At one point in my life, I would've found this data fascinating (if it were opt-in, perhaps)! Now, I find it exhausting.
A similar discussion came up recently with Siri Recap[1], where some expressed how useful they thought it was and that, as a party to the conversations it summarizes, there was very little moral or ethical concern. But I have a social concern.
If I'm having a chat with someone, there's a milieu constructed around the fact that not everything will be remembered verbatim. There are all sorts of social and contextual cues that can easily get lost if listening gets replaced by flat relentless transcription.
With Recaps and this social graph (were it to be productized), the informal fluidity in the way we relate to one another starts to go away, and the expectations that follow from that become artificial and burdensome. In the same way that autocorrect has made typos unacceptable, how could you ever be forgiven for forgetting the name of your colleague's new nephew or the Italian village that your neighbor took a spontaneous boat trip to while on vacation last year?
I too would've found this fascinating at one point, but my jaded older self now just finds this incredibly creepy, unsettling and not at all surprising.
Maybe a decade or so ago I had a discussion with a colleague about chat apps and he said, why would you want to have a conversation on snapchat if it is not saved, and I was like, why are you having this conversation now, you can't save it either.
Please stop using AI to author your articles. It was really hard to read it despite how interesting the topic was. I dropped halfway because it was just too annoying.
Yeah seriously. Article becomes unreadable immediately after the intro. I believe this author probably could actually write two years ago, it's a shame people are just giving up on that and choose to turn out stuff like this instead.
I was bothered not so much by the slopful article as by the AI-written text in the dashboard that iKnow produces. I ran the program on my Mac and it did display charts and graphs with names of people and organizations I’ve interacted with over the years, but I couldn’t figure out what it all was supposed to mean.
One example: "323 people, 122 events and other records, 200 employers and 1 stored group — as the store on this Mac links them. Read live just now. Colour is what each node is to you: the stored group in red, family and closeness by iKnow’s rules on the kinship scorecard. Size is how many rows in the store name it."
Claude etc. can be prompted to produce clearer explanations than that.
It's not possible to un-AI text, only replace it wholesale.
Suggestion for next time: "Review my prose for accuracy, grammar, appropriate strength of claims, etc. Flag any issues for my attention. Do not provide text; this work should not contain any AI-written content." That'll turn it into a very good check that doesn't take your voice away.
Complaints of AI are like a form of friction that eventually goes away. Either AI will undergo a change to create more approachable material, or in general people's tastes will change as they accommodate the style and get used to it. Or some mix of the two. As young people continue to grow up reading AI generated content they will likely begin to find that person written content from before AI grates on their nerves. More, they'll even ask AI to reword documents into preferable AI-like styles.
As an older person, AI grates on me, but I found the meat like "It is a triple store written in a vocabulary of 181 entity classes and 278 predicates: every row is a subject, a predicate and an object. The subject is an entity (a person, an app, an event); the object is either a plain value (a name, a handle, a date) or another entity, which is where the graph gets its edges. The predicate is the interesting part: it says what kind of relationship the system believes holds between the two" quite approachable.
So it was like a meal where I just have to push, say, mushrooms aside as best I can to get to what I like. And there was a time I would have refused the entire plate because I truly abhor mushrooms. I blame watching Matango as an impressionable child.[0]
If you create a strict template, like a PowerPoint deck or tightly formatted report, the LM does a really, really good job. at populating that template with good data.
If you give it an open license to just create stuff, it’s always terrible for the general audience — but it may be fine for you or a small group that is familiar with the subject.
I compare it to fast food. An uneducated teenager is quite capable of putting together a McDonald’s lunch within the constraints of the system. If you let the kid loose in a kitchen designed for trained chefs, the result would not be a good lunch.
This has not been my experience. People talk about LLM-isms (Claude-isms especially) all the time. "The absolute shape", "load bearing", "seams", "smoking gun", genuinely empty preambles, constant sloganesque phrasing. If you're seeing it less in conversation, it's not because it's improved.
Sorry your comment got dinged. I assume the reference to age triggered that. Ageism is a very, very real thing, hereabouts.
I think it was insightful and relevant. I am not a “slop enthusiast” (a great term), but I am a longtime student of human nature, and basically agree with the premise.
Like it or not, slop is here to stay. It’s up to us, to guide and shape it into a net benefit.
If you came away thinking I was enthusing about AI slop, it's probably because I was too subtle in comparing it to a food I abhor while saying there are tidbits, non-slop tidbits, that I could appreciate. Apologies for that human generated mischance!
I'm actually rather bemused to see many young folks, arguing against LLMs and ML, and seeing more experienced folks, embracing it.
I'm pretty experienced, and have found LLMs to be quite useful.
Like any new tool, there are deficits and peculiarities, but these are becoming less of a concern, and the utility of these tools is advancing rapidly.
You can’t unslop A.I. prose. Once it’s written, the content is poisoned.
“Sure there was a turd in the punch bowl but I did my best to remove it, so you should be able to drink it.”
Nope.
And the technical parts are what matter, saying you did that with A.I. tells me “I don’t care about this part” so what the heck are we even doing here?
I second this. Unfortunately you just can’t just “unslop” it, much like you can’t just “shake the dirt away” from food that fell into a bucket of… you get my point, it won’t be good to eat anyway. Bonus additional ironic parallel could magically manifest if you can guess the bucket content.
Anyway, it’s a shame, this could have been excellent work. Instead, I find reading it unpleasant and part of my attention rots away wondering if you even checked what the LLM wrote for you. You probably did, and I could also do it myself, but still…
It's like the saying beauty is only skin deep, but ugly goes clean to the bone. Bad AI writing makes an article "unreadable" - you'd have to force yourself to read it, and people don't get paid for reading stuff on HN. The HN value of an unpleasant to read article is basically zero.
If the applied style is painful, can't one just take a lightweight model and have it transcribe the content in a style that one personally finds pleasurable? Even phones should be powerfull enough to chew through such a task locally these days (even across languages), and once you define the criteria for what you enjoy you'll never have to be pained again (unless you enjoy it offcourse)
You don't have to leave a comment if you are too lazy to read it. It's not even if the words were written mirrored. Anyone who can read English including children can read the output of AI without needing to force anything. Maybe you need to see a doctor if you need to force yourself to read it.
I agree. One comment constructive is enough.
It feels like some guys pushing the topic away.
It annoys me as it takes my attention with them so with all respect: user aharmetz (and others) please keep it in one post that you’d prefer non-AI text. Point.
There was a very obvious and unnerving shift in style that happened early in the article where it went from human to LLM. I find such shifts really glaring. Agents, in their current state, simply cannot produce quality prose and are not good enough to write blog posts targeting a human audience.
I found the topic fascinating, but eventually I started skimming and then gave up on it.
It says there's 137 live/certain facts in my graph. I can see where they're sourced from/a statistical breakdown of type/the number of unique confidence values/how many citations they're using/how many ingests contributed no facts. Am I missing where I can see what the actual facts are?
A lot of these statistics aren't very useful for mining insights from the graph, from a technical standpoint of how the graph works or from a personal standpoint of how the graph is interacting with my personal data.
You wrote that other apps can't access it because of TCC Full Disk Access. Presumably the OS itself can read it, so do you have any information on whether parts of the graph (or fingerprints, or inferences made with it) are ever sent to the cloud/apple servers?
Just spent 30 minutes poking around this. It's horrifying. Absolutely horrifying. All the intelligence and Siri stuff is turned off and blocked on my Mac yet it's still doing this.
Think this might kick me over the fence finally to Debian.
>All the intelligence and Siri stuff is turned off and blocked on my Mac yet it's still doing this.
Wait until you discover that macOS entirely ignores your manually-input DNS settings, before every bootup, to reach out to security.apple.com.
Have a network-wide PiHole (allegedly) blocking the entire *.apple.com domain? No you don't, (at least) not at bootup, because APPLE DOES NOT RESPECT YOUR PRIVACY and will use any network-discoverable DNS (instead of your configured Network Settings) to [attempt a] phone-home.
----
I finally made the full switch to Linux earlier this year. Been a (primarily) Apple user since 1992.
[†] I am but an observant (on my DNS logs) bluecollar poweruser – my tip will reduce only one vector, and we probably live in a world where IPv6 DNS addresses are [about to be, if not already] being hardcoded into TVs, computer security chips, and phones
----
At an absolute minimum: LittleSnitch will NOT help you here (and isn't doing everything it claims to, under the hood); it is entirely impossible for LS to stop pre-boot DNS queries; even under normal operation (e.g: when a dialogue pops up to `ACCEPT/DENY` a DNS resolution) your ISP still sees every. single. URL resolution (even if you `DENY` the connection).
Before you write "no it doesn't" please check your own DNS logs. It do be like it is.
Speculating out loud: sure, perhaps this userprofile data is stored locally; but perhaps also (like Apple's "advertising token") this local data can help remote servers better track your entire computing humanity. While still "locally stored."
(Language doesn't mean anything when EULAs are dozens of unreadable pages long; "stored locally" is just another set of weasle words, like "then just opt-out" [why is opt-in the default, if -out is even allowed at all?!]).
Wouldn't surprise me if every single webpage has an advertising option to exploit the local user data (via Apple$$$), if only by offering on-device scrubbing; one could then look to see what advertising content the localhost is requesting (and thus have essentially root access to all userdata, "locally stored," by detecting outbound queries) with the user's own host acting as Apple's backendserver (while Apple profits, not you).
My two newest Macs are entirely offline (air-gapped) and used primarily for LLM – maybe Apple's secret (all along) has been to get people like me to purchase twice as much hardware (as would be necessary in a privacy-respecting world)?
My main desk has three Apple Silicons (two entirely airgapped, and am typing on third); sitting next to it is my beautiful 27" iMac (2009), with full Adobe CS4 (also offline); just purchased a pristine MacPro5,1 – seventeen year old computer! – for my main online Apple experience (using OS 10.14 from a decade ago), mostly because it cannot AI/LLM/Siri, even if it wanted to.
But since the beginning of this year, I am now primarily running (online) a 9800X3D with a 5070Ti – Linux - which is only possible with LLM tutorials/help (mostly offline, on my own Silicon LLM).
This is why throwing an extra DNS server on the network doesn't truly work. You have to actually remove the DNS protocol from existence entirely and perform TLS inspection if you want to filter effectively. This of course requires you to install a certificate authority on all devices connected to the network. Otherwise devices and applications will get around you with stuff like this, or DoT/DoH.
TLS inspection is starting to reach general availability but it's pricey. You could probably set something up if you own your router. The only vendor I know offering it OOTB is Ubiquiti/UniFi and only on their more expensive enterprise units that advertise "NeXT AI Cybersecurity". I have a hunch there is no actual AI or even ML involved, but this is the world we live in now.
This does work for unencrypted DNS and DoT over port 853, but it fails for DoT over any other port and DoH, which looks just like a website visit. You can't tell whether a website is acting as a DoH server without TLS inspection, and you can't differentiate between multiple domains on the same IP address either (unless a client is stupid enough to use SNI). Some clients do actually hardcode direct IP addresses to get around entirely missing DNS servers, notably Windows.
At least they're honest, using host <spynet2.microsoft.com> to replicate Apple's <smoot.> & <security.> subdomains.
>This does work for unencrypted DNS and DoT over port 853, but it fails for DoT over any other port and DoH
My task for today is to learn more about this (I'm only doing the unencryptedDNS part) and then implement an entire subnet that is whitelist-only (for all DNS). I think what you wrote will help solve this better, thanks.
The bad news is the hardware. The most important parts of a laptop, to me anyway, is that it runs silent and doesn't get hot to touch. Not just to avoid sweaty fingers but because I usually have my laptop in or on my lap. Nothing I can find could replace my m1 8gb macbook air in that regard.
I can find some pretty cool linux laptops from framework or tuxedo, but they have fans and they get hot.
I'm quite frustrated with some aspects of macOS, particularly driving it from the keyboard, so I'm probably going to get a better experience out of it in the end.
With Hammerspoon⁽³⁾ and the Accessibility API, I’ve found there’s not really much I can’t do.
For example, I use the Claude desktop app, but when I copy something from it, the Clipboard gets multiple content types, including RTF and Markdown, but I usually want the latter. So I wrote a quick snippet in Hammerspoon so that when Claude.app is the active app and the Clipboard changes, strip all the parts from the Clipboard except the “UTF8” part so I get just the Markdown (unless the Shift key is held down, which leaves it as-is). No more “formatted text” when I want to paste the Markdown.
My biggest problem now is that when I help somebody else on their Mac, I’m so used to “my controls” that I have to stop and think how to do things the normal way.
This is deeply troubling,
this data is unnecessary collected and impossible to opt out from apparently and may be read at any point in the future by an attacker or by apple itself without notice
This is context for local LLMs like Siri so that when you say “Text James and tell him …” it knows which James you mean.
Likewise, you could say “Tell my boss I’m not coming in tomorrow” and it would have to know who your boss is.
We could get super angry about this, but seeing as it’s fully local and doesn’t leave the device, it really doesn’t seem like a big deal. Is it much different to your car remembering who you call and where you go at what times for navigation purposes?
It doesnt leave your device...until an authoritarian government extracts it without a warrant. Or a dozen other entities like divorce attorneys, business rivals, legal trolls, false accusers, nation states, teenage hackers, human traffickers.
How often is this endangering info available to anyone at my work or home when someone leaves their laptop open while going to the bathroom for 5 minutes?
It's irresponsible to collect and concentrate this much info on millions of people.
It's a no win situation. All of this surveillance is happening on the internet right now. All the companies are doing this surveillance Apple happens to be doing it on device. This is the data Apple is trying to prevent from leaving the device. Apple is constantly fighting court battles to protect the data. Apple is always being accused of hampering developers freedom's while trying to safeguard access to this data.
All that said turning off Apple Intelligence should eliminate the collection.
If my threat model realistically included all of those things, then I would be very careful about keeping any sort of address book or contact log/history at all. At that point the only defense really is just to try to memorize phone numbers. And only talk when walking together, mob-style.
... I feel like if an authoritarian government could extract the family classifier info from local, they could just extract the damn messages.
And if you think that handful of megabytes is too much, they could extract a million other signals that tell them this, such as who you have pinned in the relevant apps.
What does it mean "fully local"? I have an old "kitchen Macbook" that I'm using pretty much exclusively for stuff like looking up cooking recepies. I don't have any photos there, any messaging apps, etc. But I'm logged in with my Apple account.
And guess what, that "local graph" sees everything I did on my iPhone - I see contacts from all messaging apps, including Signal. I see image-based people clusters, even though there are no local pictures to match faces against.
Im still optimistic that Apple will do all this stuff on your device as opposed to in an opaque cloud somewhere you have 0 visibility into. Meta probably has 1000x this system in a data warehouse somewhere. If this info is not leaving the device that’s fine with me.
Yes, but having some kind of documentation and way to disable it would have been reasonable. Also, the thing is that it seems that apple intelligence models can query this information. So, you know, everything is a prompt injection away in theory.
Probably, but how much meta hardware do you own? Hell how much meta software do you have installed? I know people use varied amounts, but meta doesn't really make anything worth owning. I guess Apple is speedrunning the point where they don't either, which sucks, because I have no idea where to find a fanless linux laptop that doesn't get hot but has the performance of a m1.
My understanding is that collecting these data and constructing the knowledge graph violates the European General Data Protection Directive (GDPR). According to the GDPR, personal data (such as an email's sender name and email address) are provided for a specific purpose (in the case of sending an email, for identifying the sender and facilitating email replies). Similarly, a calendar invitation is provided to organize a meeting. Processing that data for a different purpose (as described here, for constructing a knowledge graph) without the subject's explicit consent is unlawful.
I'm not sure. IMO when you buy a Mac (or use Windows or an Android phone) you have accepted so much that there is nothing which is not allowed.
But nevertheless: as far as I have understood it everything is built on your computer and is still used on your computer. You - as the user of the Mac - has took and saved the information. Your Mac is then only working for you - if Apple doesn't load the information in their enviroment.
This is going to depend on the holder's relationship with the data subject.
If you send a friend an email, they may add you to their address book and set a reminder for your birthday while they're at it. They don't need your permission to do that, so why would a device acting on their behalf?
On the other hand, if a customer of mine asks for the data I have on them, do I have to send them my Personal Knowledge Graph facts about them? That's less obvious, and I'd lean towards Apple's helpful database being a liability even if it's local.
It's on the person's Mac so it is a little more complicated. The person you send your data to is the Data Controller, and fails you by not publishing their privacy policy. But Apple isn't a Sub-Processor here, because they aren't sending the data to Apple.
I'm saying that because I wish the GDPR also covered the software providers who perform Data Processing against the will of all parties, but it isn't the case. Lastly, they're a person and not a business.
It is however clear that building a database of people which may contain race, religion or political opinions is illegal, even if it's just a person doing it.
"It's for you own good. It would be harmful to disable it. And you don't understand it so you shouldn't have an opinion. I trust Apple to do this safely, even if I don't want whatever this is."
Is how I'm generally greeted when I ask this sort of question.
Turn off your Mac and use a different operating system. Because it doesn't matter how you disable this feature in isolation. It is merely the current projection of Apple's opinion of its relationship with you. It's a hydra and you are going to lose the fight.
If Microsoft didn't this, this wouldn't be interesting research, but a cause for uproar.
And yes, I know they're probably already doing it, but in the cloud.
Apple makes money of your data as does anyone else. The just use your compute and energy for it too and convinced everyone that was privacy.
do i see that right - the data is unencrypted, the profile is built by an undocumented background profiler without consent or an option of opting out, it is consumed by cloud connected platforms like siri and apple intelligence and ANY application on my mac can access it?
Yeah I could see that it correlated my WhatsApp calls with real person names (i.e. de-anonymizing WhatsApp), inferred my race based on face clusters (illegal under AI act), built a co-appearance network ("who stands beside whom in photos"), etc. etc.
None of this shit is legal, but the laws are toothless anyway: Apple can get "fined" a couple million and that's it. The fallout from the press would be a harder hit for them.
It’s not like Apple is actually hiding this. I mean it’s sitting there in perfectly readable structured format. I’m most surprised this isn’t known earlier, it indicates to me how little MacOS users investigate their own devices if we are talking about this 4 years after it rolled out.
Ack. For what I could see, specific tables are added / removed across versions. So, some information might not be visualized (e.g., the graph is tiny with few nodes).
The output stays local on your disk (in theory). So I’m not sure that breaks any rule. But in theory, Apple Intelligence can query it. Again, close to 0 documentation, so it is hard to know how this is consumed.
1) Before macOS boots, LittleSnitch ("LS") does not have any ability to block hostnames, e.g: on every boot, any modern Apple product ignores any manually-configured DNS IP and attempts to phone-home
2) Allegedly, this user profile information is stored locally (so LS wouldn't see any outgoing connection attempt to Apple servers)
3) LS does not block DNS-resolution, only IP-connection (i.e. your ISP sees all DNS queries, whether you click `ACCEPT` or `DENY` when LS popup asks)
4) If you're using LS with intentions of primary DNS blockage, you have failed; try pi-hole (or other local hardware DNS resolver); use physically-dedicated hardware (e.g: Raspberry Pi) to run your PiHole, not a Docker/Prox/&c
5) I am investigating another user's comment on using TLS certs to DENY-all DNS, but need to research further on this (new info to me!)
Uh, am I the only one for whom my Mac apparently knows next to nothing about my social circles beyond my parents?
(MacOS Golden Gate 27.1, and yes some of the pages do say that the data isn't working, but the ones that do work are hilariously wrong once you get beyond my parents, which, since they are noted as "mom" and "dad", well, that's a freebie.
I'm definitely not your average user, but still. It has eight photos of me but since I generate random EXIF data when they get moved onto my hard drive and I use an Android phone, the identification between photo and me, - and I video chat, facetime with my mom who's on vacation in Egypt and my dad who's on business in China, among others, did not seem to get recorded. I also have a half-sister who is not mentioned anywhere, but it did identify an obviously fake entity in my contacts as my: family, partner, child, and sibling. None of that is true. It also identified a throwaway number that I control as part of my family, a friend of is dead as my family, my lawyer misses the cut at 0.47 for friends, my aunts, cousins, and uncles do not make the list. Some are mostly overseas - Canada, China, Hong Kong, but my father spends 6 months out of the year in China and makes the list. My partner, to whom I'm not married and have no plan to do so, but does live with me 6 months out of the year and works on her family's ranch in the state and we even share family plan for our phone service, is nowhere on the list, and she has an iPhone.
Johnny Appleseed features prominently in my FAF. One of my close friends is known only through an employer, even though I'm self-employed. I'm not much of a photo-taker and certainly not with my iPad, and so the photo clusters are mainly... baseball cards, which I process via a ADF scanner, since taking 5 digits worth of baseball cards, front and back, is ridiculous.But not more ridiculous than how it seems unable to fully or even partially beyond the most basic aspects of the data it can access to make connections with any sort of certainty that is correct. It identified half of my exes but only barely. It decided that obviously fake (555 area code, come on) entries in my contacts as close friends, and it seems to make the fundamental mistake of creating connections under the assumptions that seems ripped from Anna Karenina, but only made its way half through the epigraph and mistranslated it.My MacBook does not get turned off entirely unless I'm out and about with it, but it's easier to simply use my iPad Pro M4 and my Pixel - neither particularly known for privacy - instead of lugging my entire setup around unless I'm going on a trip, which happens twice a year usually.
I get the sense that this shows what my MacOS doesn't know what to do with the data it has. I'm not going to enrich it so, let it be, I guess. But just to be clear, I'm not the partner, father, and sibling of some entity that I have never had any contact with. I also do not use Apple Mail (eM Client is what I prefer). Although now I'm going to dig up an Almodovar movie or something I guess.
At one point in my life, I would've found this data fascinating (if it were opt-in, perhaps)! Now, I find it exhausting.
A similar discussion came up recently with Siri Recap[1], where some expressed how useful they thought it was and that, as a party to the conversations it summarizes, there was very little moral or ethical concern. But I have a social concern.
If I'm having a chat with someone, there's a milieu constructed around the fact that not everything will be remembered verbatim. There are all sorts of social and contextual cues that can easily get lost if listening gets replaced by flat relentless transcription.
With Recaps and this social graph (were it to be productized), the informal fluidity in the way we relate to one another starts to go away, and the expectations that follow from that become artificial and burdensome. In the same way that autocorrect has made typos unacceptable, how could you ever be forgiven for forgetting the name of your colleague's new nephew or the Italian village that your neighbor took a spontaneous boat trip to while on vacation last year?
[1]: https://news.ycombinator.com/item?id=49641401
I too would've found this fascinating at one point, but my jaded older self now just finds this incredibly creepy, unsettling and not at all surprising.
Maybe a decade or so ago I had a discussion with a colleague about chat apps and he said, why would you want to have a conversation on snapchat if it is not saved, and I was like, why are you having this conversation now, you can't save it either.
I guess that those days are soon over.
Autocorrect has had the opposite effect IMO, there’s an expectation that like 1 in 8 words from an iPhone user are not what the sender intended.
Typos have just graduated from misspellings to sentences randomly being nonsensical.
I turned autocomplete off for this reason
Did your autocorrect correct “autocorrect” to “autocomplete”?
I am the author of the research. Please, feel free to ask me anything. Also, check out the followup research: https://8rank.ai/game/apple-intelligence-knowledge-graph-in-...
Please stop using AI to author your articles. It was really hard to read it despite how interesting the topic was. I dropped halfway because it was just too annoying.
Yeah seriously. Article becomes unreadable immediately after the intro. I believe this author probably could actually write two years ago, it's a shame people are just giving up on that and choose to turn out stuff like this instead.
I was bothered not so much by the slopful article as by the AI-written text in the dashboard that iKnow produces. I ran the program on my Mac and it did display charts and graphs with names of people and organizations I’ve interacted with over the years, but I couldn’t figure out what it all was supposed to mean.
One example: "323 people, 122 events and other records, 200 employers and 1 stored group — as the store on this Mac links them. Read live just now. Colour is what each node is to you: the stored group in red, family and closeness by iKnow’s rules on the kinship scorecard. Size is how many rows in the store name it."
Claude etc. can be prompted to produce clearer explanations than that.
My bad. In my defense, only the technical parts are AI. But, thanks for the feedback; I thought I had unsloped it enough...
It's not possible to un-AI text, only replace it wholesale.
Suggestion for next time: "Review my prose for accuracy, grammar, appropriate strength of claims, etc. Flag any issues for my attention. Do not provide text; this work should not contain any AI-written content." That'll turn it into a very good check that doesn't take your voice away.
Complaints of AI are like a form of friction that eventually goes away. Either AI will undergo a change to create more approachable material, or in general people's tastes will change as they accommodate the style and get used to it. Or some mix of the two. As young people continue to grow up reading AI generated content they will likely begin to find that person written content from before AI grates on their nerves. More, they'll even ask AI to reword documents into preferable AI-like styles.
As an older person, AI grates on me, but I found the meat like "It is a triple store written in a vocabulary of 181 entity classes and 278 predicates: every row is a subject, a predicate and an object. The subject is an entity (a person, an app, an event); the object is either a plain value (a name, a handle, a date) or another entity, which is where the graph gets its edges. The predicate is the interesting part: it says what kind of relationship the system believes holds between the two" quite approachable.
So it was like a meal where I just have to push, say, mushrooms aside as best I can to get to what I like. And there was a time I would have refused the entire plate because I truly abhor mushrooms. I blame watching Matango as an impressionable child.[0]
[0] https://www.youtube.com/watch?v=yOpbTaX6HWQ
AI writing is fine in a constrained space.
If you create a strict template, like a PowerPoint deck or tightly formatted report, the LM does a really, really good job. at populating that template with good data.
If you give it an open license to just create stuff, it’s always terrible for the general audience — but it may be fine for you or a small group that is familiar with the subject.
I compare it to fast food. An uneducated teenager is quite capable of putting together a McDonald’s lunch within the constraints of the system. If you let the kid loose in a kitchen designed for trained chefs, the result would not be a good lunch.
>Complaints of AI are like a form of friction that eventually goes away.
This. Already there is evolution in AI-related comments here.
For example, a year ago they often featured remarks like "I know AI when I see it: here are unmistakeable 'tells.'"
Then there would be a lively back-and-forth about the validity of the stated giveaways.
Such "How I know it's AI"-type comments no longer appear.
This has not been my experience. People talk about LLM-isms (Claude-isms especially) all the time. "The absolute shape", "load bearing", "seams", "smoking gun", genuinely empty preambles, constant sloganesque phrasing. If you're seeing it less in conversation, it's not because it's improved.
Genius-level response. And honestly, that's valid.
Sorry your comment got dinged. I assume the reference to age triggered that. Ageism is a very, very real thing, hereabouts.
I think it was insightful and relevant. I am not a “slop enthusiast” (a great term), but I am a longtime student of human nature, and basically agree with the premise.
Like it or not, slop is here to stay. It’s up to us, to guide and shape it into a net benefit.
It is funny to me how slop enthusiasts argue from the past and the future to deny the present state.
"Slop is here to stay so you better get used to it because slop is inevitable and you will be left behind."
It is both already a cemented permanent presence and destined to arrive by an unbreakable prophecy at a date TBD.
[dead]
If you came away thinking I was enthusing about AI slop, it's probably because I was too subtle in comparing it to a food I abhor while saying there are tidbits, non-slop tidbits, that I could appreciate. Apologies for that human generated mischance!
I'm actually rather bemused to see many young folks, arguing against LLMs and ML, and seeing more experienced folks, embracing it.
I'm pretty experienced, and have found LLMs to be quite useful.
Like any new tool, there are deficits and peculiarities, but these are becoming less of a concern, and the utility of these tools is advancing rapidly.
> In my defense, only the technical parts are AI.
A defense would be "I just used it for spellchecking" not "I just used it for the meat and the important parts people care about, the rest is mine!"
So most of it?
You can’t unslop A.I. prose. Once it’s written, the content is poisoned.
“Sure there was a turd in the punch bowl but I did my best to remove it, so you should be able to drink it.”
Nope.
And the technical parts are what matter, saying you did that with A.I. tells me “I don’t care about this part” so what the heck are we even doing here?
So only the parts that matter, you ai slopped. Great work.
I second this. Unfortunately you just can’t just “unslop” it, much like you can’t just “shake the dirt away” from food that fell into a bucket of… you get my point, it won’t be good to eat anyway. Bonus additional ironic parallel could magically manifest if you can guess the bucket content.
Anyway, it’s a shame, this could have been excellent work. Instead, I find reading it unpleasant and part of my attention rots away wondering if you even checked what the LLM wrote for you. You probably did, and I could also do it myself, but still…
Is this really the top question now?
Can we add a rule that if you blame someone for using AI, you _also_ have to bring something interesting to the table?
Why? "There is gravel in the food" is a perfectly valid contribution. You don't need to write a compelling story about it.
Because the article itself is way more interesting than the question of with what style it was written.
It's like the saying beauty is only skin deep, but ugly goes clean to the bone. Bad AI writing makes an article "unreadable" - you'd have to force yourself to read it, and people don't get paid for reading stuff on HN. The HN value of an unpleasant to read article is basically zero.
I really don't want to waste more words on it here.
I think the difference might be that you find bad style easy to ignore and I don't.
If the applied style is painful, can't one just take a lightweight model and have it transcribe the content in a style that one personally finds pleasurable? Even phones should be powerfull enough to chew through such a task locally these days (even across languages), and once you define the criteria for what you enjoy you'll never have to be pained again (unless you enjoy it offcourse)
The criteria I enjoy is "written by a human".
You can pay someone to rewrite it. There's loads of college students who wouldn't mind extra spending money.
You don't have to leave a comment if you are too lazy to read it. It's not even if the words were written mirrored. Anyone who can read English including children can read the output of AI without needing to force anything. Maybe you need to see a doctor if you need to force yourself to read it.
"How dare you have a preference that is not my preference?"
Its interestingness is zero if it’s so aggravating to read that you just give up.
> Can we add a rule that if you blame someone for using AI, you _also_ have to bring something interesting to the table?
No.
I agree. One comment constructive is enough. It feels like some guys pushing the topic away.
It annoys me as it takes my attention with them so with all respect: user aharmetz (and others) please keep it in one post that you’d prefer non-AI text. Point.
Thank you.
I’d rather this over the very long tradition in this site if people bellyaching about site design and layout on otherwise u related comment threads
> Can we add a rule that if you blame someone for using AI, you _also_ have to bring something interesting to the table?
Criticism is a valid 'something interesting'. And growing numbers of us also don't want to feed the US mega-companies LLM training sets.
More and more people are keeping human-generated stuff to themselves, cause of that.
There was a very obvious and unnerving shift in style that happened early in the article where it went from human to LLM. I find such shifts really glaring. Agents, in their current state, simply cannot produce quality prose and are not good enough to write blog posts targeting a human audience.
I found the topic fascinating, but eventually I started skimming and then gave up on it.
It says there's 137 live/certain facts in my graph. I can see where they're sourced from/a statistical breakdown of type/the number of unique confidence values/how many citations they're using/how many ingests contributed no facts. Am I missing where I can see what the actual facts are?
A lot of these statistics aren't very useful for mining insights from the graph, from a technical standpoint of how the graph works or from a personal standpoint of how the graph is interacting with my personal data.
You wrote that other apps can't access it because of TCC Full Disk Access. Presumably the OS itself can read it, so do you have any information on whether parts of the graph (or fingerprints, or inferences made with it) are ever sent to the cloud/apple servers?
Just spent 30 minutes poking around this. It's horrifying. Absolutely horrifying. All the intelligence and Siri stuff is turned off and blocked on my Mac yet it's still doing this.
Think this might kick me over the fence finally to Debian.
>All the intelligence and Siri stuff is turned off and blocked on my Mac yet it's still doing this.
Wait until you discover that macOS entirely ignores your manually-input DNS settings, before every bootup, to reach out to security.apple.com.
Have a network-wide PiHole (allegedly) blocking the entire *.apple.com domain? No you don't, (at least) not at bootup, because APPLE DOES NOT RESPECT YOUR PRIVACY and will use any network-discoverable DNS (instead of your configured Network Settings) to [attempt a] phone-home.
----
I finally made the full switch to Linux earlier this year. Been a (primarily) Apple user since 1992.
> APPLE DOES NOT RESPECT YOUR PRIVACY and will use the router-supplied DNS (instead of your configured Network Settings).
is that for real? wow
Is for real.
I provided a "better DNS configuration" suggestion, a month ago, but at this point I doubt it will remain useful [†]:
<https://news.ycombinator.com/item?id=50071310#50078574> [have router's DHCP auto-issue DNS to your pi-hole].
[†] I am but an observant (on my DNS logs) bluecollar poweruser – my tip will reduce only one vector, and we probably live in a world where IPv6 DNS addresses are [about to be, if not already] being hardcoded into TVs, computer security chips, and phones
----
At an absolute minimum: LittleSnitch will NOT help you here (and isn't doing everything it claims to, under the hood); it is entirely impossible for LS to stop pre-boot DNS queries; even under normal operation (e.g: when a dialogue pops up to `ACCEPT/DENY` a DNS resolution) your ISP still sees every. single. URL resolution (even if you `DENY` the connection).
Before you write "no it doesn't" please check your own DNS logs. It do be like it is.
Wow!!
The above are facts.
----
Speculating out loud: sure, perhaps this userprofile data is stored locally; but perhaps also (like Apple's "advertising token") this local data can help remote servers better track your entire computing humanity. While still "locally stored."
(Language doesn't mean anything when EULAs are dozens of unreadable pages long; "stored locally" is just another set of weasle words, like "then just opt-out" [why is opt-in the default, if -out is even allowed at all?!]).
Wouldn't surprise me if every single webpage has an advertising option to exploit the local user data (via Apple$$$), if only by offering on-device scrubbing; one could then look to see what advertising content the localhost is requesting (and thus have essentially root access to all userdata, "locally stored," by detecting outbound queries) with the user's own host acting as Apple's backendserver (while Apple profits, not you).
¢¢
Added to the list of reasons I need to look at this seriously. Thank you!
Think I need a per-host approval mechanism for DNS as well. There are fewer domains I want my computer talking to than not.
Keep adding to your list, friend:
<https://news.ycombinator.com/item?id=50071310#50078574> [Spotlight's "helpfulness," which cannot be disabled (viaGUI)]
Apple's OS is toxic, at this point.
My two newest Macs are entirely offline (air-gapped) and used primarily for LLM – maybe Apple's secret (all along) has been to get people like me to purchase twice as much hardware (as would be necessary in a privacy-respecting world)?
Well fuck.
Just dug my old Xeon workstation out of the cupboard and am installing Debian on it.
My main desk has three Apple Silicons (two entirely airgapped, and am typing on third); sitting next to it is my beautiful 27" iMac (2009), with full Adobe CS4 (also offline); just purchased a pristine MacPro5,1 – seventeen year old computer! – for my main online Apple experience (using OS 10.14 from a decade ago), mostly because it cannot AI/LLM/Siri, even if it wanted to.
But since the beginning of this year, I am now primarily running (online) a 9800X3D with a 5070Ti – Linux - which is only possible with LLM tutorials/help (mostly offline, on my own Silicon LLM).
----
Happy restomodding!
This is why throwing an extra DNS server on the network doesn't truly work. You have to actually remove the DNS protocol from existence entirely and perform TLS inspection if you want to filter effectively. This of course requires you to install a certificate authority on all devices connected to the network. Otherwise devices and applications will get around you with stuff like this, or DoT/DoH.
TLS inspection is starting to reach general availability but it's pricey. You could probably set something up if you own your router. The only vendor I know offering it OOTB is Ubiquiti/UniFi and only on their more expensive enterprise units that advertise "NeXT AI Cybersecurity". I have a hunch there is no actual AI or even ML involved, but this is the world we live in now.
Can one not just use their router's firewall to drop all DNS queries which do not originate from your PiHole?
This does work for unencrypted DNS and DoT over port 853, but it fails for DoT over any other port and DoH, which looks just like a website visit. You can't tell whether a website is acting as a DoH server without TLS inspection, and you can't differentiate between multiple domains on the same IP address either (unless a client is stupid enough to use SNI). Some clients do actually hardcode direct IP addresses to get around entirely missing DNS servers, notably Windows.
>notably Windows
At least they're honest, using host <spynet2.microsoft.com> to replicate Apple's <smoot.> & <security.> subdomains.
>This does work for unencrypted DNS and DoT over port 853, but it fails for DoT over any other port and DoH
My task for today is to learn more about this (I'm only doing the unencryptedDNS part) and then implement an entire subnet that is whitelist-only (for all DNS). I think what you wrote will help solve this better, thanks.
The good news is that with AI, nothing stops Linux desktops from looking and feeling like (Former) osx..
The bad news is the hardware. The most important parts of a laptop, to me anyway, is that it runs silent and doesn't get hot to touch. Not just to avoid sweaty fingers but because I usually have my laptop in or on my lap. Nothing I can find could replace my m1 8gb macbook air in that regard.
I can find some pretty cool linux laptops from framework or tuxedo, but they have fans and they get hot.
I'm quite frustrated with some aspects of macOS, particularly driving it from the keyboard, so I'm probably going to get a better experience out of it in the end.
Have you looked at Karabiner Elements⁽¹⁾? It’s extremely capable; see a config like this⁽²⁾ for inspiration.
⁽¹⁾ https://karabiner-elements.pqrs.org/
⁽²⁾ https://github.com/pdelfino/karabiner-config
Yes I'm using it because I don't use a standard Apple keyboard with my Mac.
The issue is mostly the user interface design is mouse first, not keyboard first. Windows was more than a little better on that front.
With Hammerspoon⁽³⁾ and the Accessibility API, I’ve found there’s not really much I can’t do.
For example, I use the Claude desktop app, but when I copy something from it, the Clipboard gets multiple content types, including RTF and Markdown, but I usually want the latter. So I wrote a quick snippet in Hammerspoon so that when Claude.app is the active app and the Clipboard changes, strip all the parts from the Clipboard except the “UTF8” part so I get just the Markdown (unless the Shift key is held down, which leaves it as-is). No more “formatted text” when I want to paste the Markdown.
My biggest problem now is that when I help somebody else on their Mac, I’m so used to “my controls” that I have to stop and think how to do things the normal way.
⁽³⁾ https://www.hammerspoon.org/
one would think, but a lot of things stopped Linux to just fucking work with Wifi on my machine
This is deeply troubling, this data is unnecessary collected and impossible to opt out from apparently and may be read at any point in the future by an attacker or by apple itself without notice
this is a class action lawsuit in the making
This is context for local LLMs like Siri so that when you say “Text James and tell him …” it knows which James you mean.
Likewise, you could say “Tell my boss I’m not coming in tomorrow” and it would have to know who your boss is.
We could get super angry about this, but seeing as it’s fully local and doesn’t leave the device, it really doesn’t seem like a big deal. Is it much different to your car remembering who you call and where you go at what times for navigation purposes?
It doesnt leave your device...until an authoritarian government extracts it without a warrant. Or a dozen other entities like divorce attorneys, business rivals, legal trolls, false accusers, nation states, teenage hackers, human traffickers.
How often is this endangering info available to anyone at my work or home when someone leaves their laptop open while going to the bathroom for 5 minutes?
It's irresponsible to collect and concentrate this much info on millions of people.
It’s not clear whether this still happens in the “Lockdown” mode that Apple specifically added for people who expect to be in this scenario.
It's a no win situation. All of this surveillance is happening on the internet right now. All the companies are doing this surveillance Apple happens to be doing it on device. This is the data Apple is trying to prevent from leaving the device. Apple is constantly fighting court battles to protect the data. Apple is always being accused of hampering developers freedom's while trying to safeguard access to this data. All that said turning off Apple Intelligence should eliminate the collection.
>All that said turning off Apple Intelligence should eliminate the collection.
Amen.
If my threat model realistically included all of those things, then I would be very careful about keeping any sort of address book or contact log/history at all. At that point the only defense really is just to try to memorize phone numbers. And only talk when walking together, mob-style.
... I feel like if an authoritarian government could extract the family classifier info from local, they could just extract the damn messages.
And if you think that handful of megabytes is too much, they could extract a million other signals that tell them this, such as who you have pinned in the relevant apps.
So why does it keep collecting data when I have chosen to entirely disable Siri?
What does it mean "fully local"? I have an old "kitchen Macbook" that I'm using pretty much exclusively for stuff like looking up cooking recepies. I don't have any photos there, any messaging apps, etc. But I'm logged in with my Apple account.
And guess what, that "local graph" sees everything I did on my iPhone - I see contacts from all messaging apps, including Signal. I see image-based people clusters, even though there are no local pictures to match faces against.
Im still optimistic that Apple will do all this stuff on your device as opposed to in an opaque cloud somewhere you have 0 visibility into. Meta probably has 1000x this system in a data warehouse somewhere. If this info is not leaving the device that’s fine with me.
I have lost my faith in that
Yes, but having some kind of documentation and way to disable it would have been reasonable. Also, the thing is that it seems that apple intelligence models can query this information. So, you know, everything is a prompt injection away in theory.
I had faith in Apple until I saw the technology oligarchy and how they behaved around the current US administration.
Probably, but how much meta hardware do you own? Hell how much meta software do you have installed? I know people use varied amounts, but meta doesn't really make anything worth owning. I guess Apple is speedrunning the point where they don't either, which sucks, because I have no idea where to find a fanless linux laptop that doesn't get hot but has the performance of a m1.
Horrifying, but also really interesting as a student of schema design, especially for probabilistic or inferential processes, and graph data.
Woah. Still on Sequoia with no intention to upgrade, but maybe the next laptop runs Linux.
I'm on Sequoia. Check out ~/Library/IntelligencePlatform
It might depend on your needs and priorities, GhostBSD is a great option for people who like something more like Mac.
It's just a tuned FreeBSD distro using Mate.
My understanding is that collecting these data and constructing the knowledge graph violates the European General Data Protection Directive (GDPR). According to the GDPR, personal data (such as an email's sender name and email address) are provided for a specific purpose (in the case of sending an email, for identifying the sender and facilitating email replies). Similarly, a calendar invitation is provided to organize a meeting. Processing that data for a different purpose (as described here, for constructing a knowledge graph) without the subject's explicit consent is unlawful.
I'm not sure. IMO when you buy a Mac (or use Windows or an Android phone) you have accepted so much that there is nothing which is not allowed.
But nevertheless: as far as I have understood it everything is built on your computer and is still used on your computer. You - as the user of the Mac - has took and saved the information. Your Mac is then only working for you - if Apple doesn't load the information in their enviroment.
As a general rule you cannot override national law by accepting service terms.
Unlike Americans seem to think, terms of use don’t override law.
Someone should start a company with terms mentioning that whoever agrees to it will live forever.
If we are overriding federal laws, might as well make people immortal too.
https://news.ycombinator.com/item?id=47873517
This is going to depend on the holder's relationship with the data subject.
If you send a friend an email, they may add you to their address book and set a reminder for your birthday while they're at it. They don't need your permission to do that, so why would a device acting on their behalf?
On the other hand, if a customer of mine asks for the data I have on them, do I have to send them my Personal Knowledge Graph facts about them? That's less obvious, and I'd lean towards Apple's helpful database being a liability even if it's local.
It's on the person's Mac so it is a little more complicated. The person you send your data to is the Data Controller, and fails you by not publishing their privacy policy. But Apple isn't a Sub-Processor here, because they aren't sending the data to Apple.
I'm saying that because I wish the GDPR also covered the software providers who perform Data Processing against the will of all parties, but it isn't the case. Lastly, they're a person and not a business.
It is however clear that building a database of people which may contain race, religion or political opinions is illegal, even if it's just a person doing it.
How does one disable this?
you can't
"It's for you own good. It would be harmful to disable it. And you don't understand it so you shouldn't have an opinion. I trust Apple to do this safely, even if I don't want whatever this is."
Is how I'm generally greeted when I ask this sort of question.
Turn off your Mac and use a different operating system. Because it doesn't matter how you disable this feature in isolation. It is merely the current projection of Apple's opinion of its relationship with you. It's a hydra and you are going to lose the fight.
or "grandma needs this to just work, Apple knows better"
By using an OS that doesn't view you as cattle to be milked and eventually sent to the slaughterhouse.
Articles like this make me rethink my politics. I now think software tech monopolies need some 1950's style broadcast TV level regulation.
If Microsoft didn't this, this wouldn't be interesting research, but a cause for uproar. And yes, I know they're probably already doing it, but in the cloud.
Apple makes money of your data as does anyone else. The just use your compute and energy for it too and convinced everyone that was privacy.
On a long enough timeline, all closed-source software turns into spyware.
What about iphones?
this violates GDPR in so many ways, jeez
do i see that right - the data is unencrypted, the profile is built by an undocumented background profiler without consent or an option of opting out, it is consumed by cloud connected platforms like siri and apple intelligence and ANY application on my mac can access it?
Yeah I could see that it correlated my WhatsApp calls with real person names (i.e. de-anonymizing WhatsApp), inferred my race based on face clusters (illegal under AI act), built a co-appearance network ("who stands beside whom in photos"), etc. etc.
None of this shit is legal, but the laws are toothless anyway: Apple can get "fined" a couple million and that's it. The fallout from the press would be a harder hit for them.
It’s not like Apple is actually hiding this. I mean it’s sitting there in perfectly readable structured format. I’m most surprised this isn’t known earlier, it indicates to me how little MacOS users investigate their own devices if we are talking about this 4 years after it rolled out.
I'm going to be very angry if my mac turns out to be doing this
We also created a tool to inspect your graph: https://github.com/8rank/iKnow Only tested on recent MacOS versions (last 26s, and early 27s), though
Works on Sequoia too fwiw
Ack. For what I could see, specific tables are added / removed across versions. So, some information might not be visualized (e.g., the graph is tiny with few nodes).
Also works on Sonoma.
>iKnow is fully LLM-generated and may not be 100% accurate.
"We" sounds pretty damn generous if you don't even know whether it works.
Can we start a class action suit?
The output stays local on your disk (in theory). So I’m not sure that breaks any rule. But in theory, Apple Intelligence can query it. Again, close to 0 documentation, so it is hard to know how this is consumed.
In Germany, it breaks TDDDG §25 verbatim, storage of data not explicitly consented to and not technically necessary is already illegal.
can we block these two processes using little snitch?
No.
1) Before macOS boots, LittleSnitch ("LS") does not have any ability to block hostnames, e.g: on every boot, any modern Apple product ignores any manually-configured DNS IP and attempts to phone-home
2) Allegedly, this user profile information is stored locally (so LS wouldn't see any outgoing connection attempt to Apple servers)
3) LS does not block DNS-resolution, only IP-connection (i.e. your ISP sees all DNS queries, whether you click `ACCEPT` or `DENY` when LS popup asks)
4) If you're using LS with intentions of primary DNS blockage, you have failed; try pi-hole (or other local hardware DNS resolver); use physically-dedicated hardware (e.g: Raspberry Pi) to run your PiHole, not a Docker/Prox/&c
5) I am investigating another user's comment on using TLS certs to DENY-all DNS, but need to research further on this (new info to me!)
The processes aren't (as far as anyone knows) slinging data over the network.
Uh, am I the only one for whom my Mac apparently knows next to nothing about my social circles beyond my parents?
(MacOS Golden Gate 27.1, and yes some of the pages do say that the data isn't working, but the ones that do work are hilariously wrong once you get beyond my parents, which, since they are noted as "mom" and "dad", well, that's a freebie.
I'm definitely not your average user, but still. It has eight photos of me but since I generate random EXIF data when they get moved onto my hard drive and I use an Android phone, the identification between photo and me, - and I video chat, facetime with my mom who's on vacation in Egypt and my dad who's on business in China, among others, did not seem to get recorded. I also have a half-sister who is not mentioned anywhere, but it did identify an obviously fake entity in my contacts as my: family, partner, child, and sibling. None of that is true. It also identified a throwaway number that I control as part of my family, a friend of is dead as my family, my lawyer misses the cut at 0.47 for friends, my aunts, cousins, and uncles do not make the list. Some are mostly overseas - Canada, China, Hong Kong, but my father spends 6 months out of the year in China and makes the list. My partner, to whom I'm not married and have no plan to do so, but does live with me 6 months out of the year and works on her family's ranch in the state and we even share family plan for our phone service, is nowhere on the list, and she has an iPhone.
Johnny Appleseed features prominently in my FAF. One of my close friends is known only through an employer, even though I'm self-employed. I'm not much of a photo-taker and certainly not with my iPad, and so the photo clusters are mainly... baseball cards, which I process via a ADF scanner, since taking 5 digits worth of baseball cards, front and back, is ridiculous.But not more ridiculous than how it seems unable to fully or even partially beyond the most basic aspects of the data it can access to make connections with any sort of certainty that is correct. It identified half of my exes but only barely. It decided that obviously fake (555 area code, come on) entries in my contacts as close friends, and it seems to make the fundamental mistake of creating connections under the assumptions that seems ripped from Anna Karenina, but only made its way half through the epigraph and mistranslated it.My MacBook does not get turned off entirely unless I'm out and about with it, but it's easier to simply use my iPad Pro M4 and my Pixel - neither particularly known for privacy - instead of lugging my entire setup around unless I'm going on a trip, which happens twice a year usually.
I get the sense that this shows what my MacOS doesn't know what to do with the data it has. I'm not going to enrich it so, let it be, I guess. But just to be clear, I'm not the partner, father, and sibling of some entity that I have never had any contact with. I also do not use Apple Mail (eM Client is what I prefer). Although now I'm going to dig up an Almodovar movie or something I guess.