Category: 4. Technology

It’s only been a couple of months since I first gave Linux a serious shot on one of my PCs, and since then, I’ve quickly grown into a big fan of it. These days, I would rather spend more time in Linux than on Windows, but of course, there are some things that make this impossible. Missing apps like Adobe Lightroom is a significant problem, and for many people, the same sentiment applies to something like Microsoft Office.

Enter WinApps, a project I recently heard about from a friend. The purpose of WinApps is kind of akin to using GUI Linux on Windows through WSL, but reversed. With it, you can run Windows apps on Linux as if they were naturally part of the operating system. It’s a fantastic idea that could mean you don’t ever need to use Windows again, though it’s going to take some work.

How does WinApps work?

The power of VMs

Running Windows on Linux is, of course, not a new concept thanks to virtual machines, but what stands out about WinApps is how it takes that basic technology and adapts to provide a much more seamless experience. Typically, a virtual machine will make it so that you have to launch and view the entire operating system to interface with it, but WinApps can make each of your apps launch independently in the Linux desktop environment.

For this, it uses Docker, Podman, or libvirt to virtualize a Windows install, and then, it relies on FreeRDP to provide a remote desktop connection to the virtual machine. Then, WinApps itself can provide links to installed apps within your VM so that they appear in the application launcher on your Linux host. Thanks to FreeRDP, when you launch on of these apps, the remote desktop session only captures the application window, and makes it resizeable within your Linux desktop, as if it was running natively.

As I already mentioned, this is a very similar approach to how Linux apps run on Windows through WSL, though there are some limitations, such as not being able to easily pass through an integrated GPU to the guest OS. You can pass through a discrete GPU depending on the virtualization backend you’re using, though.

FreeRDP works best on Linux desktops using the X11 windowing system, which means you may experience a few more bugs if you’re using the more modern Wayland. The friend who told me about WinApps says he’s interested in creating a solution for better performance on Wayland, though, so hopefully we get some developments on that front.

Setting up WinApps

It takes a bit of time

Getting started with WinApps isn’t the easiest process, but it also won’t feel too daunting if you’re used to working in Linux, particularly with the terminal. You can find detailed instructions on the project’s GitHub page, but here’s a basic rundown. First, you’ll need to install your preferred virtualization engine (I went with Docker) and create the Windows virtual machine inside it. Then, after testing FreeRDP, you can install any apps you’d want to use inside the Windows machine.

Once you have your Windows apps installed, you can install WinApps on the Linux host, and this allows you to register your preferred apps to appear in the Linux app launcher. I did notice that some apps may not appear on the list when installing WinApps, though thankfully, you can also manually run apps that aren’t registered during the WinApps setup by using the terminal command, replacing the text in quotes with the path to the appropriate file inside Windows:

winapps manual "Path/to/Exe/file.exe"

The biggest issue I had during setup was the fact that Docker commands required me to use sudo, but the command for downloading and installing WinApps didn’t work if I preceded it with sudo. I didn’t see this on the official guidance, but I soon learned that I needed to add my Linux user account to the Docker group. The command for that is as follows:

sudo usermod -aG docker $USER

Once you have that set, WinApps should install, and you can go ahead and add your Windows apps to your Linux app launcher.

Does it actually work?

Yes… with some caveats

I primarily wanted to test WinApps with things I can’t use on Linux, and my first thoughts went to Adobe Photoshop and Lightroom. Both of these apps are officially listed as supported by WinApps, and well, they function. Adobe Lightroom works the best of the two. I could use it pretty much without major issues. I loaded up a RAW photo and applied automatic lighting corrections, which performed well enough for day-to-day use. Because I didn’t have a discrete GPU I could pass through, the AI-powered Denoise feature ran extremely slowly and took a few minutes to process a single image. However, manual noise reduction is still available and works fine. While it’s not as easy, I’d still rather use it here than something like Darktable or RawTherapee since Lightroom usually gets me better end results.

Photoshop didn’t go as well, unfortunately. Something about the way Photoshop operates results in a few bugs that break the flow. Sometimes, the UI will just freeze and stop responding, but you can briefly fix it by using the task switcher in Linux to switch to another window and back. Forcing the UI to refresh seems to fix the issues, but they’re very frequent with Photoshop, so the workflow is impacted a good bit.

Microsoft Office is also listed as officially supported, but since I don’t currently have a license for it, I went with WPS Office instead, which is my go-to suite these days. This, too, worked mostly fine, and it would be completely usable for writing documents or making presentations. The “New” button is broken, but you could still create new documents by going to the appropriate section of the Home tab.

What I’ve found is that resizing app windows, and especially maximizing/windowing them, can often make the UI unresponsive, similar to what happens when using certain features in Photoshop. For the most part, once you have your apps in the size you want, this shouldn’t be an issue, but it can be a bit frustrating when you’re not expecting it and you just want to get things done. But the use case for this shouldn’t be for you to be juggling a lot of Windows apps or anything, it’s meant to be a solution for the occasional app that doesn’t exist on Linux. Another annoying quirk is that, if you’ve already opened a Windows app, opening another one can sometimes result in the existing app disappearing, but attempting to launch the second app again will bring them both back on screen. Both apps will also have the same label on the dock, since they’re still technically just a Windows VM.

Despite the fact that WinApps is expected to work more reliably under X11, I didn’t have a ton more issues using it under Wayland. Photoshop seemed to become even less stable, but it wasn’t great either way.

It won’t be for everyone

But for the right person, it’s an amazing idea

More and more, I’ve been finding myself uninterested in using Windows when I can avoid it, and a solution like WinApps sounds like it could be a dream come true, at least in concept. Running Windows apps within Linux with decent performance is all I really need to stop using Windows as a standalone OS, and that’s exactly what WinApps is aiming for. The execution isn’t quite perfect, but if you mostly work with Microsoft Office, you’re probably completely fine.

However, not everyone is going to fit that description. An app like Lightroom can run quite well on a modern laptop with an integrated GPU, but with the virtualization used in Docker, you’re leaving a lot of performance on the table, and certain workloads may not be feasible. Something similar can be said for DaVinci Resolve. The Linux version of Resolve requires a discrete GPU, and on Windows, you can use an integrated GPU, but because the virtualization doesn’t pass through that integrated GPU, actually editing video becomes basically impossible.

If you only have a very small need for Windows-based tools, then WinApps can bridge that gap, and maybe there will be some virtualization improvements in the future that make it even better. As it stands, though, Windows on Linux is still not a full replacement for a proper Windows install. But it’s great to picture a feature where that might be the case.

This as-told-to essay is based on a conversation with Advait Maybhate, a software engineer. The following has been edited for length and clarity. Business Insider has verified his employment and academic history.

When I graduated from the University of Waterloo with my bachelor’s degree in 2023, I had done about a dozen tech internships.

Internships are a big deal at Waterloo, and students usually do six during their time there. I started doing internships before I enrolled and took some gap semesters to squeeze in a couple more stints.

To me, internships meant exploring varied fields, from gaming to fintech. I also got to intern at companies of different scales, from early-stage startups to mature Big Tech companies.

The first summer internship I did at Waterloo was at Google. Interning there was an eye-opening experience. I got to work on Google Search, a product that billions of people, including myself, use every single day.

When I took up the internship, like any freshman, I just thought it would be cool to work at a big company and ship big products. I ended up interning at Google twice, first in the summer of 2019, and then during the following summer in 2020.

During my internships at Google, I learned a lot, particularly about operating as a software engineer on large-scale products. That included learning how to write unit tests and good technical design documents. Big companies are great at that.

That said, I didn’t enjoy the bureaucracy that came with working in a Big Tech company. If you are shipping something on Google Search, you cannot break Google Search. That is just one of the underlying rules.

I understand why things have to be slow at that scale. It’s just that for someone who wants to learn fast and try out different things, it can feel limiting.

Even for my internship projects, it took a few months just for the code to get shipped. Although the projects were technically done, we still had to conduct A/B testing experiments and get sign-offs before the code could be deployed.

Going from Big Tech to startups

That experience eventually set me on the path toward working at startups. I chose to focus on AI because I wanted to be at the edge of what technology can do.

I was initially an AI skeptic. I didn’t buy into the hype of how it could change everything. It was only when I started using AI on a day-to-day basis that I began to appreciate how it could usher in a fundamental shift in the way we work.

It also helps that working on AI is fun and exciting. There are new advancements in space every week, and the frontier of what we can do just keeps going further.

I ended up doing two internships at two AI startups before I graduated. The first one was at Warp, an AI agent platform for developers, and the second one was at Ramp, a fintech startup that uses AI to automate financial operations.

I received full-time offers from both Warp and Ramp and chose to work at Warp. Both were great companies, but I wanted to work at Warp because I wanted to be part of a startup that was in a relatively early stage of development.

Ramp was at a much more mature stage than Warp at the time, and was focused on scaling up. Warp, on the other hand, was still trying to figure things out. On a personal level, I wanted to see how a startup goes through that process. I wanted to grapple with questions like, “How does pricing work? How does the business model work?”

That is harder to see at a mature startup, where all of these things have already been figured out and growth is the priority.

So far, working at Warp for the past two years has lived up to my expectations. We ship code every week. I could be working on something on Tuesday, and it gets shipped out on Thursday. I work maybe 60 to 70 hours a week. It’s a very different kind of velocity and cadence than at Big Tech.

In the near term, I want to continue to work on AI because it’s one of the most rapidly expanding areas in tech. Companies like Warp and its competitors, Cursor and Cognition, are all expanding very rapidly.

I am somewhat tempted to launch my own startup, but I think it’s difficult to gain market share in this hyper-competitive space. That’s something I will give serious thought about in the future.

Do you have a story to share about working at an AI startup? Contact this reporter at ktan@businessinsider.com.

For disabled and neurodivergent students, accessing and learning information delivered verbally or visually in the classroom or online can be difficult. As a result, they can struggle to keep up with their peers. 

In the UK, if a student declares a disability, their higher education institution has a legal obligation to make reasonable adjustments. Yet, only 34 per cent of disabled students have had support both agreed and completely implemented, according to the recent Disabled Students UK Access Insights Report. Despite their needs being recognised, and the good intentions of staff, the provision of tailored, individualised adjustments can be costly both in time and money. 

New ways of teaching and learning that leverage generative artificial intelligence (GenAI) offer an opportunity to move towards inclusive teaching and learning, by generating learning materials adjusted to the individual needs of all learners, not only those with disabilities. 

To tap into the potential of personalised learning opportunities, we need to avoid the bottom-up piecemeal application of educational technology, where adoption is often based on the specific preferences of individual staff. Instead, implementation should be guided by a top-down approach that leverages GenAI to meet clear educational goals and learning outcomes in a personalised manner. 

For example, the Universal Design for Learning (UDL) framework offers guidance for designing multiple ways of engagement, representation, action and expression. This makes teaching and learning environments more inclusive and supports the embedding of AI-generated personalised learning support, placing the student at the centre of the process. 

Remember that learning doesn’t just involve engaging with accessible resources. It’s a complex process involving a range of cognitive and socio-emotional processes such as memory, attention, executive function, metacognition, motivation and self-esteem. Effective learning, especially reflective and critical thinking, does not occur in a vacuum – it occurs in an interactive way, within a complex educational ecosystem. 

The increased use of online learning and AI is rapidly changing the traditional nature of this ecosystem, and it is essential that students have the appropriate psychological skill set to enable them to engage successfully with it. Students with disabilities often struggle in online environments – partly because of cognitive constraints that influence their ability to process information, and partly owing to social-emotional constraints such as poor self-esteem, low motivation and anxiety, which can limit their willingness to share ideas and opinions, and affect their acceptance of feedback and challenge. 

If GenAI-supported learning is going to help their participation in the wider educational ecosystem, we need to do more than provide the appropriate digital environment – we need to ensure students possess the right psychological resources and skills to enable their learning in this environment.

As highlighted in a recent Campus article, metacognition is important for learning. If confidence and self-efficacy are developed for disabled students through improved opportunities to adapt and engage with learning materials, metacognitive processes can be facilitated too. In the first instance, disabled students must be able to access learning materials in a format that suits them. Then they can begin to develop self-regulatory processes and reap the benefits of improved engagement and increased academic achievement. 

Harness personalisation at all levels

Creating an inclusive teaching space using GenAI benefits not only disabled and neurodivergent students, but all students. In fact, the degree of personalisation now available provides the best opportunity to date to place all learners at the centre of their educational journey, both now and in the future. 

GenAI supports dynamic multilevel personalisation: 

• personalisation based on the psychological profile and preferences of the learner
• personalisation based on their learning environment, course and assessment type
• personalised support for situated learning as it adapts over time to meet the evolving needs of the learner and their assessments.

For example, developments in the areas of multimodal AI can allow students to choose the medium (visual or verbal) they learn in, and also the best format to demonstrate their learning. From a cognitive perspective, students with dyslexia frequently have difficulties with memory and attention and experience cognitive load. This can be addressed by paced study strategies and the chunking and presentation of content in the most appropriate visual or verbal format. 

If self-efficacy is low, targeted scaffolding and positive feedback mechanisms would support learning of both low-level (content acquisition) and higher-order tasks (critical thinking, analysis and synthesis) and assist with well-being and engagement.

GenAI can also support learners’ organisational and metacognitive skills by assisting with time management. For example, if a student found assessed presentations particularly challenging, GenAI could allow extra time to prepare, practise and receive feedback.

As technology develops, students can access emotionally intuitive AI resources, which could support the socio-emotional facets of learning and build resilience. 

As we negotiate our way through the increasingly complicated educational technology landscape, we can learn a lot from how GenAI can be used to alleviate the challenges faced by students with disabilities, and generalise these benefits to the wider student population. Instead of becoming blinded by technology, remember the ability to offer personalisation, which entails a sharp focus on the learner. 

Unlike AI, human learning is subject to a range of psychological constraints. Although AI may help remove many physical barriers to learning, its true potential will only be fully realised when we address psychological barriers and enablers to learning and the skill set required. Using AI to deliver a learner-centric approach will help address these challenges and bring educational benefit for all. 

In 2023, the watchmaker Sylvain Berneron launched his namesake brand with a bang. His Mirage—featuring an unusual warped shape and twisty hands—became a smash hit, selling out instantaneously and earning watch-of-the-year plaudits in the industry. According to Berneron, he could up his annual production of the Mirage five times over and still comfortably find a buyer for every piece. Now, Berneron is back with the sequel—though anyone hoping for more Dalí-esque curves will likely be disappointed.

“It is the strict opposite of a Mirage,” he says of his latest creation, “which is either stupid or brave.”

Whereas his debut defied expectations aesthetically, the new Quantième Annuel is downright old-fashioned at first glance, with a symmetrical round case and elegant dial layout. Despite its more traditional appearance, however, the Quantième Annuel is far more complicated beneath the hood: While the Mirage’s movement had 135 pieces, the QA boasts 450. The goal, Berneron suggests, is to prove he’s not merely a radical designer but a masterful watchmaker as well. “When you make a new brand, people tend to box you into your first [release],” he says. “I don’t want to be a one-trick pony.”

What is most apparent in the new watch is that Berneron thought hard about how real people would actually use his creation. The dial is organized to provide relevant information with hyper–efficiency: “You start top to bottom to read the time, and if you want the date you go left to right,” he explains. “And finito, there is no drama.” The watch comes fitted with replaceable steel bumpers, laid over the soft and dentable platinum case like armor. And while he might have loved to show off his watchmaking prowess with a more sophisticated perpetual calendar (which is designed to remain accurate with no adjustments until the year 2100), he surveyed collectors and found an annual calendar (set once a year) to be more practical.

Berneron isn’t a medium, but he has a clear vision for his brand’s future. The next 10 years are meticulously plotted out in his mind: He knows the exact number of timepieces he’ll produce in 2035 (600, on the higher end for independent watchmakers), and plans for his third collection, called the Fiasco and designed to showcase high–jewelry techniques, are already well underway. It seems critical for him to proceed this way, plugging all the potential gaps to sustain his brand. Because Berneron knows what he can’t see too. “It would be naive to think that the Mirage would be the talk of the town for the next 50 years,” he says.

The sophomore release is always the most difficult in any industry. How do you convert a single hyped watch into an enduring business? That’s what Berneron is hoping to achieve with the Quantième Annuel. When he discusses his fledgling company, he doesn’t compare it to other indie watchmakers. Rather, he takes inspiration from the way Patek Philippe and Cartier have diversified their staple offerings, or the way Rolex always thinks decades ahead. “There is a huge difference between making a cool watch and making a brand,” he says. A second cool watch certainly helps with the latter.

Cam Wolf is GQ’s Watch Editor.

A version of this story originally appeared in the September 2025 issue of GQ with the title “How to Build the Next Great Watch Brand”

Sony’s newest flagship wireless noise cancelling headphones – the Sony WH-1000XM6 – has dropped to the lowest price of the year. As part of its Labor Day Sale, Adorama is offering it for only $379 after you apply coupon code “AUGU2925” during checkout (when you put in your payment info). That’s 15% off the normal $449 list price. The WH-1000XM6 builds upon our previous favorite headphone with better sound, noise cancelation, and ergonomics.

Sony WH-1000XM6 Headphones for $379 (was $449)

TEM analysis evaluated the size and morphology of NPs. Figure 2a and b shows images of α-Fe₂O₃NPs. In Fig. 2c, the corresponding histogram was plotted using TEM images of the nanoparticles and analyzed with Image J software. The average particle size was estimated to be 45 nm with a standard deviation of 11 nm. Besides, they have a somewhat spherical shape and some porosity on their surface.

The XRD pattern of α-Fe₂O₃NPs is shown in Fig. 3a, which agrees with the standard pattern data (JCPDS card No. 33–0664)³⁷. The peaks appearing in the 24.10, 33.09, 35.61, 40.82, 49.44, 54.00, 57.46, 62.41, and 63.95 degrees have been attributed to the diffraction planes (012), (104), (110), (113), (024), (116), (018), (214) and (300) of NPs crystalline respectively. All sharp and narrow peaks appearing in the range of 2θ indicate the crystalline nature and the high purity of α-Fe₂O₃NPs extracted by the combined of green synthesis and hydrothermal method. In addition, the XRD pattern of nanocomposite is shown in Fig. 3b. In this pattern, all the characteristic peaks related to α-Fe₂O₃NPs and the characteristic broad peak at the angle of 2θ = 20.26 related to CS biopolymer have been observed, confirming the successful formation of this nanocomposite³⁸.

XRD pattern of a α-Fe₂O₃ NPs, b CS-α-Fe₂O₃ nanocomposite.

The crystallite size of α-Fe₂O₃NPs can be estimated by using the Williamson–Hall method (Eq. 1) ³⁹:

In this equation, β is full width at half maximum (FWHM), θ is the diffraction angle, λ is the X-ray wavelength, d is the average crystallite size, and ε is the lattice strain. The crystallite size of α-Fe₂O₃NPs was determined by drawing a diagram (Fig. 4) and calculating the width from the origin. With these interpretations, the crystallite size was estimated to be 43 nm.

Williamson–Hall diagram of –Fe₂O₃ crystallite.

One of the important characteristics of colloidal suspensions is the tendency of their particles to stick together. In an aqueous environment, particles continuously interact, and the stability of such solutions is determined based on these interactions. In order to create a stable solution, short-range repulsive forces are needed, among which we can mention the steric interfacial forces that play an important role in stabilizing suspensions. Therefore, these electric forces create layers around the NPs, and this issue plays a significant role in the value of zeta potential and ultimately prevents particles from agglomerating⁴⁰. In this research, the stability of the CS-α-Fe₂O₃ nanocomposite solution originates from the strong bonding of CS on the surface and the charge of α-Fe₂O₃NPs, which causes repulsion between them. The CS, a biocompatible and biodegradable polymer, forms a protective layer around the α-Fe₂O₃NPs, preventing them from agglomerating and thus contributing to the stability of the solution. In order to study the stability of α-Fe₂O₃NPs and CS-α-Fe₂O₃ nanocomposite solutions with pH 5.11 and viscosity 0.9327 mPas, zeta potential analysis as a critical parameter was performed, and their results are shown separately in Fig. 5a and b. In these graphs, the distribution function of the zeta potential of individual NPs and nanocomposite in the solution is presented as a percentage, and the average zeta potential of unmodified α-Fe₂O₃NPs and CS-α-Fe₂O₃ nanocomposite was measured 29.93 and 36.67 mV, respectively. Also, by comparing the two graphs, the results indicate that adding CS to α-Fe₂O₃NPs narrows the peak width of the graph, which indicates that a large percentage of nanocomposites have the same zeta potential, and this issue was evaluated as stabilization of CS-α-Fe₂O₃ nanocomposite suspension. In an experiment, two solutions of NPs and CS-α-Fe₂O₃ nanocomposite with the same concentration of 5 mg/ml were prepared, and the role of adding CS in stabilizing the α-Fe₂O₃NPs solution after 72 h is shown in Fig. 5c. Another result obtained from the zeta potential of CS-α-Fe₂O₃ nanocomposite suspension indicates the net positive charge of these nanocomposites. Furthermore, these nanocomposites can stick and penetrate cancer cells with a negative charge through electrostatic forces⁴¹.

Zeta potential distribution function diagram of a α-Fe₂O₃ NPs, b CS-α-Fe₂O₃ nanocomposite, c comparing the stability of Fe₂O₃ NPs solution and CS-Fe₂O₃ nanocomposite solution after 72 h.

Figure 6a and b shows the UV–visible absorption spectrum of α-Fe₂O₃NPs and CS-α-Fe₂O₃ nanocomposite between 200 and 900 nm. The absorption peak for α-Fe₂O₃NPs and CS-α-Fe₂O₃ nanocomposite is 450 and 380 nm, respectively. This result is significant and reassuring, as it aligns with the data of another research³⁶ enhancing our confidence in the findings. Considering the absorption in the infrared region, this issue guarantees the wide potential of this nanocomposite to be used in cancer treatment by PTT and PDT with an 808 nm laser.

UV–Visible absorption spectrum of a α-Fe₂O₃ NPs, b CS-α-Fe₂O₃ nanocomposite.

The optical band gap of α-Fe₂O₃ NPs can be calculated by using Tack’s equation (Eq. 2)⁴².

In this equation, α is the absorption coefficient, A is a constant, hν is the photon energy, and n is a constant that depends on the nature of the electron transition, which is 2 for direct transition and 1/2 for indirect transition. Since α-Fe₂O₃ NPs have a direct energy gap, the graph of (αhν)^² in terms of energy is drawn in Fig. 7⁴³. The value of the energy gap associated with α-Fe₂O₃ NPs was approximated at 1.8 eV using the extrapolation method, which is comparable to the result of another research⁴⁴. Also, our Experimental investigation indicates that the NPs exhibit an indirect band gap of roughly 1.7 eV.

α-Fe₂O₃ NPs Tauc diagram.

FTIR analysis was performed in the 400–4000 cm^− 1 wave number range to determine the bonds and functional groups on the surface of CS powder, α-Fe₂O₃NPs, and CS-α-Fe₂O₃ nanocomposite. The FTIR spectrum of α-Fe₂O₃NPs is shown in Fig. 8a. A broad peak in the range of 3445 cm^− 1 is assigned to the stretching vibration between oxygen and hydrogen belonging to the functional groups of hydroxyl and water molecules, which proves the absorption of some water at the surface of α-Fe₂O₃NPs⁴⁵. Also, two sharp absorption peaks in the range below 1000 Cm^− 1 indicate the main characteristics of α-Fe₂O₃ NPs, which are attributed to the stretching frequencies of metallic iron. The high frequency peak in 526 cm^− 1 refers to Fe-O deformation in tetrahedral and octahedral environments. At the same time, the peak at low frequency in the range of 450 cm^− 1 refers to the Fe–O shape change in the octahedral environment of hematite⁴⁶. In the spectrum of CS powder, the peaks appearing in the 3445 cm^− 1 and 3360 cm^− 1 range are related to O–H and N–H stretching vibrations, respectively, and intra molecular hydrogen bonds. The peaks around 2925 cm^− 1 and 2854 cm^− 1 are attributed to symmetric and asymmetric C–H stretching vibrations, which are common bonds characteristic of polysaccharides. The remaining presence of N-acetyl groups was confirmed by peaks in the range of 1741 cm^− 1 related to C=O stretching vibrations. Also, C–H bending vibrations and the symmetric deformation were confirmed by the peak in 1369 cm^− 1. The peak appearing in 1020 cm^− 1 is attributed to C–O stretching vibrations. All peaks related to CS material are shown in Fig. 8b⁴⁷. By analyzing the FTIR analysis of CS-α-Fe₂O₃ nanocomposite, all the characteristic peaks of the infrared Fourier transform of α-Fe₂O₃NPs except Fe–O and all the characteristic peaks of CS were observed (Fig. 8c).

FTIR spectrum of a α-Fe₂O₃ NPs, b CS, c CS-α-Fe₂O₃ nanocomposite.

Three different concentrations of CS-α-Fe₂O₃ nanocomposite solution were prepared to study PTT effects using an ultrasonic bath. Three ml of each of these different concentrations were poured into a tube and placed in a relatively dark under the radiation of 808 nm laser with a power density of 1 W/cm². The temperature change was measured by a digital thermometer equipped with a thermal sensor for 15 min. Figure 9 shows temperature changes over 15 min for nanocomposite solutions at 0.1, 0.5, and 2 mg/ml concentrations, and DI water, with increases of 7.4, 10.9, 13.8, and 5.8 °C, respectively. Data are presented as mean ± SD (n = 3), with error bars indicating standard deviation. Detailed temperature differences for each concentration are provided in Table 1 .Due to the appropriate temperature change and low concentration, the 0.5 mg/ml concentration was determined as the optimal concentration.

Temperature change diagram according to 808 nm laser irradiation time CS-α-Fe₂O₃ nanocomposite solution in different concentrations. Data are presented as mean ± SD (n = 3). Error bars represent standard deviation.

The photothermal conversion efficiency of the CS-α-Fe₂O₃ nanocomposite was calculated using Roper’s method under 808 nm laser irradiation. Roper’s equation models the photothermal process as a balance between the heat generated by light absorption and the heat lost to the surroundings, allowing photothermal conversion efficiency to be derived from the steady-state temperature and system parameters⁴⁸. In the experiment, the temperature of the nanocomposite solution increased from 24 °C to a saturated value of 42.1 °C, while DI water under identical conditions reached only 34.2 °C. Based on this temperature rise and heat transfer analysis, the efficiency was determined to be 7%. Detailed calculations, fitting procedures, and related graphs are provided in the Supplementary Information.

When α-Fe₂O₃NPs as semiconductors are exposed to laser irradiation, the photocatalytic process is activated, in which the photon energy is used to transfer electrons from the valence band to the conduction band, and at the same time, it creates a similar number of holes in the valence band, which ultimately leads to the formation of an electron-hole pair. Continuing this separation and transfer of charges to the surface leads to oxidation and reduction reactions with the molecules around these α-Fe₂O₃NPs, which can produce ROS. By studying previous research, the ability of α-Fe₂O₃NPs to create hydroxyl ROS was determined³⁵. Therefore, methylene blue was used as a hydroxyl radical probe to detect indirectly this type of free radical. For this purpose, in an experiment, 1 mg of NPs powder was dissolved in 1.5 ml of DI water and 0.5 ml of methylene blue (0.2 mg/ml) using an ultrasonic bath. Two similar samples were prepared, and in a relatively dark room, 3 ml of each sample was placed on magnetic stirrers, and only one sample was exposed to an 808 laser with a power density of 1 W/cm² for 10 min. Then, the methylene blue absorption spectrums of two samples were measured by a spectrometer, and the reduction of the characteristic absorption peak of methylene blue in the sample under laser irradiation compared to the non-irradiated sample at the wavelength of 664 nm was observed. The results are shown in Fig. 10. The electrons and holes created by the photocatalytic process provide the conditions for producing active oxygen species around the α-Fe₂O₃NPs. In this way, the holes formed in α-Fe₂O₃NPs lead to the oxidation of the methylene blue molecule, which turns it into an active substance that is ready to react. Also, the electrons on the surface of α-Fe₂O₃NPs are transferred to oxygen molecules dissolved in water, which produces superoxide radical negative ions. These ions, after reacting with water molecules, can decompose, and consequently, hydroperoxyl radicals and hydroxyl ions are produced. In the end, after reacting hydroperoxyl radicals with water molecules and proton absorption, this process produces hydroxyl radicals as Eq. (3)⁴⁹:

Methylene blue absorption spectrum before and after 808 nm laser irradiation for α-Fe₂O₃NPs.

To assess the cytotoxicity of the CS-α-Fe₂O₃ nanocomposite, an MTT assay was carried out on AGS gastric cancer cells under in vitro conditions. This experiment includes three independent biological replicates, each containing five technical replicates. The percentage of relative cell viability was determined by calculating the ratio of treated cell viability to that of untreated control cells. Results are expressed as the mean of three biological replicates (n = 3) ± standard deviation. Statistical analysis was carried out using GraphPad Prism (version 9), with group comparisons performed via analysis of variance (ANOVA) followed by Tukey’s post hoc test. A p value < 0.05 was deemed statistically significant. These Cells were treated with two concentrations of the nanocomposite: 250 ppm (Group 1) and 500 ppm (Group 2). Each concentration was tested under two conditions—with and without irradiation using an 808 nm laser (power density: 1 W/cm², duration: 15 min). Untreated cells were used as the negative control, while a separate control group consisting of cells exposed to laser alone (without the nanocomposite) was included to isolate the impact of irradiation itself.

In Group 1, treatment with 250 ppm of the nanocomposite in the absence of laser exposure resulted in minimal cytotoxicity, with a high cell viability of 97%. However, upon laser irradiation, viability decreased slightly to 88%. In Group 2, treatment with 500 ppm of the nanocomposite without irradiation maintained a similarly high viability of 95%, indicating low inherent toxicity. In contrast, laser-irradiated cells in this group showed a significant decrease in viability to 68%, suggesting a pronounced phototoxic effect. The corresponding results are presented in Fig. 11.

Effect of CS-nanocomposite at two concentrations (250 ppm and 500 ppm), with and without 808 nm laser irradiation, on the viability of AGS cells as determined by the MTT assay. The data are presented as the mean of three independent experiments ± standard deviation. Comparisons among multiple groups were performed using one-way ANOVA followed by Tukey’s post hoc test (ns = non-significant, ** = p < 0.01, # = statistically significant compared to untreated cells).

The 250 and 500 µg/mL non-laser groups did not show any statistically significant cytotoxicity compared to the control group. Moreover, at the concentration of 250 µg/mL, there was no statistically significant difference between the laser and non-laser treated groups (ns). However, at 500 µg/mL, a statistically significant difference was observed between the laser and non-laser groups. These findings highlight the biocompatibility of the nanocomposite in the absence of laser activation, even at higher concentrations. Conversely, under laser exposure, a concentration-dependent reduction in cell viability was observed, demonstrating the nanocomposite’s potential for application in PTT and PDT therapy.

Just ahead of IFA 2025, HMD’s Pulse 2 Pro managed to leak thanks to a listing by a Swiss retailer. The listing comes with detailed specs and images confirming all there is to know about the upcoming budget device.

Pulse 2 Pro (TA-1687) is listed in Black and Green colors and features a revised camera island compared to its predecessor with a dual-tone finish. We can see two camera sensors stacked vertically alongside the LED flash. Based on the spec details, the Pulse 2 Pro is bringing a 50MP main shooter with OIS.

HMD Pulse 2 Pro in black

The front houses a 6.72-inch IPS LCD with FHD+ resolution and a punch hole cutout for the 50MP front-facing camera. We can also notice a pretty sizeable bottom chin. The device is equipped with an Unisoc T615 chip paired with 6GB RAM and 128GB storage (expandable via microSD).

HMD Pulse 2 Pro in green

Elsewhere, Pulse 2 Pro boots Android 15 and packs a 5,000 mAh battery with unspecified fast charging. The official dimensions are 166.14 x 76.58 x 8.3 mm and the phone weighs 198 grams. Pulse 2 Pro is listed at CHF 169, which is about €180/$211. The device is expected to get a formal announcement at IFA Berlin later this week.

Source

Aging demographics, labor shortages, the adoption of GenAI, and the 2023 implementation of e-invoicing are driving companies to automate finance, tax, procurement, and HR in Japan. Yet only 16% of digital transformations succeed, and that’s only 4–11% in traditional industries. The main barriers? Weak leadership commitment, a rigid culture, and a lack of digital talent. LayerX offers an AI SaaS platform to help enterprises scale back-office automation.

LayerX, a Japanese AI SaaS startup that enables businesses to cut back-office workload, has raised $100 million in a Series B round led by Technology Cross Ventures (TCV), marking the U.S. fund’s first investment in a Japanese startup.

The company declined to disclose its valuation, but said both the valuation and the size of the round are among the largest ever raised by a seven-year-old Japanese startup at the Series B stage. Other investors, including MUFG Bank, Mitsubishi UFJ Innovation Partners, JAFCO Group, Keyrock Capital, Coreline Venture, and JP Investment, also joined the Series B round, bringing the total raised to $192.2 million.

The startup’s key offerings include Bakuraku, a platform that automates corporate spending workflows, covering expense management, invoice processing, and corporate card operations — for more than 15,000 companies; Alterna, a retail digital securities investment platform developed in partnership with Mitsui & Co.; and Ai Workforce, a generative AI solution designed to streamline workflows and harness enterprise data.

Founded in 2018 by serial entrepreneur Yoshinori Fukushima, who studied machine learning at the University of Tokyo and previously launched the news app Gunosy, which was later listed on the Tokyo Stock Exchange, LayerX grew out of one of his digital transformation (DX) and blockchain projects.

The founder launched LayerX after identifying a significant bottleneck in Japan’s enterprise workflows: paper-based invoice processing. This insight prompted the team to pivot into SaaS with their AI-driven platform, Bakuraku, Fukushima told TechCrunch, adding that the platform’s AI-native user experience quickly gained traction, helping LayerX secure major strategic partnerships, including with MUFG, or Mitsubishi UFJ Financial Group, and paving the way for its latest funding round.

Despite a wave of digitalization, many Japanese companies still rely on paper and Excel for expense reimbursements and invoice processing, the CEO continued. Domestically, the startup competes with Money Forward Cloud Keihi, freee, and Rakuraku Seisan. Globally, its rivals include SAP Concur, Rippling, Brex, Ramp, Spendesk, and Airbase. In the AI Workforce space, it faces competition from Harvey, Fukushima noted.

Bakuraku differentiates itself with an AI-driven user experience. The company continuously upgrades automation features like “auto-entry and document splitting”, while also investing in AI agents and AI-enabled business processing outsourcing (BPO). Its team includes “more than 12 former CTOs and a Kaggle Grandmaster.” Bakuraku offers a comprehensive, integrated platform covering “expense management, invoice processing, corporate cards, workflows, e-ledger compliance, attendance, and receivables—all in one solution,” Fukushima added.

The startup has closed its Series B funding less than two years after securing its Series A in November 2023. Its signature platform, the Bakuraku Suite, has seen significant growth, the seven-year-old company said.

“We passed 10,000 customers in February 2024 and reached 15,000 by April 2025, with more enterprise clients coming on board,” the CEO said. “Headcount has also grown from about 220 employees in October 2023 to around 430 as of the end of July 2025.”

LayerX is on track to reach $68 million, equivalent to ¥10 billion, faster than any SaaS company in Japan’s history, according to the company. “The growth benchmark known as T2D3 was achieved ahead of schedule, and we expect to surpass the previous domestic record, which took eight years from product launch, in under five years,” Fukushima said.

Ai Workforce counts Mitsui & Co. and MUFG Bank among its clients, while Bakuraku serves customers such as Ippudo, IRIS Ohyama, the Imperial Hotel, and Sekisui Chemical.

Looking forward, the company targets approximately $680 million (¥100 billion) in annual recurring revenue by fiscal year 2030, with roughly half expected to come from its AI agent business. It also plans to grow its workforce to around 1,000 employees by 2028.