When you imagine a retail employee, you probably think of someone at the register or helping a customer. However, employees also spend a lot of time searching warehouses and production floors, fulfilling orders or online orders, and generally trying to keep track of all their inventory.

Tracking inventory takes a long time, in part because retailers don’t always know where everything is. That’s why when you ask a salesperson to check if they have a shirt in your size, it may take them 20 minutes to respond.

Cartesian helps retailers track inventory with technology developed at MIT. The system uses wireless signals from radio frequency identification (RFID) tags attached to items to determine their exact location in the store, from the warehouse to the production floor.

Last year, Cartesian conducted a study with the retailer and found that its platform delivered significant annual savings at the store level by improving inventory tracking, optimizing workflows and improving customer experiences.

“The big problem we’re solving is that about 50 percent of retail hours are spent on inventory management,” says co-founder Fadel Adib SM ’13, PhD ’17, an associate professor at MIT. “This is an approximately $15 billion problem in the United States alone. We use algorithms to decipher indoor locations using wireless signals. The underlying technology enables a new level of indoor location.”

Cartesian is already implemented in over 700 stores in 15 countries and cooperates with one of the largest fashion groups in the world, Inditex, which is the parent company of brands such as ZARA, Pull&Bear and Oysho.

In addition to retailers and warehouses, the Cartesian platform can also improve indoor location tracking for manufacturers, logistics operators and robotics companies.

“The broad vision of what we do is spatial artificial intelligence,” says Adib. “Today, AI is doing exceptionally well in the digital world. Now it needs to move into the physical world. That means enabling machines to perceive their environment in a way that they can interact with it. This is where spatial AI comes in and where the Cartesian comes in.”

From technology to product

Adib, who works jointly in the Media Lab and MIT’s Department of Electrical Engineering and Computer Science, has been researching wireless signals at the Institute for more than 15 years, dating back to his graduate studies.

“My group today is exploring how to use wireless signals to sense the world in ways that weren’t possible before,” Adib says. “We develop the core technology and then build systems around it. Our goal is to deploy these systems in the real world to achieve impact.”

When Adib joined the MIT faculty, the first project he worked on was indoor localization using RFID tags. Isaac Perper ’20, MEnG ’21 later joined his lab as a graduate student, and together they developed machine learning algorithms to process RFID data to translate it into localization patterns, initially focusing on helping robots locate RFID indoors.

In 2021, Adib participated in the National Science Foundation’s I-Corps program, which requires researchers to interview potential customers to find relevant problems to solve with their technology. That’s when he realized what a gigantic problem inventory management was for retailers.

Cartesian was officially founded by Adib and Perper in early 2023 after receiving a diminutive business award from the National Science Foundation. The two worked with MIT’s Office of Technology Licensing to license patents from the Adiba lab. They also received support from MIT’s Venture Mentoring Service.

“Our goal was to reduce the cost of the technology to make it scalable,” Adib recalls. “Isaac focused on simplifying the product, leveraging advances in machine learning, and accelerating performance. At first, this required a lot of iteration and testing.”

For many reasons, retail workers spend most of their time locating items. They may receive an online order to fulfill, restock store shelves, or receive an inquiry from a customer about items in the back.

Stores vary in how they organize inventory. Most separate items by category on specific shelves and bins and then operate barcodes or inventory systems that tend to become dated quickly.

“This is a big problem for stores because customers may simply leave before asking an employee to check their size, or customers may become frustrated and leave if it takes too long,” Adib says. “The co-worker also wastes time searching for items that could be used to perform higher-value work.”

The Cartesian platform works with retailers’ handheld RFID readers that store associates already operate to manage inventory. Each store installs Cartesian software into their existing warehouse applications or uses a custom application that employees have direct access to.

“Thanks to RFID readers, stores inform what is in stock and what is not” – Perper says. “We found a way to take the same scans they already use in the reader, put the generated data into our machine learning algorithms, and generate maps of where everything is located.”

Customers can create analytics powered by Cartesian technology to track inventory levels, show customers maps of where each item is located, and create other services.

“They use our location intelligence platform and create different products based on it,” says Adib. “We can work with any device, any store, any type of RFID. It’s a simple interface. All the sophisticated location algorithms are in the cloud.”

Outside of retail

Cartesian signed its first gigantic contract in 2025 and soon expanded its operations to several hundred stores. One of the advantages of Cartesian is the ability to scale quickly. Perper says he can add a store in about a minute. The Cartesian team doesn’t even have to go to a recent store to turn on the system if they already have a relationship with the company.

“It’s as simple as flipping a switch, preparing the data and sending it to our customers,” says Perper. “One of our first big bets was, ‘Can we build this entirely on existing hardware?’ That bet is starting to pay off.

Cartesian models can also work with Wi-Fi and Bluetooth signals, which the company plans to use with customers in other industries.

“Right now we’re focusing on retail applications, but the technology has a lot of relevance in manufacturing, warehouses and other locations,” Adib says.

The Cartesian team aims to deploy to tens of thousands of stores over the next year and then begin expanding beyond retail into industries such as manufacturing and robotics.

“What’s most exciting to me about Cartesian is that we’ve built a lot of the technology foundation, and now that we have the foundation in place, we hope to build specific application layers,” Perper says. “Then we can ask customers across industries about their problems and apply our technology in different ways to solve them.”

The post The startup helps retailers track their products in real time appeared first on AI SCKOOL.

The Institute for Artificial Intelligence and Fundamental Interactions (IAIFI), led by MIT, has received renewed support from the National Science Foundation (NSF) for an additional five years, increasing annual funding from $4 million to $4.98 million. The renewal marks a up-to-date phase for IAIFI, which has spent the first five years of its existence building a research model and interdisciplinary community around a central premise: that artificial intelligence can open up up-to-date ways of doing physics, while physics can assist shape better artificial intelligence systems.

Launched in 2020 as part of the National Institutes for Research on Artificial Intelligence program, IAIFI brings together researchers from MIT and Harvard, Northeastern, Tufts and Boston universities. His work has shown that machine learning can accelerate discoveries in physics, and that insights from physics can make artificial intelligence systems more rule-based and easier to interpret.

“From the beginning, IAIFI has been about a two-pronged effort: AI enabling better physics and physics enabling better AI,” says Jesse Thaler, director of IAIFI and professor of physics at MIT. “Over the past five years, we have witnessed this virtuous cycle in many areas of physics and artificial intelligence. The exchange brings not only new results, but also truly new ways of doing science.”

Research in the field of physics and artificial intelligence

IAIFI’s research spans particle physics, nuclear physics, astrophysics and basic artificial intelligence, and many advances result from collaboration in these areas.

In particle physics, IAIFI researchers have developed artificial intelligence techniques to handle massive amounts of data from the Enormous Hadron Collider in real time, helping to transform the firehose of collision data into actionable physics. In nuclear physics, IAIFI researchers operate artificial intelligence-based generative methods to model the interactions of quarks and gluons in lattice quantum chromodynamics, creating up-to-date ways to study the structure of matter from first principles. In astrophysics, machine learning is being used to discover up-to-date cosmic phenomena and escalate the sensitivity of MIT’s LIGO gravitational wave experiment.

At the same time, ideas from physics influence the development of up-to-date artificial intelligence methods. IAIFI researchers are developing learning algorithms and up-to-date model architectures that embed physical knowledge and best practices – including symmetries, geometric structures, accuracy guarantees and statistical methodologies – directly into neural networks, creating systems that are more reliable, interpretable and data proficient.

“Artificial intelligence has begun to change the way physicists tackle some of the most difficult problems in the field,” says Mike Williams, interim director of IAIFI and professor of physics at MIT. “More importantly, the frontier of problems we can realistically address is beginning to expand, making it possible to address issues that were once completely beyond our reach.”

Training the next generations

A characteristic feature of IAIFI is investing in people. The IAIFI Postdoctoral Fellowship Program supports early-career scientists conducting research at the intersection of physics and artificial intelligence, connecting each fellow with mentors in both fields and fostering collaboration between institutions.

To date, eight scholarship recipients have completed the program. Three have secured teaching positions; others have taken up research roles at leading artificial intelligence companies or joined start-ups, reflecting how widely the skills cultivated at IAIFI translate.

“The IAIFI fellowship shows what can happen when early-career scientists are given the freedom and support to work across traditional boundaries,” says Phiala Shanahan, interim deputy director of IAIFI and professor of physics at MIT. “Our fellows don’t just make separate contributions to physics or artificial intelligence — they help shape a growing field at the intersection.”

The annual IAIFI Summer Doctoral School has become a focal point for a growing community of “centaur scientists” with expertise in both physics and artificial intelligence. The 2026 edition of the program received nearly 600 applications for approximately 100 spots, with approximately 300 additional participants expected to join virtually. Previous participants have highly recommended this school to their colleagues due to its combination of lectures, hands-on tutorials, coding sprints, and networking events.

At MIT, IAIFI has helped shape up-to-date educational pathways, including an interdisciplinary Ph.D. program in physics, statistics, and data science—a collaboration between the Department of Physics and the Center for Statistics and Data Science—that will award 20 Ph.D.s as of 2021. IAIFI members Phil Harris and Isaac Chuang also developed a course in computational data analytics in physics, offered both on campus (Course 8.16) and within the program free online course through MITx.

A growing community

In addition to its core research and training programs, IAIFI organizes annual summer workshops for researchers, which will be held this year at the MIT Schwarzman College of Computing. The Institute also engages the broader public through partnerships with the MIT Museum, the Boston Science Museum, hackathons, and widely viewed online content on artificial intelligence and physics.

“IAIFI shows what becomes possible when researchers in physics, computation, statistics, and data science organize around shared science questions,” says Nergis Mavalvala, dean of the MIT School of Science and the Curtis and Kathleen Marble Professor of Astrophysics. “This type of sustained, interdisciplinary collaboration is essential to the future of scientific discovery.”

IAIFI is hosted by the MIT Nuclear Science Laboratory, chaired by Director Jesse Thaler (currently on sabbatical), interim director Mike Williams, interim deputy director Phiala Shanahan, and managing director Marisa LaFleur, as well as steering committee members Lisa Barsotti, Isaac Chuang, Will Detmold, Bill Freeman, Phil Harris, Lina Necib, Tess Smidt, and Marin Soljacic (and steering committee members from other IAIFI universities).

Looking to the future

As a member of the National Institutes for Artificial Intelligence Research program, IAIFI participates in nationwide efforts to advance artificial intelligence-based discoveries and innovations.

“The connections between NSF AI Institutes are as valuable as the work performed within them and are constantly evolving,” says Marisa LaFleur, managing director of IAIFI. “We provide management strategies and resources for training, community building and collaboration that strengthen the entire network.”

For IAIFI, the renewed funding is an opportunity to delve deeper into what the institute calls the “physics of artificial intelligence” – the operate of physical reasoning, physical challenges and physical tools not only to apply artificial intelligence, but also to understand and improve it. This program, along with a growing community of researchers trained to work in a variety of disciplines, is driving the next phase of the institute’s activities.

“The first phase of IAIFI established a model: interdisciplinary research, early-career talent, and a dynamic community organized around the idea that AI and physics are mutually reinforcing,” says Thaler. “Now we have the foundation – and the entrepreneurial spirit of our centaur scientists – to push this model into new territory and raise our ambitions.”

The post NSF resumes support for MIT-led Artificial Intelligence and Physics Institute, expanding up-to-date discovery model appeared first on AI SCKOOL.

In 2026, the hype around AI agents will be louder than ever before. These semi-autonomous programs can “think” and perform well-defined tasks in areas such as customer service and software development, typically using language models (LMs). However, fields such as medical diagnostics and scientific discovery require them to seek a wide range of solutions in the uncertain environments that LMs face.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Harvard University’s School of Engineering and Applied Sciences (SEAS) took a closer look at LMs to understand their main problems in high-stakes situations. Their test: “Battleship” is a classic guessing game that has helped cognitive scientists study how people search for information.

Scientists from CSAIL and SEAS added a twist by changing the game to ask and answer questions in natural language. In Collaborative Battleship, one participant is the “captain” who asks where the hidden ships are, while his or her teammate acts as the “observer,” answering these questions in real time.

First, the researchers asked more than 40 people to play together, collecting their yes and no questions and answers to build the “BattleshipQA” dataset. These results provided a helpful point of comparison as the team tested state-of-the-art LMs (such as GPT-5) and smaller models (such as Llama 4 Scout) in their game. Without training the models in advance, they found that the best LMs could “beat” humans at “Battleship” – that is, complete the game in fewer turns – but smaller systems were much less rational.

The main problem was that many models are simply not adept at asking useful questions. To get LM to ask questions in a way that reveals more information about the hidden ships, the researchers gave each model a Monte Carlo inference strategy that precisely measures the probability that different options will be correct for each answer. The result: AI models that can beat regular Battleship players at any scale.

Perhaps the most striking result was the gains of the Lamy 4 Scout. As a relatively diminutive LM, it only defeats humans 8 percent of the time. However, by improving its inference strategy, the model achieved an 82% win rate for “Battleship” compared to humans. This careful and proficient style of questioning also enabled the model to outperform the frontier model (GPT-5), at a cost of approximately 1 percent of its cost.

In addition to this improvement, the researchers reduced the gap between humans and LMs in answering questions. While GPT-5 was a reliable tracking tool, helping models complete games faster, smaller systems had a bad habit of giving incorrect answers about where ships were hidden. The models saw an average 15 percent raise in accuracy when they started turning questions into code that explicitly told them how to verify their answers (for example, allowing the model to quickly scan an area when asked if there was a ship).

“Modern language models are primarily optimized for answering complex queries, but it is less clear whether they learn to ask good questions,” says MIT graduate student and CSAIL researcher Gabriel Grand SM ’23, who is the lead author of the study paper about work. “Our work shows that asking information questions depends on the ability to predict and simulate the world. We found that when we give agents access to a ‘world model,’ they ask better questions and make discoveries more effectively.”

A huge change for the Champions League

The team focused primarily on getting LMs to ask better questions. When implementing Monte Carlo inference strategies, LMs consider potential guesses as individual particles. Those that seem more exact with each observer’s response will have more weight, sort of like game balls that inflate or deflate each turn. With this more calculated and adaptive approach, the captain was able to ask questions, which allowed much more information to be obtained from the observer.

The researchers then turned to the widely used Python programming language to support AI observers. Each question asked by the captain was automatically converted into a coded command. For example, a question like “Is there a ship in the first column that spans two rows?” turns into instructions for the LM observer to search the area and assess how wide the digital game element is. By giving the model clear directions in a language it understood particularly well, each system was much more likely to provide correct answers. For example, the lightweight GPT-4o-mini system saw a nearly 30 percent raise in performance, and even the gigantic Claude 4 Opus model jumped by about eight points.

“In this field, ‘automatic formalization’ strategies, in which LMs generate code to validate their solutions, have been very successful,” says senior author Jacob Andreas, associate professor of electrical engineering and computer science at MIT and principal investigator of CSAIL. “What’s most exciting about this work is that it opens the possibility of using these techniques to generate better solutions, primarily by improving the information exploration and gathering capabilities of LMs. We’re excited to extend this work from scientific fields to applications such as coding and math problem solving.”

Let’s play something else

But how would this approach work in other board games? The team tested their newly equipped LMs on “Guess Who?”, where gigantic and diminutive models skillfully whittled down 100 options to correctly guess which hidden character was chosen. Llama 4 Scout was successful 30 percent of the time, but after adjustments by Grand and his colleagues, it completed the task in more than 72 percent of runs. Meanwhile, GPT-4o increased from 62 percent to 90 percent. GPT-5 watched every game to ensure questions were answered as accurately as possible.

While LM has made promising progress in both games, there is still a lot of work to be done. For example, models still have difficulty answering elaborate questions compared to humans. OpenAI researcher, recent Harvard graduate and co-author Valerio Pepe adds that “GPT-5 can beat the average battleship player, and with our methods it does a hair better. However, for all models it is still difficult to beat experienced players, unlike chess, where even the best players cannot cope with artificial intelligence systems.”

The researchers’ findings show that AI agents have untapped potential for needle-in-a-haystack discoveries – navigating a expansive space of options to find a infrequent solution to scientific challenges. While improved information retrieval skills would make them excellent research assistants, for example identifying the molecular structure of a compound, scientists caution that the “Common Battleship” is a fairly basic testbed. They would like to test LM in more elaborate settings where systems need to consider many more options.

Grand also plans to collaborate with humans and AI models to see if they work better together. The models could also benefit from more refined game simulations, and with greater computational power, LMs would have more advanced inference capabilities to predict game evolution.

“As AI systems become increasingly agentic, the most difficult problems turn out to be social: tracking common ground, resolving misunderstandings, and adapting to different partners over time,” says Robert Hawkins, an assistant professor of linguistics at Stanford University, who was not involved in the paper. “This work elegantly captures these phenomena in a controlled, collaborative environment and provides compelling evidence that the real bottleneck for AI agents is not just the computation of optimal questions, but the pragmatic reasoning needed to get the most out of the answers.”

Grand and Pepe co-wrote the paper with two CSAIL principal investigators: MIT associate professor Jacob Andreas and MIT professor Joshua Tenenbaum. Their work was supported in part by the MIT Siegel Family Quest for Intelligence, the MIT-IBM Watson AI Lab, the FinTechAI@CSAIL initiative, a Sloan research fellowship, Intel, the Air Force Office of Scientific Research, the Defense Advanced Research Projects Agency, the Office of Naval Research, and the National Science Foundation. They presented their paper as an oral presentation at the International Conference on Learning Representations (ICLR) in April.

The post Teach AI agents to ask better questions by playing ‘Battleship’ appeared first on AI SCKOOL.

To accelerate and improve decision-making in a active global marketplace, enterprises can deploy generative artificial intelligence models that assist summarize and interpret the charts that often populate market summaries and financial reports.

However, even the latest vision-linguistic models sometimes struggle with this task because it requires a model that integrates visual, numerical, and linguistic understanding. A company investing in a cutting-edge model may still receive misleading or incomplete information.

To fill this performance gap, researchers at MIT and the MIT-IBM Computing Research Lab have developed multi-faceted resources for AI users, designed specifically to teach vision-language models (VLM) how to effectively interpret graphs.

They used a novel data generation method to build a state-of-the-art dataset with over a million different graphs. The dataset also encodes many of the visual, linguistic, and numerical components of each graph image, which enables models to reason reliably about the information in the graph.

By enabling open source models to better leverage their commercial counterparts, ChartNet can enable compact businesses with narrow budgets to more easily leverage AI. The open-source dataset can be used to improve the capabilities of AI models for tasks such as analyzing business trends and interpreting scientific data.

“We designed ChartNet as a one-stop-shop for graphing, covering essentially everything an AI model and the practitioners that train that model might need. We hope our work will motivate researchers to achieve state-of-the-art performance using smaller models that don’t require infinite computation,” says Jovana Kondic, an electrical engineering and computer science (EECS) graduate student at MIT and lead author of the book ChartNet article.

She was joined on the paper by a number of co-authors from MIT, the MIT-IBM Computing Research Lab and IBM Research, including Pengyuan Li, a research associate at IBM Research; Dhiraj Joshi, senior scientist at IBM Research; Isaac Sanchez, software engineer at IBM Research; Aude Oliva, director of strategic industry engagement at the MIT Schwarzman College of Computing, director of the MIT-IBM Computing Research Lab, and senior research fellow at the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Rogerio Feris, principal scientist and manager at the MIT-IBM Computing Research Lab. The research results will be presented at the IEEE Computer Vision and Pattern Recognition Conference.

Bottleneck in the dataset

Scientists have made great strides in developing generative artificial intelligence models that excel in natural language processing and reasoning about natural images. However, less work has been focused on interpreting the convoluted multimodal data contained in charts, Kondic says.

But for companies gigantic and compact in almost every industry, understanding charts is a critical task.

“The financial industry thrives on graphs. If vision language models can extract information from graphs, such as trend descriptions, it facilitates many of the workflows that take place downstream,” says Joshi.

The lack of high-quality training data is a major bottleneck hindering the development of VLMs that can accurately interpret graphs. Many datasets contain narrow graph images downloaded from the Internet and often lack the necessary scale and additional information to assist the model interpret the underlying data.

“A visual language model, unlike our brains, may need to see thousands of examples during training to reliably recognize something in the form of a line graph,” Kondic says.

Scientists have tried to overcome these shortcomings by generating synthetic data. Synthetic data is artificially generated by algorithms to mimic the statistical properties of real data.

The ChartNet dataset contains over one million high-quality chart images along with the appropriate code used to generate each chart, a text description, and a table containing numerical information. Additionally, each data point contains question and answer pairs that train the model to correctly answer questions about the graph image.

“These additional data modes guide the model to combine and align the various information encoded in the graph image,” says Kondic.

Data generation

To build ChartNet, researchers created a two-step process for generating synthetic data.

First, their automated system translates any pre-existing set of chart images into code. The system then iteratively extends this code to change various aspects of each chart, such as the chart type, data values, topic, colors, etc.

“We can start with a single graph, which we use as a seed, and develop hundreds of extensions to it. In this way, we managed to build a dataset with over a million different images,” explains Kondic.

They also implemented an automated quality control process to ensure the high quality of synthetic data. This process verifies that the code is executable and that the rendered graph images are right and tidy.

“We don’t just want to generate a variety of samples. We also want the information to be presented in an understandable way,” he says.

ChartNet also includes a selection of data points in charts annotated by experts. This provides access to additional chart types and supporting data that are guaranteed to be valid.

A trainee can operate the annotated data to tune an existing VLM, further improving performance in a specific application, adds Joshi.

With ChartNet, compact open source models consistently outperformed much larger commercial models.

“Many previous training datasets focused solely on answering simple graph questions. At ChartNet, we have tried to go beyond that by generating data that supports all aspects of solid graph understanding,” says Kondic.

In the future, researchers plan to further expand ChartNet by including data with additional levels of complexity. They also want to benefit from feedback from the research community.

This research was funded in part by the MIT-IBM Computing Research Lab.

The post MIT researchers teach AI models to interpret graphs appeared first on AI SCKOOL.

MIT economist Whitney Newey Ph.D.’83, Ford Professor Emeritus of Economics, received the 2026 Year Award Erwin Plein Nemmers Prize in Economics.

The two-year Nemmers Awards from Northwestern University recognizes top scholars for their lasting contributions to modern knowledge, outstanding achievements, and the development of significant modern ways of analysis.

The university cited Newey, whose research focused on econometrics, for creating “a body of work that shaped the field of semiparametric econometrics, guided both econometricians and empirical researchers for several decades, and helped lay the foundation for modern machine learning inference.”

Newey will work with Northwestern faculty and students on programs scheduled for the 2026–27 academic year. The award also includes a $300,000 prize.

“I am delighted, deeply honored and very grateful,” Newey says of the honor. “I am excited about the opportunity and am currently working on ideas and approaches that are important to modern machine learning inference and modern empirical economics more generally, much of it with such capable collaborators. This award will accelerate that work.”

Newey has been a leading figure in econometric theory for over forty years, shaping both research and training in the field. He has done groundbreaking work in variance estimation, nonparametric simultaneous equations, consumer surplus estimation under general heterogeneity, and unbiased machine learning.

“My colleagues and I are thrilled that Whitney’s incredible career has been recognized with this high honor,” says Jonathan Gruber, Ford Professor of Economics and chair of the Department of Economics. “His research gave birth to many of the econometric methods that are now second nature to economists, and those of us in his orbit also know him as a source of wise, comprehensive, and generous advice. Whitney, through both his groundbreaking research and great generosity, symbolizes what has made MIT economics so great for so many years.”

Newey is a distinguished fellow of the American Economic Association, a fellow of the American Academy of Arts and Sciences, and a fellow of the Econometric Society. He is also a member of the International Association for Applied Econometrics CEMP at Jinan University and an international member of the Center for Microdata Methods and Practice (CEMMAP) at University College London. He earned bachelor’s and doctorate degrees in economics from Brigham Adolescent University and MIT, respectively.

Newey was named a 2020 Distinguished Fellow by the American Economic Association. He was a member of the Center for Advanced Study in the Behavioral Sciences and received a research fellowship from the Sloan Foundation. He was co-editor of the journal published by the Econometric Society and program co-chair of the Econometric Society World Congress. He also served on the executive committee of the Econometric Society. He previously taught economics at Princeton University and MIT, and previously served as chair of the MIT Department of Economics. He has been a guest researcher, professor and lecturer at institutions around the world.

The post MIT economist Whitney Newey awarded the Erwin Plein Nemmers Prize in Economics appeared first on AI SCKOOL.

It is estimated that of all possible chemical compounds from 10²⁰ and 10⁶⁰ may have potential as diminutive molecule drugs.

Experimentally evaluating each of these compounds would be far too time-consuming for chemists. Therefore, in recent years, researchers have begun to exploit artificial intelligence to identify compounds that could be good drug candidates.

One of those researchers is MIT Associate Professor Connor Coley PhD ’19, a 1957 career development associate professor with joint appointments in the departments of chemical engineering, electrical engineering and computer science, and the MIT Schwarzman College of Computing. His research transcends the boundary between chemical engineering and computer science, as he develops and implements computational models to analyze the extensive number of possible chemical compounds, design recent compounds, and predict the reaction pathways that may generate these compounds.

“It’s a very general approach that can be applied to any application of organic molecules, but the main application we’re thinking about is small-molecule drug discovery,” he says.

The intersection of artificial intelligence and science

Coley’s interest in science runs in the family. In fact, he says, there are more scientists than non-scientists in his family, including his father, a radiologist; his mother, who earned a degree in molecular biophysics and biochemistry before going to MIT Sloan School of Management; and his grandmother, a mathematics professor.

As a high school student in Dublin, Ohio, Coley competed in Science Olympiads and graduated from high school at the age of 16. He then went to Caltech, where he chose chemical engineering as a major because it allowed him to combine his interests in science and mathematics.

During his undergraduate studies, he also developed an interest in computer science, working in a structural biology lab using the Fortran programming language to solve the crystal structure of proteins. After graduating from Caltech, he decided to pursue a degree in chemical engineering and came to MIT in 2014 to begin his Ph.D.

Under the advice of professors Klavs Jensen and William Green, Coley worked on ways to optimize automated chemical reactions. His work focused on combining machine learning and cheminformatics – the application of computational methods to analyze chemical data – to plan reaction pathways that could produce recent drug molecules. He also worked to design equipment that could be used to perform these reactions automatically.

Some of this work was done as part of a DARPA-funded program called Make-It, which focused on using machine learning and data analytics to improve the synthesis of drugs and other useful compounds from elementary building blocks.

“That was my real starting point for thinking about cheminformatics, about machine learning, and how we can use models to understand how different chemicals can be made and what reactions are possible,” Coley says.

Coley began applying for faculty positions while still an undergraduate and accepted an offer from MIT at the age of 25. He received various pieces of advice for and against accepting a job at the same school where he was studying, until he finally decided that the position at MIT was too tempting to turn down.

“MIT is a unique place in terms of resources and fluidity between departments. MIT seemed to be doing a really good job supporting the intersection of AI and science, and it was a vibrant ecosystem worth staying in,” he says. “The caliber of the students, their enthusiasm and just the incredible power of collaboration far outweighed any potential concerns about staying in the same place.”

Chemical intuition

Coley deferred his teaching position for a year to pursue a postdoctoral fellowship at the Broad Institute, where he sought more experience in chemical biology and drug discovery. There, he worked on ways to identify diminutive molecules from the billions of candidates in DNA-encoded libraries that might have binding interactions with mutant disease-related proteins.

Upon returning to MIT in 2020, he built a lab group whose mission was to exploit artificial intelligence not only to synthesize existing compounds with therapeutic potential, but also to design recent molecules with desired properties and recent ways of producing them. Over the past few years, his lab has developed a variety of computational approaches to achieve these goals.

“We try to think about how best to combine a challenge in chemistry with a potential computational solution. Often this combination motivates the development of new methods,” says Coley. One of the models developed in his lab, known as ShEPhERD, was trained to evaluate potential recent drug molecules based on their interactions with target proteins, based on the three-dimensional shapes of the drug molecules. This model is currently used by pharmaceutical companies to support them discover recent drugs.

“We’re trying to give the generative model more of a medicinal chemistry intuition so that the model is aware of the right criteria and considerations,” Coley says.

In another project, Coley’s lab developed a generative artificial intelligence model called FlowER that can be used to predict the reaction products that result from combining different chemicals.

To design this model, scientists used knowledge of basic physical principles, such as the law of conservation of mass. They also forced the feasibility model to include intermediate steps that must take place along the path from reactants to products. The researchers found that these constraints improved the accuracy of the model’s predictions.

“Thinking about these intermediate steps, the mechanisms involved, and how reactions evolve is something that chemists do very naturally. That’s how you learn chemistry, but it’s not something that models inherently think about,” Coley says. “We spent a lot of time thinking about how to ensure that our machine learning models are based on an understanding of reaction mechanisms, in the same way an experienced chemist would.”

Students in his lab also work on a wide variety of areas related to chemical reaction optimization, including computer-assisted structure elucidation, laboratory automation, and optimal experimental design.

“Through many different research strands, we hope to expand the boundaries of artificial intelligence in chemistry,” says Coley.

The post Creating artificial intelligence models that understand chemical principles appeared first on AI SCKOOL.

Justin Solomon, associate professor in MIT’s Department of Electrical Engineering and Computer Science (EECS), has been named associate dean for engineering education at the MIT School of Engineering, effective July 1.

In this modern role, Solomon will focus on supporting innovation in engineering education across the school. It will facilitate shape modern pedagogical approaches in the context of an AI-enabled world and explore experiential, hands-on and other ways of learning. Working closely with academic departments, Solomon will serve as a thought partner in integrating AI into curricula and will facilitate create interdisciplinary and collaborative learning opportunities between departments and other schools. He will also play a key role in helping the school implement relevant recommendations from the Commission on the apply of AI in teaching, learning and research training.

Solomon will explore opportunities to build industry collaborations, including modern models of on-campus internships and industry-led learning. Working with department chairs and the School of Engineering leadership team, he will also support faculty in designing modern courses and developing existing programs to meet emerging opportunities in engineering.

“Justin’s interdisciplinary approach will be especially valuable as we continue to advance engineering education to meet new opportunities and challenges. His extensive experience applying artificial intelligence across a wide range of fields will help every academic department thoughtfully integrate artificial intelligence and new educational models into their curricula,” says Paula T. Hammond, dean of the School of Engineering and professor at the institute. “I look forward to the vision and perspective he will bring to the school’s leadership team.”

As a dedicated educator, Solomon played a key role in shaping computer science education at MIT. He is a key contributor to Common Ground for Computing, where he co-teaches Primary 6.C01 (Modeling with machine learning: from algorithms to applications) with Regina Barzilay, Delta Electronics Professor in MIT’s Department of Electrical Engineering and Computer Science and faculty member of the Institute for Medical Engineering and Science. Within EECS, he teaches 6.7350 (numerical algorithms in computer science and machine learning) and 6.8410 (shape analysis). He is also the founder Summer Geometric Initiativea six-week program that introduces students to geometry processing through intensive training, collaboration, and research experiences.

Solomon’s commitment to teaching and helping students was recognized with various awards, including the EECS Outstanding Educator Award and the Burgess Award (1952) and the Elizabeth Jamieson Award for Excellence in Teaching. He is the author of “Numerical Algorithms” – a textbook presenting a newfangled approach to numerical analysis for computer science students.

Solomon is a principal investigator at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), where he leads the Geometric Data Processing Group. His research focuses on the intersection of geometry and computation, with applications including computer graphics, autonomous navigation, political boundary transition, physical simulation, 3D modeling, and medical imaging. He is also a core faculty member of the MIT-IBM Watson AI Lab, contributing to research on the fundamentals and applications of artificial intelligence.

His scholarly contributions have been recognized with numerous honors, including the 2023 Harold E. Edgerton Faculty Achievement Award for outstanding contributions to teaching, research and service. In 2025 he received the title Schmidt Polymathsupporting interdisciplinary research in areas such as acoustics and climate that relies on large-scale simulations of physical systems.

Solomon joined the MIT faculty in 2016. Previously, he was the recipient of an NSF Mathematical Sciences Postdoctoral Research Fellowship at Princeton University’s Applied and Computational Mathematics program. He received bachelor’s, master’s and doctorate degrees from Stanford University. While attending Stanford University, he also worked as a research assistant at Pixar Animation Studios.

The post Justin Solomon named associate dean for engineering education appeared first on AI SCKOOL.

The post Universal AI is “a path to AI fluidity that is accessible and accessible to anyone, anywhere.” appeared first on AI SCKOOL.

Gabriele Farina grew up in a compact town in the hilly wine region of northern Italy. Neither of his parents had a college degree, and although they both believed they “didn’t understand math,” Farina says, they bought him the technical books he asked for and didn’t discourage him from attending a science-oriented high school rather than a classical high school.

Around the age of 14, Farina focused on an idea that would prove to be the foundation of his career.

“I was fascinated very early on by the idea that a machine could predict and make decisions much better than humans,” he says. “The fact that human-made mathematics and algorithms can create systems that are in some ways superior to their creators, while still relying on simple elements, has always been a major source of awe for me.”

At age 16, Farina wrote code for a board game he played with his 13-year-old sister.

“I used game by game to calculate the optimal move and prove to my sister that she had already lost long before either of us could see it with our own eyes,” Farina says, adding that his sister was less thrilled with his modern system.

Currently, as an assistant professor in MIT’s Department of Electrical Engineering and Computer Science (EECS) and principal investigator in the Laboratory for Information and Decision Systems (LIDS), Farina combines concepts from game theory with tools such as machine learning, optimization, and statistics to advance the theoretical and algorithmic foundations of decision making.

Starting her studies at the Politecnico di Milano, Farina studied automation and control engineering. Over time, however, he realized that what sparked his interest was not “just applying known techniques, but understanding and expanding their fundamentals,” he says. “Gradually I turned more and more towards theory, but I was still very concerned with demonstrating concrete applications of that theory.”

Farina’s advisor at the Politecnico di Milano, Nicola Gatti, a professor and researcher in computer science and engineering, introduced Farina to research topics in computational game theory and encouraged him to pursue a Ph.D. The first in his immediate family to earn a college degree and living in Italy, where doctorates are treated differently, Farina says he didn’t even know what a doctorate was.

Nevertheless, a month after completing his bachelor’s degree, Farina began a PhD program in computer science at Carnegie Mellon University. There, he won honors for his research and dissertation, as well as a Facebook Fellowship in Economics and Computation.

While completing her PhD, Farina worked for a year as a research associate in Meta’s basic artificial intelligence research labs. One of his major projects was helping develop Cicero, an artificial intelligence that could beat human players in a game of forming alliances, negotiating, and detecting when other players were bluffing.

Farina says, “When we built Cicero, we designed him so that he wouldn’t agree to an alliance if it wasn’t in his interest, and similarly understood if a player was likely to lie because following through on his proposal would go against his own incentives.”

A 2022 paper in the aforementioned Cicero could mark progress toward artificial intelligence that can solve convoluted problems that require trade-offs.

After a year at Meta, Farina joined the MIT faculty. In 2025, he was awarded the CAREER Award of the National Science Fund. His work – based on game theory and its mathematical language of describing what happens when different parties have different goals and then quantifying an “equilibrium” where no one has a reason to change their strategy – aims to simplify huge, convoluted real-world scenarios in which such an equilibrium could take a billion years to calculate.

“I’m researching how we can use optimization and algorithms to actually find these stable points effectively,” he says. “Our work attempts to shed new light on the mathematical foundations of the theory, provide better control and prediction of these complex dynamic systems, and uses these ideas to compute good solutions to large interactions between multiple agents.”

Farina is particularly interested in environments with “imperfect information”, meaning that some agents have information unknown to other participants. In such scenarios, information has value, and participants must strategically approach the information they have so as not to reveal it and reduce its value. An everyday example occurs in the game of poker, where players bluff to hide information about their cards.

According to Farina, “we currently live in a world where machines are much better at bluffing than humans.”

The “enormous amount of imperfect information” situation took Farina back to the origins of the board game. Stratego is a military strategy game that has inspired a research effort that has cost millions of dollars to create systems that can defeat human players. Requiring a convoluted calculation of risk and misrepresentation or bluffing, it was perhaps the only classic game where the greatest efforts did not produce superhuman results, Farina says.

Thanks to modern algorithms and training costing less than $10,000 rather than millions, Farina and his research team were able to beat the greatest player of all time – by 15 wins, four draws and one loss. Farina says he’s excited to be able to obtain such results in such a cost-effective way and hopes “these new techniques will be incorporated into future pipelines,” he says.

“We’ve seen continued progress in building algorithms that can reason strategically and make sound decisions despite large operating spaces or imperfect information. I’m excited about the opportunity to incorporate these algorithms into the broader AI revolution happening around us.”

The post Games humans—and machines—play: Unraveling strategic reasoning to advance artificial intelligence appeared first on AI SCKOOL.

The human brain remains one of medicine’s most fascinating and perplexing mysteries. Scientists still struggle to match neurological activity to brain function and detect problems early, slowing efforts to treat neurological disorders and other diseases.

Beacon Biosignals works to understand the brain by monitoring its activity during sleep. The company, founded by Dr. Jake Donoghue ’19 and former MIT researcher Jarrett Revels, has developed a lightweight headband that uses electroencephalogram (EEG) technology to measure brain activity while people sleep normally at home. This data is processed by machine learning algorithms to monitor the effects of novel therapies, find up-to-date signs of disease progression, and create patient cohorts for clinical trials.

“When you remove the sleep lab and bring clinical-grade EEG into the home, there’s a dramatic change,” says Donoghue, who serves as Beacon’s CEO. “It transforms sleep from a limited facility-based test to a scalable source of high-quality data for diagnostics, drug development and long-term assessment of brain health.”

Beacon partners with pharmaceutical companies to accelerate its path to patients. The company’s FDA 510(k)-cleared medical device has already been used in more than 40 clinical trials around the world in studies to treat conditions such as major depressive disorder, schizophrenia, narcolepsy, idiopathic hypersomnia, Alzheimer’s disease and Parkinson’s disease.

With each implementation, Beacon learns more about how the brain works – gaining knowledge that it uses to create a “base model” of the brain.

“We believe the dataset that will transform brain health does not exist yet, but we are building it quickly,” Donoghue says. “Our platform can characterize the heterogeneity of disease progression, generating dynamic insights that cannot be fully captured by static methods such as sequencing or imaging. The brain is an electrical organ and changes through synaptic plasticity, so tracking brain function across multiple diseases at scale will allow us to discover new disease subsets and map them over time.”

Brain lighting

Donoghue was educated in the Harvard-MIT Health Sciences and Technology Program, earning a doctorate in neuroscience from MIT under the supervision of Professor of Cognitive Sciences Earl K. Miller, and also completed clinical training as a physician. During the program, Donoghue trained at Massachusetts General Hospital and Boston Children’s Hospital, where he helped care for patients, including oncologists, during the development of genome sequencing to guide precision cancer therapies. He later worked in neurology and psychiatry, where care often relied on more iterative approaches, highlighting the ability to provide similar data-driven precision in brain health.

“What struck me most was the inability to measure brain function in the way cardiologists can monitor heart function in patients at home,” Donoghue says. “At MIT, I built a belief that processing large amounts of brain data and working to link it to brain function would impact how we identify and treat these neurological diseases.”

Toward the end of his training, Donoghue began to develop his ideas further, working with mentors including HST and Harvard Medical School professors Sydney Cash and Brandon Westover. He met Revels, who was working as a research software engineer at MIT’s Julia Lab during his Ph.D., and convinced him to start Beacon with him in 2019.

“We decided that starting a company to understand an organ of interest — the brain — would be a great start in understanding heterogeneous neuropsychiatric diseases and developing better treatments,” Donoghue recalls.

Beacon started as a computing and analytics company building wearable devices to expand clinical impact and reach. Since its inception, Beacon has partnered with huge pharmaceutical companies conducting clinical trials to offer a less invasive way to monitor brain activity and gain insight into how their drugs affect the brain as well as patients’ sleep.

“It was clear that sleep was the right window into understanding the brain,” Donoghue says. “Neural activity during sleep can be orders of magnitude greater and more structured, almost like language. This is a great surface for understanding how the brain functions and the effects of different drugs on the brain.”

Donoghue says Beacon devices can collect laboratory data on each patient for multiple consecutive nights, ensuring higher quality assessments. The company uses machine learning to draw conclusions such as the time patients spend in different sleep stages and the number of compact awakenings that occur throughout the night. It can also detect subtle changes in sleep architecture that can lead to cognitive decline.

“We are starting to analyze the characteristics of sleep activity and link them to outcomes in a way that has never been done before with such precision,” Donoghue says.

To date, Beacon has participated in clinical trials for sleep and mental health disorders, as well as neurodegenerative diseases, where sleep changes can occur years before symptoms appear.

“We work a lot on diseases like Alzheimer’s and Parkinson’s, which affected my grandfather,” Donoghue says. “We are analyzing the characteristics of rapid eye movement and slow-wave sleep to detect early changes that precede clinical symptoms. This is an opportunity to move these diseases from late diagnosis to much earlier data-driven detection.”

Improving brain therapies for millions

Last year, Beacon acquired an at-home sleep apnea testing company that serves more than 100,000 patients annually across the U.S., accelerating access to high-quality, comprehensive at-home testing and expanding the reach of its platform. The company then raised $97 million in November to accelerate that expansion.

“The vision has always been to reach patients and help them at scale,” Donoghue says. “The most important thing is that we are creating a longitudinal record of how the brain functions over time,” Donoghue says. “A patient can come in for a sleep apnea screening, but if they develop Parkinson’s disease years later, the earlier data provides insight into the disease before symptoms appear. This makes routine testing the basis for entirely new predictive biomarkers and a path to detecting and preventing brain diseases earlier, potentially before symptoms even appear.”

The post Beacon Biosignals maps the brain while you sleep appeared first on AI SCKOOL.