Q&A: What is agentic AI today, and what do we want it to be?

The deployment of automated software systems called AI agents has recently exploded. A November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35 percent of surveyed businesses had already deployed AI agents, while another 44 percent planned to implement agentic AI soon. 

To understand the fundamentals and potential impacts of these increasingly popular tools, MIT News spoke with Phillip Isola, an associate professor in the Department of Electrical Engineering and Computer Science (EECS) and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL), who studies the intelligence AI agents possess, as well as the underlying models and mechanisms that power agentic AI systems.

Q: What is agentic AI and how is it different from generative AI models like ChatGPT and Claude?

A: Agentic AI is AI that takes actions in the world. These actions could be a physical action, like robotic manipulation, or a digital action, like booking a flight. On the other hand, we think of generative AI as making up stories, poems, art, and images, rather than taking actions for us. 

The word “agent” is just a brand name. It usually means AI that is going to help people interact with an application, a website, or the physical world. Most agents we encounter today are digital agents, like customer service agents you can talk with about product complaints. 

Most companies that offer agents use the same few AI models under the hood and give them the ability to take actions and remember what happened. An agent starts with a fundamental generative AI system, like Claude, at the core. Then companies put different wrappers around that foundation model for their product or application. Those wrappers might be specific tools that agent can use, and those tools depend on the application. Maybe the agent has access to a calculator so it can solve math problems, or maybe it has access to a more complicated hard drive and operating system so it can remember a firm’s financial data and past business negotiations. 

The biggest challenge in developing agentic AI comes from a lack of training data. If I want to create a system that can go online and book a flight for me, that seems pretty simple. But we don’t have a lot of data that spells out exactly how to do that — where to move the mouse, which buttons to click on, what to do if something goes wrong, or how to call somebody and negotiate about the price of the airline ticket. One way to train a system like this is to have the AI agent visit airline websites, try things out, and see what works and what doesn’t work. These environments are hard to model, so often the agent must learn by trial and error.

Q: What are some promising applications of agentic AI?

A: I think the area where we’ve seen the most success has been with coding agents. This is something that evolved from generative AI. People trained language models on code, and then they can predict what a human would do to solve a coding problem. In addition, an agent can learn to do this by going through a feedback loop where it tries out different solutions and checks to see if it got the answer right. As long as it can check the answer, the AI agent can perform this trial-and-error loop until it figures out a good strategy.

But there is always a balance between automating decision making versus simply assisting and informing humans. Analytical AI methods, like the systems that help predict possible outcomes of decisions, are not agentic in nature, but are very informative to human decision-makers. For cases that are either high-stakes or safety-critical, like medicine, security, high-level business policies, etc., the technology might not be ready for AI to completely automate those processes, or we might not even be comfortable with that.

Q: Are there risks we should be thinking about when using AI agents?

A: One big risk area comes from the fact that it is often very easy to get agents to do certain types of work for you. With coding agents, you can “vibe code” and just ask the agent to make a code for you, so you don’t have to do the hard work yourself. There is a big risk that, because it is so easy, people will not put enough effort into verifying that it is doing the right thing. Bugs will be introduced, private data will get leaked — this is already happening.

Agents aren’t perfect, in the sense that they might make mistakes because they are not well-trained and don’t know what to do. But even if they are very competent, if a human doesn’t use them appropriately or gives them an instruction that is too vague, the AI agent could make a mistake because the human made a mistake. If humans are less involved in thinking through all the consequences, I think we might be more prone to making those mistakes. 

An additional aspect is the risk of de-skilling. It is unclear how far this will go, but when we are relying on agents to do our homework, our coding, and our math, we might lose the ability to do that ourselves, and we might lose that ability too soon because the technology is not yet ready to fully automate those processes.

Q: What does the future hold for agentic AI?

A: What we think of now as agentic AI refers to large language models using tools to interact with digital and physical systems. One obvious limitation is that, under the hood, these have the architecture of a language model and are trained on text data. To make even more powerful AI agents, we might need to model videos, physical forces, time series, radar scans, and other modalities. We might need to have models with fundamentally different architectures that can handle continuous data, high-dimensional data, stochastic data, and so on. 

But, on the other hand, maybe an extremely good coding model could act as a puppeteer to interface with sensors, actuators, and web APIs? Perhaps, once you have a super-smart reasoning system that understands math, language, and code, you can give it a camera and a keyboard and it will figure out what to do in the spatial domain. Is the next wave of AI just going to be Claude with sensors, actuators, and tools, or is it going to be something built in a new way from the ground up? That’s the big question a lot of people in AI are grappling with right now.

de MIT News https://ift.tt/z4Hn1RM

Scientists find ozone depletion began decades before discovery of ozone hole

The Antarctic ozone hole was discovered in 1985, when scientists observed a severe depletion in the Earth’s protective layer of stratospheric ozone. Industrial chemicals known as chlorofluorocarbons (CFCs), then widely used as refrigerants, propellants, foam-blowing agents, and solvents, were at the root of the ozone depletion. After concerted global effort to phase out the use of CFCs, ozone today is recovering, especially in the Antarctic. 

The discovery of the ozone hole was possible thanks, in part, to the measurement tools that were available at the time. Advances in those tools, along with satellites and other monitoring technologies, have since allowed scientists to track ozone’s recovery. 

But what if today’s tech was available much earlier? Would scientists have been able to spot even earlier signs of human-induced ozone depletion? And if so, when would those first signs have popped up, and where? 

MIT scientists now have some answers. The team, led by atmospheric chemist Susan Solomon, has carried out a thought experiment in which they consider a hypothetical world where today’s atmospheric monitoring capabilities were available throughout the last century. In this scenario, they simulated the atmosphere’s chemistry through history and discovered not only when the earliest sign of ozone depletion would have been detectable, but also where, and why. 

In a study appearing today in the Proceedings of the National Academy of Sciences, the scientists suggest that the first signs of ozone depletion appeared as early as 1957 — about 30 years before the ozone hole was discovered. And, this first signal of ozone loss popped up not in the Antarctic, but in the upper stratosphere of the tropics. What’s more, the cause of this early depletion was not due to CFCs, but to another industrial chemical: carbon tetrachloride. 

“What we’ve learned from textbooks is that CFCs result in ozone depletion,” says the study’s first author, Jian Guan, a graduate student in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “It turns out there was another compound that caused ozone depletion much earlier than CFCs. This was a big surprise.”

For Solomon, who was an early pioneer in the study of ozone’s effects on the atmosphere, and who was the first to show that CFCs were the main agent eroding Antarctic ozone, the new results were a complete shock. 

“The fact that ozone depletion would have happened as early as the late 1950s, which is much earlier than I would have thought, just absolutely blew my mind,” says Solomon, the Lee and Geraldine Martin Professor of Environmental Studies and Chemistry at MIT. “This study shows it’s really important to keep monitoring so that we can fully understand how the atmosphere responds and recovers.”

The study’s MIT co-authors include Peidong Wang, Yaowei Li, and Kane Stone; along with Benjamin Santer of the University of East Anglia; Qiang Fu of the University of Washington; Rolando Garcia, Douglas Kinnison, and Jun Zhang of the National Center for Atmospheric Research; Jean-Francois Lamarque of Climate Modeling and Analysis LLC; and Gabriel Chiodo of the Spanish National Research Council. 

Chlorine connection

Ozone is a highly reactive molecule, made from three oxygen atoms, that exists naturally in the upper layers of the atmosphere. In the stratosphere, ozone acts as a shield, absorbing the sun’s rays and reducing the harmful ultraviolet radiation that can reach the Earth’s surface. 

In the late 1980s, after scientists first observed signs of ozone depletion in the Antarctic, Solomon led expeditions to the region to measure the stratosphere’s composition. Those measurements confirmed that ozone’s agent of destruction was CFCs — the chemicals which were used globally in refrigeration, air conditioning, and aerosol propellants, among other uses. 

Specifically, Solomon measured higher-than-expected levels of chlorine dioxide in the Antarctic stratosphere. The presence of this molecule, in the same place where ozone depletion was observed, had only one chemical explanation: Ozone was being broken apart by rogue atoms of chlorine. At the time, chlorine-heavy CFCs were in wide use, and MIT chemist Mario Molina proposed that if CFCs drifted up to the stratosphere, photons from the sun could break apart the molecules and release atoms of chlorine, which would then be free to break apart ozone’s oxygen atoms. 

Molina’s work, and Solomon’s measurements, were key in showing that CFCs could deplete ozone — a discovery that earned Molina a share of the 1995 Nobel Prize in Chemistry. Soon after, nearly every country in the world signed the Montreal Protocol, which ultimately led to the successful phase-out of CFCs and other ozone-depleting substances. In recent years, as a result of that global cooperation, scientists have observed initial signs of ozone recovery.

“We know what we have now, and ozone is starting to recover,” Solomon says. “But no one has ever really documented where and when and why the first ozone depletion would have happened.”

Signal over noise

For their new study, Solomon, Guan, and their colleagues took a “what-if” approach, posing the question: What if the past had the monitoring capabilities of the present? When would we have been able to detect the earliest sign of human-induced ozone depletion? 

Today’s monitoring tools are sensitive to a certain signal to noise, meaning they can identify patterns of ozone loss that are more likely a “signal” of human-induced depletion (such as from CFCs), versus ozone loss that is due to “noise,” such as random fluctuations from weather and natural phenomena. 

With this in mind, the team looked to reproduce the chemistry of the atmosphere over the last century to see whether they could see a signal over the noise, based on the sensitivity of today’s monitoring tools. 

The team used 16 different model runs, each of which simulates varying conditions and dynamics of the atmosphere at various latitudes and altitudes, as well as the concentrations and interactions of ozone and other molecules. Ozone is affected by not only human-caused chemicals but also natural phenomena such as volcanic eruptions and El Niño weather patterns. Each model run simulates ozone’s response to these natural phenomena, which the team combined to establish a range of “noise,” or ozone depletion that likely is due to natural variability.

They added to each model the various industrial chemicals that were known to have been produced at various times over the last century. 

“Year by year, we have estimates from industry of how much of these chemicals were made and sold globally, and the emissions of all these chemicals, which the models include,” Solomon explains. “And in the case of carbon tetrachloride, the really cool thing is, we also have ice core data.”

Ice cores are drilled-out cylinders of deeply buried ice, that had formed in the Antarctic and Arctic from the falling and layering of snow over hundreds of years. Ice cores contain the remnants of snow, as well as whatever trace chemicals in the atmosphere the snow originally fell through. Scientists can therefore use ice cores to estimate the composition of the atmosphere through history. 

“We actually see in the ice cores that carbon tetrachloride starts increasing already by the 1940s,” Solomon notes. 

The team incorporated industrial and ice core data into their models, then looked to see whether a signal of human-induced ozone loss stood out from the noise of natural fluctuations. Their analysis revealed that a signal did appear, as early as 1957. Not only did they see when the signal appeared, but also where: in the tropics, rather than the Antarctic. 

The researchers say that human-induced ozone loss was likely occurring globally, but was easier to spot in the tropical upper stratosphere, since that is the region where the range of natural fluctuations is the smallest, and therefore where a signal can stand out better.

Finally, the analysis indicated that carbon tetrachloride, and not CFCs, was the cause of the earliest ozone depletion. 

“That’s the only ozone-depleting substance that was increasing that early,” Solomon says. “We started using carbon tetrachloride in the 1930s as a dry-cleaning agent, and as a degreasing solvent. We didn’t start using CFCs until quite a bit later.”

Carbon tetrachloride has since been phased out of use in most of the world, initially due to its health concerns; the chemical can cause nervous system disorders with prolonged exposure and is a suspected carcinogen. Since the Montreal Protocol began to tightly limit its use in the 1990s, the molecule’s concentrations in the atmosphere have been on a decline. Still, Solomon says the new study highlights the need for vigilance in monitoring carbon tetrachloride, CFCs, and other ozone-depleting substances that may have been phased out but can still linger for decades.

“We’ve gone through a big effort to get rid of these chemicals,” Solomon says. “Don’t we have an obligation to keep monitoring to make sure the atmosphere responds the way we think it should?”

This research was supported, in part, by the National Science Foundation, the National Oceanic and Atmospheric Administration, and the European Commission.

de MIT News https://ift.tt/Xpa9PoN

3 Questions: Beyond data-driven aesthetics

“Beyond Data-Driven Aesthetics,” by MIT Architecture alumnus and researcher Alexandros Haridis, on view at the MIT Keller Gallery through June 30, examines 20th- and 21st-century efforts to transform computing into a medium for creative production and aesthetic judgment in architecture and the applied arts. Drawing on philosophy, mathematics, computer science, and design computation, the exhibition translates algorithms, theories, and machine-learning systems into physical installations and interactive visualizations.

Q: What inspired “Beyond Data-Driven Aesthetics,” and what questions does it explore?

A: The conceptual origins of “Beyond Data-Driven Aesthetics” emerged from three intersecting lines of research.

First, while completing my PhD in design and computation in the MIT Department of Architecture around 2022, I observed in real time how advances in data-driven machine learning — systems such as ChatGPT and Stable Diffusion — were rapidly entering public discussions about creativity, aesthetic judgment, design, and even high-profile art auctions.

At the same time, my own research was already focused on aesthetic judgment and evaluation, and it became increasingly clear to me that many of the questions presented publicly as “new” in relation to AI actually have a much longer history across the 20th century. For example, in the 1956 Dartmouth Summer Research Project, a foundational event for the field of AI, creation and evaluation processes were identified as one of seven key dimensions of human intelligence that future AI research should address.

Second, the exhibition was influenced by research in design computation and shape grammars that investigates relationships between human insight and computation through rule-based methods, rather than purely data-driven learning. More recent interpretative studies of aesthetic theories — drawing from figures such as Samuel Taylor Coleridge, Oscar Wilde, and even John von Neumann — have been especially important to me. These studies examine whether theories of aesthetic value and comparison articulated in philosophical and literary texts may reveal possibilities or limitations in contemporary models of digital computation and AI in architecture and design.

Finally, the exhibition was motivated by the use of design, fabrication, and data visualization as methods for interpreting mathematical concepts, algorithms, and “black box” machine-learning systems. Across disciplines, researchers increasingly use reconstruction and visualization techniques to make computational systems more tangible and interpretable — from neural network visualization in computer science to software reconstruction and digital fabrication in architecture and curatorial practice.

Q: How do you translate research on computation and aesthetics into an exhibition?

A: The approach of the exhibition is to ask what exactly in a particular research paper or book captures its most salient idea, and then use design to interpret that idea in a visual, spatial, and experiential format. Drawing on design techniques such as software reconstruction, physical making, and data visualization, the exhibition takes written sources that are dense with algorithmic ideas, abstract concepts, and mathematical formulas, and translates them into stories in space that include interaction, material form, and digital visualization.

The exhibition itself is organized around five thematic areas: Aesthetic Measure, Aesthetic Guidelines, Algorithmic Aesthetics, Aesthetic Appropriation, and Aesthetic Novelty. Each theme functions as a selective “window” into a distinct computational approach to aesthetic judgment drawn from a specific publication — a book or research paper. The titles of these themes are derived from concepts central to each publication. For example, “measure” refers to mathematician George Birkhoff’s work in the 1930s to quantify aesthetic value mathematically, while “novelty” examines how the machine learning system AICAN judges generated images according to a theory in cognitive aesthetics that balances familiarity and deviation from known artistic styles.

Across all five cases, the key insight is that design itself can function as a method of interpretative translation — a way of making visible, tangible, and experiential what traditional academic scholarship in technical domains typically communicates only through words and word-like representational devices, such as scientific diagrams and tables.

Q: What questions are you hoping to explore next?

A: “Beyond Data-Driven Aesthetics” is conceived both as a research exhibition and as an ongoing platform for investigating how computational systems participate in processes of aesthetic judgment, generation, and transformation across architecture and the applied arts.

One of the central questions of the exhibition — and one that researchers across architecture, design, and engineering are increasingly focusing on — is computational evaluation beyond purely performative or functional requirements. This applies to many different design spaces, whether buildings, structural forms, or everyday products. The exhibition’s case studies suggest that many of these questions long predate current interest in computing and AI, and have been approached through a range of computational and theoretical models of evaluation since at least the early 20th century.

At the same time, I’m increasingly interested in how these ideas can move into broader applications related to the built environment. In particular, I am interested in how research connected to “Beyond Data-Driven Aesthetics” can help designers and engineers better understand how computation — whether rule-based or data-driven — can inform us about what contributes positively to human experience in relation to the spaces and objects people inhabit and use.

Finally, a direction I continue to explore is the methodological role of design itself as an interpretative device. Through software reconstruction, visualization, and physical making, the exhibition uses design to translate opaque computational systems into more legible, tangible, and experiential artifacts. More broadly, this opens questions not only about mechanizing “beauty” or “taste” (the traditional preoccupation of aesthetic formalism in the 20th century), but also about how traditional forms of research scholarship and communication may evolve through spatial, visual, and public-facing formats.

de MIT News https://ift.tt/pDaukcL

Graphene can hold multiple states of superconductivity, a new study finds

The ordinary graphite in pencil lead is proving to be surprisingly multifaceted at the microscale. 

In a study appearing today in the journal Nature, MIT researchers report that a certain microscopic structure found in natural graphite can host multiple superconducting states. Superconductivity is an electronic state of matter in which electrons pair up and glide through a material with zero resistance. 

While there are thousands of materials that are known to be superconductors, it is rare for one material to host multiple forms of superconductivity. 

The researchers discovered the multiple superconducting states in atomically thin exfoliations of graphite, known as graphene. Specifically, graphene is a single-atom-thin sheet of carbon atoms arranged precisely in a microscopic lattice. The team made its discoveries in samples of rhombohedral graphene, which is a natural structure within graphite consisting of a stack of four or five graphene layers. 

Interestingly, the researchers found that several of the new superconducting states in rhombohedral graphene are able to persist in the presence of a magnetic field, which normally kills superconductivity. 

And in a further surprise, these superconducting states even get stronger when exposed to a magnetic field. 

Overall, the findings reveal a new family of unconventional superconducting states in one seemingly simple material. 

“People might assume that this is a simple, boring carbon material,” says Long Ju, the Lawrence C. and Sarah W. Biedenharn Associate Professor of Physics at MIT. “But we can control this material by tuning certain experimental ‘knobs,’ such as electrical voltages. This is how a simple physical material can exhibit so many different superconducting properties.” 

It’s still unclear exactly how each of the multiple superconducting states arise, or how they are able to persist under a magnetic field, when normally superconductivity should fade.

“From a fundamental physics point of view, it’s very exotic that a magnetic field doesn’t kill superconductivity, and instead it boosts it,” Ju says. “We have provided a lot of experimental results and provided the nutrition that people can absorb to try to think about what’s going on here.” 

The study’s MIT co-authors include co-first authors Junseok Seo and Shenyong Ye, together with Tonghang Han, Zhenghan Wu, Wei Xu, Jixiang Yang, Emily Aitken, Prayoga Liong, Phatthanon Pattanakanvijit, Zach Hadjri, and Mingda Li. External collaborators are co-first author Armel Cotten and members of Dominik Zumbuhl’s group at the University of Basel in Switzerland, plus others at Florida State University, the University of Florida, Gainesville, and the National Institute for Materials Science in Japan. 

Natural steps

Graphene and other atomically thin, two-dimensional materials can exhibit unexpected electronic, magnetic, thermal, and physical properties. And when two or more sheets of graphene are stacked and twisted at precise orientations, the “magic-angle” structure can suddenly host weird and exotic phenomena. 

Ju’s group has been probing the exceptional properties of graphene. But rather than artificially stacking and twisting layers, they have looked for interesting behavior in naturally occurring graphene structures. In recent years, they have unearthed surprising electronic properties in rhombohedral graphene. This particular configuration consists of graphene layers stacked on top of each other, each one slightly offset from the last, similar to the steps in a staircase. 

Rhombohedral graphene can be found naturally in ordinary graphite. But to find it first requires exfoliating a block of graphite (usually with Scotch tape), then searching the exfoliated sample for the telltale staircase-like pattern, which researchers can then isolate for further experimentation. 

Using this approach, Ju and his colleagues have been able to isolate and probe samples of four- and five-layer rhombohedral graphene. They have so far discovered that the structure can host a rare, “chiral” form of superconductivity, as well as fractional electron charge, among other behavior. 

In the flow

For their new study, the team took a slightly different approach in studying rhombohedral graphene. Previously, they electrically “doped” their samples, progressively adding electrons as they passed a separate electric current into the material. They then measured the voltage, or essentially the force that pushes the current through the material, and looked for instances when the voltage dropped to zero, indicating that the current was passing through without resistance.

In this way, the team has observed superconductivity when adding electrons to rhombohedral graphene. So they wondered: What might happen if they did the opposite, and took electrons away? 

In their new study, the team looked for signs of superconductivity as they carefully removed electrons from rhombohedral graphene, progressively lowering the material’s electron density, as they applied a separate, external electric current to measure the electrical resistance. In these experiments, they also applied external magnetic field along directions parallel and perpendicular to the graphene plane. These experiments were carried out in collaboration with Zumbuhl’s group in Switzerland, who provided access to a laboratory setup in which graphene samples could be exposed to high magnetic fields and ultracold temperatures. 

In these experiments, the researchers found that at certain electron densities, four different superconducting states emerged. What’s more, three of the states persisted in the presence of a relatively high magnetic field. 

Normally, magnets destroy superconductivity by severing the bond between the paired electrons gliding through the material. 

But in Ju’s experiments, the team observed three superconducting states that survived in a magnetic field up to around 9 tesla, which is about 180,000 times stronger than the Earth’s magnetic field. In these instances, the magnetic field they applied was in a parallel orientation with respect to the plane of the material. When they switched the magnetic field to a perpendicular orientation, they discovered another surprise: At a certain electron density, superconductivity not only persisted, but increased. The material was able to continue superconducting, at higher temperatures than predicted. 

Every superconducting material has a critical temperature below which electrons can conduct without resistance, and above which superconductivity cannot persist. But the team found that, at a certain electron density, and in the presence of a perpendicular magnetic field, superconductivity in rhombohedral graphene was able to survive beyond the material’s critical temperature that corresponds to zero magnetic field. 

“The superconductivity actually is enhanced, as in, the transition temperature goes from 55 millikelvin to probably 90 millikelvin,” Ju explains. “At the same time, the material can take another 50 or 60 percent extra current before superconductivity gets destroyed. And that is very unusual.”

The researchers are unsure of what microscopic behavior is enabling multiple and unconventional superconducting states, though they propose one idea. Conventional superconductivity emerges when electrons pair up. These “Cooper pairs” consist of electrons with opposite spin, and it’s thought that a magnetic field can pull the spins out of their opposite configurations, and as a result, break up superconductivity. 

Instead, the team proposes that perhaps in rhombohedral graphene, and at certain electron densities, electrons can pair up with aligned spins. Any magnetic field would still pull on the spins, but in the same direction, preserving their alignment, and their superconductivity. 

The researchers acknowledge that the idea needs much more investigation, both experimentally and theoretically. For now, they see the results as a demonstration of what new and exotic phenomena can emerge in a seemingly simple material, with the right measurements and controls. 

“We can control the simplest of chemicals — carbon — and structurally alter the material, which is part of our fun,” says lead author Junseok Seo, who is a graduate student in Ju’s group. “We’re not only dealing with what nature gives us, but we’re applying additional controls to change it to something that nature does not give us, but that can exist in the same material.”

This work was supported, in part, by the U.S. Office of Naval Research. Device fabrication was carried out, in part, at MIT.nano.

de MIT News https://ift.tt/A15JMzf

Students from across the Northeast step inside MIT.nano’s cleanroom

“Illuminating.” “Spectacular.” “Compelling.” This is how community college students described the two days they spent at MIT.nano learning about the complex tools inside the cleanroom and building and packaging their own functional photonic chips.

“Integrated photonics is an essential part of semiconductor packaging today,” says Anu Agarwal, principal research scientist in the Materials Research Laboratory at MIT. “But there is no single, standardized university curriculum for integrated electronics-photonics packaging. We need to create educational materials to teach this subject across the talent pipeline from K-12 and beyond, which is exactly what we’re doing at the Initiative for Knowledge and Innovation in Manufacturing (IKIM) and MIT.nano.”

As leader of the Lab for Education and Application Prototypes (LEAP) facility located on MIT.nano’s fifth floor, Agarwal stresses the importance of hands-on learning when studying integrated photonics, the science of guiding and manipulating light on a semiconductor chip. Through the Northeast Consortia for Advanced Integrated Silicon Technologies (NCAIST) program, she’s bringing community and four-year college students to MIT.nano for experimental boot camps that teach how to use semiconductor tools for electronic-photonic packaging and testing.

“Having a workforce skilled in resource-efficient semiconductor manufacturing, including electronic-photonic packaging, is critical to maintain the exponential growth of the chip industry and build national security,” says Agarwal. “MIT.nano, through programs like NCAIST, are helping to train more people in STEM.”

Working closely with AIM Photonics, a U.S. Manufacturing Innovation Institute, NCAIST coordinates and accelerates the transition of technician education content and teaching methodologies from key AIM-affiliated U.S. universities to community, technical, and four-year colleges in the Northeast. Through NCAIST, in Massachusetts, the Massachusetts Bay Community College (MBCC) is paired with MIT, North Shore Community College (NSCC) with Stonehill College, and Springfield Technical Community College (STCC) with Western New England University.

“The NCAIST program offers a transformative opportunity for our community college students to experience hands-on training at MIT.nano’s LEAP facility,” says Marina Bograd, professor and chair of the engineering department at MassBay Community College. “For many of them, this is their first time stepping into a cleanroom or seeing semiconductor manufacturing up close. The experience helps open doors that might otherwise feel out of reach, builds confidence, and inspires our students to see themselves pursuing careers in emerging technologies.”

The most recent MIT.nano boot camp, held on May 20-21, expanded participation to include not only those from MBCC, but also students from NSCC, Stonehill College, and SUNY Polytechnic Institute, where NCAIST is headquartered. Twelve students spent two full days at MIT.nano operating a die saw, die bonder, wire bonder, and flip chip tool to build and test a packaged chip.

“I found the combination of hands-on activities, lectures, and informal discussion with the MIT.nano team and fellow students fostered an awesome learning environment,” says Cari Caudill, a student at NSCC. “As a mechanical engineering student, I was most interested in packaging and the machines themselves, so I loved getting direct experience with the tools and discussing with our instructors how procedural and technological development has impacted precision, efficiency, and scalability in the semiconductor industry.”

"The NCAIST boot camp was an exciting and illuminating experience!” adds MassBay Community College student Wyatt Maurer. “I really appreciated getting the chance to work with semiconductor manufacturing tools and to learn about the future of photonics from leaders in the field.”

Students attended lectures on cleanroom safety by Kristofor Payer, assistant director of operations at MIT.nano; electronic-photonic packaging by Agarwal; and photonic integrated circuit sensing by Department of Materials Science and Engineering graduate student Lizzie Gower. They were also offered virtual reality (VR) simulation exercises by Sajan Saini, the director of education at IKIM, to help build intuition about photonic devices and semiconductor packaging tools. These VR simulations serve as a foundational tool to help students visualize photonic devices and complex tool mechanics, as well as run digital process steps and deepen their technical understanding. By bridging physical fabrication with advanced simulation resources, the LEAP students are mastering highly specialized manufacturing, assembly, and testing pipelines required to build the future of electronic-photonic integration.

“The experience at this boot camp not only strengthens our student technical skills, it helps them see themselves as future contributors to a rapidly evolving field,” says Mary Beth Steigerwald, professor and engineering department chair at North Shore Community College. “It also enriches their professional portfolios and gives them a stronger, more compelling story to share during internship and transfer interviews.”

The students will use this training to secure summer internships at hard technology companies. Several have also been accepted to four-year degree programs to continue their education in the fall.

de MIT News https://ift.tt/S3P5kLH

Past participants are now the leaders of MIT’s dynaMIT Club

Every summer for the past 13 years, students in MIT’s club dynaMIT have taught STEM principles to Boston-area middle school students at no cost, all in an effort to inspire the next generation of innovators.

In August, dynaMIT will welcome two cohorts of budding scientists and engineers to campus. First, 40 middle schoolers in grades 6–7 will dive into hands-on STEM learning through creative activities like solar s'mores and paper rockets. The following week, another 40 students in grades 8–9 will join in, exploring innovative experiments that spark curiosity and creative problem-solving. Each day, a new topic is covered, exposing attendees to chemistry, machine learning, physics, math, biology, and earth and space science.

Several of the program's attendees have gone on to apply and be accepted to MIT, including the club’s co-director, Dominique Dang. When the Quincy, Massachusetts, native saw the club’s table at the Midway Fair, she knew she wanted to join to give back.

“I didn’t receive a lot of STEM exposure in middle school, but then I saw online about the STEM program offered by dynaMIT, and I was really interested. I had so much fun, and it introduced me to creating things, and not just reading about them in a textbook. I knew I wanted to be a scientist, but I didn’t know what type of science I wanted to study, so having dynaMIT expose me to a different STEM topic each day was a transformative experience,” says Dang, who is now studying computer science and molecular biology.

Megan Zhu, the club’s other co-director, was immediately drawn to the organization’s educational mission. A biology major with plans to pursue an MD/PhD program, Zhu is passionate about advancing science education and aspires to teach at the university level upon completing her degree.

“I happened to stop by the dynaMIT table at the club fair, and it seemed really cool. I spoke to a couple of the club leaders, and they talked about how they help support education in the Boston area. Education has always been something that I was passionate about in my hometown in Rapid City, South Dakota, and I wanted to emphasize giving back to the community,” says Zhu.

Lukeman Nouri, who grew up in Saugus, Massachusetts, attended dynaMIT as a sixth grader. “I barely knew what MIT was, or even what STEM meant, so I wasn't particularly excited to go. However, that changed after the very first day of the program! I remember extracting DNA from a strawberry, making elephant toothpaste, and gathering fingerprints from various surfaces. However, my biggest highlight was learning Scratch and creating my very first game,” says Nouri, who is majoring in computer science and engineering. “After dynaMIT, MIT became my dream college, and I spent the next six years learning more about STEM and MIT.”

Erick Liang, who grew up in Boston’s Chinatown and Roslindale neighborhoods and is now majoring in nuclear science and engineering and physics, had a similar experience after attending dynaMIT. “As a first-generation, low-income student, having a meaningful and engaging program like dynaMIT to participate in over the summer was really important for me. DynaMIT exposed me to different fields of science I had not encountered yet in elementary or middle school and helped spark my interest in STEM,” says Liang.

Zhu says this year they are adding a new activity related to climate change and clean water that they hope will create an interest in these two important areas. “This summer, one of our activities is called Sponge City. It’s about runoff water and clean, reusable water. We’ll have the students build a city that can withstand a storm. They will be given a budget and have to decide how to spend the resources after we pour water all over the tray containing their city — all in an effort to show them how important climate change and clean drinking water are.”

The club is also partnering with the Koch Institute for Integrative Cancer Research at MIT and will tour lab space and work on a fun experiment about cell heterogeneity and cancer tumor formation. Attendees will then be able to talk to scientists and ask them questions.

“I’m looking forward to giving this cohort the same great experience that I had six summers ago. DynaMIT was so much fun, and I learned so much from it that I feel a responsibility to help make it just as impactful for future students,” says Nouri.

Liang adds, “I am excited to return and help set up the plasma demo kits for the program’s physics day!”

“It’s a great full-circle moment,” says Dang. “That’s just one of the reasons why I joined the club.”

“Watching the students work on the activities is always the most rewarding part of the two weeks, and that makes the entire year of planning worth it,” says Zhu, adding, “the club is also an excellent community at MIT.”

Students interested in joining dynaMIT or volunteering for this summer’s program can find more information on the club’s website.

de MIT News https://ift.tt/vks98ea

LLMs help robots understand vague instructions and focus on key details

Imagine working at a warehouse or office sometime in the near future, and you’re asked to help a new trainee learn the basics of their job. The catch: It’s a robot. To teach them, you might want to play a game of “show and tell” — that is, physically showing how to do something a few different ways, while also explaining what you’re doing.

Let’s say you asked the robot to place some coffee on your desk without disturbing you during a Zoom call. You’ll prefer that the robot doesn’t get too close to you and the laptop so that it doesn’t interrupt your meeting. To enable this behavior, the robot should be trained with data that clearly demonstrates the full task. Computer scientists have attempted to explain manipulation tasks to robots by recording lots of physical demonstrations or writing extensive directions. But if you don’t have both, the machine is likely to misunderstand what it needs to do.

It’s laborious for humans to do all that showing and telling, so researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have automated the process of teaching a robot, while clarifying instructions automatically and using nearly five times less demonstration data. Their “Masked Inverse Reinforcement Learning” (Masked IRL) approach uses a large language model (LLM) to elaborate on ambiguous prompts based on the data collected from a user’s demo. Another LLM then narrows down which details an algorithm should incorporate into a motion plan, so that a robot can safely complete chores in homes, offices, and factories.

“Our approach could come in handy when a human interacts with a robot but doesn’t want to spell out all the details of a task,” says MIT PhD student and CSAIL researcher Minyoung Hwang, who is a lead author on a paper presenting the project. “We’re minimizing human effort by enabling machines to get to the bottom of what users really want.”

According to Hwang, Masked IRL can help robots safely maneuver in settings where there are elements a human might not describe in a prompt, but that are crucial nonetheless. For example, a machine grabbing you a snack from the kitchen may not know to avoid bumping into your laptop. Likewise, a factory robot placing items into different boxes must carefully navigate around shelves.

To learn new tasks in these situations, Masked IRL uses the robot’s sensors to capture information about its surroundings. These components also log each movement of a kinesthetic demonstration — a training approach where a human physically moves a robot to do a specific action. It’s sort of like being the machine’s physical therapist, bending joints in a particular direction to show a robot how to grab, move, and place objects.

MIT’s system then calls on an LLM to compare this sequence of motions (called a trajectory) to the shortest possible path. The model also elaborates on what might be unclear in a prompt, turning a request like “stay close” into “stay close to the surface of the table.” Using the trajectory comparison and clarified directions, the LLM begins to understand why the motions it was trained on are important to the task. 

A second LLM then evaluates details of the environment, such as the position of obstacles and the shape of the robot’s target object. During this process, it “masks” (in other words, ignores) the elements it deems irrelevant to the task at hand, scoring each one as either a “1” (important) or “0” (not so much). For example, whether or not a user was leaning on a table during a demonstration would be a “0,” making it irrelevant. Any detail considered a “1” is incorporated into the final action plan by an algorithm.

These masks gave Masked IRL a key advantage over comparable baselines in both 3D and real-world demos because it taught a robot which information to prioritize. Thanks to the researchers’ system, virtual and real robots alike were able to skillfully maneuver objects around obstacles, such as moving a coffee mug around a laptop to different spots on a table. In these tasks, Masked IRL correctly identified users’ preferences, which they didn’t explicitly state in their prompts, up to 15 percent more often than comparable baselines.

During simulation experiments, CSAIL researchers also found that Masked IRL was a fast learner. It required fewer demos to understand how to move the mug than its baselines. They also found that the robots performed better when an LLM cleared up instructions, instead of having the machine try to follow a vague request.

This more focused approach also translated well to a real robotic arm, executing prompts the system hadn’t seen during its training phase. After being trained on 50 kinesthetic demonstrations, the robot carefully moved a cup toward a human while avoiding colliding with a user’s computer — an obstacle it learned to avoid by elaborating on a more general request to “stay away.” It also wiped a table down while “staying close” to it, and handed a user a bag of chips while “staying away” from both a human and a table.

Masked IRL senses and explains what users leave unsaid, but soon, it might “see” it too. CSAIL researchers plan to make their approach more dynamic by equipping it with cameras, allowing a robot to take images of its surroundings. Then it could highlight and focus on specific elements nearby. For example, if you asked the machine to pick up a toy, it might see some bananas nearby and ignore them before handling its target object.

Hwang wrote the paper with three CSAIL colleagues: PhD student Alexandra Forsey-Smerek ’20, SM ’22; postdoc Nathaniel Dennler; and MIT Assistant Professor Andreea Bobu, who is a member of the Department of Aeronautics and Astronautics and CSAIL. Their work was supported, in part, by the Tata Group via the MIT Generative AI Impact Consortium Award, and the Department of Defense. They’ll present the project at the 2026 IEEE International Conference on Robotics and Automation in June.

de MIT News https://ift.tt/iADEbWp