MIT’s delta v accelerator receives $6M gift to supercharge startups being built by student founders

With the impact artificial intelligence is having on how companies operate, the environment for how MIT students are learning entrepreneurship and choosing to create new ventures is seeing rapid changes as well. To address how these student startups are being built, the Martin Trust Center for MIT Entrepreneurship undertook a months-long series of discussions with key stakeholders to help shape a new direction for delta v, MIT’s capstone entrepreneurship accelerator for student founders.

Two of Boston’s most successful tech entrepreneurs have stepped forward to fund this growth of new MIT ventures through a combined $6 million gift that supports the delta v accelerator run out of the Trust Center. Ed Hallen MBA ’12 and Andrew Bialecki, co-founders of Boston-based customer relationship management firm Klaviyo, are providing the donation to support the next wave of innovation-driven entrepreneurship taking place at MIT.

“In the early days of Klaviyo, we learned almost everything by building, testing assumptions, making mistakes, and figuring things out as we went,” Hallen says. “MIT delta v creates that same learning-by-doing environment for students, while surrounding them with mentorship and resources that help founders build with clarity and momentum. We’ve seen the difference delta v can make for founders, and we’re excited to help the Trust Center extend that opportunity to the next generation of students.”

“We’ve always believed the world needs more entrepreneurs, and that Boston should be one of the places leading the way,” adds Bialecki. “Boston is a hub of innovation with ambitious students and a strong community of builders. MIT delta v plays a critical role in developing founders early, not just helping them start companies but helping them build companies that last. Supporting that mission is something Ed and I care deeply about.”

The Martin Trust Center plans to “accelerate the accelerator” with the funding. Recognizing the opportunity that exists as AI impacts how students are able to build companies, along with the increased interest being shown by students to learn about entrepreneurship during their time on campus, is a major driver for these changes. One of the main impacts will be the ability of delta v participants to earn up to $75,000 in equity-free funding during the program, an increase from $20,000 in years past. 

Also, delta v will be introducing a partner model composed of leading founders from companies such as HubSpot, Okta, and Kayak, C-suite operators, subject matter experts, and early-stage investors who will all be providing significant guidance and mentorship to the student ventures.

“Core to MIT’s mission is developing the innovative technologies and solutions that can help solve tough problems at global scale,” says MIT Provost Anantha Chandrakasan. “The AI revolution is creating exciting new opportunities for MIT students to build the next wave of impactful companies, and the delta v accelerator is a perfect vehicle to help them make that happen.”

In recent years MIT-founded startups such as Cursor and Delve who use AI as a core part of their business have seen explosive growth in both customers and revenue as well as valuation. In addition, delta v alumni entrepreneurs and their companies such as Klarity and Reducto are providing software-as-a-service (SaaS) platforms using AI tools while Vertical Semiconductor is growing thanks to providing the energy solutions that data centers need to power today’s computing demands. These are just some of the businesses MIT students are looking to as models they can follow to build and launch successfully, whether they are working on solutions in health care, climate, finance, the future of work, or another global challenge.

“MIT Sloan is the place for entrepreneurship education, part of a unique ecosystem of collaboration across MIT to solve problems," says Richard M. Locke, the John C Head III Dean at the MIT Sloan School of Management. “The delta v program is a great example of how MIT students dedicate their energy to starting a venture, connect with mentors, and incorporate proven frameworks for disciplined entrepreneurship. This gift from Ed Hallen and Andrew Bialecki will provide additional funding for this important program, and I’m so grateful for their support of entrepreneurship education at MIT.” 

“I remember when Ed and Andrew were giving birth to Klaviyo at the Trust Center,” says Bill Aulet, the Ethernet Inventors Professor of the Practice and managing director of the Trust Center. “Through their ingenuity and drive, they have created an iconic tech company here in Boston with the support of our ecosystem. Through their willingness to give back, many more students will now be able to follow their path and become entrepreneurs who can create extraordinary positive impact in the world.”

Applications for the next delta v cohort will open on March 1 and close on April 1. Teams will be announced in May for the summer 2026 accelerator.

“MIT delta v is about creating belief in our most exceptional entrepreneurial talent — and turning that belief into consequential impact for the world. By supporting early-stage founders who take bold ideas from improbable to possible, we help them build companies that matter,” says Ana Bakshi, the Trust Center’s executive director. “Our students are the next generation of job creators, economic drivers, and thought leaders. To realize this potential, it is critical that we continue to invest in and scale startup programs and spaces so they can build at unprecedented levels. Ed and Andrew’s generosity gives us a powerful opportunity to change velocity—and make that future possible.”

Founded in 1991, the award-winning Martin Trust Center for MIT Entrepreneurship is today focused on teaching entrepreneurship as a craft. It combines evidence-based entrepreneurship frameworks, used in over a thousand other organizations, with experiential learning, experiences, and community building inside and outside the classroom to create the next generation of innovation-driven entrepreneurs. Alumni who have gone through Trust Center programs have started companies including Cursor, Delve, Okta, HubSpot, PillPack, Honey, WHOOP, Reducto, Klarity, and Biobot Analytics, and thousands more in industries as diverse as biotech, climate and energy, AI, health care, fintech, business and consumer software, and more. 

In the first 10 years of delta v, the program's alumni have helped create entrepreneurs who have gone on to experience extraordinary success. The five-year survival rate of their companies has been 69%, and they have raised well over $3 billion in funding while addressing the world’s greatest challenges — evidenced by the fact that 89% are directly aligned with the UN Sustainable Development goals.

de MIT News https://ift.tt/SgsR4l2

More trees where they matter, please

One of the best forms of heat relief is pretty simple: trees. In cities, as studies have documented, more tree cover lowers surface temperatures and heat-related health risks.

However, as a new study led by MIT researchers shows, the amount of tree cover varies widely within cities, and is generally connected to wealth levels. After examining a cross-section of cities on four continents at different latitudes, the research finds a consistent link between wealth and neighborhood tree abundance within a city, with better-off residents usually enjoying much more shade on nearby sidewalks.

“Shade is the easiest way to counter warm weather,” says Fabio Duarte, an MIT urban studies scholar and co-author of a new paper detailing the study’s results. “Strictly by looking at which areas are shaded, we can tell where rich people and poor people live.”

That disparity is evident within a range of cities, and is present whether a city contains a large amount of tree cover overall or just a little. Either way, there are more trees in wealthier spots.

“When we compare the most well-shaded city in our study, Stockholm, with the worst-shaded, Belem in northern Brazil, we still see marked inequality,” says Duarte, the associate director of MIT’s Senseable City Lab in the Department of Urban Studies and Planning (DUSP). “Even though the most-shaded parts of Belem are less shaded than the least-shaded parts of Stockholm, shade inequality in Stockholm is greater. Rich people in Stockholm have much better shade provison as pedestrians than we see in poor areas of Stockholm.”

The paper, “Global patterns of pedestrian shade inequality,” is published today in Nature Communications. The authors are Xinyue Gu of Hong Kong Polytechnic University; Lukas Beuster, a research fellow at the Amsterdam Institute for Advanced Metropolitan Solutions and MIT’s Senseable City Lab; Xintao Liu, an associate professor at Hong Kong Polytechnic University; Eveline van Leeuwen, scientific director at the Amsterdam Institute for Advanced Metropolitan Solutions; Titus Venverloo, who leads the MIT Senseable City Amsterdam lab; and Duarte, who is also a lecturer in DUSP.

From Stockholm to Sydney

To conduct the study, the researchers used satellite data from multiple sources, along with urban mapping programs and granular economic data about the cities they examined. There are nine cities in the study: Amsterdam, Barcelona, Belem, Boston, Hong Kong, Milan, Rio de Janeiro, Stockholm, and Sydney. Those places are intended to create a cross-section of cities with different characteristics, including latitude, wealth levels, urban form, and more.

The scholars looked at the amount of shade available on city sidewalks on summer solistice day, as well as the hottest recorded day each year from 1991 to 2020. They then created a scale, ranging from 0 to 1, to rate the amount of shade available on sidewalks, both citywide and within neighborhoods.

“We focused on sidewalks because they are a major counduit of urban activity, even on hot summer days,” Gu says. “Adding tree cover for sidewalks is one crucial way cities can pursue heat-reduction measures.”

Duarte adds: “When it comes to those who are not protected by air conditioning, they are also using the city, walking, taking buses, and anybody who takes a bus is walking or biking to or from bus stops. They are using sidewalks as the main infrastructure.”

The cities in the study offer very different levels of tree coverage. On the 0-to-1 scale the researchers developed, much of Stockholm falls in the 0.6-0.9 range, with some neighborhoods being over 0.9. By contrast, large swaths of Rio de Janeiro are under the 0.1 mark. Much of Boston ranges from 0.15 to 0.4, with a few neighborhoods reaching 0.45 on the scale.

The overall pattern of disparities, however, is very consistent, and includes the more affluent cities. The bottom 20 percent of neighborhoods in Stockholm, in terms of shade coverage, are rated at 0.58 on the scale, while the top 20 percent of Belem neighborhoods rate at 0.37; Stockholm has a greater disparity between most-covered and least-covered. To be sure, there is variety within many cities: Milan and Barcelona have some lower-income neighborhoods with abundant shade, for instance. But the aggregate trend is clear. Amsterdam, another well-off place on average, has a distinct pattern of less shade in lower-income areas.

“In rich cities like Amsterdam, even though it’s relatively well-shaded, the disparity is still very high,” Beuster says. “For us the most surprising point was not that in poor cities and more unequal societies the disparity would be notable — that was expected. What was unexpected was how the disparity still happens and is sometimes more pronounced in rich countries.”

“Follow transit”

If the tree-shade disparity issue is quite persistent, then it raises the matter of what to do about it. The researchers have a basic answer: Add trees in areas with public transit, which generate a lot of pedestrian mileage.

“In each city, from Sydney to Rio to Amsterdam, there are people who, regardless of the weather, need to walk,” Duarte says. “And it’s those people who also take public transportation. Therefore, link a tree-planting scheme to a public transportation network. And secondly, they are also the medium-and low-income part of the population. So the action deriving from this result is quite clear: If you need to increase your tree coverage and don’t know where, follow transit. If you follow transit, you will have the right shading.”

Indeed, one takeaway from the study is to think of trees not just as a nice-to-have part of urban aesthetics, but in functional terms.

“Planners and city officials should think about tree placement at least partly in terms of the heat-mitigating effect they have,” Beuster says.

“It’s not just about planting trees,” Duarte observes. “It’s about providing shade by planting trees. If you remove a tree that’s providing shade in a pedestrian area and you plant two other trees in a park, you are still removing part of the public function of the tree.”

He adds: “With increasing temperatures, providing shade is an essential public amenity. Along with providing transportation, I think providing shade in pedestrian spaces should almost be a public right.”

The Amsterdam Institute for Advanced Metropolitan Solutions and all members of the MIT Senseable City Consortium (including FAE Technology, Dubai Foundation, Sondotécnica, Seoul AI Foundation, Arnold Ventures, Sidara, Toyota, Abu Dhabi’s Department of Municipal Transportation, A2A, UnipolTech, Consiglio per la Ricerca in Agricoltura e l’Analisi dell’Economia Agraria, Hospital Israelita Albert Einstein, KACST, KAIST, and the cities of Laval, Amsterdam, and Rio de Janeiro) supported the research.

de MIT News https://ift.tt/Xu4IQJj

Study reveals climatic fingerprints of wildfires and volcanic eruptions

Volcanoes and wildfires can inject millions of tons of gases and aerosol particles into the air, affecting temperatures on a global scale. But picking out the specific impact of individual events against a background of many contributing factors is like listening for one person’s voice from across a crowded concourse.

MIT scientists now have a way to quiet the noise and identify the specific signal of wildfires and volcanic eruptions, including their effects on Earth’s global atmospheric temperatures.

In a study appearing this week in the Proceedings of the National Academy of Sciences, the researchers report that they detected statistically significant changes in global atmospheric temperatures in response to three major natural events: the eruption of Mount Pinatubo in 1991, the Australian wildfires in 2019-2020, and the eruption of the underwater volcano Hunga Tonga in the South Pacific in 2022.

While the specifics of each event differed, all three events appeared to significantly affect temperatures in the stratosphere. The stratosphere lies above the troposphere, which is the lowest layer of the atmosphere, closest to the surface, where global warming has accelerated in recent years. In the new study, Pinatubo showed the classic pattern of stratospheric warming paired with tropospheric cooling. The Australian wildfires and the Hunga Tonga eruption also showed significant warming or cooling in the stratosphere, respectively, but they did not produce a robust, globally detectable tropospheric signal over the first two years following each event. This new understanding will help scientists further pin down the effect of human-related emissions on global temperature change.

“Understanding the climate responses to natural forcings is essential for us to interpret anthropogenic climate change,” says study author Yaowei Li, a former postdoc and currently a visiting scientist in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS). “Unlike the global tropospheric and surface cooling caused by Pinatubo, our results also indicate that the Australian wildfires and Hunga Tonga eruption may not have played a role in the acceleration of global surface warming in recent years. So, there must be some other factors.”

The study’s co-authors include Susan Solomon, the Lee and Geraldine Martin Professor of Environmental Studies and Chemistry at MIT, along with Benjamin Santer of the University of East Anglia, David Thompson of the University of East Anglia and Colorado State University, and Qiang Fu of the University of Washington.

Extraordinary events

The past several years have set back-to-back records for global average surface temperatures. The World Meteorological Organization recently confirmed that the years 2023 to 2025 were the three warmest years on record, while the past 11 years have been the 11 warmest years ever recorded. The world is warming, due mainly to human activities that have emitted huge amounts of greenhouse gases into the atmosphere over centuries.

In addition to greenhouse gases, the atmosphere has been on the receiving end of other large-scale emissions, including sulfur gases and water vapor from volcanic eruptions and smoke particles from wildfires. Li and his colleagues have wondered whether such natural events could have any global impact on temperatures, and whether such an effect would be detectable.

“These events are extraordinary and very unique in terms of the different materials they inject into different altitudes,” Li says. “So we asked the question: Do these events actually perturb the global temperature to a degree that could be identifiable from natural, meteorological noise, and could they contribute to some of the exceptional global surface warming we’ve seen in the last few years?”

In particular, the team looked for signals of global temperature change in response to three large-scale natural events. The Pinatubo eruption resulted in around 20 million tons of volcanic aerosols in the stratosphere, which was the largest volume ever recorded by modern satellite instruments. The Australian fires injected around 1 million tons of smoke particles into the upper troposphere and stratosphere. And the Hunga Tonga eruption produced the largest atmospheric explosion on satellite record, launching nearly 150 million tons of water vapor into the stratosphere.

If any natural event could measurably shift global temperatures, the team reasoned, it would be any of these three.

Natural signals

For their new study, the team took a signal-to-noise approach. They looked to minimize “noise” from other known influences on global temperatures in order to isolate the “signal,” such as a change in temperature associated specifically with one of the three natural events.

To do so, they looked first through satellite measurements taken by the Stratospheric Sounding Unit (SSU) and the Microwave and Advanced Microwave Sounding Units (MSU), which have been measuring global temperatures at different altitudes throughout the atmosphere since 1979. The team compiled SSU and MSU measurements from 1986 to the present day. From these measurements, the researchers could see long-term trends of steady tropospheric warming and stratospheric cooling. Those long-term trends are largely associated with anthropogenic greenhouse gases, which the team subtracted from the dataset.

What was left over was more of a level baseline, which still contained some confounding noise, in the form of natural variability. Global temperature changes can also be affected by phenomena such as El Niño and La Niña, which naturally warm and cool the Earth every few years. The sun also swings global temperatures on a roughly 11-year cycle. The team took this natural variability into account, and subtracted out the effects of these influences.

After minimizing such noise from their dataset, the team reasoned that whatever temperature changes remained could be more easily traced to the three large-scale natural events and quantified. And indeed, when they pinned the events to the temperature measurements, at the times that they occurred, they could plainly see how each event influenced temperatures around the world.

The team found that Pinatubo decreased global tropospheric temperatures by up to about 0.7 degree Celsius, for more than two years following the eruption. The volcanic sulfate aerosols essentially acted as many tiny reflectors, cooling the troposphere and surface by scattering sunlight back into space. At the same time, the aerosols, which remained in the stratosphere, also absorbed heat that was emitted from the surface, subsequently warming the stratosphere.

This finding agreed with many other studies of the event, which confirmed that the team’s approach is accurate. They applied the same method to the 2019-2020 Australian wildfires, and the 2022 underwater eruption — events where the influence on global temperatures is less clear.

For the Australian wildfires, they found that the smoke particles caused the global stratosphere to warm up, by up to about 0.77 degree Celsius, which persisted for about five months but did not produce a clear global tropospheric signal.

“In the end we found that the wildfire smoke caused a very strong warming in the stratosphere, because these materials are very different chemically from sulfate,” Li explains. “They are particles that are dark colored, meaning they are efficient at absorbing solar radiation. So, a relatively small amount of smoke particles can cause a dramatic warming.”

In the case of the Hunga Tonga, the underwater eruption triggered a global cooling effect in the middle-to-upper stratosphere, of up to about half a degree Celsius, lasting for several years.

“The Australian fires and the Hunga Tonga really packed a punch at stratospheric altitudes, and this study shows for the first time how to quantify how strong that punch was,” says Solomon. “I find their impact up high quite remarkable, but the ongoing issue is why the last several years have been so warm lower down, in the troposphere — ruling out those natural events points even more strongly at human influences.”

de MIT News https://ift.tt/8UcqRCl

Fragile X study uncovers brain wave biomarker bridging humans and mice

Numerous potential treatments for neurological conditions, including autism spectrum disorders, have worked well in mice but then disappointed in humans. What would help is a non-invasive, objective readout of treatment efficacy that is shared in both species. 

In a new open-access study in Nature Communications, a team of MIT researchers, backed by collaborators across the United States and in the United Kingdom, identifies such a biomarker in fragile X syndrome, the most common inherited form of autism.

Led by postdoc Sara Kornfeld-Sylla and Picower Professor Mark Bear, the team measured the brain waves of human boys and men, with or without fragile X syndrome, and comparably aged male mice, with or without the genetic alteration that models the disorder. The novel approach Kornfeld-Sylla used for analysis enabled her to uncover specific and robust patterns of differences in low-frequency brain waves between typical and fragile X brains shared between species at each age range. In further experiments, the researchers related the brain waves to specific inhibitory neural activity in the mice and showed that the biomarker was able to indicate the effects of even single doses of a candidate treatment for fragile X called arbaclofen, which enhances inhibition in the brain.

Both Kornfeld-Sylla and Bear praised and thanked colleagues at Boston Children’s Hospital, the Phelan-McDermid Syndrome Foundation, Cincinnati Children’s Hospital, the University of Oklahoma, and King’s College London for gathering and sharing data for the study.

“This research weaves together these different datasets and finds the connection between the brain wave activity that’s happening in fragile X humans that is different from typically developed humans, and in the fragile X mouse model that is different than the ‘wild-type’ mice,” says Kornfeld-Sylla, who earned her PhD in Bear’s lab in 2024 and continued the research as a FRAXA postdoc. “The cross-species connection and the collaboration really makes this paper exciting.”

Bear, a faculty member in The Picower Institute for Learning and Memory and the Department of Brain and Cognitive Sciences at MIT, says having a way to directly compare brain waves can advance treatment studies.

“Because that is something we can measure in mice and humans minimally invasively, you can pose the question: If drug treatment X affects this signature in the mouse, at what dose does that same drug treatment change that same signature in a human?” Bear says. “Then you have a mapping of physiological effects onto measures of behavior. And the mapping can go both ways.”

Peaks and powers

In the study, the researchers measured EEG over the occipital lobe of humans and on the surface of the visual cortex of the mice. They measured power across the frequency spectrum, replicating previous reports of altered low-frequency brain waves in adult humans with fragile X and showing for the first time how these disruptions differ in children with fragile X.

To enable comparisons with mice, Kornfeld-Sylla subtracted out background activity to specifically isolate only “periodic” fluctuations in power (i.e., the brain waves) at each frequency. She also disregarded the typical way brain waves are grouped by frequency (into distinct bands with Greek letter designations delta, theta, alpha, beta, and gamma) so that she could simply juxtapose the periodic power spectra of the humans and mice without trying to match them band by band (e.g., trying to compare the mouse “alpha” band to the human one). This turned out to be crucial because the significant, similar patterns exhibited by the mice actually occurred in a different low-frequency band than in the humans (theta vs. alpha). Both species also had alterations in higher-frequency bands in fragile X, but Kornfeld-Sylla noted that the differences in the low-frequency brainwaves are easier to measure and more reliable in humans, making them a more promising biomarker.

So what patterns constitute the biomarker? In adult men and mice alike, a peak in the power of low-frequency waves is shifted to a significantly slower frequency in fragile X cases compared to in neurotypical cases. Meanwhile, in fragile X boys and juvenile mice, while the peak is somewhat shifted to a slower frequency, what is really significant is a reduced power in that same peak.

The researchers were also able to discern that the peak in question is actually made of two distinct subpeaks, and that the lower-frequency subpeak is the one that varies specifically with fragile X syndrome.

Curious about the neural activity underlying the measurements, the researchers engaged in experiments in which they turned off activity of two different kinds of inhibitory neurons that are known to help produce and shape brain wave patterns: somatostatin-expressing and parvalbumin-expressing interneurons. Manipulating the somatostatin neurons specifically affected the lower-frequency subpeak that contained the newly discovered biomarker in fragile X model mice.

Drug testing

Somatostatin interneurons exert their effects on the neurons they connect to via the neurotransmitter chemical GABA, and evidence from prior studies suggest that GABA receptivity is reduced in fragile X syndrome. A therapeutic approach pioneered by Bear and others has been to give the drug arbaclofen, which enhances GABA activity. In the new study, the researchers treated both control and fragile X model mice with arbaclofen to see how it affected the low-frequency biomarker.

Even the lowest administered single dose made a significant difference in the neurotypical mice, which is consistent with those mice having normal GABA responsiveness. Fragile X mice needed a higher dose, but after one was administered, there was a notable increase in the power of the key subpeak, reducing the deficit exhibited by juvenile mice.

The arbaclofen experiments therefore demonstrated that the biomarker provides a significant readout of an underlying pathophysiology of fragile X: the reduced GABA responsiveness. Bear also noted that it helped to identify a dose at which arbaclofen exerted a corrective effect, even though the drug was only administered acutely, rather than chronically. An arbaclofen therapy would, of course, be given over a long time frame, not just once.

“This is a proof of concept that a drug treatment could move this phenotype acutely in a direction that makes it closer to wild-type,” Bear says. “This effort reveals that we have readouts that can be sensitive to drug treatments.”

Meanwhile, Kornfeld-Sylla notes, there is a broad spectrum of brain disorders in which human patients exhibit significant differences in low-frequency (alpha) brain waves compared to neurotypical peers.

“Disruptions akin to the biomarker we found in this fragile X study might prove to be evident in mouse models of those other disorders, too,” she says. “Identifying this biomarker could broadly impact future translational neuroscience research.”

The paper’s other authors are Cigdem Gelegen, Jordan Norris, Francesca Chaloner, Maia Lee, Michael Khela, Maxwell Heinrich, Peter Finnie, Lauren Ethridge, Craig Erickson, Lauren Schmitt, Sam Cooke, and Carol Wilkinson.

The National Institutes of Health, the National Science Foundation, the FRAXA Foundation, the Pierce Family Fragile X Foundation, the Autism Science Foundation, the Thrasher Research Fund, Harvard University, the Simons Foundation, Wellcome, the Biotechnology and Biological Sciences Research Council, and the Freedom Together Foundation provided support for the research.

de MIT News https://ift.tt/Xni1Tlw

Chip-processing method could assist cryptography schemes to keep data secure

Just like each person has unique fingerprints, every CMOS chip has a distinctive “fingerprint” caused by tiny, random manufacturing variations. Engineers can leverage this unforgeable ID for authentication, to safeguard a device from attackers trying to steal private data.

But these cryptographic schemes typically require secret information about a chip’s fingerprint to be stored on a third-party server. This creates security vulnerabilities and requires additional memory and computation.

To overcome this limitation, MIT engineers developed a manufacturing method that enables secure, fingerprint-based authentication, without the need to store secret information outside the chip.

They split a specially designed chip during fabrication in such a way that each half has an identical, shared fingerprint that is unique to these two chips. Each chip can be used to directly authenticate the other. This low-cost fingerprint fabrication method is compatible with standard CMOS foundry processes and requires no special materials.

The technique could be useful in power-constrained electronic systems with non-interchangeable device pairs, like an ingestible sensor pill and its paired wearable patch that monitor gastrointestinal health conditions. Using a shared fingerprint, the pill and patch can authenticate each other without a device in between to mediate.

“The biggest advantage of this security method is that we don’t need to store any information. All the secrets will always remain safe inside the silicon. This can give a higher level of security. As long as you have this digital key, you can always unlock the door,” says Eunseok Lee, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this security method.

Lee is joined on the paper by EECS graduate students Jaehong Jung and Maitreyi Ashok; as well as co-senior authors Anantha Chandrakasan, MIT provost and the Vannevar Bush Professor of Electrical Engineering and Computer Science, and Ruonan Han, a professor of EECS and a member of the MIT Research Laboratory of Electronics. The research was recently presented at the IEEE International Solid-States Circuits Conference.

“Creation of shared encryption keys in trusted semiconductor foundries could help break the tradeoffs between being more secure and more convenient to use for protection of data transmission,” Han says. “This work, which is digital-based, is still a preliminary trial in this direction; we are exploring how more complex, analog-based secrecy can be duplicated — and only duplicated once.”

Leveraging variations

Even though they are intended to be identical, each CMOS chip is slightly different due to unavoidable microscopic variations during fabrication. These randomizations give each chip a unique identifier, known as a physical unclonable function (PUF), that is nearly impossible to replicate.

A chip’s PUF can be used to provide security just like the human fingerprint identification system on a laptop or door panel.

For authentication, a server sends a request to the device, which responds with a secret key based on its unique physical structure. If the key matches an expected value, the server authenticates the device.

But the PUF authentication data must be registered and stored in a server for access later, creating a potential security vulnerability.

“If we don’t need to store information on these unique randomizations, then the PUF becomes even more secure,” Lee says.

The researchers wanted to accomplish this by developing a matched PUF pair on two chips. One could authenticate the other directly, without the need to store PUF data on third-party servers.

As an analogy, consider a sheet of paper torn in half. The torn edges are random and unique, but the pieces have a shared randomness because they fit back together perfectly along the torn edge.

While CMOS chips aren’t torn in half like paper, many are fabricated at once on a silicon wafer which is diced to separate the individual chips.

By incorporating shared randomness at the edge of two chips before they are diced to separate them, the researchers could create a twin PUF that is unique to these two chips.

“We needed to find a way to do this before the chip leaves the foundry, for added security. Once the fabricated chip enters the supply chain, we won’t know what might happen to it,” Lee explains.

Sharing randomness

To create the twin PUF, the researchers change the properties of a set of transistors fabricated along the edge of two chips, using a process called gate oxide breakdown.

Essentially, they pump high voltage into a pair of transistors by shining light with a low-cost LED until the first transistor breaks down. Because of tiny manufacturing variations, each transistor has a slightly different breakdown time. The researchers can use this unique breakdown state as the basis for a PUF.

To enable a twin PUF, the MIT researchers fabricate two pairs of transistors along the edge of two chips before they are diced to separate them. By connecting the transistors with metal layers, they create paired structures that have correlated breakdown states. In this way, they enable a unique PUF to be shared by each pair of transistors.

After shining LED light to create the PUF, they dice the chips between the transistors so there is one pair on each device, giving each separate chip a shared PUF.

“In our case, transistor breakdown has not been modeled well in many of the simulations we had, so there was a lot of uncertainty about how the process would work. Figuring out all the steps, and the order they needed to happen, to generate this shared randomness is the novelty of this work,” Lee says.

After finetuning their PUF generation process, the researchers developed a prototype pair of twin PUF chips in which the randomization was matched with more than 98 percent reliability. This would ensure the generated PUF key matches consistently, enabling secure authentication.

Because they generated this twin PUF using circuit techniques and low-cost LEDs, the process would be easier to implement at scale than other methods that are more complicated or not compatible with standard CMOS fabrication.

“In the current design, shared randomness generated by transistor breakdown is immediately converted into digital data. Future versions could preserve this shared randomness directly within the transistors, strengthening security at the most fundamental physical level of the chip,” Lee says.

“There is a rapidly increasing demand for physical-layer security for edge devices, such as between medical sensors and devices on a body, which often operate under strict energy constraints. A twin-paired PUF approach enables secure communication between nodes without the burden of heavy protocol overhead, thereby delivering both energy efficiency and strong security. This initial demonstration paves the way for innovative advancements in secure hardware design,” Chandrakasan adds.

This work is funded by Lockheed Martin, the MIT School of Engineering MathWorks Fellowship, and the Korea Foundation for Advanced Studies Fellowship.

de MIT News https://ift.tt/o0VeDxa

Study: AI chatbots provide less-accurate information to vulnerable users

Large language models (LLMs) have been championed as tools that could democratize access to information worldwide, offering knowledge in a user-friendly interface regardless of a person’s background or location. However, new research from MIT’s Center for Constructive Communication (CCC) suggests these artificial intelligence systems may actually perform worse for the very users who could most benefit from them.

A study conducted by researchers at CCC, which is based at the MIT Media Lab, found that state-of-the-art AI chatbots — including OpenAI’s GPT-4, Anthropic’s Claude 3 Opus, and Meta’s Llama 3 — sometimes provide less-accurate and less-truthful responses to users who have lower English proficiency, less formal education, or who originate from outside the United States. The models also refuse to answer questions at higher rates for these users, and in some cases, respond with condescending or patronizing language.

“We were motivated by the prospect of LLMs helping to address inequitable information accessibility worldwide,” says lead author Elinor Poole-Dayan SM ’25, a technical associate in the MIT Sloan School of Management who led the research as a CCC affiliate and master’s student in media arts and sciences. “But that vision cannot become a reality without ensuring that model biases and harmful tendencies are safely mitigated for all users, regardless of language, nationality, or other demographics.”

A paper describing the work, “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users,” was presented at the AAAI Conference on Artificial Intelligence in January.

Systematic underperformance across multiple dimensions

For this research, the team tested how the three LLMs responded to questions from two datasets: TruthfulQA and SciQ. TruthfulQA is designed to measure a model’s truthfulness (by relying on common misconceptions and literal truths about the real world), while SciQ contains science exam questions testing factual accuracy. The researchers prepended short user biographies to each question, varying three traits: education level, English proficiency, and country of origin.

Across all three models and both datasets, the researchers found significant drops in accuracy when questions came from users described as having less formal education or being non-native English speakers. The effects were most pronounced for users at the intersection of these categories: those with less formal education who were also non-native English speakers saw the largest declines in response quality.

The research also examined how country of origin affected model performance. Testing users from the United States, Iran, and China with equivalent educational backgrounds, the researchers found that Claude 3 Opus in particular performed significantly worse for users from Iran on both datasets.

“We see the largest drop in accuracy for the user who is both a non-native English speaker and less educated,” says Jad Kabbara, a research scientist at CCC and a co-author on the paper. “These results show that the negative effects of model behavior with respect to these user traits compound in concerning ways, thus suggesting that such models deployed at scale risk spreading harmful behavior or misinformation downstream to those who are least able to identify it.”

Refusals and condescending language

Perhaps most striking were the differences in how often the models refused to answer questions altogether. For example, Claude 3 Opus refused to answer nearly 11 percent of questions for less educated, non-native English-speaking users — compared to just 3.6 percent for the control condition with no user biography.

When the researchers manually analyzed these refusals, they found that Claude responded with condescending, patronizing, or mocking language 43.7 percent of the time for less-educated users, compared to less than 1 percent for highly educated users. In some cases, the model mimicked broken English or adopted an exaggerated dialect.

The model also refused to provide information on certain topics specifically for less-educated users from Iran or Russia, including questions about nuclear power, anatomy, and historical events — even though it answered the same questions correctly for other users.

“This is another indicator suggesting that the alignment process might incentivize models to withhold information from certain users to avoid potentially misinforming them, although the model clearly knows the correct answer and provides it to other users,” says Kabbara.

Echoes of human bias

The findings mirror documented patterns of human sociocognitive bias. Research in the social sciences has shown that native English speakers often perceive non-native speakers as less educated, intelligent, and competent, regardless of their actual expertise. Similar biased perceptions have been documented among teachers evaluating non-native English-speaking students.

“The value of large language models is evident in their extraordinary uptake by individuals and the massive investment flowing into the technology,” says Deb Roy, professor of media arts and sciences, CCC director, and a co-author on the paper. “This study is a reminder of how important it is to continually assess systematic biases that can quietly slip into these systems, creating unfair harms for certain groups without any of us being fully aware.”

The implications are particularly concerning given that personalization features — like ChatGPT’s Memory, which tracks user information across conversations — are becoming increasingly common. Such features risk differentially treating already-marginalized groups.

“LLMs have been marketed as tools that will foster more equitable access to information and revolutionize personalized learning,” says Poole-Dayan. “But our findings suggest they may actually exacerbate existing inequities by systematically providing misinformation or refusing to answer queries to certain users. The people who may rely on these tools the most could receive subpar, false, or even harmful information.”

de MIT News https://ift.tt/Nri2kad

Exposing biases, moods, personalities, and abstract concepts hidden in large language models

By now, ChatGPT, Claude, and other large language models have accumulated so much human knowledge that they’re far from simple answer-generators; they can also express abstract concepts, such as certain tones, personalities, biases, and moods. However, it’s not obvious exactly how these models represent abstract concepts to begin with from the knowledge they contain.

Now a team from MIT and the University of California San Diego has developed a way to test whether a large language model (LLM) contains hidden biases, personalities, moods, or other abstract concepts. Their method can zero in on connections within a model that encode for a concept of interest. What’s more, the method can then manipulate, or “steer” these connections, to strengthen or weaken the concept in any answer a model is prompted to give.

The team proved their method could quickly root out and steer more than 500 general concepts in some of the largest LLMs used today. For instance, the researchers could home in on a model’s representations for personalities such as “social influencer” and “conspiracy theorist,” and stances such as “fear of marriage” and “fan of Boston.” They could then tune these representations to enhance or minimize the concepts in any answers that a model generates.

In the case of the “conspiracy theorist” concept, the team successfully identified a representation of this concept within one of the largest vision language models available today. When they enhanced the representation, and then prompted the model to explain the origins of the famous “Blue Marble” image of Earth taken from Apollo 17, the model generated an answer with the tone and perspective of a conspiracy theorist.

The team acknowledges there are risks to extracting certain concepts, which they also illustrate (and caution against). Overall, however, they see the new approach as a way to illuminate hidden concepts and potential vulnerabilities in LLMs, that could then be turned up or down to improve a model’s safety or enhance its performance.

“What this really says about LLMs is that they have these concepts in them, but they’re not all actively exposed,” says Adityanarayanan “Adit” Radhakrishnan, assistant professor of mathematics at MIT. “With our method, there’s ways to extract these different concepts and activate them in ways that prompting cannot give you answers to.”

The team published their findings today in a study appearing in the journal Science. The study’s co-authors include Radhakrishnan, Daniel Beaglehole and Mikhail Belkin of UC San Diego, and Enric Boix-Adserà of the University of Pennsylvania.

A fish in a black box

As use of OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and other artificial intelligence assistants has exploded, scientists are racing to understand how models represent certain abstract concepts such as “hallucination” and “deception.” In the context of an LLM, a hallucination is a response that is false or contains misleading information, which the model has “hallucinated,” or constructed erroneously as fact.

To find out whether a concept such as “hallucination” is encoded in an LLM, scientists have often taken an approach of “unsupervised learning” — a type of machine learning in which algorithms broadly trawl through unlabeled representations to find patterns that might relate to a concept such as “hallucination.” But to Radhakrishnan, such an approach can be too broad and computationally expensive.

“It’s like going fishing with a big net, trying to catch one species of fish. You’re gonna get a lot of fish that you have to look through to find the right one,” he says. “Instead, we’re going in with bait for the right species of fish.”

He and his colleagues had previously developed the beginnings of a more targeted approach with a type of predictive modeling algorithm known as a recursive feature machine (RFM). An RFM is designed to directly identify features or patterns within data by leveraging a mathematical mechanism that neural networks — a broad category of AI models that includes LLMs — implicitly use to learn features.

Since the algorithm was an effective, efficient approach for capturing features in general, the team wondered whether they could use it to root out representations of concepts, in LLMs, which are by far the most widely used type of neural network and perhaps the least well-understood.

“We wanted to apply our feature learning algorithms to LLMs to, in a targeted way, discover representations of concepts in these large and complex models,” Radhakrishnan says.

Converging on a concept

The team’s new approach identifies any concept of interest within a LLM and “steers” or guides a model’s response based on this concept. The researchers looked for 512 concepts within five classes: fears (such as of marriage, insects, and even buttons); experts (social influencer, medievalist); moods (boastful, detachedly amused); a preference for locations (Boston, Kuala Lumpur); and personas (Ada Lovelace, Neil deGrasse Tyson).

The researchers then searched for representations of each concept in several of today’s large language and vision models. They did so by training RFMs to recognize numerical patterns in an LLM that could represent a particular concept of interest.

A standard large language model is, broadly, a neural network that takes a natural language prompt, such as “Why is the sky blue?” and divides the prompt into individual words, each of which is encoded mathematically as a list, or vector, of numbers. The model takes these vectors through a series of computational layers, creating matrices of many numbers that, throughout each layer, are used to identify other words that are most likely to be used to respond to the original prompt. Eventually, the layers converge on a set of numbers that is decoded back into text, in the form of a natural language response.

The team’s approach trains RFMs to recognize numerical patterns in an LLM that could be associated with a specific concept. As an example, to see whether an LLM contains any representation of a “conspiracy theorist,” the researchers would first train the algorithm to identify patterns among LLM representations of 100 prompts that are clearly related to conspiracies, and 100 other prompts that are not. In this way, the algorithm would learn patterns associated with the conspiracy theorist concept. Then, the researchers can mathematically modulate the activity of the conspiracy theorist concept by perturbing LLM representations with these identified patterns. 

The method can be applied to search for and manipulate any general concept in an LLM. Among many examples, the researchers identified representations and manipulated an LLM to give answers in the tone and perspective of a “conspiracy theorist.” They also identified and enhanced the concept of “anti-refusal,” and showed that whereas normally, a model would be programmed to refuse certain prompts, it instead answered, for instance giving instructions on how to rob a bank.

Radhakrishnan says the approach can be used to quickly search for and minimize vulnerabilities in LLMs. It can also be used to enhance certain traits, personalities, moods, or preferences, such as emphasizing the concept of “brevity” or “reasoning” in any response an LLM generates. The team has made the method’s underlying code publicly available.

“LLMs clearly have a lot of these abstract concepts stored within them, in some representation,” Radhakrishnan says. “There are ways where, if we understand these representations well enough, we can build highly specialized LLMs that are still safe to use but really effective at certain tasks.”

This work was supported, in part, by the National Science Foundation, the Simons Foundation, the TILOS institute, and the U.S. Office of Naval Research. 

de MIT News https://ift.tt/3K9rJov