Bass is foundational. A dedicated speaker capable of reproducing convincing bass and sub-bass not only makes music and films feel more authentic, but can even improve the detail and dynamics of the higher registers by freeing up sonic space below. So attempting to jam a real subwoofer into a handheld speaker is by no means a crazy idea—it’s just crazy difficult.
That’s what makes the Brane X so impressive. Over the years, I’ve heard lofty claims about low-extension “woofers” in virtually every kind of compact speaker, but the Brane X is the first to hit all the notes literally. It does so with a specialized RAD (Repel Attract Driver) that seems to perform physics-defying magic from within the speaker’s stout frame. That’s matched by multiple high-frequency drivers above for clear and surprisingly expansive performance across registers.
To land this kind of breakthrough bass in a package you can take virtually anywhere, Brane charges a fee that will make most casual buyers spit up their beverages. But if your budget is negotiable, the Brane X is a fabulous wireless speaker that comes as close as I’ve heard to putting a full home audio system in a pint-sized box.
Surreal Sound
Do you listen to music with percussion or drums? Bass guitar, synth, or strings? If so, you’ll benefit from what the Brane X can do. What’s most notable about the speaker’s potent punch isn’t how much bass it pumps out, but the kind and quality. This is pure, full, and unadulterated low-frequency performance from the source.
Photograph: Ryan Waniata
As you might guess, the Brane X is most viscerally impressive when hammering hip hop and electronic grooves, where the kick hits with clean authority well below 50 Hz, without overpowering the other instruments. Yet, listening through my catalog, it became increasingly clear that the RAD subwoofer’s pointed touch is just as impressive when more subtly deployed.
The haunting vocal and guitar intro in Springsteen’s “The Ghost of Tom Joad” sounds clear and airy, as you’d expect from a quality portable. Then the bass guitar kicks in, and everything is elevated. The rich natural tone the Brane X reproduces adds depth and gravity to the song, bringing its full emotional poignancy to life.
I didn’t have a suitable portable that could stand up to the Brane X, so I called in some of my favorite homebound Wi-Fi speakers, including the Sonos Era 100 (9/10, WIRED Recommends) and Era 300 (9/10, WIRED Recommends). Even the thrilling 300, which serves up impressive bass in its own right, couldn’t match the Brane in the heaviest tracks, sketching out more of an impression of the lowest tones than the full monty.
Brane says its Repel Attract Driver can deliver this kind of authenticity by using a special magnet and spring system to cancel the “internal air pressure forces” that hinder other speakers. This results in a claimed “tenfold increase in sub-bass sound and a hundredfold increase in sub-bass efficiency.” It sounds like so much techno mumbo jumbo out there, but with the Brane X, you can actually hear it working, and not just from the performance. Tapping the power key evokes little whirring sounds, apparently adjusting its complex internal mechanism ahead of playback.
The Mac cursor doesn’t have to be boring black-and-white. Image: D. Griffin Jones/Cult of Mac
The Mac pointer (also called the cursor) is black with a white outline by default, but you can change the pointer color if you want to make it more prominent. Plus, you can even change the Mac pointer size.
You might lose track of the small cursor if you have low vision — or just a really big monitor. Making the pointer larger can make it easier to see.
Changing the pointer color to, for instance, bright red can make it easier to see, too. You can choose any color, like light purple or green, to fit the aesthetic of your setup. (This is especially cool if you have a matching colorful iMac.)
Here’s how to change your Mac pointer size and color.
Change pointer color and size on Mac
Time needed: 5 minutes
How to change pointer color and size on Mac
Open Accessibility Settings
On your Mac, go to the menu bar in the top left and click > System Settings… to open the settings panel. In the sidebar, scroll down to click on Accessibility, then click on Display.
Change Mac pointer size
To make your Mac’s cursor bigger, scroll down to the Pointer section. Move the Pointer size slider to the right to embiggen it. Just the first notch might be enough to make the pointer easier for you to see, as it doubles in size. You can make it up to 6× bigger if you need or want to.
Change your Mac pointer color
Look a little further down in the Accessiblity Display settings and you can edit the pointer color — both the outline and its fill color. Click the color, and a pop-up window will let you pick a color. I recommend clicking the crayon tab for a simple color palette. You can pick from shades of gray or all the colors at three different levels of brightness. If you pick one of the lighter colors, I recommend you set the outline color to black.
Some people go all-in on creating a themed Home Screen for their iPhones. Similarly, I think if you have a colorful aesthetic setup on your Mac, you might like changing the pointer color so you have a cursor to match.
Alternatively, if you’re a lifelong PC user who recently made the switch to a Mac, you can go with a simple black outline and white fill to match the color scheme used on Windows.
Another setting you should make sure you have checked in this settings panel is Shake mouse pointer to locate. Once enabled, if you lose track of your pointer’s location on the screen, you can shake your mouse on your desk (or wave around on the trackpad) to enlarge the cursor temporarily.
Hello Nature readers, would you like to get this Briefing in your inbox free every day? Sign up here.
The most dimorphic species was the northern elephant seal (Mirounga angustirostris), where males had a mean mass 3.2 times that of females (Michele and Tom Grimm/Alamy)
An analysis of more than 400 mammal species is overturning Charles Darwin’s assertion that in most, males are larger than females. In fact, this is only the case for 45% of species, with 39% showing no sexual size dimorphism and 16% having larger females than males. “There’s been this really strong inertia toward the larger male narrative, but it was just based on Darwin’s hand-wavy statement, and the evidence doesn’t really support it,” says evolutionary biologist and study co-author Kaia Tombak.
US President Joe Biden has proposed modest increases in federal spending on science and innovation for the 2025 fiscal year, including a 20% increase for the Environmental Protection Agency and more than 5% for the Centers for Disease Control and Prevention. Similar proposals in previous years have run up against opposition among lawmakers in Congress, who decide how much the government will spend. Biden’s budget “is nothing more than a showcase for the policies and the spending that the White House would like to pursue if it had the ability to do so, which it doesn’t,” says Michael Lubell, a physicist who tracks science-policy issues. Meanwhile, negotiations continue over the 2024 budget: much of the federal government — including the National Institutes of Health, the world’s largest public funder of biomedical research — is poised to shut down in less than two weeks unless lawmakers act.
The release of OpenAI’s Sora text-to-video AI tool last month was met with a mix of excitement and trepidation from researchers who are concerned about misuse of the technology. Some say the result will be a deluge of fake but compelling video and audio, which poses particular problems for upcoming elections. But such tools could be used for good too, such as communicating findings to a lay audience and enabling researchers to wade through huge datasets.
Sora is one of several AI tools that generates video from text promptsCredit: OpenAI
This footage isn’t real — it was created by the AI tool Sora. The system seems to be able to produce more photorealistic and compelling content from simple text prompts than other video-generating algorithms. (OpenAI)
Features & opinion
WebAssembly can compile code written in dozens of programming languages, including C, C++ or Rust, and works inside a browser — no installation required. Researchers are using it to build analysis tools for scientific data, which can then be run even on low-power laptops or phones. But users must brace themselves for complications. “You often have to modify the original code to get around things that WebAssembly doesn’t support,” says bioinformatics software engineer Robert Aboukhalil. Check out a demo from Nature technology editor Jeffrey Perkel.
Most PhDs are assessed on their final dissertation — something that hasn’t changed for at least half a century. Reform is urgently needed, argues a Nature editorial. Part of the problem is that few doctoral supervisors have teaching qualifications or awareness of education innovations. “At a minimum, doctoral supervisors should take the time to engage with the research that exists in the field of PhD education, and how it can apply to their interactions with students.”
Last week, The Lancet published a conversation-sparking editorial arguing that “commercial companies and individuals with vested interests have over-medicalized menopause”. The oversimplified idea of menopause as a health problem to be solved by hormone replacement therapy (HRT) is not based on evidence, argued a series of papers and perspectives in the journal. “Menopause does not herald the start of a period of decay and decline but is a developmental life stage,” says the editorial.
In UnHerd, physician Margaret McCartney and health journalist Deborah Cohen dig into the evidence for the benefits of HRT and the therapy’s cultural and historical context. “There was a time when women were expected to shut up and put up” with symptoms such as hot flashes, they write. Now, in some ways, the pendulum has swung too far in the other direction, says public health researcher Brooke Nickel, who studies the language used to sell HRT products to women. “They’re really spinning that feminist narrative around,” she says. “If you don’t do something, you’re not taking control of your body or you’re being oppressed by other people.”
These adornments made from the teeth of animals including bear (1), elk (4) and fox (6,7) were crafted by Gravettian hunter-gatherers — the culture responsible for the iconic Venus of Willendorf figurine. Researchers analysed thousands of such personal ornaments, along with genetic data. They found that the variety in the jewellery styles couldn’t be fully explained by how far apart groups lived, and identified nine distinct cultural groups existing in Europe between 34,000 and 24,000 years ago. These groups mostly align with genetic data, but also reveal more nuanced patterns, indicating that culture and genetics are interconnected but not perfectly aligned. (Nature Human Behaviour | 43 min read) (Jack Baker et al/Nature Human Behaviour)
Quote of the day
The science is clear, argue epidemiologists Richard Wilkinson and Kate Pickett: places with large gaps between rich and poor tend to have worse health statistics, more violence and worse pollution than do more-equal countries. (Nature | 9 min read)
I’m paralyzed with jealousy that I won’t be witnessing the total solar eclipse visible in parts of the United States, Canada and Mexico on 8 April. But I can still enjoy the fact that researchers will be observing how animals react to the unusual experience. During a 2017 eclipse that passed over a South Carolina zoo, “to our astonishment, most of the animals did surprising things,” says biologist Adam Hartstone-Rose.
I hope you’ll eclipse my inbox with your feedback on this newsletter. Your e-mails are always welcome at [email protected].
When venturing into the world of language models, it’s tempting to think that the bigger the model, the better it will perform. This notion is rooted in the belief that more data and more parameters mean a model can do more. But the reality is not so straightforward. The ideal size for a language model depends on what you need it to do and the limitations you’re working with. This guide will help you figure out the best size for your language model by considering the context, the influences, and the requirements of the application.
It’s important to understand that a larger model isn’t always the best choice. While having more parameters can help a model process and generate text that sounds like a human wrote it, there’s a point where adding more data doesn’t improve the results. This happens because of overfitting, where a model gets too tuned to the data it was trained on and can’t handle new information well.
The context in which you use a language model is key to deciding the best size for it. If you need a model for simple text predictions, like finishing a sentence, you might not need as many parameters as you would for more complex tasks, like translating languages or creating original content. It’s crucial to know what you want your language model to do so you can find the right balance between size and usefulness.
What is the optimal LLM size
Here are some other articles you may find of interest on the subject of large language models :
There are several things to think about when picking the size of a language model. These include the computational resources you have, the variety and quality of the training data, what you want the model to do, and the model’s design. Bigger models need more computing power and memory, which can be costly and might not be necessary for every project. The quality of the training data is just as important; a model trained on a big but low-quality dataset might not do as well as a smaller model trained on high-quality data.
Areas to consider when choosing a large language model
To figure out the right size for your language model, you need to consider the trade-offs between the model’s complexity and what you need it to do. Start by defining the goals of your language model. What tasks should it handle? How accurate and flexible does it need to be? Once you have a clear set of requirements, you can start to think about the right size. Looking at existing models that do similar things can give you a starting point. Testing and refining your model will help you fine-tune its size to get the best balance, making sure it’s not too weak or unnecessarily big.
Purpose and Complexity of Tasks:
Different tasks require different levels of language understanding and generation capabilities. A model designed for simple text predictions (like autocomplete features) may not need as many parameters as one intended for complex activities such as generating coherent long-form content or understanding nuanced conversations.
Overfitting Risks:
Larger models, with their vast number of parameters, can become too finely tuned to the training data. This overfitting makes them less adaptable to new, unseen data, reducing their generalization capabilities.
Computational Resources:
Running larger models requires significant computational power, including advanced GPUs and substantial memory. This necessitates a cost-benefit analysis, as the expenses (both financial and energy-related) can be considerable.
Training Data Quality and Variety:
The diversity and quality of the training data are crucial. A model trained on a vast but poorly curated dataset might perform worse than a smaller model trained on well-selected, high-quality data.
Model Design and Architecture:
The efficiency of a model isn’t just a function of its size; it’s also about its design. Innovations in model architecture can lead to more efficient processing, potentially reducing the need for a larger number of parameters.
Balance Between Size and Usefulness:
It’s essential to strike a balance where the model is adequately sized for its intended tasks without being unnecessarily large, which could lead to inefficiencies and increased costs.
Testing and Refinement:
Rigorous testing helps in understanding the actual performance of the model. Continuous refinement based on these results can lead to optimizing the model size, ensuring it’s neither too small (underperforming) nor too large (wasteful).
Context of Use:
The environment in which the model operates is a key consideration. For instance, a model used in real-time applications may need to be smaller and more efficient, whereas size may be less of a constraint in non-real-time, research-focused applications.
Cost vs. Performance Trade-Offs:
Larger models generally come with higher operational costs. It’s important to evaluate whether the performance improvements justify these additional costs.
Benchmarking Against Existing Models:
Examining similar models in the field can provide insights into the necessary size and capabilities for specific tasks. This benchmarking can serve as a guideline for setting initial expectations and goals.
Goal Definition:
Defining clear, quantifiable goals for what the model needs to achieve helps in determining the optimal size. This includes setting specific targets for accuracy, response time, adaptability, and any other relevant performance metrics.
Choosing the perfect size for a language model is a complex decision that requires careful consideration of many factors. It’s not just about how many parameters there are, but also the context, the quality of the data, and what you need the model to do. By taking a thoughtful approach to these aspects, you can customize your language model for its specific purpose, finding a good balance between how well it works and how efficient it is. The goal is to find the sweet spot where the model’s size and performance match your unique needs.
Filed Under: Guides, Top News
Latest timeswonderful Deals
Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, timeswonderful may earn an affiliate commission. Learn about our Disclosure Policy.
