies Office and the U.S. Army Test and Evaluation Command. WiSPR technology has since been transitioned to a vendor for production. By adopting WiSPR, Army units will be able to conduct their missions more effectively.

"We're anticipating that this system will be used in the not-too-distant future," Lyons says. "Our work has pushed the state of the art in MMW radars and communication systems for both military and commercial applications."

"This is exactly the kind of work Lincoln Laboratory is proud of: keeping an eye on the commercial sector and leveraging billions-of-dollars investments to build new technology, rather than starting from scratch," says Lincoln Laboratory assistant director Marc Viera.

This effort supported the U.S. Army Rapid Capabilities and Critical Technologies Office. The team consists of additional members from the laboratory's Airborne Radar Systems and Techniques, Integrated RF and Photonics, Mechanical Engineering, Advanced Capabilities and Systems, Homeland Protection Systems, and Transportation Safety and Resilience groups.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a novel artificial intelligence model inspired by neural oscillations in the brain, with the goal of significantly advancing how machine learning algorithms handle long sequences of data.

AI often struggles with analyzing complex information that unfolds over long periods of time, such as climate trends, biological signals, or financial data. One new type of AI model, called "state-space models," has been designed specifically to understand these sequential patterns more effectively. However, existing state-space models often face challenges — they can become unstable or require a significant amount of computational resources when processing long data sequences.

To address these issues, CSAIL researchers T. Konstantin Rusch and Daniela Rus have developed what they call “linear oscillatory state-space models” (LinOSS), which leverage principles of forced harmonic oscillators — a concept deeply rooted in physics and observed in biological neural networks. This approach provides stable, expressive, and computationally efficient predictions without overly restrictive conditions on the model parameters.

"Our goal was to capture the stability and efficiency seen in biological neural systems and translate these principles into a machine learning framework," explains Rusch. "With LinOSS, we can now reliably learn long-range interactions, even in sequences spanning hundreds of thousands of data points or more."

The LinOSS model is unique in ensuring stable prediction by requiring far less restrictive design choices than previous methods. Moreover, the researchers rigorously proved the model’s universal approximation capability, meaning it can approximate any continuous, causal function relating input and output sequences.

Empirical testing demonstrated that LinOSS consistently outperformed existing state-of-the-art models across various demanding sequence classification and forecasting tasks. Notably, LinOSS outperformed the widely-used Mamba model by nearly two times in tasks involving sequences of extreme length.

Recognized for its significance, the research was selected for an oral presentation at ICLR 2025 — an honor awarded to only the top 1 percent of submissions. The MIT researchers anticipate that the LinOSS model could significantly impact any fields that would benefit from accurate and efficient long-horizon forecasting and classification, including health-care analytics, climate science, autonomous driving, and financial forecasting.

"This work exemplifies how mathematical rigor can lead to performance breakthroughs and broad applications," Rus says. "With LinOSS, we’re providing the scientific community with a powerful tool for understanding and predicting complex systems, bridging the gap between biological inspiration and computational innovation."

The team imagines that the emergence of a new paradigm like LinOSS will be of interest to machine learning practitioners to build upon. Looking ahead, the researchers plan to apply their model to an even wider range of different data modalities. Moreover, they suggest that LinOSS could provide valuable insights into neuroscience, potentially deepening our understanding of the brain itself.

Their work was supported by the Swiss National Science Foundation, the Schmidt AI2050 program, and the U.S. Department of the Air Force Artificial Intelligence Accelerator.

In the dim light of the lab, friends, family, and strangers watched the image of a pianist playing for them, the pianist’s fingers projected onto the moving keys of a real grand piano that filled the space with music.

Watching the ghostly musicians, faces and bodies blurred at their edges, several listeners shared one strong but strange conviction: “feeling someone’s presence” while “also knowing that I am the only one in the room.”

“It’s tough to explain,” another listener said. “It felt like they were in the room with me, but at the same time, not.”

That presence of absence is at the heart of TeleAbsence, a project by the MIT Media Lab’s Tangible Media group that focuses on technologies that create illusory communication with the dead and with past selves.

But rather than a “Black Mirror”-type scenario of synthesizing literal loved ones, the project led by Hiroshi Ishii, the Jerome B. Wiesner Professor of Media Arts and Sciences, instead seeks what it calls “poetic encounters” that reach across time and memory.

The project recently published a positioning paper in PRESENCE: Virtual and Augmented Reality that presents the design principles behind TeleAbsence, and how it could help people cope with loss and plan for how they might be remembered.

The phantom pianists of the MirrorFugue project, created by Tangible Media graduate Xiao Xiao ’09, SM ’11, PhD ’16, are one of the best-known examples of the project. On April 30, Xiao, now director and principal investigator at the Institute for Future Technologies of Da Vinci Higher Education in Paris, shared results from the first experimental study of TeleAbsence through MirrorFugue at the 2025 CHI conference on Human Factors in Computing Systems in Yokohama, Japan.

When Ishii spoke about TeleAbsence at the XPANSE 2024 conference in Abu Dhabi, “about 20 people came up to me after, and all of them told me they had tears in their eyes … the talk reminded them about a wife or a father who passed away,” he says. “One thing is clear: They want to see them again and talk to them again, metaphorically.”

Messages in bottles

As the director of the Tangible Media group, Ishii has been a world leader in telepresence, using technologies to connect people over physical distance. But when his mother died in 1998, Ishii says the pain of the loss prompted him to think about how much we long to connect across the distance of time.

His mother wrote poetry, and one of his first experiments in TeleAbsence was the creation of a Twitterbot that would post snippets of her poetry. Others watching the account online were so moved that they began posting photos of flowers to the feed to honor the mother and son.

“That was a turning point for TeleAbsence, and I wanted to expand this concept,” Ishii says.

Illusory communication, like the posted poems, is one key design principle of TeleAbsence. Even though users know the “conversation” is one-way, the researchers write, it can be comforting and cathartic to have a tangible way to reach out across time.

Finding ways to make memories material is another important design principle. One of the projects created by Ishii and colleagues is a series of glass bottles, reminiscent of the soy sauce bottles Ishii’s mother used while cooking. Open one of the bottles, and the sounds of chopping, of sizzling onions, of a radio playing quietly in the background, of a maternal voice, reunite a son with his mother.

Ishii says sight and sound are the primary modalities of TeleAbsence technologies for now, because although the senses of touch, smell, and taste are known to be powerful memory triggers, “it is a very big challenge to record that kind of multimodal moment.”

At the same time, one of the other pillars of TeleAbsence is the presence of absence. These are the physical markers, or traces, of a person that serve to remind us both of the person and that the person is gone. One of the most powerful examples, the researchers write, is the permanent “shadow” of Hiroshima Japanese resident Mitsuno Ochi, her silhouette transferred to stone steps 260 meters from where the atomic bomb detonated in 1945.

“Abstraction is very important,” Ishii says. “We want something to recall a moment, not physically recreate it.”

With the bottles, for instance, people have asked Ishii and his colleagues whether it might be more evocative to fill them with a perfume or drink. “But our philosophy is to make a bottle completely empty,” he explains. “The most important thing to let people imagine, based on the memory.”

Other important design principles within TeleAbsence include traces of reflection — the ephemera of faint pen scratches and blotted ink on a preserved letter, for instance — and the concept of remote time. TeleAbsence should go beyond dredging up a memory of a loved one, the researchers insist, and should instead produce a sense of being transported to spend a moment in the past with them.

Time travelers

For Xiao, who has played the piano her whole life, MirrorFugue is a “deeply personal project” that allowed her to travel to a time in her childhood that was almost lost to her.

Her parents moved from China to the United States when she was a baby — but it took eight years for Xiao to follow. “The piano, in a sense, was almost like my first language,” she recalls. “And then when I moved to America, my brain overwrote bits of my childhood where my operating system used to be in Chinese, and now it’s very much in English. But throughout this whole time, music and the piano stayed constant.”

MirrorFugue’s “sense of kind-of being there and not being there, and the wish to connect with oneself from the past, comes from my own desire to connect with my own past self,” she adds.

The new MirrorFugue study puts some empirical data behind the concept of TeleAbsence, she says. Its 28 participants were fitted with sensors to measure changes in their heart rate and hand movements during the experience. They were extensively interviewed about their perceptions and emotions afterward. The recorded images came from pianists ranging in experience from children early in their lessons to professional pianists like the late Ryuichi Sakamoto.

The researchers found that emotional experiences described by the listeners were significantly influenced by whether the listeners knew the pianist, as well as whether the pianist was known by the listeners to be alive or dead.

Some participants placed their own hands alongside the ghosts to play impromptu duets. One daughter, who said she had not paid close attention to her father’s playing when he was alive, was newly impressed by his talent. One person felt empathy watching his past self struggle through a new piece of music. A young girl, mouth slightly open in concentration and fingers small on the keys, showed her mother a past daughter that wasn’t possible to see in old photos.

The longing for past people and past selves can be “a deep sadness that will never go away,” says Xiao. “You’ll always carry it with you, but it also makes you sensitive to certain aesthetic experiences that’s also beautiful.”

“Once you’ve had that experience, it really resonates,” she adds, “And I think that’s why TeleAbsence resonates with so many people.”

Uncanny valleys and curated memory

Acutely aware of the potential ethical dangers of their research, the TeleAbsence scientists have worked with grief researchers and psychologists to better understand the implications of building these bridges through time.

For instance, “one thing we learned is that it depends on how long ago a person passed away,” says Ishii. “Right after death, when it’s very difficult for many people, this representation matters. But you have to make important informed decisions about whether this drags out the grief too long.”

TeleAbsence could comfort the dying, he says, by “knowing there is a means by which they are going to live on for their descendants.” He encourages people to consider curating “high-quality, condensed information,” such as their social media posts, that could be used for this purpose.

“But of course many families do not have ideal relationships, so I can easily think of the case where a descendant might not have any interest” in interacting with their ancestors through TeleAbsence, Ishii notes.

TeleAbsence should never fully recreate or generate new content for a loved one, he insists, pointing to the rise of “ghost bot” startups, companies that collect data on a person to create an “artificial, generative AI-based avatar that speaks what they never spoke, or do gestures or facial expressions.”

A recent viral video of a mother in Korea “reunited” in virtual reality with an avatar of her dead daughter, Ishii says, made him “very depressed, because they’re doing grief as entertainment, consumption for an audience.”

Xiao thinks there might still be some role for generative AI in the TeleAbsence space. She is writing a research proposal for MirrorFugue that would include representations of past pianists. “I think right now we’re getting to the point with generative AI that we can generate hand movements and we can transcribe the MIDI from the audio so that we can conjure up Franz Listz or Mozart or somebody, a really historical figure.”

“Now of course, it gets a little bit tricky, and we have discussed this, the role of AI and how to avoid the uncanny valley, how to avoid deceiving people,” she says. “But from a researcher’s perspective, it actually excites me a lot, the possibility to be able to empirically test these things.”

The importance of emptiness

Along with Ishii’s mother, the PRESENCE paper was also dedicated “in loving memory” to Elise O’Hara, a beloved Media Lab administrative assistant who worked with Tangible Media until her unexpected death in 2023. Her presence — and her absence — are felt deeply every day, says Ishii.

He wonders if TeleAbsence could someday become a common word “to describe something that was there, but is now gone.”

“When there is a place on a bookshelf where a book should be,” he says, “my students say, ‘oh, that’s a teleabsence.’”

Like a sudden silence in the middle of a song, or the empty white space of a painting, emptiness can hold important meaning. It’s an idea that we should make more room for in our lives, Ishii says.

“Because now we’re so busy, so many notification messages from your smartphone, and we are all distracted, always,” he suggests. “So emptiness and impermanence, presence of absence, if those concepts can be accepted, then people can think a bit more poetically.”