Steps that either would not work between architectures or that Rosetta would handle (thereby answering your own question)

What? They would compile the code that they control for the M-series architectures. The target architecture is M-series ARM chips running on MacOS.

Rosetta is for translating x86 to ARM. Which wouldn’t be necessary because there are no x86 binaries involved at all.

I don’t understand why Rosetta has to be involved in any way at all.

Apple provides APIs for programming low level GPU instructions, including AI workloads, on their chips, using the Metal API. Anyone interested in using Apple hardware to its full potential can just write Mac-native software, same as any other MacOS native software directly compiled as binaries to run on Mac M series chips. No Rosetta required for translating x86 instructions to ARM, when the whole thing compiles for Apple’s native instructions in the first place.

Wait can tokens just be directly compared like that? My impression is that token cost can vary by 2 orders of magnitude, depending on model, because the actual work of computation varies by that much between models.

Reportedly, the app uses an on-device machine learning model to classify the content.

Using your shiny new phone to train AI…

I would think the models are already trained, and then deployed to phones to do on-device inference. The models should be fixed and static by the time they end up on a phone (in large part because phone hardware is just not technically suitable for training, whether pre training or post training).

aren’t just wrong — they’re a bridge too far

Haven’t the PR flacks learned by now to sound less obvious about using LLM generated text?

But it’s not the company itself putting those assets up for sale. It’s the court-approved or court-appointed bankruptcy trustee of that company, executing transactions approved by the court. That’s why bankruptcy has to go through court, but it also provides a level of immunity (commonly called “exculpation” in bankruptcy cases) to the individuals actually doing the things ordered by the court.

If it’s illegal to do, a court shouldn’t be ordering it, so the affected people should have to fight the court order before it happens. But it’s not illegal to sell anonymized user data like this, and not illegal to sell employee work-related data, either, in the U.S.

It needs to be a system and a contractual framework that complies with the medical privacy regulations of the place where it’s being used (in the US, HIPAA). I have some doubts about whether specific companies or people will actually comply, or the enforcement of those laws by certain governmental authorities, but I don’t think it’s conceptually impossible to have this kind of tech be compliant with those types of regulations.

on average peoples’ computers

Average people don’t have self-administered computers anymore.

People have work laptops administered by their work IT departments, or they have phones and tablets running iOS/Android. Some children have school-administered Chromebooks, and may never own a normal laptop running a traditional desktop OS.

The main people operating self-administered traditional computers are PC gamers (a dying breed during the current price crisis), freelance/independent workers in the fields who still benefit from a mouse and keyboard for productive work and aren’t provided a computer by an employer, and tech nerds.

Probably not. The article credits Shutterstock submitter Vladislav Noseek, whose online presence with photographs of this style (soft lighting, bright lighting, shallow depth of field, a focus on breakfast foods) long predates generative AI, much less of this quality. It’s stock photographers like this guy whose inoffensive visual style and identity have been most directly incorporated into the AI defaults.

Fairphone is not about the latest tech. They tend to go for well-supported hardware to ensure longterm support.

The USB 3.0 standard was finalized and published in 2008. USB 3.1 specification superseded it in 2013, and then USB 3.2 superseded that in 2017. Each version is fully backward compatible with every standard version from before, so there’s nothing stopping people from releasing a 3.0 controller and certifying it works in the 3.2 mode that is a drop in replacement for 3.0.

It hasn’t felt t like there’s been much significant performance increases or development in RAM in the last… decade?

In memory? There’s been a ton of improvement, even if most of the coolest stuff isn’t making it into DIMMs that are installed in user laptops/desktops.

Advanced packaging technology has allowed chip manufacturers to put different silicon dies together with increasingly high performance (high bandwidth, low latency) connections in the same package, including with some three dimensional stacking. That way they can mix and match different silicon dies for greater cost effectiveness, yield, performance, etc.

This also means that in-package memory is now the standard in certain chips. Apple’s M-series silicon has its memory packaged right into the CPU/GPU package, as a system-in-a-package, so that the connection between the logic and memory is comparatively much higher performance, several times higher bandwidth than desktops or laptops that don’t follow that kind of architecture.

Similarly, in data centers, the AI boom has caused all the memory manufacturers to switch their production lines to high bandwidth memory, where they vertically stack a bunch of DRAM chips on each other, with ultra-fast, high bandwidth connections, so that they can shove terabytes of memory into these data center servers. These recent generations have been improving speed and bandwidth in ways that make consumer level DDR5 RAM look like child’s play.

So they’re improving things. Just not in ways that really show up in DIMM sticks.

Not strictly, there are usualy hurdles to overcome for home usage of datacentre tech, but it’s possible.

The hurdles are basically insurmountable with the hardware released after 2024.

The NVL72 for the Blackwell generation cost about $3 million and takes up a single server rack. The power consumption is about 130 kW, and most configurations require dedicated plumbing for the liquid cooling.

To put things in perspective, a residential electrical hookup is usually 50A or 100A for a house, with recommendations that anyone who is going to be charging electric cars should have 100A service. 100A at 240V is 24 kW.

So one server rack uses as much power as the maximum electrical capacity of 5 homes. You’ll never be able to pull that off in an actual residential environment.

Oh, and the newest 2026 generation, the Rubin NVL72s, use something like 230 kW of electrical power, almost twice as much as the previous 2024 generation.

There’s always going to be a robust used market for phones that were purchased outright, to be resold on a different cycle than every 2 years (plenty of rich people changing phones every year, and plenty of people replacing on a 3, 4, 5, or 6 year cycle). You can expect the market to basically settle on a curve where it depreciates along a predictable rate.

Leases don’t really change that, any more than leases changed the market for used cars, or even certified pre-owned by the same dealers and organized by the same manufacturers who sell new cars.

There will be times that the predefined lease terms will unexpectedly prove to be either beneficial or detrimental to the consumer. Sometimes external factors will affect the entire used market, like currency issues, or component pricing issues (imagine if RAM prices dramatically swing again for new devices in a way that affects the value of the already-sold devices out in the world), where the predefined lease prices turn into a windfall for someone. Like in 2021 or so when expiring car leases allows the lessee to buy out the car at the end of the lease for much cheaper than the car itself was worth.

It’s generally going to be a less than ideal financial decision to lease, but it also won’t collapse the used device market and it won’t be that far off the practice of selling your old phone when you buy a new one.

It’ll only make a difference during the times in which it is being hit by a beam that wouldn’t otherwise hit the earth, since whatever it reflects will be offset by some shadow behind it. I wonder if, as a matter of orbital mechanics, it would be possible to make the satellite orbit in such a way so that it never casts a shadow and is always in the sunlight.

“Technical debt” is a phrase that was invented to explain the phenomenon in a way that bean-counting business bosses could understand: we can do it this way in a way that costs us less today but will incur a debt that will slow us down in the future. It was always an attempt to quantify the concept in a way that we have to think about tradeoffs between our current resources and our future resources.

the growth itself is hella juiced because the GPUs are only relevant for about 3 years till the new ones are out and make more AI for less power. And they depreciate them over 7 years. More than twice as long as they can or should use the GPUs for.

We don’t actually know this for sure, yet. I had expected the A100 generation (released in 2020) to no longer be profitable to run by now, but the backlog in new data centers being turned on and the high demand from Anthropic and OpenAI still leaves those chips useful for inference. You can rent those 2020 chips out today at some price above what they cost to continue running (300W, so electricity prices of USD $0.20 per kWh would translate into about 6 cents per hour. Prevailing spot prices appear to be about $2/hour right now.

But just because I was wrong on 2020 chips, originally sold for about $15,000 in a low interest rate environment, doesn’t mean that I’m wrong about 2024 chips, the B100s that use 1000W and were sold for $35,000, requiring a ton more specialized cooling, power, and network infrastructure. Or the 2026 R100s that use 2000W, and whose prices I can’t seem to find published anywhere, but were set after the memory companies basically locked in their record breaking prices for their HBM. That’s an unsustainable path and at some point, data centers start struggling to find users willing to pay the bare minimum necessary to continue turning a profit on GPU usage.

I doubt the 2024 chips stay in service to 2031. And I’m really, really skeptical that the 2026 chips stay in service to 2033, especially after NVIDIA switches to yearly release cycles next year.

But it isn’t encoding knowledge, it’s encoding word correlations.

I’m saying that humans do this a lot, too. Qualitatively, it’s different, in that this particular batch of frontier LLMs will get things wrong in ways that most human brains wouldn’t, but as a category of error it’s not unique to LLMs.

I know a ton of facts that I learned only through reading, and have no actual firsthand knowledge/experience or ability to test it: Jupiter is larger than Saturn, the atmosphere during the Carboniferous period was high in oxygen, cigarettes cause cancer, Thomas Jefferson owned slaves, the capital of Norway is Oslo. At best, I can cross reference other sources and see that things are consistent with each other. Is my belief in those facts “knowledge,” or is it merely recognizing from my training data that those particular words can validly be presented in that order?

If you ask average people on the street whether FAT32 is a good filesystem for a 64GB removable drive, most of them won’t know, but there are a handful of bullshitters who might confidently parrot back things they can Google but not understand. That’s part of the human condition, too.

I’m by no means an AI booster/enthusiast. I suspect LLMs/transformers are actually a dead end, and expect the upcoming crash to be economically and financially devastating to the tech and financial sectors. But I also have a pretty dim view of human intelligence, too, and see way too many parallels in LLMs as bullshit artists to humans as bullshit artists, too.

It modifies the prompt, aka the input, not the output. It is smuggling 3 bits of secret user/session data in a wrapper that doesn’t look like it contains that data. As the article explains:

So the marker becomes part of the system context sent to the model.

This is a normal timestamp on a prompt:

Today's date is 2026-07-11.

But if your system timezone is a Chinese mainland timezone, it looks like:

Today's date is 2026/07/11.

Then, if your base URL includes a keyword like “deepseek,” it silently replaces the apostrophe from a ' to a ʼ:

Todayʼs date is 2026-07-11.

Or if the base URL has one of the domains on the list, like any .cn domain, it replaces the apostrophe with another apostrophe character:

Today’s date is 2026-07-11.

And if it has both a URL and a keyword on the watchlist, the prompt context includes:

Todayʹs date is 2026-07-11

That’s 3 bits of information: does this system have a mainland Chinese time zone, does the base URL contain a known keyword (associated with Chinese AI competitors) or a known domain (associated with mainland China or its major tech companies). And it sneaks it on by without making it obvious.

That’s steganography.

up to 150kW at full load.

That’s the last generation. They’re moving from Blackwell to Rubin chips now, and the 72-GPU Rubin servers use up to 230 kW.

The typical residential connection in the U.S. has a 24 to 48 kW electrical connection. A block of houses might not have enough power infrastructure to power just one of these server racks.

It’s that they are trying to use statistics to encode entire thought processes into hidden variables from conversation snippets. They want to use statistics to go from many individual interactions to a large model, and then use that model to predict individual interactions again.

Has it been shown that the human brain doesn’t model the world in a similar way, though? A huge portion of human knowledge is both stored and transmitted in the form of language. Lots of human knowledge also follows the garbage in, garbage out theory, where you can have entire areas of knowledge that aren’t actually true but might be internally consistent, at least within certain scopes: conspiracy theories, belief in the supernatural, entire academic disciplines built on a religion or theology that not everyone believes, etc. Or even world building in fiction, the words on a page can be enough to convey ideas such that it “tricks” human brains into filling in the gaps so that they internally see a rich, fleshed out world that is entirely fictional and where specific details might not find strong direct support in the underlying text.

it has no concept of correctness

But statistical weight on what is more or less likely to be correct still makes a difference to objective quality of the outputs. If the model weights are trained on the reality that high quality university texts describe something and reflect some sort of underlying model of what is described using language, then can’t the model itself learn as much as a human could from those words on a page?

All models are wrong, but some can be useful. And different models have different quality in different domains. So although I don’t believe LLMs will overtake the hump of getting ahead of human knowledge, I also don’t believe that any given LLM can be evaluated on quality, and that Facebook’s LLMs are significantly behind other LLMs we see.

And that maybe a huge part of it is its internal process of preparing the model to evaluate the quality of its inputs, such that the output it produces can also score high on quality.