Techin' Geek

As I approach the conclusion of my 20s later this year, I’ve officially hit the point where sleeping in an awkward position or stretching a bit too much can lead to discomfort. I’ve always had my doubts about massage guns, largely because I’ve tested a handful of off-brand models and figured they’d just end up gathering dust.

However, after several weeks of trying out the Therabody Theragun Sense (2nd Gen), I was taken aback. It turns out massage guns aren’t the fads I had assumed they were.

The Theragun Sense, priced at $299, is a massage gun focused on wellness that aims to alleviate everyday discomfort, relieve muscle tightness, encourage relaxation, and ease soreness. It serves more as a gentle massage and relaxation device than a robust recovery massage gun.

If you’re in search of a more intense deep-tissue massage, consider looking into the Theragun PRO or Prime. But if a heavily featured, powerful massage gun—including heated therapy and deep-tissue functionality—isn’t a necessity for you, the Theragun Sense is an excellent choice for regular use. 

What I appreciate most about the massage gun is its built-in visually guided routines. It offers four routines, and with the Therabody app, you can access even more, syncing your preferred ones to the device via Bluetooth.

The Theragun Sense features an LCD display that guides you through routines step by step. After selecting a routine, the screen instructs you on where to apply the massage gun, how much time to spend on each area, and the pressure to use.

As someone who had little idea of how to utilize a massage gun other than directing it at a tight muscle and hoping for relief, I found the Therabody Sense made the process much simpler.

The routine I enjoyed the most was the six-minute “Sleep” routine, which provides a quick full-body massage to help prepare you for bedtime. The app offers a variety of other routines, including “Pre-Walk Warm-up,” “Desk Relief,” “Lower Back Recovery,” and many more. 

The massage gun includes five speed options to adjust the pressure, allowing for a lighter touch, medium pressure, or a more intense release of tension. I appreciated the ability to alter the intensity based on how sore I felt that day.

As the device is made to be lightweight, I found it easy to handle and access those areas on my back and shoulders that are less accessible.

For those seeking more customization, the Therabody app elevates the guided experience with Coach, the company’s AI feature that formulates tailored recovery suggestions based on your objectives and activity. Coach can provide recommendations on which areas to focus on, what settings or attachments to utilize, and the optimal timing for your recovery sessions. 

I also valued the relatively quiet motor, allowing me to use the device while watching television or relaxing without being disturbed by the noise of the massage gun. 

Included with the Theragun Sense is a USB-C charging cable, a travel pouch, and two massage attachments: the Standard Ball and the Dampener, designed for targeting knots and sensitive areas. For an additional $35, you can purchase the “Supersoft” attachment for extra delicate areas of the body. 

In summary, the Theragun Sense (2nd Gen) won me over. It’s user-friendly, comfortable to hold, and notably eliminates the uncertainty involved with using a massage gun. While priced at $299, it’s not inexpensive, but I consider it a fantastic choice for those seeking a daily wellness device that aids in relaxation and recovery.

Prior to this week, the prevailing narrative about Nvidia was essentially this: Throughout the initial years of the AI surge, Nvidia stood as the sole provider of premium GPUs, which became extraordinarily lucrative as the sector expanded. Recently, hyperscalers like Amazon and Google have begun developing their own chips, resulting in Nvidia no longer being the sole option available, prompting investors to question the sustainability of its edge.

This is a captivating narrative, and largely accurate. After experiencing a tenfold increase in its market capitalization from early 2023 to mid-2025, Nvidia’s stock has followed a more tempered path over the last year, influenced by worries regarding GPU rivalry.

A fresh narrative has emerged following the company’s earnings report on Wednesday, and investors are starting to recognize that Nvidia’s strengths extend far beyond GPUs. As AI’s computational needs escalate to gigawatt levels, orchestration has transformed into a progressively intricate responsibility. Unsurprisingly, Nvidia has developed much of the cutting-edge hardware required to manage it, providing the company with a substantial advantage in the systems surrounding the GPU, despite facing intensified competition in GPU production itself. 

For all the discussions regarding computing as a commodity, managing a megascale data center at optimal efficiency remains extraordinarily challenging — and as deployments become larger and more rapid, this challenge only intensifies.

Rack by Rack

You can observe some of this simply by examining the specifics of what Nvidia is offering. The company is currently introducing its Vera Rubin architecture, which combines the Rubin GPU with various other units, such as the Vera CPU, the Groq 3 LPX inference accelerator, and similar racks for storage and networking.

Over the past week, I’ve engaged with representatives at Nvidia about the functions of these systems, and the findings have been quite revealing. Like the Rubin GPU itself, they are highly specialized systems, but rather than solely processing tokens, they ensure that everything surrounding the GPU operates as efficiently as possible. If the GPU serves as the engine, these components function as the rest of the vehicle.

Particularly, the Vera CPU is concentrated on the challenge of data orchestration. “Vera is significant because there’s a limit to the memory you can fit within a single server or any compute platform,” Jason Hardy, Nvidia’s VP of storage technology, explained to me. 

As data centers have escalated their computing capabilities, the memory capacity has likewise increased, which is why businesses like Micron have thrived in the recent infrastructure boom’s second wave. However, delivering that data to the GPU at the appropriate moment is not trivial — and as firms strive to lower tokens-per-watt, they are coming to realize the critical importance of directing that traffic.

“We observed improvements exceeding 3x in these operations, where the Vera CPU enables acceleration,” Hardy remarked. “Thus, we can now fully utilize our flash without creating bottlenecks.”

Similar versions of this issue can be identified beyond Nvidia. When OpenAI created its Jalapeño chip, a primary goal was to entirely circumvent these obstacles by minimizing data movement.

“We designed Jalapeño to reduce data movement and communication delays,” the company mentioned in a blog entry earlier this month. “Its expansive domain allows the entire workload to remain within a single connected system, thereby minimizing data movement and ensuring that the complete request remains rapid and efficient from start to finish.”

This presents an alternative approach, eliminating data movement by executing a workload within a single integrated chip. Yet the underlying principle is consistent, boosting efficiency through more intelligent traffic management instead of merely relying on increased processor cycles. This, in turn, paves the way for a new tier of infrastructure for companies to vie over.

This new emphasis on data orchestration does not guarantee a victory for Nvidia. The firm will need to contend with competing chipmakers and hyperscalers just as it has with GPUs. However, the competition has evolved to a new dimension, where creating a rival GPU is less pertinent than ensuring the efficiency of the entire system. 

And at least during the initial stages, Nvidia appears to have a significant advantage.

The excitement surrounding humanoid robots is not particularly recent. Credit Tesla’s CEO Elon Musk and his Optimus robot, as well as the numerous videos showcasing Boston Dynamics’ Atlas robot, for that.

However, beneath that excitement lies significant advancement. The physical skills of robots are steadily advancing, and researchers now believe that the AI methods utilized in large language models can enable sophisticated robots to learn virtually any task.

These favorable conditions have motivated a new wave of companies to seize the potential profits from humanoid robots. Many of the newest players are automakers from China.

Earlier this week, Xpeng’s robotics division secured over $900 million at a post-money valuation exceeding $6.3 billion. The funding round, spearheaded by IDG Capital with contributions from Gaorong Ventures, Tencent, and Alibaba, was characterized by the company as the largest single-round private financing ever documented in China’s “embodied AI” sector (AI systems integrated directly into physical devices).

This month, AiMOGA, the robotics division of China’s Chery Automobile, reportedly started preparations for an IPO, while BYD introduced a humanoid robot named Xiao Di. Other Chinese automotive manufacturers, including Changan, GAC, Li Auto, SAIC, and Seres, are also in the process of developing humanoid robots.

Among these, Xpeng stands as the Chinese automaker that closely monitors and adopts Tesla’s strategies, according to Michael Dunne, CEO of advisory firm Dunne Insights, based in San Diego and Singapore.

“It’s the most dedicated to autonomy, and it’s the first to significantly invest in humanoid robots,” Dunne stated to TechCrunch, adding that Xpeng’s founder He Xiaopeng is a tech billionaire noted for his adaptability and swift changes. “He perceives razor-thin profits in cars on the immediate horizon. Robots appear much more promising.”

Xiaopeng and Xpeng co-president Brian Gu are optimistic enough to have invested their own money into the robotics division. According to the Wall Street Journal, the duo contributed around $100 million to the recent funding round.

Xpeng is focusing on Iron, a humanoid robot that has a lifelike human form and is designed for commercial deployment.

Chinese automakers such as Xpeng do offer a manufacturing advantage.

“They possess all the hardware necessary to complete the task,” Dunne remarked. “The question is whether they can keep pace with Tesla in AI.”

Naturally, many other firms are also developing humanoid robots, including Agility Robotics, Apptronik, and Figure, all pursuing the same aim: large-scale commercial deployment.

Hyundai-owned Boston Dynamics is nearing that objective. Hyundai plans to integrate Boston Dynamics’ Atlas humanoid robot into its factory in Georgia this year, with the goal of employing the robots for tasks like parts sequencing by 2028. The Korean automaker, which has teamed up with Google’s AI research lab DeepMind to expedite the development of Atlas, is opening a U.S. facility this year called a Robot Metaplant Application Center, which will instruct robots on how to navigate movements like lifts and turns.

Other automotive companies are also joining the trend, including supplier Mobileye, which purchased humanoid robot startup Mentee Robotics earlier this year for $900 million. Even Rivian is exploring robotics with its Mind Robotics spinout — although its robots are not anticipated to resemble the humanoids under development elsewhere.

I’m a massive film enthusiast. I consume an excessive number of movies and constantly seek out new formats to enjoy them. So, when XREAL, the company behind smart glasses, sent me the a01 — a recent release from May of this year — I was keen to try them out as the latest means of indulging my media obsession.

The a01 doesn’t boast a particularly advanced design. Unlike software-intensive AR glasses such as the Meta Orion or Snap’s Specs, it functions mainly as an external display. There’s no battery or internal power source. Instead, a straightforward USB-C cable connects the glasses to your chosen device, which acts as the headset’s power supply. Additionally, it’s quite affordable, priced at approximately $300.

The a01 is primarily tailored for gaming, being compatible with devices like the Steam Deck or other handheld consoles. However, XREAL markets it for watching films and TV shows as well — and since that aligns more with my interests, I plugged it into my laptop and started streaming the Criterion Channel. I found myself engrossed in David Lynch’s “Wild at Heart,” thoroughly enjoying the moment when Nicolas Cage, clad in a snakeskin jacket, throws punches at a guy in a bar and then serenades with an Elvis track.

I must admit: The visuals are quite impressive. Equipped with dual mini OLED displays, the glasses deliver a sharp image (with a 1080p resolution and up to 1,600 nits of brightness) featuring vibrant colors. Pointing the glasses at a wall gives a sensation akin to watching a bright home projector — or perhaps as if you’re at a drive-in theater. In a dimly lit environment, you’re closer to a true cinema experience.

Nevertheless, the overall experience raised some questions. Specifically, why would I wear glasses while sitting right next to my computer watching a movie that is merely 14 inches away? XREAL claims the glasses offer a more immersive experience (they equate the projections to viewing on a 147-inch screen). Still, the redundancy of watching a film while it also plays nearby makes one wonder about the device’s actual purpose.

XREAL has indicated that the glasses can serve as a “second monitor” (they’ve been dubbed a “wearable display”), but the practicality of this is questionable. It’s quite challenging to focus on anything besides what the glasses are showing — making it tricky to, for example, work on a laptop while wearing them.

The glasses can also link to your phone. Unfortunately, I own an older iPhone, meaning the cable for the a01 doesn’t fit my device’s port. An adapter would be necessary to make the connection.

However, it’s easy to envision the user experience. Linking the glasses to a phone frees you from being confined to indoor environments, unlike bulkier devices such as a laptop. You could watch a film while flying or on a train (or even while strolling down the street — though I’d advise against this option unless you don’t mind risking bumping into traffic).

However, there are still troublesome restrictions when using the device this way. For starters, the glasses function solely as a screen-mirroring device — which means your phone’s screen must remain active while using the glasses. This brings us back to the redundancy issue. You’re viewing a video on a screen in front of your face while the same video plays on a different screen that is mere inches away. You could partially address this by stowing the phone in your pocket, but any movement might disrupt the device’s playback capabilities.

It’s also important to note that the device can be paired with an additional accessory called the Beam Pro, essentially a mini-tablet serving as a standalone streaming hub. Users can download films and shows onto the Beam, connect it to the glasses, and enjoy. However, this accessory will set you back another $200.

Then there’s the issue of heat. The a01 heats up fairly quickly — generating a peculiar tingling sensation on the bridge of your nose and over your eyes. This isn’t a problem unique to XREAL’s devices; it’s a common complaint with most XR glasses. You can only fit so much computing power into a compact plastic frame before the electrical components begin to generate heat. Still, it’s somewhat unsettling and not ideal for an accessory worn on your face.

That said, aside from the heat, the a01 is relatively light and comfortable — not nearly as cumbersome as other smart glasses I’ve tried. (After experiencing Snap’s Specs at CES earlier this year, I can assure you those weigh significantly more — though it’s also a completely different type of device than the a01.)

XREAL is continuously enhancing its product offerings, with each new release appearing to advance beyond the last. Indeed, some of the core dilemmas found in the a01 and earlier XREAL headsets seem to be addressed in the company’s latest (and yet-to-be-released) model: Project Aura — which I glimpsed during my visit to Google I/O this year — promises a much more immersive and seamless experience.

The Aura operates on Android XR, an extended reality operating system developed by Google and Samsung. It features native hand tracking (missing in the a01), as well as access to the Google Play Store, significantly enhancing the glasses’ interactive functions and AR capabilities. Additionally, it comes with a puck connected to the glasses that serves as the charging source and computing unit. The puck’s compact design allows you to easily carry it in your pocket, which provides greater mobility compared to the a01, allowing you to use it without being tethered to a separate device.

Returning to the a01, though, it’s unfortunately not ideal for cinephiles to watch movies through a pair of sunglasses. Generally, film lovers prefer a larger screen — the bigger, the better, in fact. In a landscape populated by 4K OLEDs of varying enormous dimensions, consumers have numerous choices. My TV — a 55-inch TCL S-series — isn’t even particularly high-end, yet it delivers a homely viewing experience far more comfortable and satisfying than the a01 can offer. To me, watching a film at home on a sizable flat screen is second only to visiting a theater. Meanwhile, experiencing a film on miniature screens just inches from your eyes offers an intriguing sense of immersion but is far from optimal.

The a01 serves as an engaging insight into a hardware industry that is continuously evolving and still finding its place with consumers. I’m intrigued to observe how the user experience will transform with XREAL’s upcoming Aura, and I’m open to revisiting my movie-watching preferences when that time arrives.

Training AI systems utilizing other AI frameworks has emerged as a highly sought-after objective for neolabs — and now, an investigator in Anthropic’s fellows initiative has offered us an initial glimpse at how this could manifest in real-world applications.

On Friday, Anthropic released a new study titled “Automated Researchers Can Reliably Mitigate Alignment Failures,” explaining how AI systems might consistently enhance a model’s performance against a series of alignment criteria. When presented with 10 measures for particular misaligned actions, the automated systems succeeded in boosting performance on each one without compromising overall efficacy.

Headed by Anthropic fellow Chen Yueh-Han, the system emulates much of the conventional methodology in research. Each automated entity scans the existing literature, suggests a technique, and trains the model using that technique for 30 minutes, steadily increasing the benchmark over multiple iterations. Successful methods are retained while those that are ineffective are eliminated, enabling the system to function rapidly and on a large scale.

“Overall, these findings offer preliminary proof that automated alignment post-training could be feasible in the near future,” states the paper.

The study is a move towards recursive self-enhancement, which many view as the next critical advancement in AI development. If models are capable of refining their own alignment training, it’s likely they could enhance training methodologies more broadly — at which stage, human AI researchers might soon be rendered unnecessary.

The paper openly confronts this notion, directly contrasting the Automated Alignment Researcher (AAR) with its human counterpart. “The best AAR method outperforms what seasoned humans propose, on average within six hours,” notes the paper. “Human-guided research directions do not yield superior results.”

There’s even a financial comparison, should anyone remain skeptical. “An AAR incurs a cost of approximately $4 per hour in API inference, compared to the $150 per hour allotted for our human researchers.”

In fairness, the paper also acknowledges certain limitations of this method. The automated framework only functions effectively to the extent that the benchmarks accurately align with the genuine alignment objectives, and even then, considerable effort is needed to establish and uphold those benchmarks — not to mention the necessity of maintaining and expanding the literature from which the automated researchers derive their information.

An appealing reason to switch to Brave, an alternative to Chrome, has just been introduced: email aliases. This feature, revealed this week, enables Brave users to register on various websites and online services without disclosing their actual email addresses.

To utilize this feature, you must start by creating an account with Brave, providing your genuine email address. Once logged in to the browser, a pop-up will appear when you click on an email field on a website, offering the choice to utilize an alias instead. (If the pop-up does not show, simply right-click on the field to see the option.) The email alias will automatically fill in the website’s form, and any emails sent to that address will be redirected to your main email account.

The company states that utilizing an alias can safeguard your privacy, especially since websites use your email as a unique identifier, which allows them to target you with advertisements. Advertising technology giants such as Meta can also match the email you submitted to a retailer’s site, for example, with the address they have on file for you, enabling these firms to monitor your purchases or views.

Additionally, if a website where you shared your email is breached, hackers would then have access to your primary email address, potentially leading to the leaking of this information to data brokers.

Brave claims it developed this new feature to address this privacy gap, asserting that it does not examine the content of emails sent to an alias — it merely processes them for spam and virus detection. After the email is forwarded, Brave removes it from its servers. All data in users’ Brave accounts is also encrypted while stored, and any notes saved alongside an alias are stored locally on your device, unless you activate the sync feature. If syncing is active, notes will be end-to-end encrypted.

Since the email alias feature is still in its early stages, Brave has warned that some of the forwarded emails might initially land in your spam folder, but this should improve as Brave enhances its reputation as a mail provider.

For now, Brave users are granted five complimentary email aliases. Additional aliases will be made available for Premium plan subscribers in the future.

All eyes are on Nvidia as it is expected to announce this week’s most intriguing tech acquisition: A purported $13 billion purchase of Hugging Face, a platform dedicated to the sharing of open-weight AI models and benchmarks.

Currently recognized as a prime target for a squad of reward-hacking OpenAI agents, Hugging Face sits at the core of the community of developers focused on creating and implementing LLMs outside the realm of frontier labs. Consider it the GitHub equivalent for the AI age.

Speculations regarding this acquisition follow Nvidia’s $6 billion deal with Poolside, a builder of open-weight models, which will result in the majority of its workforce transitioning to the semiconductor titan. Additionally, two weeks prior, Stripe obtained OpenRouter, the leading supplier of open-weight models for businesses, for upwards of $7 billion.

This influx of investment into a sector that thrives on sharing resources highlights emerging trends in the AI landscape.

For Nvidia, minimizing reliance on partnerships with major hyperscalers and frontier labs is crucial. This is especially relevant as significant AI model developers like OpenAI and Google are simultaneously creating their own inference chips, such as OpenAI’s Jalapeño, which was unveiled this week. If model developers are producing chips, Nvidia aims to secure a piece of the model development pie.

Nvidia has its own Nemotron line of open-weight models, but their adoption has been limited. By annexing the largest developer community in the U.S. focused on open models, the firm will gain access to a large user base that can be directed toward its chips and standards.

Moreover, there are escalating concerns about AI inference costs, prompting companies to investigate more affordable models crafted by Chinese companies such as Moonshot, DeepSeek, and Alibaba. Although current adoption rates stand at a modest level, with only 6% of companies leveraging open-weight models, according to spending data compiled by Ramp, and just 2% of software engineers assessed by Jellyfish, which develops tools for coders.

Nik Albarran, the AI product lead at Jellyfish, informed TechCrunch that open-weight models are mainly utilized by firms whose offerings depend on repeated inference tasks, like customer service chat functionalities. Given that these tasks involve high volumes and frequent repetition, an open-weight model can be optimized for cost-effective responses.

This is notably the framing Stripe has adopted in discussing its OpenRouter acquisition. “Tokens serve as the main currency for companies developing AI solutions, and it’s evident that the tangible economic potential hinges on efficiently managing limited computing resources,” Patrick Collison, Stripe’s co-founder and CEO, remarked in a statement.

In contrast, for coding and agent-related tasks, varying requests and deeper reasoning indicate that frontier models often excel, partly because proprietary labs provide easier access, and sometimes offer token subsidies. Albarran notes that as organizations refine their AI workflows, migrating to open models will become more feasible. Still, the primary motive for companies exploring these models now is for control and adaptability, rather than cost concerns.

“There aren’t many companies that find themselves in that situation yet… [but] if prices continue to rise from frontier labs, an increasing number of companies will have no choice but to at least contemplate it,” Albarran shared with TechCrunch. “When your AI-driven processes are significantly more developed, that’s when investing in self-hosted models truly makes sense.”

Lin Qiao, the CEO of Fireworks, a prominent router and host for open-weight models aimed at corporate clients, is frequently mentioned as a possible acquisition target for a major tech firm. Qiao stated that her company manages 40 trillion tokens daily, surpassing both Gemini’s and OpenAI’s APIs in volume.

Fireworks focuses on model diversity: As LLMs expand and enhance, it will become increasingly simpler for organizations to tailor them to their specific requirements. “Every single app company should contemplate bringing on an internal researcher,” she remarked to TechCrunch last week. “They can utilize their product and the data it generates to develop their own model. The future will truly be about specialized intelligence. Every company ought to have its own model tailored for each use case, and this will occur organically.”

It’s easy to overlook how early we are in AI’s evolution as both a tool and a commercial venture. However, the supremacy of OpenAI and Anthropic is not a foregone conclusion. As tech giants seek to hedge their investments in the largest labs, the charm of open technology is proving to be difficult to resist.