• Xiaomi reveals its first extended-range electric vehicle
• SkyNomad will sit separately from its pure EV Xiaomi Auto company
• The 1.5-liter engine is manufactured by Changan’s subsidiary Harbin Dongang

Xiaomi is set to enter the hotly contested luxury SUV sector with an all-new business that it has dubbed SkyNomad.

Fresh off the success of both the SU7 and YU7, the former of which has outsold the Tesla Model 3 in the Chinese market, smartphone-maker Xiaomi sees a gap in the market for its first extended-range electric vehicle (EREV), which sees a gasoline engine act as a generator to charge battery packs on the move.

Try the new Siri AI and system-wide performance improvements.

Siri and Apple Intelligence

Apple

Speed boosts and more features

watchOS 27, macOS 27 Golden Gate and iPadOS 27

Apple

How to install

A pattern is emerging among people who’ve already made it big. They’re rolling up their sleeves again, seemingly out of fear of missing AI’s defining moment and, presumably, the irresistible allure of making even more money — potentially a lot more.

Tom Blomfield, who co-founded GoCardless and Monzo before spending 4.5 years mentoring founders as a Y Combinator Group Partner, announced on Monday that he is taking a leave of absence to join Anthropic’s compute team — not as an executive, but as a member of technical staff.

He’s not alone in making that kind of move. Instagram co-founder Mike Krieger joined Anthropic as Chief Product Officer in 2024, and Andrej Karpathy, a founding member of OpenAI who went on to lead AI at Tesla and start his own company, Eureka Labs, joined Anthropic’s pre-training team in May, framing the decision almost identically to Blomfield’s, writing that “the next few years at the frontier of LLMs will be especially formative.”

Not everyone is joining someone else’s lab. Chamath Palihapitiya, the “SPAC King” who has mostly stuck to boardrooms and all things “All In” since leaving Facebook in 2011, just took his first full-time operating role in over a decade as CEO of 8090 Labs, his enterprise AI coding startup, which he announced a couple of weeks ago along with a $135 million Series A led by Salesforce Ventures. Wrote Palihapitiya on X, “I am convinced that what we are building now is even more important, so there was no decision to make except to be all in.”

Similarly, Eric Wu, who ran Opendoor for a decade before stepping back in 2023, recently launched NavigateAI, an AI “copilot” for construction workers, with $25 million in seed funding. Wu told me directly on a recent call about his decision to dive into an AI startup, “I knew if I looked back in 10 years and didn’t do something related to it, I would probably regret that.”

The clearest sign of how keen people who’ve already “made it” are to work on what they view as the still-early-innings of AI might be the job title itself. “Member of technical staff” is the deliberately flat, non-hierarchical label that Anthropic and OpenAI use for nearly everyone on their technical teams, regardless of seniority. It’s the same title Blomfield is taking.

It’s also the title that Peter Bailis took this March, just months after becoming Workday’s CTO, a role overseeing AI strategy across an $8 billion-revenue business. Bailis lasted less than a year before trading it for a spot at Anthropic.

from the alarm-bells dept

I’m not sure who out there is in RFK Jr.’s corner anymore, beyond some unfortunately powerful people in seats of federal power at the moment. That Kennedy’s tenure at HHS has lasted even this long is as absurd as it is dangerous, given the mountains of chaos he’s created in a mere year and change thus far. All of this anti-vaxxer nonsense, the seemingly random attacks on Tylenol of all things, an ongoing measles outbreak he’s mismanaging, and an inability to follow proper governmental procedure has produced a sample size of sucking that really should have been enough to get him booted from office at this point. Whatever you might think of Kennedy’s conspiracy theories and policies, there is simply no arguing that he doesn’t completely suck at his job.

The public polling around Kennedy has reflected this reality. Spineless senators who once supported him in the role are turning their backs. And then there are the warning bells being rung from people who were very recently insiders at CDC, such as its former Chief Medical Officer.

Dr. Debra Houry, the former chief medical officer at the Centers for Disease Control and Prevention (CDC), decried the direction of the agency under Department of Health and Human Services (HHS) Secretary Robert F. Kennedy Jr. 

“I think the secretary has caused a lot of irreparable harm, and when you look at many of the polls out there, the trust in public health, specifically CDC, has decreased dramatically, over 20 points in many polls,” Houry told host Margaret Brennan in an interview that aired Sunday on CBS News’s “Face the Nation.”

“That’s really difficult to recover from, and when states are removing links to the CDC website and following other medical organizations, I don’t know how you build back that trust overnight,” she added.

Advertisement

You absolutely don’t and this is a point I’ve been making for many months. It doesn’t take much skill or time to destroy the trust the public has in federal health officials. That part is very fast and very easy, as Kennedy is demonstrating. But to rebuild that trust, to win back the faith of the public, is going to take years, or decades, or perhaps may never really happen at all. The consequences of the idiotic placement and confirmation of RFK Jr. to lead HHS is going to span decades. The nihilists who managed to put this current cadre of clowns into federal office may not understand that, or may simply not care. But that is the reality.

A poll conducted by Harvard University’s T.H. Chan School of Public Health and the de Beaumont Foundation’s Public Health Listening Lab from March 19 through April 1 found that 50 percent of 2,205 U.S. adults said they trust health recommendations from the CDC. 

In spring 2025, 77 percent of respondents to a similar survey conducted by the joint pollsters said they trust recommendations from the agency.

Whenever this country moves past the MAGA era, it’s going to have what might be the Sisyphean task of repairing all of this damage. And not just in terms of reestablishing good, sane health policies. That’s just part of the task. The other will be the public messaging that must go along with it. That is equally, if not more important to repairing all of the damage Kennedy has and is doing. It’s not enough to have good policy built on science. Someone has to actually get the public to buy into and trust in those policies.

And the public is going to be in a very reasonable place when they ask why they should trust the next government to not be anymore idiotic than this one.

Filed Under: cdc, debra houry, health, health & human services, rfk jr.

Google Home Speaker

MSRP $99.99

Advertisement

“A delightful smart speaker that’s easy to love.”

Pros

• Gemini makes voice control feel genuinely conversational
• Strong smart home integration
• Better sound than previous compact Google speakers
• Modern, compact design
• Excellent value at $99

Cons

• Advanced features require a subscription
• Best experience requires buying into Google’s ecosystem.
• No display for visual controls or information

Quick Recap

Google just introduced a new Home Speaker powered by Gemini, and it may represent the biggest shift that the company’s smart home lineup has seen in years. This isn’t simply a hardware refresh with improved sound or a new design. Instead, Google is positioning Gemini as the foundation of a smarter home assistant, one that understands natural conversation instead of relying on rigid voice commands. At $99, the new Home Speaker also enters one of the most competitive segments of the smart home market, where it will inevitably be compared with devices like Apple’s HomePod mini.

Having used the speaker for the past couple of weeks, it quickly becomes clear that Gemini is the real upgrade. The hardware itself is a step forward over Google’s previous compact speakers, but the biggest difference comes from how naturally the speaker understands requests and carries on conversations. Rather than forcing you to think about the right command, it adapts to the way you naturally speak.

Google Home Speaker specs: What’s inside the round shell?

Google Home Speaker design and setup: It’s clean and breezy

Measuring 3.44 inches tall, 4.2 inches in diameter, and weighing just 0.9 pounds, the new Google Home Speaker is compact enough to blend into almost any space while still feeling more premium than Google’s older Nest speakers. Setup is straightforward. Inside the box, Google includes the speaker, a 30W USB-C power adapter, and a 1.5-meter power cable, with everything configured through the Google Home app.

Google is offering the speaker in four colors: Hazel, Porcelain, Jade, and Berry, with Jade and Berry currently exclusive to the U.S. The Porcelain model used for this review has a clean finish that should fit comfortably into most homes without drawing attention to itself. Overall, the design feels minimal, modern, and understated. Rather than becoming the focal point of a room, it blends naturally into a desk, shelf, or living space. Like much of Google’s recent hardware, recycled materials are used throughout the design, including the 3D-knit fabric exterior, which immediately brings Apple’s HomePod to mind.

One of my favorite details is the dynamic light ring around the base. It changes depending on whether the speaker is listening, processing a request, or responding, so instead of wondering what it’s doing, you always have visual feedback. It sounds like a small addition, but it makes the entire experience feel more alive. Omnidirectional microphones round things out, allowing the speaker to hear requests clearly from across the room without constantly asking you to repeat yourself.

Google Home Speaker interactions: Gemini changes how you use a smart speaker

The biggest shift with Google’s new Home Speaker has very little to do with the hardware. Instead, it comes from replacing Google Assistant with Gemini, and that fundamentally changes how you interact with the speaker.

Previous Google smart speakers worked best when you thought in commands. You would ask it to turn off the lights, set the thermostat, or play music, usually one request at a time. Gemini moves away from that approach. Rather than remembering specific phrases, you simply talk to it the way you would another person. Saying something like, “Set the house up for bedtime,” is enough for Gemini to understand the intent behind the request and carry out the necessary actions. It can also handle multiple requests in a single sentence and even adapt if you change your mind halfway through speaking.

Reasoning is where Gemini begins to separate itself from Google’s previous assistants. During testing, asking whether it would rain during a baseball game didn’t just produce the day’s weather forecast. Rather than simply pulling information, Gemini reasons through the request by combining context from multiple sources. Google calls this real reasoning rather than simply retrieving data, and it is one of the biggest differences between Gemini and the Google Assistant experience that came before it.

Gemini is just as useful for everyday tasks. It can help with reminders, calendar events, shopping lists, and planning your day while supporting natural follow-up questions that build on the conversation instead of starting over each time. Users can also choose from 10 different voice options to personalize how the assistant sounds.

Google is making the core Gemini experience available without an additional subscription. Buyers who purchase the Home Speaker before the end of September also receive six months of Google Home Premium, which unlocks Gemini Live. Instead of issuing commands, you can simply say, “Let’s chat,” and have a full back-and-forth conversation with your home assistant. Premium also introduces more advanced smart home features, including camera search history that lets you ask questions like whether someone left the garage open or whether the dog jumped on the couch, along with Home Briefs, which summarize everything that happened while you were away. Google Home Premium is available in two tiers, with the standard version included in Google AI Pro and the advanced tier bundled with Google AI Ultra

Google Home Speaker sound quality: It’s okay for the mission

Audio has received meaningful upgrades alongside Gemini. Google says the new Home Speaker features improved microphone processing for better voice pickup, true 360-degree sound, and 2.5 times stronger bass than the Nest Mini. Compared to Google’s previous compact smart speakers, the difference is immediately noticeable, making this a worthwhile upgrade for anyone coming from an older Nest Mini.

Stereo pairing is available when using two Home Speakers together, while integration with Google TV Streamer, Nest devices, and Cast-enabled products allows it to become part of a whole-home audio setup. Rather than existing as a standalone smart speaker, it slots neatly into Google’s broader ecosystem. As far as the raw audio quality goes, it avoids the expected pitfall of overtly-processed and synethetic tunes. It can fill a small to medium-sized room with punchy audio without any jarring distortion at high volume levels.

The sound profile is pleasant, in general, it’s sufficiently clear for listening to podcasts and audiobooks. Notably, it excels at mids, which means vocal-heavy tracks and classic music will please you ear canals. On the flip side, don’t expect delicate instrumental separation and at high volumes, complex tracks definitely get a tad muddy. If you seek that kind of audio nirvana, you might want to pay up a little bit and get the Sonos Era 100, or the bigger Google Nest Audio.

Google Home Speaker vs. HomePod mini

Comparison with Apple’s HomePod mini is almost unavoidable since both speakers occupy the same $99 price point. Both feature compact, fabric-covered designs that are meant to blend into a room and are available in multiple colors. The difference lies in what each product is trying to be. HomePod mini feels like a music-first accessory for the Apple ecosystem, while Google’s new Home Speaker is designed as a display-less AI hub powered by Gemini.

Gemini is also where Google creates the biggest distinction. Natural conversations, follow-up questions, and multi-step requests all feel more fluid than the command-driven interactions that have traditionally defined smart speakers. Apple has introduced Siri AI, but the current HomePod mini will not support those new capabilities. Anyone looking for Apple’s next-generation AI assistant will likely need to wait for future HomePod hardware.

Buyers already invested in HomeKit will still find plenty to like about the HomePod mini, but Google’s approach feels more flexible today. Gemini supports multi-action commands, context-aware conversations, and AI-driven automation, making interactions feel less reactive and more conversational.

Audio performance is another area where the two speakers differ. HomePod mini delivers clean, balanced sound for its size, while Google’s new Home Speaker focuses on stronger bass, wider 360-degree room-filling audio, and stereo pairing. Neither is intended to replace a dedicated speaker system, but Google places greater emphasis on creating a more immersive listening experience throughout a room.

Should you buy

Stepping back, this is more than a speaker upgrade. Google is rethinking what a smart home assistant should actually feel like. The future is not about commands or repeating yourself. It is about natural interactions, and that is ultimately the biggest selling point of the new Google Home Speaker. For its size, the sound quality is surprisingly punchy, and it performs pretty well if you’re more into listening to podcasts or music without deep audiophile expectations.

It, however, excels with arguably the smartest on-device AI assistant out there. Voice interactions are natural, the cross-device interplay is rewarding, and it can actually get work done across different apps and services, if you have linked the respective accounts with your Google account. The biggest caveat is that some of the smartest capabilities are locked behind a subscription, but if you merely need a no-frills tiny smart speaker, the Google Home Speaker’s latest avatar is as good as it gets.

Why not try

Amazon Echo Dot Max — If audio quality is your top preference, the two-way speaker fitted on the latest Amazon Echo Dot Max is right up your alley. Packing a dedicated woofer and tweeter, it delivers a surprising amount of bass and clarity. The cool Omnisense system brings presence awareness to the table, and it also offers support for multiple smart home protocols, including Matter and Thread. On the audio side, it even adapts to the layout of your room or home space. Plus, the new Alexa+ assistant is a meaningful upgrade.

Apple HomePod mini — The direct rival to Google’s speaker, Apple’s HomePod mini offers a similar design and build profile. It offers a signature audio output that is pleasing, though not hte loudest or room-filling kind. Where it wins is the deep Apple ecosystem integration, and finally, a much smarter Siri AI that is now ready to pull intelligence from — and get work done across — third party apps. But if you have an Android device in your hands, it’s not the best bet because a healthy bunch of features get locked.

Sonos Era 100 SL — In case you’re chasing a true audio pedigree, the Sonos Era 100 SL is arguably the best bet, even though it’s slightly more expensive. It delivers more refined audio, deeper bass, wider soundstage, and stereo separation, thanks to the combination of dual-angled tweeters and a bigger mid-woofer. It seamlessly allows multi-room audio playback and offers meaningful EQ tuning, as well. The assistant situation takes a hit, however, and the Sonos app still needs some work.

How we tested

We tested the Google Home speaker for a couple of weeks. In that span, it was linked to a personal Google account for accessing all the Gemini smart home features and automations. The audio quality was tested standalone, and to get a better perspective, it was also compared against the latest Apple HomePod mini. While testing, we focused on three core areas for qualitative evalauation, and they include raw sound quality, responsiveness and accuracy of the onboard AI assistant, and the wider cross-device weaved around it.

For audio quality evaluation, we played a variety of songs across difference genres to gauge how it handles different frequency ranges. Additionally, the onboard AI assitant was tested by throwing natural language queries its way, ranging from day-to-day smart device controls to knowledge delivery. We focused on accuracy and latency as the key metrics to assess the digital assistant’s efficacy. For the overall setup, we tested it across different positions under varying network and cross-device syncing environments.

An AI rewrite of a popular Anthropic-owned JavaScript runtime and toolchain has sparked praise for the speed of its execution, but also criticism of the coding practices behind the project itself. 

Last week, Bun creator Jarred Sumner announced that he ported Bun from the Zig programming language to Rust in only 11 days, using a fleet of Claude agents running in parallel. The work cost an estimated $165,000 at API pricing, suggesting that software revisions previously considered too large to undertake could actually be feasible now with AI.

Sumner said the port was necessary given the growing number of bugs Bun users were finding, including one implicated in the recent Claude Code source leak.

But the creator of Zig, Andrew Kelley, didn’t want his project to be seen as the culprit behind Bun’s woes, which he blames on Sumner’s bad programming practices. 

For Kelley, the move to Rust was not about the feature differences between the two languages, or even the use of AI, but rather “the diverging value systems of the two projects,” he wrote. 

Bun in the oven

Bun is a JavaScript suite consisting of a runtime, package manager, bundler and test runner. Some developers like it because it is a fast one-stop shop that plays well with Node.js.  

To make Bun speedy, Sumner used Apple’s low-memory fast-start WebKit JavaScriptCore (JSC) engine, rather than Google’s stock V8 engine. He used the up-and-coming Zig because he appreciated its performance and low-level control.

Anthropic acquired Bun in December 2025. The company built its core state machine on Bun. 

By then, Sumner had also grown to appreciate AI’s coding abilities, and was using it heavily in the upkeep of Bun.  By the time of acquisition, a Claude Bot called RoboBun had been doing a lot of the heavy lifting in the Bun repo. It supplied the most merged PRs of any contributor, fixing bugs and remediating test failures. 

But as Bun’s user base grew, more cracks started appearing in the code. Users found issues across the software. Anthropic’s 512,000-line code leak in March? That was Bun’s fault, thanks to a bug in the bundler that generated source maps during builds even when told not to, NodeSource reported. 

All these bugs weren’t Zig’s fault, Sumner explained in a blog post last week detailing the migration. Bun’s architecture mixed garbage collection and application-driven memory management. Sumner admitted that Zig wasn’t designed for that task. Rust was just better at automating memory management.

The Rustification of Bun

Rewriting 500,000 lines of Zig into another language would be a gargantuan undertaking if done by hand. “A rewrite in another language would take a small team of engineers a full year. It would mean freezing bugfixes, security fixes or feature development for that time,” Sumner wrote.

Instead, Sumner went with Claude. He spun up about 50 dynamic Claude Code workflows, reaching a peak of about 1,300 lines of code per minute and generating over a million lines of Rust code. The job took 11 days and cost about $165,000 at API pricing. Claude Fable did most of the heavy lifting. 

The Rust-based Bun was then subjected to Bun’s exhaustive test suite of more than one million assertions. According to Sumner, it passed 100 percent of those tests across all supported platforms without skipping or deleting any.

“There’s absolutely no way an engineer with that salary would’ve been able to achieve the milestones Claude did in 11 days,” an impressed HashiCorp co-founder Mitchell Hashimoto noted on X. 

Zig zags

But does Bun’s speed of execution betray the core tenets of good software development? 

One person not impressed has been Zig’s Kelley, who shared his misgivings in an impassioned post entitled “My Thoughts on the Bun Rust Rewrite.”

Even before the Anthropic acquisition, “we became increasingly horrified at the programming practices we saw in Bun’s codebase,” Kelley wrote. Bun was one of the largest and highest profile projects using Zig and, up until the Anthropic acquisition, a regular financial contributor to The Zig Software Foundation. 

In Kelley’s view, the project aggressively released new features, resulting in piled-up bugs, bad error-handling code, and technical debt. 

Sumner “was already writing slop well before he had access to LLMs,” Kelley quipped. He speculated that Sumner may have been under pressure to meet business objectives rather than technical ones, a pressure that increased with Anthropic’s acquisition. 

In fact, Bun’s codebase had grown so suspect in Kelley’s estimation that Bun parting with Zig was good news.  As he put it, no longer would “the publicly presumed poster child for Zig programming language actually [be] the prime example of How Not To Write Zig Code,” he wrote. 

The Bun team also tried to upstream some of its AI-assisted work to Zig, to no avail. Leading up to the Bun rewrite, the team maintained a fork of Zig that it said improved debug compilation speed fourfold, as eagle-eyed Reg reporter Tim Anderson revealed in May. But the Zig project would not accept Bun’s changes, citing a policy of not accepting AI-based contributions. 

Zig had been getting an influx of LLM-generated submissions, most of dubious quality. This lack of engineering oversight around AI-generated code would lead to countless problems down the road, Kelley reasoned.  

Kelley pointed out that if Bun’s tests missed these bugs in Zig, how would they be caught in unsupervised Rust code?

“The argument for shipping all the million lines of unreviewed code is that the test suite is good enough to catch everything,” he wrote. “It’s not sufficient to catch bugs in Zig code but it is sufficient to catch bugs in [a] million lines of unreviewed slop?”  ®

Meta seems to be having a bit of an identity crisis. On Monday, the social networking singularity said it would spend $50 billion to expand its Hyperion datacenter project in Richland Parish, Louisiana, from 2.2 to 5 gigawatts.

The news comes less than a week after a report broke claiming that Meta was actively exploring options to offload its excess compute capacity to other AI labs.

So, which is it, Zuck? Did you invest too much or too little in AI?

The easy answer is that Meta overcommitted. Inspired by the early success of Llama, it made a huge bet on the AI gold rush. Offloading spare compute to the highest bidder is just a hedge in case its Superintelligence team turns out to be another pipe dream, like the Reality Labs Metaverse that utterly failed to spark enthusiasm for immersive environments accessible through Meta’s Quest cybergoggles.

The more pragmatic read is that Zuckerberg has woken up to the fact he’ll never be as cool as OpenAI boss Altman or Anthropic’s Amodei, and renting out spare compute is just the natural progression for any sufficiently large hyperscaler.

Dawn of the Meta cloud?

Meta’s business model is closer to Google’s than those operated by OpenAI and Anthropic.

Both Meta and Google offer various services which generate revenues by connecting users with advertisers. For Google it’s a search and entertainment empire. For Meta it’s enabling an endless feed of content generated by friends, family, influencers, and yes, bots. 

Both are immensely profitable, earning $132.2 billion and $60.5 billion in profits last year, respectively. That’s profit, not revenue.

But both are now plowing over $100 billion a year into AI infrastructure to power large language and image and video generation models. As we learned from Meta’s recent earnings calls, the most commercially potent of those models get the right ads in front of the right eyeballs.

The open secret is Meta was already one of the most successful AI companies long before ChatGPT debuted. Except, it’s not large language models (LLMs) that make Meta money, at least not in the conventional sense. Instead, Meta’s most profitable AI models are the recommender systems that mine profiles for context and use it to infer your needs. Meta’s devs evolved those models considerably over the past few years, and their architectures now look a lot more like an LLM than the now-pedestrian neural networks on which Zuckerberg built his empire.

Google is in a similar situation. It’s investing heavily in AI to feed its fast-growing and profitable cloud business, even as advertising still pays most of the bills. But unlike Google, Meta hasn’t yet made the leap from hyperscaler to cloud provider.

Amazon, Google, Microsoft, even Oracle got there eventually, and it seems AI may be the catalyst that turns Meta into a cloud, too.

Recent reports suggest that Zuckerberg is warming to the idea. 

“I think that’s certainly a thing that we could do and that I think would make sense to consider,” he said in an interview with Bloomberg last week. “As a backstop, even if for whatever reason we don’t need all the compute ourselves or for any number of reasons, there’s a very large amount of demand that I think you could sell it long-term like AWS or Azure or Google Compute.” 

But while the demand may be there, Zuckerberg emphasized the compute capacity is not readily available.

But as Ben Thompson of Stratechery put it, cashing in on this compute may be more than a backup plan. In a post channeling an imaginary Zuckerberg, Thompson suggested that becoming a neocloud would force Meta to stop chasing pipe dreams and pet projects. His logic is that if Meta can’t make money with infrastructure it buys for AI ventures, the social networking giant can lease the orphaned hardware to the highest bidder.

The takeaway for investors — should Meta follow its fellow hyperscalers-turned-cloud-providers down this road — is that the profitability of its hardware investments would no longer be tied to its ability to commercialize them.

Seizing the means of production

If history tells us anything, scale matters. Building a cloud like Amazon Web Services (AWS) is next to impossible unless you’ve already figured out how to profit from those same resources.

Meta’s scale puts it in a position to acquire compute in volumes impossible for smaller players. Its ability to capitalize on infrastructure demand relies entirely on having something others want but can’t get anywhere else.

For what it’s worth, Zuckerberg wouldn’t be the first to come to this conclusion. Earlier this year Musk-owned xAI surprised many when it announced plans to rent out its Colossus supercluster in Memphis to rival model dev Anthropic.

The calculus here is the same. Making a profit off LLMs, like Grok, isn’t easy — just ask OpenAI — but selling the means of AI production to those that haven’t yet figured that out is enormously lucrative.

The logic appears to have gotten Zuck’s attention.

“The SpaceX model I think is quite interesting in terms of just making these short-term deals that are at a big premium,” Zuckerberg told Bloomberg. “So we get offers for all kinds of stuff like this and we’ll evaluate them and see what makes sense.”

Reports suggest Meta is seriously considering two strategies for commoditizing its compute assets. The first would be a usage-based compute platform similar to Amazon Web Services’ Bedrock.

The service would allow customers to run models and serve them through APIs — interfaces that abstract operational complexity. To be clear, Meta already offers API access to its homegrown models, at least the ones it didn’t pull after realizing the way they’d been implemented could be abused. So, from what we gather the difference would be allowing customers to run third party models as well.

The second scheme reportedly being explored would involve selling raw compute resources available to end customers — similar to CoreWeave or Lambda.

All the right ingredients 

Meta’s silicon strategy may help as well. One thing all the major cloud providers have in common is a growing catalogue of custom cloud silicon.

Once they’ve identified a core use case, Amazon, Google, and Microsoft all rolled their own silicon to maximize their margins. AWS Trainium, Microsoft Maia, and Google TPUs are all examples of AI accelerators originally built for internal workloads but later made available to the broader public.

Meta has been building its own AI chips for years. The first few Meta Training and Inference Accelerators (MTIA) were designed to speed up its recommender models. But new designs, developed in collaboration with Broadcom, are far better suited to running LLMs like Llama and Muse Spark, and whatever else its customers are willing to pay for access to.

More importantly, this mix of compute means that Meta can take advantage of the fact GPUs are extremely flexible to bring new products to market quickly. Then once they’ve proven performers, Meta could transition those workloads to its custom chips and offload spare GPU compute to its cloud customers.

Meta has all the ingredients, compute, scale, and capital necessary to become a major cloud provider.  ®

Rumor mill: Apple is overhauling its Mac chip roadmap to place artificial intelligence at the center of its hardware strategy. Instead of completing the M6 lineup with the usual Pro, Max, and Ultra variants, the company is reportedly moving directly to the M7 generation and planning significantly larger Neural Engine upgrades, according to people familiar with the plans.

This fall’s Macs are still expected to debut with a base M6 chip. Under Apple’s traditional release pattern, that chip would be followed by M6 Pro and M6 Max variants for higher-end laptops, along with an M6 Ultra for desktop-class machines.

This time, that sequence will reportedly stop after the entry-level chip. The company has already moved on to the M7 design, taping it out only about six months after the M6 reached the same stage. The compressed timeline highlights how urgently Apple wants its Macs to handle increasingly demanding AI workloads.

According to the internal roadmap, the first M7 Macs are scheduled to launch in the first half of 2027. Higher-end M7 Pro and M7 Max systems are expected later that year, while the M7 Ultra is targeted for 2028.

Apple has skipped an Ultra chip before – the M4 family did not include one – but abandoning every high-end M6 variant at once would mark a first. People briefed on the plans say Apple determined that the M7’s AI-focused upgrades were significant enough to justify skipping further M6 development.

At the center of those changes is the Neural Engine, Apple’s dedicated on-chip AI hardware that powers on-device generative models, accelerated inference, and Apple Intelligence features. Apple has refined the Neural Engine with every Mac chip generation since the M1 debuted in 2020, and the M4 represented one of the biggest improvements to date.

The company now wants the M7, particularly the M7 Ultra, to approach the level of performance developers expect from dedicated AI accelerators rather than traditional general-purpose desktop processors.

Memory support is a major part of that effort. The M7 Ultra is being designed to support up to 1.5 terabytes of unified memory. That is roughly double the capacity planned for the upcoming M5 Ultra server chip and matches the maximum RAM configuration available on Apple’s 2019 Intel-based Mac Pro.

With that much memory, significantly larger AI models can remain in memory, reducing bottlenecks and limiting the need to rely on external storage or cloud-based compute. Whether Apple ships Macs with the full 1.5TB configuration will depend on memory availability, as supply constraints and elevated prices remain concerns.

These desktop plans tie directly into Apple’s server strategy. The company is preparing a more powerful AI server based on the M5 Ultra, known internally as J246. Engineers are already working on a successor built around an M7 Ultra-derived server chip, with a launch window targeted for around 2029.

In other words, the same architecture expected to power Apple’s highest-end Macs could also underpin the next generation of servers running Apple Intelligence in the cloud.

Beyond the M7 family, Apple is developing an M8 generation with even more AI-focused silicon. The lineup reportedly includes a processor code-named Soko, targeted for 2028, along with other chips for high-end Macs under the Cardinal name.

The 2028 chips are planned for a 1.4-nanometer process, which should deliver another leap in efficiency. The shift comes as AI chips encounter increasing power and cooling constraints, pushing Apple to prioritize more transistors for neural processing units and memory bandwidth rather than simply expanding CPU and GPU cores.

All of this hardware development comes alongside a more mixed picture on the software side. Apple has struggled to deliver AI-powered services at the pace many expected. Apple Intelligence and the redesigned Siri have rolled out more slowly than planned, forcing the company to adjust expectations along the way.

Still, Apple’s hardware teams have spent more than a decade building the foundation for this moment, often through projects that never reached consumers.

The most notable example is the company’s abandoned self-driving car initiative. From the beginning, Apple targeted full Level 5 autonomy and invested heavily in machine learning and custom silicon capable of processing massive amounts of AI workloads in real time.

The car project never reached the market, but the chip development work directly contributed to the Neural Engine architecture that first appeared in the iPhone X in 2017 before expanding to Macs and other devices. That same hardware foundation now sits at the center of Apple’s AI strategy.

With the M7 and M8 families, Apple is treating AI as the primary driver of its chip designs rather than simply an additional feature. AI workloads are now shaping which chips get developed, when they launch, and how their architectures are built.

from the the-last-gasps-of-antitrust-reform dept

I think I’ve extensively explained at this point why the $111 billion merger between Paramount/CBS and Warner Brothers is a gargantuan pile of shit that will indisputably harm labor, consumers, markets, creatives, and potentially even national security. It doesn’t matter the company names; every single major media merger of this type ends badly for everyone but the trust fund brunchlords at the top.

Not only that, every single merger involving this particular company (Time Warner, Warner Brothers) in the last quarter century has resulted in nothing but layoffs, price hikes, shittier product, and a lot of whimpering. And there are ample signs that the Paramount folks are even less competent than past suitors, including the AT&T executives, who quickly got too far out over their skis.

While the Trump DOJ has unsurprisingly rubber stamped Larry Ellison’s clumsy effort to dominate what’s left of U.S. corporate media, states keep hinting at the fact they’ll file a collective antitrust lawsuit.

That’s certainly the case in Oregon, where Attorney General Dan Rayfield is asking for a 60 day pause in deal finalization while his office investigates both the deal — and apparently the Trump cronyism that has helped enable it. Rayfield, for one, accuses Paramount of refusing to adequately respond to state AG requests for information about the deal’s impact:

“We’re not going to let Paramount Skydance play hide the ball so they can rush through their massive merger. Oregonians have a real stake in this deal – in our film industry, in our economy, in the choices they’ll have as consumers. Paramount had every opportunity to hand over records and answer a few basic questions. Instead, it is trying to run out the clock and evade scrutiny. We’re asking the court to make sure Oregonians get the answers they’re owed before this deal closes, not after.”

Rayfield says that Paramount has been particularly cagey when asked for data on its interactions with the Trump administration and Trump DOJ. Including details on a federal government influence campaign Paramount internally calls “project warrior”:

“Paramount has not complied. According to court papers, the company declined to accept service of the request, waited weeks to respond, and ultimately sent objections on the day its documents were due – objections the state dismisses as a baseless tactic to avoid turning over the records. Paramount has told Oregon it does not intend to close the deal before July 16 but has not agreed to hold off any longer while the state’s investigation continues.”

So while the $111 billion deal is abjectly terrible, it’s not quite yet a done deal yet. I’d suspect that a joint antitrust lawsuit featuring the handful of states that still care about such a thing will arrive sometime in the next month or two. While it might not succeed in scuttling the deal, it could extend the timeline in a way that could prove costly for Larry Ellison, David Ellison, and their debt-riddled proposal.

Larry is comically over-leveraged on AI, and should an AI bubble pop coincide in significant regulatory delays for his media domination bid, things could get very interesting. As an aside, I’d like you to notice how when the state lawsuits come, all the Republicans who spent last election season pretending they were suddenly interested in antitrust reform will be absolutely nowhere to be found.

Filed Under: antitrust reform, consolidation, Dan Rayfield, hollywood, larry ellison, media mergers, oregon, regulatory approval

X has made a “tweak” to its algorithm to boost the visibility of posts to users’ “mutuals” — the people they follow who follow them back, head of product, Nikita Bier, said Monday.

“We noticed this data was missing from the algo and it made your friends appear less in your replies. This resulted in the reply section feeling more like a battleground with people you don’t recognize.”

The change may not drastically revamp the site’s user experience, but may make X feel a little bit more like a community rather than a torrent of disparate voices shouting into the digital abyss.

Bier noted that the change would also “help clusters form around interests more easily, which many people have asked for.”

X has introduced a number of changes lately — many of which seem designed to make the site a bigger hub for creators. Earlier this year, the site changed how it compensates accounts in an effort to incentivize original content rather than mere aggregation, and, earlier this month, it also introduced a video editor designed to make it easier for users to work on the platform.

This tweak follows changes that Meta’s Threads has been making to its algorithm aimed at creating communities, largely as a differentiation from its main rival X. For instance, last month Threads rolled out a Your Algo feature, which lets users privately control what they see in their feed. It also reached 500 million monthly active users.