Blogroll

Pine64, the creator of the PinePhone line of Linux-focused smartphones, is apparently killing off the Pro edition of the PinePhone. You can still get the original PinePhone though.

Life with kids can be chaotic. Trust me, I know. So when I find something that works, I want to tell everyone about it. As a mom to a toddler and a preschooler, these are the Amazon finds that I use and rely on every single day.

Do you remember vibrating with excitement at the thought of a new video game when you were a child? What about being planted in front of a screen for a whole weekend, much to your parents’ disapproval?

Meta is quickly becoming the forerunner in Big Tech's chorus against "Woke AI," as it announces a collaboration with known conservative influencer Robby Starbuck to ensure its AI tools are free of "ideological bias."

SEE ALSO: Truth Social's new AI search is 'woke' enough to disagree with Trump

Starbuck, to clarify, is not an expert in developing or training AI tools or LLMs. Instead — as explained in an August 8 statement posted to the X accounts of Starbuck and Meta Chief Global Affairs Officer Joe Kaplan — he will provide guidance to the company's developers on how to further "mitigate" political bias in AI tools. Meta's stance is that removing such "DEI bias" makes its models "more accurate." The advisory role is part of a deal struck in a settlement between Meta and Starbuck, who sued the company for $5 million, alleging the company's AI chatbot falsely claimed he was at the Jan. 6, 2021 Capitol insurrection.

“We don’t want a future where you’ve got AI putting the thumb on the scale when it comes to politics,” Starbuck said in an interview with CNBC, explaining his goal was "ethical" and "neutral" AI. Starbuck has previously advised the Federal Communications Commission (FCC) and head Brendan Carr on eliminating DEI and equal opportunity initiatives in telecommunications, a strategy that includes withholding FCC approvals for companies that don't comply with the Trump administration's mandates.

This Tweet is currently unavailable. It might be loading or has been removed.

The move comes after President Trump announced a new federal AI Action Plan and executive order targeting what conservatives have deemed "Woke AI," or Large Language Models with ideological or political "bias" that promote liberal beliefs, including the concept of Diversity, Equity, and Inclusion. Many considered the move to be a capitulation to Big Tech's leaders, most of whom funded and now advise the Trump administration.

Starbuck, a former music video director, has become a leading figure among conservative social media influencers, gaining popularity for videos that attack companies with diversity initiatives and galvanize his followers to boycott or pressure them to capitulate to right wing demands by removing DEI policies. Like many other conservative influencers, Starbuck often shares alarmist comments about gender affirming healthcare for trans youth, refers to liberal commentators as "mentally challenged" and "lunatics," and has declared DEI policies a "war on men." He's also a visiting fellow at the Heritage Foundation, the conservative think tank behind that crafted the controversial Project 2025 movement.

SEE ALSO: Metadata on U.S. government memos reveals authors linked to Project 2025

Even before the president's sweeping anti-DEI orders, Meta had already struck down its DEI mandates and related policies, including protections for LGBTQ communities, women, and people of color. Meta founder Mark Zuckerberg has been gradually shifting the company's policies and priorities to the right in the name of "free speech" — echoing the sentiments of the current administration — including donating $1 million to the Trump campaign in the company's name.

"Since engaging on these important issues with Robby, Meta has made tremendous strides to improve the accuracy of Meta AI and mitigate ideological and political bias," the company said in a statement.

Privacy-focused phones are not a new concept, even as they might be a little gimmicky and silly. You need to make a lot of uncomfortable compromises to go truly locked down. Another one of those privacy-first devices has just popped up, but you should probably not buy this one.

Training AI models requires choosing an optimizer and for nearly a decade, Adam( (opens in new tab)–W) (opens in new tab) has been the optimizer of choice. Given that durability and success, it was fair to doubt that any further improvement was possible. And yet, last December, a new optimizer called Muon (opens in new tab) showed serious promise by powering a nanoGPT speedrun (opens in new tab). This proved out, with multiple AI labs (e.g., Kimi-AI (opens in new tab) and Essential-AI (opens in new tab)) reporting 2x scale improvements and the release of the 1T parameter Kimi K2 (opens in new tab) model. Restated: you can train a model to similar performance with half as many GPUs.

There’s one fly in the ointment: Muon requires large matrix multiplications in the optimizer, which requires heavy communication in large models at the scale where FSDP (opens in new tab) and TP (opens in new tab) parallelization becomes desirable. Going back to the inspiration for Muon, (opens in new tab) the key idea is an orthonormal update, which sparked the search for more scalable alternative linear algebras realizing the same goal. That’s exactly what Dion is. We have open-sourced this new optimizer to enable anyone to train large models more efficiently at scale.  

What’s an orthonormal update? Figure1. Illustration of matrix parameters

At the core of Transformers, a set of input activations is multiplied by a learned weight matrix to produce a new set of output activations. When the weight matrix is updated during training, the resulting change in the output activations generally depends on the direction of the input activations. As a result, the learning rate must be chosen conservatively to accommodate the input direction that induces the largest change. Orthonormalized updates alter this behavior by (approximately) making the change in output activations invariant to the direction of the input. This is achieved by enforcing orthonormality (opens in new tab) on the update matrix, thereby equalizing its effect across all input directions.

What is Dion?

While Muon has shown strong empirical results, scaling it to very large models poses challenges. As reported by Essential AI (opens in new tab), applying Muon to large architectures like LLaMA-3 becomes compute-bound—and potentially communication-bound—due to the cost of the Newton–Schulz orthonormalization steps (opens in new tab).

Figure 2. Pseudocode of the centralized version of Dion

This is where Dion enters. At a high level, Dion introduces a new axis for scalability: the rank. Specifically, for a given rank r, Dion orthonormalizes only the top r of the singular vector space, reducing communication and compute overhead while preserving performance. Empirically, we observe that the necessary rank for good performance grows much more slowly than the number of parameters in larger models.

Download Dion optimizer 

Dion implements orthonormalization using amortized power iteration (opens in new tab). Power iteration typically pulls out the largest singular value by repeated matrix multiplication. By amortizing this process over optimization steps—applied to the slowly-evolving momentum matrix—we reduce the cost to just two matrix multiplications per step. Incorporating a QR decomposition allows us to extract an approximate orthonormal basis spanning the top singular directions, rather than just the leading one. This amortized power iteration is fully compatible with standard distributed training techniques such as FSDP and tensor parallelism. Here, we show a simple centralized version, but the technique works for more complex forms of parallelization as presented in the paper. In other words, we can orthogonalize a matrix without ever seeing a full row or column of it. 

Low-rank approximation would ordinarily introduce error, but Dion overcomes this through an error feedback mechanism. This keeps the residual of low rank approximation in the momentum matrix so that any systematic gradient structure not initially captured accumulates to eventually be applied in a future update.

PODCAST SERIES

The AI Revolution in Medicine, Revisited

Join Microsoft’s Peter Lee on a journey to discover how AI is impacting healthcare and what it means for the future of medicine.

Listen now Opens in a new tab How does it work?

Something very strange happened in our experiments. Usually, adding an extra constraint on the way an algorithm works can be expected to decrease overall performance. And indeed, at the 120M parameter scale of the speedrun, we see Dion’s update taking more time than Muon, while not yielding any significant gains. But at larger scales, we observed a different trend: Dion began to outperform Muon.

Figure 3. Wall-clock time speedup of Dion for 3B model training

Why would adding a constraint improve the update rule? The answer lies in what the constraint enforces. Dion achieves a much closer approximation to true orthonormalization than Muon. This precision, initially subtle, becomes increasingly important as the number of singular vectors grows. Over increasing model scale and training steps, this small advantage accumulates—leading to a measurable improvement in performance.

This edge further grows with batch size—with larger batches the update quality tends to degrade, but notably more slowly with Dion than Muon (and Muon is already a significant improvement over AdamW).

Figure 4. Scaling of Dion across different batch sizes

Here you can see how the number of steps to reach a pretraining loss compared to AdamW varies as batch size grows with full rank and ¼ rank Dion (in orange) and Muon (in blue).   

In our experiments, these benefits extend to various post-training regimes as well.

We also experimented with rank, discovering empirically that larger models tolerate smaller rank well.

Figure 5. Low-rank Dion across different model sizes

Projecting this trend out to the scale of the LLaMA-3 (opens in new tab) 405B parameter models suggests that Dion is fully effective even with rank fractions as low as 1/16 or 1/64 for large dense models like LLaMA-3.    

Using hardware timings of the individual update steps suggests a story that looks this:

Figure 6. Estimated wall-clock time of each optimizer step for Llama 3 405B. Lower is better. Muon is highlighted in orange as our baseline, next to Dion with varying rank fractions. Suggested rank fractions for a 405B parameter model are shown in blue. Using Dion with rank fraction 1/16 or lower offers an order-of-magnitude speedup over Muon.

We’ve open-sourced a PyTorch FSDP2 + Tensor Parallel (TP) implementation of Dion, available via a simple pip install. Our goal is to make faster training with Dion accessible to everyone. As a bonus, the repository also includes a PyTorch FSDP2 implementation of Muon.

Dion optimizer Acknowledgements

We thank Riashat Islam and Pratyusha Sharma for their helpful feedback on the writing and presentation.

Opens in a new tab

The post Dion: the distributed orthonormal update revolution is here appeared first on Microsoft Research.

Although macOS isn’t known for being the most configurable operating system, the System Settings app provides many options. You name it; everything, from wallpaper slideshow timings to Dock magnification, is customizable. But more settings are hiding just below the surface, and this app can help you uncover them.

The current Nissan Leaf is in its final lap, with a next-gen crossover set to take its place soon. It’s been around since 2018 in this form and, while it never lit up the sales charts, it’s quietly remained one of the most important and most affordable electric cars out there.

Razer has released the Razer Wolverine V3 Pro 8K PC, a brand new wireless controller designed for competitive PC gaming. This controller is packed with top-tier features that you expect from a brand name like Razer, but the price tag matches the features.

Finding a good budget phone can be a challenge, but there are some neat options out there. T-Mobile's Revvl line is a pretty good option for those wanting to save a buck, and the newest member is here—and while I think there are still some better options, there are a few reasons why this could also be fine for some.

Syncthing, the cross-platform file synchronization tool that doesn’t need cloud storage, has finally released its big 2.0 update. The new version has some great performance and reliability improvements, updated terminal commands, and much more.

From the widely memed "Labubu Matcha Dubai Chocolate" slang overload to viral brands that everybody on your timeline seems to go crazy about, the ripple of social media influence can be widely felt in every app you use.

The AI wars roll on — this time with Anthropic stepping right in behind OpenAI by offering its Claude chatbot to the federal government for the bargain price of just $1 a year.

The San Francisco–based company becomes the latest AI player to pitch its flagship LLM to Washington, a move widely seen as a bid to win favor with President Donald Trump’s administration. The announcement on Tuesday comes less than a week after OpenAI revealed a nearly identical deal, making ChatGPT available to the General Services Administration for the same token fee.

SEE ALSO: Anthropic reportedly cut OpenAI access to Claude

According to the Financial Times, the Claude agreement clearly states that federal agencies aren’t obligated to use the chatbot at all. Even if they do, Claude’s use will be limited to sensitive but unclassified work.

Anthropic, Google, and OpenAI have only recently been cleared to supply their chatbots to the US government. According to the Financial Times, Google is already working on a similar arrangement to offer its Gemini AI to federal agencies at a steeply discounted rate.

Josh Gruenbaum, commissioner of the Federal Acquisition Service — which oversees procurement for US agencies — told the FT that the goal is simple: “get widespread adoption [of AI tools] in the federal government.”

So far, multiple federal agencies have already begun experimenting with AI tools. The Pentagon has awarded $200 million in contracts to Anthropic, Google, OpenAI, and xAI. Wired has also reported on AI being used inside agencies like the GSA and HUD to identify redundant federal regulations, though according to their reporting, the results have been, at best, mixed.

Gruenbaum told the Financial Times that the government has no official preference for one AI provider over another. Still, it’s worth noting that President Trump has made it clear the White House will refuse to do business with what it calls "woke AI" — a label applied to any chatbot it deems to be pushing "partisan bias or ideological agendas."

Disclosure: Ziff Davis, Mashable’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.

National Public Data — the online background check and fraud prevention service targeted by hackers in what became one of the biggest social security breaches ever — is back. It may still pose a security threat.

SEE ALSO: AOL will pull the plug on dial-up internet, '90s nostalgia ensues

In 2024, the private company announced a malicious actor had gained access to its systems in a Dec. 2023 breach, leaking information for several months to the web's black market. It took National Public Data even longer to disclose details of the breach to the public, which exposed nearly three decades of Social Security records — it was estimated that the hackers accessed hundreds of millions of records. Data included Social Security numbers, full names, addresses, emails, and phone numbers, prompting many to freeze their credit and go on the defense against identity theft.

The site shut down last year following the scandal, but now it appears that National Public Data is back in the personal information game. PC Mag reports that the domain is once again active and running with a new, unnamed owner. The site, which is operating as a people search engine, appears to still store personal data pulled from publicly available sources, including federal databases.

At the bottom of the homepage, the new National Public Data links to a statement on the 2023 breach. "The security of our users’ data is the most important factor for us. That’s why we protect it by implementing robust encryption protocols, regularly updating our security systems, and complying with all relevant data protection regulations," the site's FAQ reads.

How to opt out of National Public Data's search

Like other people search engines, individuals must request their personal information not be displayed on the site. To do so, users have to fill out the company's Opt Out Form. Individuals will need to first find and copy the link to their personal data profile on the site itself.

Individuals can also use a personal data removal service to scrub their information from multiple sites at once or manually remove their information from search sites like Google.

Elon Musk, as is his way, fired off an accusatory tweet Monday night. The target of his ire? Apple.

He posted on X: "Apple is behaving in a manner that makes it impossible for any AI company besides OpenAI to reach #1 in the App Store, which is an unequivocal antitrust violation. xAI will take immediate legal action."

Unfortunately for Musk, his own social media site quickly fact-checked his statement. Community notes on his post noted that both Deepseek and Perplexity had hit the top of the App Store charts, even after Apple and OpenAI announced a partnership in 2024.

This Tweet is currently unavailable. It might be loading or has been removed.

Still, Musk kept posting about Apple and his belief that it was tamping down on the popularity of xAI and its signature AI chatbot, Grok.

"Unfortunately, what choice do we have?" he posted on X. "Apple didn’t just put their thumb on the scale, they put their whole body!"

Musk has not yet provided evidence that Apple has suppressed his apps. As of this writing, ChatGPT is the top app on the Apple App Store charts; Grok is in the fifth spot, lagging behind Threads.

OpenAI CEO Sam Altman didn't take the accusation kindly. He quote-tweeted Musk's post and wrote, "This is a remarkable claim given what I have heard alleged that Elon does to manipulate X to benefit himself and his own companies and harm his competitors and people he doesn't like."

This Tweet is currently unavailable. It might be loading or has been removed.

The conflict this week between Musk, Altman, and Apple is yet another example of the billionaire men running the world getting into petty, public spats. Remember the Musk, Trump feud? Or Zuck and Musk? Or previous Altman vs. Musk fights?

Still, that's not to say there haven't been controversies and troubles on Apple's end regarding its App Store. Epic Games, for instance, has been locked in a years-long legal battle with Apple over listing its popular game Fortnite in the App Store. The judge in the case actually found in May that Apple violated a court order prohibiting the company from engaging in anti-competitive behavior.

It's not clear if Musk will actually take legal action, but it's interesting to see the tech giants going at one another in a public space.

AOL is officially ending its dial-up service after 34 years. How much were users paying for the internet via phone line service with the iconic beep-boop-screeching sound?

According to an announcement last Friday, AOL said it its dial-up service would be discontinued on Sept. 30. For those curious about how much it cost for a paltry 0.056 megabits per second (compared to today's standard of 500 Mbps) we looked into the cost of AOL dial-up.

How much are AOL users paying for dial-up?

At this point, even signing up for AOL dial-up is tricky, and there's not much information about how to get it. According to AOL's dial-up page, the necessary software for their dial-up service to function is only available for Windows Desktop. For an "optimal browsing experience," AOL recommends installing the AOL Dialer, which connects the AOL network and your computer, and the AOL Shield Browser, which is optimized for older operating systems.

The AOL Dialer requires users to purchase an AOL Advantage Plan. Trying to find the cost of the Advantage Plan on the AOL website is a fruitless series of rabbit holes that eventually takes you to a page offering various bundles for customer support and identity theft services, ranging from $11.99 to $16.99 a month.

However, internet comparison site Allconnect says the AOL Advantage Plan costs $9.99 a month. An old CNET article from 2006 (via the Wayback Machine) said AOL increased the cost of its dial-up plan to $25.90 per month in a effort to push customers to its DSL plan, which cost the same at that time. So it looks like the cost came way down as dial-up customers dwindled. According to 2022 US Census data (via Ars Technica) around 175,000 households in the U.S. still used dial-up.

AOL dial-up is dead, but dial-up lives on elsewhere

AOL dial-up might be dead, but there are still providers offering internet through phone lines. For some people in the rural U.S., dial up or WiFi hotspots are the most viable options because of the high cost of an internet subscription or connectivity dead zones. The same Allconnect story lists the pricing for several different dial-up providers. NetZero offers 10 free monthly hours of dial-up, and then charges $11.95 per month. Earthlink's dial-up plan is $9.95 a month for the first three months then $24.95 a month after that.

Compared to the current costs of fiber optic internet plans which range from $50 to $250 a month, these prices are a steal. But you pay for what you get. So the real cost is a frustratingly slow browsing experience.

After a very Hollywood-heavy fourth season, Only Murders in the Building returns firmly to New York in Season 5. And this season, it looks like the podcasting trio of Charles-Haden Savage (Steve Martin), Oliver Putnam (Martin Short), and Mabel Mora (Selena Gomez) are taking on their most New York foes yet: the literal mob.

SEE ALSO: 'Only Murders in the Building' Season 4 ending explained: Who killed Sazz and why?

According to Season 5's first trailer, Charles, Oliver, and Mabel believe that the mob may have had something to do with the murder of the Arconia's doorman Lester (Teddy Coluca) in the Season 4 finale. The mob angle would certainly make sense, given that the Season 4 finale also introduced new player Sofia Caccimelio (Téa Leoni), whose missing husband Nicky — "the dry cleaning king of Brooklyn" — reportedly had ties to the Caputo crime family.

But the crime families of New York aren't the only suspects on display in the trailer. Only Murders in the Building also introduces a trio of billionaires — played by new guest stars Christoph Waltz, Renée Zellweger, and Logan Lerman — who keep sniffing around crime scenes. According to Oliver, they're "the new mob of New York." How are they connected to the "old mob," or to Lester?

Waltz, Zellweger, and Lerman are just a few of the new additions to Only Murders in the Building's star-studded list of guest stars. Bobby Cannavale, Beanie Feldstein, Jermaine Fowler, Keegan-Michael Key, and Dianne Wiest are also joining up. Meanwhile, Meryl Streep returns, alongside Da'Vine Joy Randolph, Richard Kind, Nathan Lane, and Michael Cyril Creighton.

Will the Only Murders in the Building crew solve yet another murder and avenge poor Lester? And how does Nicky Caccimelio's disappearance tie into Lester's demise? We'll just have to wait until Season 5 to find out.

Only Murders in the Building premieres Sept. 9 on Hulu.

Your laptop, smartphone, tablet, and gaming handhelds don’t need power to function, but the same can’t be said about your internet connection. In the event of a power outage, being able to get your network back up quickly can allow you to resume work or play while the issue gets fixed.

TL;DR: Grab Microsoft Office Pro 2021 at just A$76 (reg. A$338) and enjoy lifetime access to essential productivity tools.

Grab Microsoft Office Pro 2021 for Windows at just A$76 (reg. A$338) and enjoy lifetime access to essential productivity tools.

What you get with your purchase

• One-time purchase, lifetime access — no monthly fees, just essential Office apps

• Includes must-have tools like Word, Excel, PowerPoint, Outlook, Access, Teams, and Publisher

• Designed for Windows users — perfect for work, school, and personal projects

• Boost productivity with reliable tools for documents, data management, and communication

• Quick and easy setup that allows you to download it instantly and activate with your product key

• No hidden fees — pay once and use Microsoft Office Professional 2021 forever

Skip the subscription headache — don’t miss this deal and get a lifetime license to Microsoft Office Professional 2021 for Windows for A$76.

StackSocial prices subject to change.

Opens in a new window Credit: Microsoft Microsoft Office Professional 2021 for Windows: Lifetime License AU$76
AU$338 Save AU$262 Get Deal

Taylor Swift announced her 12th studio album Tuesday in a clip from a coming episode of Travis Kelce's New Heights podcast, revealing its blurred-out cover art and a title: The Life of a Showgirl. The full episode will drop late Wednesday, marking Swift's first major interview in nearly two years. Long story short, we're so back.

This Tweet is currently unavailable. It might be loading or has been removed.

As a pop music junkie, the clip had me bracing myself for a new, orange-drenched era of lore and easter eggs. (For the non-Swifties, Taylor uses a different color to represent each of her albums; this one got a bright tangerine hue.) But as a shopping reporter who covers tech gadgets, I was also on my detective shit. I couldn't help but wonder: Which wired headphones is Taylor wearing in her New Heights appearance?

A quick reverse Google Images search led me to the $109.99 Sony MDR-7506s, and a rep for Sony confirmed that they're a match. Coincidentally, Amazon currently has them on sale for $10 off. (That's Tayvoodoo in action.)

Opens in a new window Credit: Sony Sony MDR-7506 Professional Stereo Monitor Headphones $99.95
$109.99 Save $10.04 Shop Now

It's worth pointing out that there's a good chance these headphones were provided to Taylor by the podcast or studio where she recorded, and they're not a pair she owns herself. That said, they are different from the ones Travis and his brother/co-host Jason are wearing in the clip.

In any case, the MDR-7506s feature a foldable, closed-ear design with 40mm drivers. Mashable's headphones expert Bethany Allard, who dabbles in podcasting in her free time, hasn't tried them herself but tells me they're quite popular in the audio world.

"They have great sound quality for editing, are comfortable to wear for longer periods of time, and aren't super expensive," Allard said, "so they're accessible to people at all levels of audio." She pointed out that they're mentioned in a 2024 gear guide compiled by Transom, an audio storytelling nonprofit. They've also been name-dropped by one of NPR's former Tiny Desk engineers.

Taylor's full New Heights episode will be available at 7 p.m. ET on Aug. 13. While The Life of a Showgirl doesn't have an official release date yet, her online store notes that it will ship on vinyl before Monday, October 13.

You can watch Swift's New Heights appearance on YouTube or listen to it on Apple Podcasts, Amazon Music, Audible, Pandora, Spotify, and Wondery.

UPDATE: Aug. 13, 2025, 12:45 p.m. EDT This story has been updated to note that the Sony MDR-7506s are now discounted on Amazon.