AI’s Latest Breakthroughs: What’s New and What’s Next?

Artificial intelligence (AI) continues to grow in importance, thanks to its high potential to revolutionize industries.

This simulation of human intelligence in machines, after all, brings benefits such as increased efficiency and productivity, improved accuracy and decision-making, new job opportunities, and enhanced customer experience.

As a result, AI has become a key driver of corporate growth, with more than 70% of businesses adopting it. Estimates show the AI global market can reach trillions by 2030, driven by increased investment and technological advancements.

Just this year alone, we have seen massive developments: OpenAI introduced a deep research agent with a 25%+ accuracy rate on Humanity's Last Exam benchmark; DeepSeek released DeepSeek-R1, which utilizes chain-of-thought reasoning; Baidu launched Ernie X1 and Ernie 4.5; and Stability AI unveiled a new model, Stable Virtual Camera, that enables photos to be turned into 3D scenes.

Meanwhile, Google is collaborating with Taiwan's MediaTek to develop the next generation of Tensor Processing Units (TPUs) to reduce production costs and reliance on current partners. At the same time, a consortium comprising OpenAI, Oracle, SoftBank, and MGX announced The Stargate Project in collaboration with the US government.

Last week, Nvidia announced a new lineup of “AI personal supercomputers” powered by its Grace Blackwell chip platform. CEO Jensen Huang unveiled DGX Spark and DGX Station at GTC 2025 to allow users to prototype, fine-tune, and run AI models of various sizes at the edge.

So, now, let's take a look at some of the most recent key breakthroughs researchers achieved through AI in different sectors.

Click here to learn if AI is all hype or if there is a substance to its growth.

AI Solving Its Own Learning Challenges

AI is not only transferring other sectors but even contributing to itself by enabling the development of more sophisticated AI systems through advancements in data analysis, machine learning, and automation, leading to faster processing, more accurate predictions and better decision-making. 

This month, researchers introduced novel memristive components that are far more robust, can operate in both analog and digital modes, and function across a wider voltage range. 

These characteristics can help address the “catastrophic forgetting,” which occurs when artificial neural networks forget what they previously learned. This typically happens when deep neural networks (DNNs), designed to mimic the human brain, are trained for a new task, and the new optimization overwrites what was learned before.

Our brain doesn't have this problem, thanks to its ability to adjust the degree of synaptic change. Also, it is suspected that the different degrees of plasticity enable our brain to permanently learn new things without forgetting old ones. Researchers have accomplished something similar with their new memristor.

Memristors behave much like brain cells and consume extremely little power. A type of resistance-switching memory device, memristors can change their resistance based on the applied voltage. Even after the voltage is turned off, their resistance value remains due to their ability to undergo structural changes.

According to Ilia Valov from the Peter Grünberg Institute (PGI-7) at Forschungszentrum Jülich, who led the research:

“Memristive elements are considered ideal candidates for learning-capable, neuro-inspired computer components modeled on the brain.”

While substantial progress has been made, components' commercialization is happening at a much slower pace due to the high failure rate in production, the product's short lifespan, and sensitivity to mechanical influences that can cause malfunctions during operation.

This requires research into better controlling nanoscale processes as well as new materials and switching mechanisms to reduce systems' complexity and increase their range of functionalities, said Valov.

For this very purpose, German and Chinese scientists came together and “discovered a fundamentally new electrochemical memristive mechanism that is chemically and electrically more stable” that will expand the horizon of neuroscience applications.

Click here to learn how AI-enabled brain chips will pioneer the next leap in humanity's evolution.

Memristor Breakthrough for AI's Continual Learning

Published in the journal Nature Communications¹, the study identified two primary mechanisms for the functioning of bipolar memristors:

1. Electrochemical Metallization (ECM)
2. Valence Change Mechanism (VCM)

ECM memristors form a metallic filament between the two electrodes. This tiny “conductive bridge” changes electrical resistance and dissolves upon reversing voltage. While allowing for low switching voltages and fast switching times, the generated states here are variable and short-lived.

VCM memristors change resistance through the movement of oxygen ions at the interface between the electrode and electrolyte by changing the Schottky barrier. While comparatively stable, this requires high switching voltages.

With each process having its own pros and cons, the team designed a memristor by combining the benefits of both types, which wasn't believed to be possible. 

So, the new memristor uses a filament made of metal oxides (instead of being entirely metallic like ECM), which is formed by oxygen and tantalum ions' movement. It is very stable and never fully dissolves. 

“You can think of it as a filament that always exists to some extent and is only chemically modified.” 

– Valov

The new, robust switching mechanism is referred to as a filament conductivity modification mechanism (FCM). Components based on this mechanism are more resistant to high temperatures, need lower voltages to produce, have a wider voltage window, and are chemically and electrically more stable. Their lifespan is also longer because fewer components burn out during production, leading to a lower rejection rate.

The different oxidation states further allow the memristor to operate in both binary and analog modes, which can help overcome the problem of “catastrophic forgetting.”

Valov explained that the new two-terminal ohmic memristor's “unique properties” “allow the use of different switching modes to control the modulation of the memristor in such a way that stored information is not lost.”

Researchers have already implemented the new memristive component in an artificial neural network model in a simulation, achieving high pattern recognition accuracies on multiple image datasets.

The team will now explore other materials that may even be better and more stable. Valov is confident that their “results will further advance the development of electronics for ‘computation-in-memory' applications.”

Now, one of the leading companies in the AI sphere is Microsoft (MSFT +0.53%), which has invested $12 billion in ChatGPT maker OpenAI and offers AI products like Copilot, Azure AI Studio, and AI-powered tools within Microsoft 365, Dynamics 365, and Power Platform. The $2.9 trillion market cap tech giant's shares are currently trading at $394.58, down 7.17% YTD. For the quarter ended December 31, 2024, it reported a 12% increase in revenue to $69.6bln and diluted EPS of $3.23.

According to Microsoft CEO Satya Nadella:

“We are innovating across our tech stack and helping customers unlock the full ROI of AI to capture the massive opportunity ahead. Already, our AI business has surpassed an annual revenue run rate of $13 billion, up 175% year-over-year.”

Latest on Microsoft

AI In Healthcare, Improving Parkinson's Diagnosis

In the healthcare sector, AI is helping reduce costs and achieve better patient outcomes by enhancing diagnostics, treatment, and patient care through disease detection, personalized medicine, and improved efficiency in administrative tasks.

Now, AI is going to help clinicians differentially diagnose Parkinson's disease and related conditions faster and with greater accuracy.

The diagnosis of Parkinson's disease currently stands between 55% and 78% in the first five years of assessment, in part due to its sibling movement disorders sharing similarities with it, which can make definitive diagnosis initially difficult. 

However, a new AI model developed by researchers at the University of Florida and the UF Health Norman Fixel Institute for Neurological Diseases boasts a precision rate beyond 96%.

As the study noted, applying an imaging-based approach to diagnose and differentiate between Parkinson's disease (PD), progressive supranuclear palsy (PSP), and multiple system atrophy (MSA) Parkinsonian variants has been particularly challenging. However, recent data has shown that it is possible with diffusion-weighted MRI paired with disease-specific machine learning (ML) algorithms.

The study, funded by the National Institutes of Health and published in JAMA Neurology², detailed the Automated Imaging Differentiation for Parkinsonism (AIDP), an automated MRI processing and ML software featuring a noninvasive biomarker technique. 

MRI manufacturers, David Vaillancourt, a professor in the UF Department of Applied Physiology and Kinesiology, noted they do not communicate with each other because of marketplace competition and use their own software and sequences. 

So, the researchers “developed novel software that works across all of them.” The tool will help physicians increase their diagnostic efficacy between different disorders, said Vaillancourt.

While diffusion-weighted MRI in AIDP helps identify where neurodegeneration is occurring by measuring how water molecules diffuse in the brain, the algorithm analyzes the brain scan to provide the results. 

The algorithm has been rigorously tested against in-person clinic diagnoses and conducted across 21 sites, which includes two in Canada and 19 in the US.

“This is an instance where the innovation between technology and artificial intelligence has been proven to enhance diagnostic precision, allowing us the opportunity to further improve treatment for patients with Parkinson's disease.”

– Michael Okun, M.D., director of the Norman Fixel Institute for Neurological Diseases at UF Health. 

Researchers are now aiming to get approval from the US Food and Drug Administration (FDA).

Using AI to Predict Cardiovascular Risk through Mammograms

Yet another deep learning model showed potential in predicting cardiovascular risk based on mammogram images, an important cancer screening tool. This separate study presented at the American College of Cardiology's (ACC) Annual Scientific Session combines the power of AI with mammograms to analyze the buildup of calcium in the arteries within breast tissue.

Calcium buildup in blood vessels is a sign of cardiovascular damage associated with aging or early-stage heart disease. Women with calcium buildup in the arteries, according to previous studies, face a 51% higher risk of heart disease and stroke.

Despite being a leading cause of death in the US, heart diseases in women actually remain underdiagnosed, which can change with AI-powered mammogram screening tools that take better advantage of tests many women routinely receive. 

In the US, about 40 million mammograms are performed each year. These X-rays are used to screen for breast cancer. While these images show breast artery calcifications, the information is not reported to patients or their clinicians.

The new study used an AI image analysis technique not previously used on mammograms to show that the technology can help fill this gap by automatically assessing calcification and then providing a cardiovascular risk rating. According to the study's lead author, Theo Dapamede, MD, PhD, and a postdoc fellow at Emory University in Atlanta:

“We see an opportunity for women to get screened for cancer and also additionally get a cardiovascular screen from their mammograms. Our study showed that breast arterial calcification is a good predictor for cardiovascular disease, especially in patients younger than age 60. If we are able to screen and identify these patients early, we can refer them to a cardiologist for further risk assessment.”

To build the tool, researchers trained an AI model to segment calcified vessels in mammogram images. The model then uses these bright pixels to calculate the risk of cardiovascular events in the future based on the data of over 56,000 patients. 

“Advances in deep learning and AI have made it much more feasible to extract and use more information from images to inform opportunistic screening.” 

– Dapamede

The study findings demonstrated that the new model performed well at identifying the cardiovascular risk of patients at low, moderate, or severe risk based on mammogram images. 

As per them, the rate of serious cardiovascular events increases with the level of breast arterial calcification in women under 60 and those between 60 and 80 after calculating the risk of death from an acute heart attack, stroke, or any cause at two and five years.

This makes this tool specifically appropriate for providing early warning of heart disease risk in younger women, allowing for early interventions.

The study's results also demonstrate that those with the highest level of breast arterial calcification, above 40 mm2, had a considerably lower five-year rate of event-free survival than those with the lowest level, below 10 mm2.

Researchers' next step is to obtain approval from the FDA and make it available for wider use. In the future, the plan is also to look into using similar AI models to analyze biomarkers for other conditions, like kidney disease, that might be extracted from mammograms.

Now, when it comes to companies advancing AI in healthcare, GE HealthCare Technologies (GEHC +0.93%) is heavily invested in the tech and integrating it across its products to improve patient care and enhance clinician workflows. The $37.26 bln market cap company's stocks are currently trading at $81.49, up 4.23%. In 4Q24, its revenue increased by 2% to $5.3 billion while diluted EPS was $1.57.

Latest on GE Healthcare Technologies

AI in Material Science: Optimizing Thin-film Growth Processes

Material science is critical for modern advances as it enables the development of new materials and improves existing ones. AI in this field is contributing by accelerating the discovery and design of new materials, optimizing existing ones, and improving manufacturing processes.

Researchers from Tokyo University of Science (TUS), led by Professor Masato Kotsugi, have now utilized AI to optimize thin-film growth processes by predicting dendritic growth in them.

These processes play a key role in the development of semiconductor devices, communication technologies, and sensor technologies.

Now, thin films are grown by laying down small layers of materials on a substrate. However, to fully utilize the performance of multilayer film devices, the substrate must be precisely fabricated as it has a considerable influence on the structural configuration, which then affects their function and performance.

However, the structures are significantly influenced by growth process conditions such as composition, atmosphere, and surface defects. During the growth process of large-area fabrication, which is important for its commercial application, a major obstacle arises in the form of dendritic structures.

These tree-like branching patterns reduce thin films' flatness and are commonly observed in materials like graphene, copper, and borophene, especially in the early growth stage and multilayer films. 

Given that microstructure directly influences the device's performance, it is critical to reduce dendritic formation. For that, we must first understand the condition that causes dendritic branching. Existing methods of studying dendrites, however, depend on simple visual analysis and subjective interpretation, requiring considerable trial and error.

To tackle these issues, researchers built an AI model to analyze dendritic structures. The innovative method integrates machine learning (ML) and persistent homology (PH) with energy analysis to bridge structure and process in dendritic growth.

“Our approach provides new insights into growth mechanisms and offers a powerful, data-driven pathway for optimizing thin-film fabrication.”

– Prof. Kotsugi

Combining AI, PH, & Energy Analysis to Study Dendritic Growth

Published in Science and Technology of Advanced Materials: Methods³, the study details the model that uses persistent homology (PH) to capture the complex topological features of dendrite microstructures, often overlooked by traditional image processing techniques.

While PH allows for the multiscale assessment of holes and connections within geometric structures, ML technique principal component analysis (PCA) helped reduce the essential dendrite morphology features extracted via PH to a 2D space.

This helped the team quantify structural changes in dendrites and establish a relationship between these changes and the free energy in the material that impacts the way dendrites are formed.

Upon assessing this relationship, the team discovered the particular conditions and hidden growth mechanisms that affect dendritic branching. 

“Our framework quantitatively maps dendritic morphology to Gibbs free energy variations, revealing energy gradients that drive branching behavior.” 

– Kotsugi

The team then studied dendrite growth in a hexagonal copper (Cu) substrate and compared the results with data from phase-field simulations to validate their approach. Kotsugi noted:

“By integrating topology and free energy, our method offers a versatile approach to material analysis. Through this integration, we can establish a hierarchical connection between atomic-scale microstructures and macroscopic functionalities across a wide range of materials, paving the way for future advancements in material science.”

This method, according to him, can lead to the development of high-quality thin-film devices for high-speed communication beyond 5G and advancements in high-performance materials, nonequilibrium physics, and sensor technology.

Now, if we look at a prominent name in the material science field that is utilizing AI to innovate, Applied Materials (AMAT -0.85%) is among the top. It leverages AI to enhance semiconductor chip manufacturing through its AI(x) platform, ExtractAI, and SEMVision™ H20 System. The $123.6 bln market cap company's shares are currently trading at $151, down 6.5% YTD. For 1Q25, it reported a 7% increase in revenue to $7.17bln while GAAP EPS was $1.45 and non-GAAP EPS was $2.38.

“The industry drive to accelerate the development of advanced compute and more sophisticated AI is gaining momentum. Applied Materials is enabling the major device architecture inflections critical for energy-efficient AI, and our focus on high-velocity co-innovation creates unique collaboration opportunities with our customers and partners, positioning Applied for continued growth and outperformance in the years to come.” 

– CEO Gary Dickerson last month

Latest on Applied Materials

Conclusion

AI has already been transforming businesses across industries. The rapid pace of innovation is now offering groundbreaking solutions to its long-standing challenges. With these advancements, ranging from material science, medical diagnostics, and personal supercomputers to even solutions for AI's own learning challenges, we are stepping into an era that promises unprecedented transformation.

As AI continues to evolve, its ability to drive productivity, improve decision-making, and unlock new frontiers makes it one of the most powerful forces shaping the future.

Click here to learn all about investing in artificial intelligence (AI).