Never let a 3rd party LLM be the core of your product or it can be changed or taken away at any moment.

What you do is use the frontier models to build a deterministic set of tools that does what you want and MAYBE put in a small core of LLM for the ambiguous stuff you can't make deterministic (yet).

And make sure you can swap that LLM core to any other provider (even local) and have a playbook ready for that.

I should be able to put in a Linux DVD or memory stick and install Linux on it.

Or at the _very_ least an alternative app store.

Game consoles had a higher import tax than "computers" -> allow linux, save money.

IIRC they did a similar thing with the PS2 with some janky-ass BASIC interpreter being available.

They’re 100% commodity hardware and fully locked down from any user freedom. Weirdly everyone focuses on Apple with all their might instead of gaming consoles.

Or just that it's so much cheaper that the cost/benefit ratio is better?

Also "finish a task" is also subjective. I can "finish the task" of building a table, but it will be a shitty table. Are you also measuring the quality of the result - which is subjective again?

My favourite example is that if you use "timestamp" when using an LLM to process video you get worse results than if you'd use "timecode".

AV professionals always say "timecode" - timestamp is a programming term.

Using the right word pushes the model closer to the correct spot in the cloud of vectors that is it's "brain".

There is no true objective measure, can you mathematically determine which song is the best for everyone for example? Or which painting different people feel is the nicest to look at or what emotion it gives them.

Yea, you can do the fucking strawberry tests or carwash trick questions, but that doesn't really measure anything useful.

You can also do benchmarks but how do you measure the output of those?

The easiest way is just to use them all and get the feels of which of them works best for you. For me it's Claude first, pi.dev + gpt5.5 second. Plain Codex is a distant third and Gemini exists - it's pretty good at finessing web UIs as it does aria labels and usability better than other, but I wouldn't write backend code with it.

How I understood it that they want _actual_ security researchers, not some random dude who once installed Kali Linux and ran nmap.

It's state of the art private compute according to actual experts and everyone will be wasting their time if the "researchers" need to be coached through the process and explained the basics of the system's operation.

Let's go through the process in detail:

- IF you had iCloud syncing on (which meant your photos were on Apple's servers anyway) they would download a pre-generated set of perceptual hashes of KNOWN CSAM. Like actual child porn, vetted and verified by humans. Perceptual hashing is a fun tech, look it up, super useful outside of this context too.

- Then IF your phone, independently, with zero connections to the outside matched MULTIPLE (they didn't specify the exact count) known child pornography image phashes those would be uploaded to Apple's servers. There was some cryptographic boogaloo at this point too that mathematically prevented Apple from decrypting the photos without enough matches.

- Then an actual human would look at "reduced quality" (can't remember the exact term) versions of the matched photos and determine whether it's some dipshit / bad actor forcing fake matches with kitten pictures or actual CSAM material.

- If the human verifier confirms that the person's phone has actual confirmed child porn, the authorities would be contacted.

---

And the basic arguments:

"but muh privacy!" - You needed to have iCloud syncing enabled for this feature to be enabled at all. Your data is already on Apple's servers, they could just scan them there. But to do that they'd have to be able to decrypt them, which means the authorites can force them to do it too. Apple didn't want this. Thus: local scanning. You don't want your photos scanned, turn of iCloud sync. Done. Bad Tim Apple can't look at any of your photos.

"But what if I get multiple fake phash match files sent to me and I save all of them to my photos!" - then you'll inconvenience a human verifier for maybe 30 seconds. You won't get swatted or put on a list anywhere.

Oh and Apple already "hashes your entire iCloud" - also fully on device. That's why your iOS device runs hot for a week or so after installing a new major iOS revision. They run the fully local ML models on your pictures in the background and store them - also locally - to a database where you can go and search for "dog" or "receipt". Adding a phash to match to a local CSAM database is a non-issue at this point.

Fancois Normal installing a 3rd party AI service which turns out to have zero security and actively just harvesting private data.

Tell me which company in your opinion would be in the LOUD headlines, Apple or the random 3rd party?

This is about Super Private Benoau AI being available for any user to install. How can they know whether it respects their privacy or not? The home page says that they're the best and mostest private ever of course, has animations generated by Claude and everything.

But actually it runs on servers bought from Hetzner's server auction and stores all logs in plain text in open S3 buckets and the owner actively sells the user data to the highest bidder.

This is what Apple is worried about and EU either doesn't care or doesn't understand the issue.

If you want it to, for example, summarise your HRV or menstrual cycle you can't anonymise it or you don't have any data to analyse. It'd be just "wink wink nudge nudge" with zero context.

Let's remember that a tiny company called Meta had a "VPN" they provided for users that just happened to spy on them: https://news.ycombinator.com/item?id=39881962

And that went on for a long while before it was noticed.

Now imagine the same situation but an infinite whack-a-mole of alternative AI providers and just regular folk who will install mobile games from a frozen baby ad...