Apple Intelligence and Privacy @ WWDC '24

June 11, 2024 1:33 AM

Yesterday at WWDC 2024 Apple announced its long-anticipated machine learning effort, a Siri overhaul dubbed "Apple Intelligence." The new system employs LLMs and diffusion model image generation while attempting to maintain a uniquely high level of privacy by splitting queries across three tiers of increasing anonymity and capability: on device, private cloud compute servers, and anonymized opt-in-only ChatGPT calls. Ars coverage on Apple Intelligence, and the ChatGPT integration.

The system will debut in the pending iOS 18, iPadOS 18, and macOS Sequoia releases and is composed of three separate layers:

1) On device, the primary mode which draws upon all personal information across apps, contacts, conversations, etc. to create a highly-detailed, user-specific context Apple refers to as a "semantic index." In addition to the ability to parse information currently displayed on screen when requested, there is a new developer API so that third-party applications can specify what kind of information Siri can draw from them, and request appropriate generated text and images. The specific information gathered and any derived data or personalized fine-tuning remains on your device, with the limited exception of difficult queries which are handed off to...

2) Private Cloud Compute, a semi-anonymous cloud-based neural network inference service hosted by Apple with exposure of personal data limited specifically to the query at hand, using a cryptographically-signed software stack and operating with a no-data-retention policy. The segment on Private Cloud Compute featured an unusually candid critique of the data harvesting common to machine learning systems by competing tech giants, without specifically naming...

3) OpenAI's ChatGPT, which will be available later this year and only with explicit user opt-in (on each individual query) for queries the new Siri detects as likely to benefit from scale beyond both on-device hardware and Private Cloud Compute. Data sent to OpenAI is heavily anonymized and multi-modal (meaning combined text and images) for asking questions about an image. Apple mentioned that other models may later become available, but did not specify whether this meant Google's Gemini, Facebook's Llama-3, or potentially even self-hosted endpoints based on open source models like Mistral 8x7b.

posted by Ryvar (117 comments total)

It's worth noting that Tim Cook has maintained a public position of slow, careful rollout of any machine learning software - one of the only tech giant CEOs publicly insisting on gradual adoption that must provide clear user benefit: “what all of us have to do is to make sure that we are using AI in a way that is for the benefit of humanity, not to the detriment of humanity.”

For my own part - given Microsoft's ramping up of data collection, intrusive ads, haphazard Copilot rollout and the security nightmare of Recall (since rolled back, but the fact that it was ever even considered...) - over the next month I will be transitioning my desktops from Windows to Linux (NixOS + a heavily locked down Windows 10 VM).

Given the enormous amount of thought and care Apple has put into this implementation and the privacy considerations in particular, I will be remaining within the iOS ecosystem despite a certain amount of shock at the scope of what the new Siri will be drawing upon.

As a final consideration: no company or research group has yet managed to produce anything even remotely approaching a state of the art language model without training on copyrighted works, without the consent of the authors. Whether or not you believe this rises to the level of outright theft, it can at least be safely assumed the creation of Apple's models was equally problematic in nature.
posted by Ryvar at 1:46 AM on June 11 [30 favorites]

The only thing I am interested in here is if it can be disabled. I am primarly on Linux currently and need to understand if it's neccessary to abandon the mac ecosystem
posted by lescour at 1:52 AM on June 11 [15 favorites]

The big 2 OS vendors pushing AI enshitification into their OS is the best advertisement for linux!
posted by neonamber at 1:57 AM on June 11 [17 favorites]

Want to give this thread a chance to breathe a little, but I missed an important link and it's relevant to what lescour just posted: Report: New “Apple Intelligence” AI features will be opt-in by default. So, it sounds like it can be disabled. Not clear whether that means literally off by default at launch or if it's one of those "hey, you haven't setup Siri yet!" perpetual Settings nags. Potentially it could be roped into a hard requirement for using CarPlay like current Siri.
posted by Ryvar at 2:01 AM on June 11 [8 favorites]

This spring, tech companies have made clear that AI will be a defining feature of online life, whether people want it to be or not. First, Meta surprised users with an AI chatbot that lives in the search bar on Instagram and Facebook. It has since informed European users that their data are being used to train its AI—presumably sent only to comply with the continent’s privacy laws. OpenAI released GPT-4o, billed as a new, more powerful and conversational version of its large language model. (Its announcement event featured an AI voice named Sky that Scarlett Johansson alleged was based on her own voice without her permission, an allegation OpenAI’s CEO Sam Altman has denied. You can listen for yourself here.) Around the same time, Google launched—and then somewhat scaled back—“AI Overviews” in its search engine. OpenAI also entered into new content partnerships with numerous media organizations (including The Atlantic) and platforms such as Reddit, which seem to be operating on the assumption that AI products will soon be a primary means for receiving information on the internet.

from This Is What It Looks Like When AI Eats the World [The Atlantic]
posted by chavenet at 2:19 AM on June 11 [10 favorites]

Hey, Siri, where was JustSayNoDawg on June Flernteenth, Twenty-Blix?

Siri proceeds to tell the lawyer, the police and even a passing priest everything because Siri's voice recognition just listens for "Hey, Siri", not specifically my "Hey, Siri" uttered in a commanding, sexy baritone, and also because Siri has no actual concept of Fourth or Fifth Amendments.

Thanks a bunch, Siri.

And for the record, I was nowhere near that house of raccoon-juicers.
posted by JustSayNoDawg at 3:35 AM on June 11 [8 favorites]

Watching the unveil was SO FRUSTRATING. Because the early bits and pieces they announced were either straightforwardly neat—natural-language processing paired with semantic understanding of in-app actions? sign me up!—or carefully constrained, like that dopey emoji-maker that exclusively does a tiny, stupid, silly thing. Then they unveiled the "analyze and rewrite your writings" and the "generate sketches based on context clues" and... okay, no way they're doing that completely ethically, but it's still only happening in a tightly-curtailed context, right? And then, in the last minute, they go "oh also use ChatGPT to generate bedtime stories for your daughter." Literal demon shit.

Honestly, some of the implementations they came up with are sincerely very neat, especially when they leaned on "analysis" rather than "wholesale creation." Reading an email to determine what components ought to be included in a reply (like "I'm bringing a plus-on" and "I'm driving and/or taking a rideshare") is legitimately cool and useful! Making Siri anything less than dogshit: very welcome! And it feels like they did a good job of compartmentalizing the different pieces, making it possible to use the less-grody features while avoiding the grotesque ones outright. (Though "less-grody" still means "using training data that was mostly obtained unethically," so, bleh.)

But... man, it makes me feel gross that all that other stuff will be baked into the OS. Apple had an opportunity to draw a clear ethical line here, and the fact that they drew some amount of line only makes it more frustrating that they didn't push it further. Especially since "we made Siri actually work" would have been enough to net them a standing ovation.
posted by Tom Hanks Cannot Be Trusted at 3:51 AM on June 11 [9 favorites]

(The real star of the show, imo, was the iPad's new calculator app, which was legitimately jaw-dropping. God, that feels like such a silly thing to say. BUT IT'S TRUE.)
posted by Tom Hanks Cannot Be Trusted at 3:51 AM on June 11 [15 favorites]

The only thing I am interested in here is if it can be disabled.

I believe it’s entirely opt-in by default.
posted by Thorzdad at 4:10 AM on June 11 [6 favorites]

My test of new Siri is going to be asking it to play the jazz standard Autumn Leaves. Up until now, if I say “hey Siri play autumn leaves”, I get a horrific version by Ed Sheeran. Will Siri realise that I mostly listen to jazz? Will Siri realise that I have clicked “suggest less like this” for every ed Sheeran album, and this song? I’m not holding my breath. It’s like Ryan Broderick said yesterday: do we need this? Mostly, the answer is turning out to be “no”
posted by The River Ivel at 4:12 AM on June 11 [16 favorites]

"Hey, Siri, how do I permanently disable all 'AI' features on this computer?" That's the only context in which I'll use it.

But seriously, it's a good time to go reread William Gibson's Pattern Recognition so I can remember how the Internet used to be weird and beautiful and wonderful before half a dozen shitty right-wing billionaires took control of it. Gibson's billionaire is way more interesting than the real ones, too.
posted by outgrown_hobnail at 4:14 AM on June 11 [21 favorites]

The big 2 OS vendors pushing AI enshitification into their OS is the best advertisement for linux!

True, but we're starting to really need the year of FOSS on mobile.
posted by trig at 4:17 AM on June 11 [7 favorites]

David Runciman reporting from 2018:

One day, perhaps, humans will build the intelligent robots capable of supplanting us. But for now, the biggest threat to our collective survival comes from the corporate machines we first started building hundreds of years ago to make our lives easier and have never really learned how to control. If we do end up manufacturing killer robots, it won’t be because individual humans made it happen. It will be because the corporations for which they worked didn’t know how to stop it.

I think of the ML that Big Tech is currently pitching as "AI" not so much as corporate intelligence as corporate digestive juices: it's like fly saliva, indiscriminately vomited onto any old shit and then sucked back up to get the dissolved nutrients extracted.

over the next month I will be transitioning my desktops from Windows to Linux (NixOS + a heavily locked down Windows 10 VM).

NixOS looks to me like tidy and well-conceived tech and at some point I fully intend to take it for a spin. That said, the reason Debian has been my daily driver for the past fourteen years has a lot less to do with its technical merit than its social design. Debian takes democracy seriously to a far greater extent than any other software curation organization I'm currently aware of, and in a world increasingly infested with corporate houseflies I really value that.
posted by flabdablet at 4:36 AM on June 11 [10 favorites]

I guess part of my problem with this world is that I don’t really want my phone or pad to be exciting; I wasn’t it to be familiar and predictable. My needs are simple enough that any kind of digital assistant is overkill. Half the time I look for something to listen to, I notice something else I’d rather experience, so anything that interferes with my process of serendipity is a bug, not a future. I don’t really want or need my phone to have functions that require the cloud.” I do need it to make phone calls occasionally. I guess having autocorrect mess up fewer comments would be nice, but most of this just doesn’t seem to be worth the effort.
posted by GenjiandProust at 4:40 AM on June 11 [27 favorites]

My Siri requests usually consist of turning on my alarm for a specific time or starting a timer so as not to burn that burnable thing. I am actually kind of curious as to what I can get Siri to do. But, on the whole, I don't think this is going to end well. Alexa has already been used in a lawsuit. How much worse will it be when there's some sort of "intelligence" behind it?

Hey, Siri! Call Elon and convince him that 58 billion is a bit much.
posted by JustSayNoDawg at 4:48 AM on June 11 [2 favorites]

For technological topics, I recommend reading the commentariat of Ars Technica.

IMHO, they are one of the smartest and most informed discussion groups on the internet (after MeFi of course) and they vote down nonsense mercilessly.

Here are the discussions for the three current relevant topics (sort by votes).
posted by fairmettle at 4:55 AM on June 11 [10 favorites]

Other than playing around with it, I have only been able to use Siri successfully in my regular life to: set timers for cooking, and (only several times): text my wife while I was driving.

Other than that, Siri has been a frustrating silly thing that I have ignored. It was a remarkable feature when it premiered, and I certainly played with it a bit then! But having it screw up a few times just makes the UX so annoying that I quickly abandoned it like I did the various Health Apps, etc. I don't even like the speech-control features on my TV remote.

Yeah I'm an old person at 53, but I grew up with computers, and I used to look forward to new features on phones, on laptops, tablets, etc. The days of new, interesting features are gone. Now it's a camera upgrade. I guess these manufacturers are hoping these "AI" toys will be the next round of "Ooohs!" and "Ahhhs!" at their reveal exhibitions, but I'm not feeling it.
posted by SoberHighland at 5:09 AM on June 11 [14 favorites]

59. Serious nerd (2 Newtons!) and I turned off Siri and all other malarkey. The tech has gone past useful in terms of iPhone features and imo Apple et al are coming up with problems to solve with existing AI-ish tools as opposed to creating new categories of products. The industry needs a breath of fresh air and LLMs are not it.
posted by whatevernot at 5:16 AM on June 11 [5 favorites]

Logitech is advertising a new mouse that has an "AI" button on it. Because "Add AI to it somehow!" is the mandate across every corporation.

I'd love to hear a single reason Apple's AI is any less pointless than that.
posted by mmoncur at 5:16 AM on June 11 [7 favorites]

Just specifically about the image generation stuff -- I feel like Apple has done right by artists. They have severely circumscribed what styles of art can be generated. They have likely seeded it with public domain and paid-for photos and artwork, and have a whole lot of documentation about exactly where their models got everything. Because the last thing Apple is going to open itself up ti is a class action lawsuit from a cadre of artists who feel they've been ripped off. So I'm not mad about the image generation.

The free-and-effortless gateway to ChatGPT which will happily rip anyone and everything off makes me mad, though. They should not be doing that. At least, however, they have made it fairly clear that if you want to violate copyright and fuck over artists and authors, you have to do it intentionally. I bet there's an ironclad clickwrap EULA on your first use of the ChatGPT integration that indemnifies Apple from anything and everything it might generate.
posted by seanmpuckett at 5:24 AM on June 11 [8 favorites]

I'm 57 and I use Siri mostly for timers, and asking it trivia, like "is a horse an ungulate," which it has always provided correct answers to (yes). I don't ask it to play music because Apple's music ecosystem is utter shit at surfacing/discovery, and I will never get the artist I want, much less the song. I sometimes use Siri to dictate texts on my watch, and it's usually okay for that.

I'm pretty excited about using text to interact with Siri on my desktop, though. I do not want to speak to my computer when I'm working, but I would be really happy to be able to just tell it to find things or look things up with my fingers.

My current phone is way too old for any of this. I am still using a 6S+ (on its fourth battery). I'm hoping the 2025 iPhone SE will have the cores to support the ML stuff. I really don't want a fancier phone than that; I certainly don't believe a phone needs three cameras and a LIDAR scanner. I guess that comes from being 57. Or maybe thinking a phone shouldn't cost $1500CAD.
posted by seanmpuckett at 5:32 AM on June 11 [4 favorites]

NYT says 'might actually be useful.'

I don't know that generative emoji will be all that useful, but it sounds like fun. And some of the other iOS 18 stuff seems neat. And, while Apple isn't perfect, they've been a lot more concerned with privacy than most of the big tech companies. I'm going with 'cautiously optimistic.'
posted by box at 5:45 AM on June 11 [5 favorites]

It’s so weird that when they unveiled the whole AI bit of the presentation, they started out with the first impression being Just Pour AI Sauce On It and installing Slop Generators (I guess arguably the text rewriter feature might be useful, especially for nonnative speakers), and then after horrifying the audience with that ghoulish ML-generated image, they’re just casually like “also we turned Siri into the Star Trek ship computer,” but not until after we go “oh no”

Why would you present in that order though
posted by DoctorFedora at 5:50 AM on June 11 [2 favorites]

Because you show the best thing last so people remember it fresh. They started the whole thing off with Craigy showing us how, in 2024, you can now slightly reskin your phone screen, maybe add a colour tint to your icons.
posted by seanmpuckett at 5:53 AM on June 11 [1 favorite]

Does any of this, or anything else announced, justify getting the new iPad Pro over the iPod Air?

I run multitasking frequently, and I want it to be smoother and faster. I can’t see how “AI” would help there at all.

I guess if I thought any of these AI tricks could take the set of diagrams I’ve been drawing and *correctly* turn them into a specific set of numbers and symbols, then maybe I’d use it. But checking that it’s been done correctly is almost as hard as just doing it in the first place, and I do really need it to be correct.
posted by nat at 5:54 AM on June 11 [1 favorite]

Random thought. If all Apple devices are going to have the same image generation model, then the specific images it generates could be precisely compressed into just the text prompt + the random number seed, because any Apple device with that model would be able to recreate the image from the prompt and seed. (Really that's all these models are, is enormously complicated, extremely lossy compression algorithms.)
posted by seanmpuckett at 6:05 AM on June 11 [4 favorites]

As always, Star Trek got there first
posted by rhymedirective at 6:17 AM on June 11 [1 favorite]

Don't the AI features require new hardware? I thought that some of them were only available on the 15 Pro and not the 15 (for phones), and M1 or later Macs.
posted by Spike Glee at 6:17 AM on June 11

I wish they had a little scene with a mid-aged college prof interacting with an iPad to do some LLM-moderated online research about deforestation in Brazil affecting weather in Africa or somesuch.

The $20/mo I punt on my ChatGPT license is the best money I spend; if this technology can scale in capability like GPUs did in the last 30 years, hold on to your butts.
posted by torokunai at 6:18 AM on June 11 [4 favorites]

I wonder if Apple going to skip the M5 release name because of Star Trek (TOS episode The Ultimate Computer [Dailymotion]). The M-5 multitronic unit, created by Richard Daystrom (portrayed by amazing Black actor William Marshall [who also played Blacula]). Because it didn't end well for the M5.
posted by seanmpuckett at 6:23 AM on June 11 [6 favorites]

Their technical blog post on how they secure it is pretty breathtaking. It’s flex after flex after flex compared with anything anyone else is doing. If you don’t want to use these tools, fine, but if you do it’s pretty clear Apple is the place to do it. For one thing, if I’m reading it right, it implies their use of ChatGPT is with them hosting the model on their own hardware, away from OpenAI. Which also makes it swappable, but either way your data is going nowhere. It seems incredibly thought through.
posted by DangerIsMyMiddleName at 6:27 AM on June 11 [13 favorites]

I'm "we'll see." But Siri fails 90% for me. It works for "Call Tricia." I however am pleased I can pause a podcast (or book or music) when I'm driving and using CarPlay using Siri. That alone makes it pretty swell for me. But AI? Hummmmmm.
posted by cccorlew at 7:01 AM on June 11

if I’m reading it right, it implies their use of ChatGPT is with them hosting the model on their own hardware

Due respect, but so what? What actual use does ChatGPT have that makes life better? It churns out obviously shitty half-baked text to a prompt. How does that improve anything? I don't want to read something written by it.
posted by outgrown_hobnail at 7:07 AM on June 11 [9 favorites]

And Screen Time remains completely unusable and woefully unfixed. Thanks, Apple.
posted by JoeZydeco at 7:08 AM on June 11 [1 favorite]

>It churns out obviously shitty half-baked text to a prompt.

It does?

https://chatgpt.com/share/5434079c-b486-44d5-99e6-2f6207a42bff

posted by torokunai at 7:11 AM on June 11 [1 favorite]

I wish they had a little scene with a mid-aged college prof interacting with an iPad to do some LLM-moderated online research about deforestation in Brazil affecting weather in Africa or somesuch.

The $20/mo I punt on my ChatGPT license is the best money I spend . . .

This is a joke, right? The joke is that the iPad would tell him his "LLM-moderated online research" is currently very likely the largest contributor to catastrophic climate change on the planet. That's why it's funny, right Anankin? Right?
posted by The Bellman at 7:22 AM on June 11 [10 favorites]

One thing that gets me about this kind of demo/launch is how few legitimate consumer use cases they have (even with the multimillion R&D money they're throwing at it). It's always, like, "make a picture of my mom as a superhero. "
I know there are serious, usefull applications for ML in general and LLMs in particular, but this is not one of them.
posted by signal at 7:41 AM on June 11 [7 favorites]

Due respect, but so what? What actual use does ChatGPT have that makes life better? It churns out obviously shitty half-baked text to a prompt. How does that improve anything? I don't want to read something written by it.

Improved natural language processing so Siri is actually useful for more than really basic stuff, presumably.

The real use case for AI is in giving verbal instructions to a computer and having it understand you, not in asking it to write a cover letter or w/e.
posted by rhymedirective at 7:42 AM on June 11 [8 favorites]

The real use case for AI is in giving verbal instructions to a computer and having it understand you

But like, why? For hands-free control of music and maps while you're driving, sure. But every other time I use a computer, I don't want to be talking to it. And honestly 70% of the time I decide I'd rather drive in silence than speak out loud to someone who isn't there and who doesn't understand me most of the time anyway.

I acknowledge that it could be helpful for folks who, for whatever reason, aren't able to use traditional means of interacting with their computer. But generally I wish tech companies would focus on developing things people actually want instead of making up bizarre tasks and pretending they represent useful applications because they're the only thing the tech is good for - and then foisting them on everyone unasked.

"Apple Intelligence" is the latest U2 album.
posted by nickmark at 7:51 AM on June 11 [10 favorites]

AI is projected to soon burn what bitcoin currently is (~100TWh). This is the power consumption of ~10 million US households, or $13m/day at 5c/kWh.

Beyond the cute generated cartoon pics, I 100% believe it is going to have much more of a positive impact on personal computing than the WWW revolution did in the 1990s.

it's doing things literally right out of science fiction now,
posted by torokunai at 7:58 AM on June 11 [1 favorite]

> foisting

I've been using Macs for 35 years solid now. The future is cursorless ("if you see a cursor they're doing it wrong").
posted by torokunai at 8:02 AM on June 11

I feel like Apple has done right by artists...They have likely seeded it with public domain and paid-for photos

Reminds me of what they did with Garageband for audio.
posted by doctornemo at 8:10 AM on June 11

The joke is that the iPad would tell him his "LLM-moderated online research" is currently very likely the largest contributor to catastrophic climate change on the planet.

LLMs are not that now.
posted by doctornemo at 8:10 AM on June 11 [3 favorites]

It’s too bad the folks come out with their pitchforks in all these threads. Like I wish there was space for them and space for folks who are curious and not interested in a rehash of the existing criticism of tech, capitalism, and AI biz. I agree with some of that criticism, it’s just every thread and tiresome.
posted by creiszhanson at 8:11 AM on June 11 [34 favorites]

Good thread here by Matthew Green, cryptography expert at Johns Hopkins, digging into some of the integrity and privacy implications of the secure private cloud:

https://ioc.exchange/@matthew_d_green/112597849837858606

FWIW I was part of a team that looked into providing this kind of infrastructure at $GIANTTECHCO a few years back. We concluded that the cost would be prohibitive, and that was before LLMs hit the scene. The sheer cost of what they’re talking about here is gobsmacking.
posted by graphweaver at 8:20 AM on June 11 [8 favorites]

Yeah, I'm like genuinely excited if Siri will be able to dig into my email and messages so I can just _ask it_ what the current state of my family's Father's day planning is, rather than trying to dig through the last 300 messages from last month and like three or four different chats of varying historical lengths.

I mean, will it actually work? God knows. It'll probably fall on its face once those decisions hit some arbitrary boundary or good ol' fashioned meatspace - phone calls and direct conversations won't index, but ... will Siri be able to dig into my facebook event calendar to see that I'm currently planning on going to a restaurant? Or is that going to be a line it can't cross for whatever reason?

But yeah - "hur hur siri can't even pause my music" isn't exactly a useful comment in a thread about an _all new Siri_, y'all, but metafilter is gonna be metafilter, I know.
posted by Kyol at 8:27 AM on June 11 [9 favorites]

"Opt-in by default" is needlessly confusing. Opt-in is opt-in. When you upgrade your phone OS or get your new device, you will have to choose to turn on the AI stuff, not turn it off.

Apple should get Kevin Costner back for a Knowledge Navigator reboot. 1987 Apple thought we'd all still be chained to our desks.
posted by emelenjr at 8:47 AM on June 11 [2 favorites]

But like, why? For hands-free control of music and maps while you're driving, sure. But every other time I use a computer, I don't want to be talking to it.

Computers encompass a range of devices now--it's not just a machine sitting on your desk. It's your phone, it's your TV, it's your car, it's your refrigerator. Many, many use cases for interacting verbally with a computing device.

And yes, it's good for accessibility, as well.

I remain relatively ambivalent about natural language processing, but I can definitely see that it has the potential to radically transform how we interact with computers.
posted by rhymedirective at 8:52 AM on June 11 [2 favorites]

I think the natural language stuff is tres cool, it’s the making up stuff I grapple with.
posted by whatevernot at 8:58 AM on June 11 [1 favorite]

But, like, my interactions with my fridge are already pretty optimal: I open it up, and eat the plums my housemate was probably saving for breakfast, just like the next person. I don't see how any mediation would really improve that.
posted by Hermione Dies at 9:05 AM on June 11 [35 favorites]

I agree that accessibility is a great use case for verbal interaction with Siri or other virtual assistant types of features. Accessibility should be discussed more, and the needs of people who benefit from it should be centered more in these conversations.

Other use cases for verbal interaction with my devices are largely null, however, because of the Ed Sheeran problem described above, and also because I have a moderate--not even that thick--regional accent that's apparently incomprehensible to Siri and all the rest. It's a maddening experience to shout at a disembodied voice that you just need to call the pizza place down the street from you, that you know exists and is open, and get addresses for a dental office and some place called "Thrash Dam" instead. I've tried this for every iteration of voice assistant and none can get it right.
posted by knotty knots at 9:06 AM on June 11 [5 favorites]

Folks--this natural language processing stuff is about as far removed from traditional voice recognition as a wagon is from an airplane
posted by rhymedirective at 9:09 AM on June 11 [3 favorites]

Hermione now fridge tell on you.
posted by whatevernot at 9:12 AM on June 11 [4 favorites]

Ned Ludd was right... Sorry to those who are excited about this technology, but it seems like incompetent garbage that steals from everyone just to make the aforementioned billionaires richer. Great that apple's version seems slightly less harmful, but the climate impact alone is awful, and all again to make some obscenely rich guys richer. Yay humanity. Lets try putting the massive effort into helping the greatest in need instead of this...
posted by WatTylerJr at 9:12 AM on June 11 [10 favorites]

Yeah, NLP as a search tool or text generation from small prompts is still pretty lousy for me - there's just enough poison in the sources to make me greatly distrust blindly accepting anything without significant editorial oversight. Using it to generate BLUFs or recharacterize existing bodies of text where it's easy to make sure the machine isn't changing your meaning? I'm a little bit enthusiastic about that.

The art, though, can all basically go straight to the bin. Either it has obvious 8 fingered tells, or it's so constrained that it looks like those goofy facebook avatars your non-technical distant relation drops in every thread. Yay? But people seem to like it, so it's probably more of a me thing than a society thing...

But every other time I use a computer, I don't want to be talking to it.

Good news! You can also type to nuSiri. I mean, I get it, Siri pissed me off when I changed my general playlist from a descriptive "100 gigs of music" title to "A1" to get it up towards the top of my playlists in the music app, and for the following week it was incapable of recognizing that I had a new playlist called A1 until I finagled it into the right place for the right shard of my personal Siri ecosystem to realize it was a new thing now. But it sounds like it's a whole new ballgame this fall, so maybe it'll work better? Or it'll blow up in more amusing ways.
posted by Kyol at 9:15 AM on June 11 [2 favorites]

I sometimes want to complain when new technology isn't perfect or doesn't exactly match my existing use case.

But I often can't tell--is this complaint going to be received like I'm a traveler from the future who's been sent back with a warning, or is it more of an oldmanyellsatcloud.jpg kind of situation?
posted by box at 9:20 AM on June 11 [6 favorites]

LLM-moderated online research" is currently very likely the largest contributor to catastrophic climate change on the planet

I thought it was going to be a joke about environmental impact, too, but there’s no way this assertion about impact is correct. Not yet, anyway.
posted by atoxyl at 9:20 AM on June 11 [4 favorites]

there’s no way this assertion about impact is correct

Yes, before anyone else comments, I apologize for the hyperbole. Please put replace "the largest contributor" with "a major contributor" and consider me adequately chastised.
posted by The Bellman at 9:29 AM on June 11 [3 favorites]

I found your original point well-taken, especially if all the power is just being used to generate fancy stickers for iPhone users . . . I'm not sure what the energy split is between one-time construction and processing power consumed by these GPT models but I think the way solar is going in the US west we've got a lot of power to spare for the former each day at least.
posted by torokunai at 9:50 AM on June 11 [1 favorite]

re: climate impact. There is very, very little in the way of hard numbers on the environmental impact from OpenAI or Google Deep Mind, beyond Sam Altman's "it costs us single-digit cents per query." Even the handful of recent papers on the subject are full of extreme speculation (I have been digging heavily on arxiv for anything solid). Best ballpark figures I've found are the leaks on GPT-4's training process and specifications, all of which fall neatly in the range of what one would expect given the output and are broadly assumed genuine by the ML community. If those are correct, the above estimate is off by about two orders of magnitude.

The Bellman wrote:
Please put replace "the largest contributor" with "a major contributor" and consider me adequately chastised.

No chastisement needed. The models planned for 2026~2028 may actually meet or exceed cryptocurrency consumption if Q* isn't amenable to some pretty serious efficiency gains beyond the initial proposal. I wrote an overview of that here. TL;DR: Microsoft isn't speccing out a $100 billion ML supercomputer for 2028 because they think it will reduce carbon emissions. There is every possibility you're correct, just a few years early.

flabdablet wrote:
That said, the reason Debian has been my daily driver for the past fourteen years has a lot less to do with its technical merit than its social design. Debian takes democracy seriously to a far greater extent than any other software curation organization I'm currently aware of

FWIW Debian is the only other distro I've seriously considered, and it's what'll fall back on if I find NixOS to be a PITA in actual practice. But Nix's whole one-user-config / one-hardware-config paradigm looks like a major, major QoL upgrade for me that provides the necessary motivation to actually switch, beyond the ethical and somewhat abstract privacy reasons.
posted by Ryvar at 9:54 AM on June 11 [3 favorites]

the weird/funny thing is I want an OS entirely hosted on ChatGPT man
posted by torokunai at 10:06 AM on June 11

What I find kind of funny about this is that the AI features sort of cover over the shortcomings of IOS. Like, I already use Siri as my primary interface for reminders, because it's much quicker than poking through the Reminders app. Now I'll be able to use it in place of the horribly organized Settings app as well! I wonder if it'll also make icon organization more reasonable, because as it stands now, even after 17 iterations, Apple just can't seem to figure out how to let you move an icon from one screen to the next without painfully dragging the wiggling, struggling icon around while every other icon dances and laughs at you. And it will explain features! Because I suppose we've finally reached the point where we do need a user manual for our phones, and the hopes that it would be completely intuitive didn't quite work out.

That sounds grouchy, but I'm actually optimistic about this. Now, I won't get to use any of it for ages, since all my devices are just a little too old, and all the new ones are way too expensive, but having a phone smart enough that I can actually use it, is a nice aspirational goal!
posted by mittens at 10:23 AM on June 11 [6 favorites]

I wonder if it'll also make icon organization more reasonable, because as it stands now, even after 17 iterations, Apple just can't seem to figure out how to let you move an icon from one screen to the next without painfully dragging the wiggling, struggling icon around while every other icon dances and laughs at you.

I'm half hoping that maybe the iPhone Screen Sharing to macOS helps with that, although I suspect it's going to involve too many touchpad standard gestures to really work. But yeah, I keep sort of figuring that everything past the third screen of apps just needs to get dumped into the All Apps page, never to be seen again until I use Siri / Spotlight to search for it.
posted by Kyol at 10:41 AM on June 11

Rhymedirective: Folks--this natural language processing stuff is about as far removed from traditional voice recognition as a wagon is from an airplane

Not sure what you mean by "traditional" but modern speech recognition uses essentially the same techniques as LLMs do. Whisper, from Meta, is a transformer model. So is ChatGPT. There are "older" (as in < 10 years ago) models that are still deep-learning/NN based. The big jump came from older modeling approaches such as HMMs. And there's a lot of benefit to be had from putting all of your language representation in one learning bucket.

We've gotten a lot of mileage out of transformers. But if history is any guide, in about 5-7 years something will come along that will change everything.
posted by scolbath at 10:43 AM on June 11 [1 favorite]

I though AI meant 'Asshole Investor' and then most of the headlines made sense tbh.
posted by mazola at 10:50 AM on June 11 [3 favorites]

The big question that they'll never directly answer is how/whether Apple Math does (or is tripped up by) locale-specific order of operations, or if order of operations will be tied to weird non-obvious things. Like my Australian Siri, in addition to telling me to turn into the car park instead of the parking lot, will occasionally show their hand by acknowledging that they created a calendar entry for me in dd/mm format instead of mm/dd. Every other locale-specific thing on my devices knows I'm in a mm/dd/yyyy locale, but because my Siri is an Aussie, they've gotta be dd/mm, I guess? (I mean, I assume y'all Australians follow that date order, but it's an assumption to be sure.)
posted by Kyol at 10:57 AM on June 11 [2 favorites]

re: climate impact. There is very, very little in the way of hard numbers on the environmental impact from OpenAI or Google Deep Mind, beyond Sam Altman's "it costs us single-digit cents per query."

Yes. The point I was trying to make (ham-handedly, for which, again, I'm sorry) is that we don't know how serious the problem is -- beyond that it's serious and getting more so -- and Apple won't tell us. Certainly Sam Altman won't! But I think it's particularly important for Apple to be transparent about this, given that it sells itself as a climate friendly, and that's part of the pitch to its customer base. As The Verge noted (just now!):

Apple didn’t immediately respond to questions from The Verge about exactly how much energy it expects these new features to use or how this all jives with its climate goals. But each of those moves comes with energy and climate costs.

Data centers used to train and run AI models have an even more voracious appetite for electricity than other data centers used to send and store the world’s digital messages and cat photos. Research from AI firm Hugging Face and Carnegie Mellon University found that large image generation models used far more energy and created more than greenhouse gas pollution than others used to sort information or generate text. (The study didn’t include generative AI video models.)

Apple’s new custom emoji come with climate costs. - The Verge (6/11/24)
posted by The Bellman at 11:15 AM on June 11 [4 favorites]

lol, in the "professional" auto-rewritten email at 1:24:40, the generated text insists on maintaining "naiming" conventions.
posted by lucidium at 11:51 AM on June 11 [2 favorites]

there's a _small_ part of me that wants to redo the Apple keynotes with more realistic photos instead of professional (if still shot on iPhone) pictures with lighting and makeup and whatnot
posted by Kyol at 12:02 PM on June 11

lucidium: lol, in the "professional" auto-rewritten email at 1:24:40, the generated text insists on maintaining "naiming" conventions.

I see "naming" in both the original and "professional" versions.
posted by RichardP at 12:04 PM on June 11

Kyol said, "so I can just _ask it_ what the current state of my family's Father's day planning is, rather than trying to dig through the last 300 messages from last month and like three or four different chats of varying historical lengths"

You can do this today. You will get an answer. Will it be true? Maybe.

This is what people mean when they say that AI is useless. It's an impressive technical demo, but you usually have to verify its work. At that point, why bother? The cost of verifying the work is usually the cost of doing it yourself in the first place.

That's before we get into the costs of training and running these models.
posted by i like crows very much at 12:18 PM on June 11 [10 favorites]

Maybe they've fixed it on the sly, I've had the window sitting open all day, but I got a screencap.
posted by lucidium at 12:24 PM on June 11 [7 favorites]

re: climate impact. There is very, very little in the way of hard numbers on the environmental impact from OpenAI or Google Deep Mind, beyond Sam Altman's "it costs us single-digit cents per query."

"Single-digit cents" is pretty scary. That would put in in the general ballpark of between 0.5 and 1 kilowatt-hours.

Like, every time you run a query, it's like leaving a hair dryer running for 20-40 minutes, or bringing 5-10 liters of water up to a boil.
posted by GCU Sweet and Full of Grace at 12:28 PM on June 11 [12 favorites]

lucidium, fascinating! It looks like the error you saw is still there. I watched closer and the "professional" text is correct during the demo of the feature in action, but is wrong a second later after the demo in the static "tilted" screenshot.
posted by RichardP at 12:37 PM on June 11 [1 favorite]

Metafilter: I agree with some of that criticism, it’s just every thread and tiresome
posted by dephlogisticated at 12:44 PM on June 11 [3 favorites]

https://www.caiso.com/todays-outlook/prices is saying the California grid will currently pay you 8c/kWh to take power off their hands in C California
posted by torokunai at 12:55 PM on June 11 [1 favorite]

If AI and crypto based their entire business models on only soaking up excess power and getting paid for that service, we would be in a very different world and I would welcome it. But they’re not. The fact that California has excess power today is apparently of little weight to these folks, based on the fact that most data centers are not getting built in California but rather in places where power prices are lower (which is most places).
posted by nickmark at 1:14 PM on June 11 [1 favorite]

I agree with some of that criticism, it’s just every thread and tiresome.

I think it’s important to try and stick with the known facts and reasonable, informed extrapolation. To give people time and space to feel their feelings until they start running wildly counterfactual, and to expect that particularly with an older, more creative Extremely Online crowd this topic is always going to be a tough sell.

My future planned posts are:
1) a deep dive into last month’s departures of the OpenAI safety team right as the first Q* model began its safety & compliance testing, plus the Dwarkesh Patel interviews with Ilya Sutskever (outgoing chief scientist) and John Schulman (incoming alignment team lead, longtime Altman ally). I’ve been trying to put something together for weeks but haven’t found the right hook, and TBH have difficulty finishing the Schulman interview because of how much I fucking hate OpenAI under Altman-Microsoft.
2) a comprehensive rundown on deploying open source models at home as a hobbyist, and even rolling your own. Andrej Karpathy just yesterday dropped a four hour “Let’s reproduce GPT2” step by step, in a destined-to-be-classic followup to his earlier “Let’s Build GPT from scratch” from last year. I just haven’t found the perfect “deploying Mistral 8x7b / Llama-3” centerpiece for this.

But FWIW that’s what’s in my mental queue. I am absolutely terrible at initiating new posts, however, vastly prefer chiming in on existing so if anyone wants to run with these please don’t stop on my account.
posted by Ryvar at 1:24 PM on June 11 [13 favorites]

One of my coworkers gave a presentation on AI today, since he's been playing around with it. He tried using ChatGPT to comment some javascript programs we have, where the existing documentation's lacking. It would do a really good job, until it eventually (inevitably) went off the rails. The prompt that he ended up using was dozens of lines long. That sounds like a lot, but when you have thousands of lines of code per program, with 4-5 programs, it's not that bad.
He later fed it one of the programs he was maintaining, and asked it which functions would need to be modified to fix an issue. It not only found the one he was expecting, but a couple of others that he didn't know about. So, that's a nice use case.
I've also seen people use GPT to figure out what parameters to pass to ffmpeg to do something, and then explain what each of the parameters does.
posted by Spike Glee at 1:26 PM on June 11 [4 favorites]

Ryvar your comments on how the folks here react is well said. I love when the conversations get good and interesting (for me) and so wanted to complain.

I have never initiated a post. I have done some tinkering with the open source tooling and models on my own. Maybe it's something for me to try and do this summer to actually contribute more.
posted by creiszhanson at 1:29 PM on June 11 [6 favorites]

I like the conclusion of The Atlantic article linked above by chavenet:

AI is eating the world is meant, by the technology’s champions, as a triumphant, exciting phrase. But that is not the only way to interpret it. One can read it menacingly, as a battle cry of rapid, forceful colonization. Lately, I’ve been hearing it with a tone of resignation, the kind that accompanies shrugged shoulders and forced hands. Left unsaid is what happens to the raw material—the food—after it’s consumed and digested, its nutrients extracted. We don’t say it aloud, but we know what it becomes.

Rapid, forceful colonization indeed.
posted by fimbulvetr at 1:36 PM on June 11 [7 favorites]

It looks like the error you saw is still there. I watched closer and the "professional" text is correct during the demo of the feature in action, but is wrong a second later after the demo in the static "tilted" screenshot.

Yeah, and there's another typo a couple lines below it ("inefficent"), also not incorrect in the prior shot.

As much as I'd like to blame the technology, this looks to me like errors probably introduced by a person retyping the original text by hand to create the static graphic (or, ironically enough, perhaps choosing the wrong, pre-proofreading file version thanks to precisely the sort of naming convention mess the fake email was pretend-warning against?)

But in any case, these sorts of missing-/extra-letter typos aren't the kind of mistakes LLMs tend to make, right?
posted by nobody at 2:14 PM on June 11 [1 favorite]

" with explicit user opt-in (on each individual query)"
🍎: do you want to use AI?
me: no
🍎: are you sure you don't want to use AI
me: yes
🍎: are you really sure?
me: yes.

🍎: GREAT! let me annoy the everliving hell out of you by asking again in 5 seconds!
me: *sighs*

Of course, there will be no opt out of that hell, because they're just making sure, right?
posted by gusset at 2:21 PM on June 11 [3 favorites]

I imagine it'll be similar to the beloved "ask app not to track" dialog.
posted by knotty knots at 2:30 PM on June 11 [1 favorite]

But like, why? For hands-free control of music and maps while you're driving, sure. But every other time I use a computer, I don't want to be talking to it. And honestly 70% of the time I decide I'd rather drive in silence than speak out loud to someone who isn't there and who doesn't understand me most of the time anyway.

Having vented my frustration about the parts of this I didn't like, let me circle back and gush about the parts I actually like!

The tl;dr is that computers are historically very, very bad with conceptual imprecision—which is a shame because actual human language is precise as hell. The reason Siri is so bad at doing basic things like "play X song by Y artist" half the time is that, in order for it to process that command correctly, it has to identify X as a song, Y as an artist, X as a predicate to Y, and also it has to parse both X and Y correctly. Back when it was introduced, Siri's handling of this was drastically better than any other mainstream implementation, but it wasn't great. And while voice assistants like Alexa are generally quite a bit better, that's partly because their applications are more constrained: Siri is a system-wide service, and as such has a lot of variables to consider. (I gather that Google's voice assistant is better in part because Google designed deep-implementation handlers for it to tie into individual Google software platforms, because Google's and Apple's approaches to vertical hardware-software integration are wildly different from each other's none of this matters shutting up sorry)

Where LLMs stand out is that they specialize in semantic complexity. They're built to understand linguistic relationships: how different terms relate to one another in a fluid range of contexts. The "intelligence" part of "artificial intelligence" refers to how LLMs actively make sense of what different things mean within an imprecise, ambiguous context: they work out what a given sentence is trying to say, not in the sense of "knowing what it means" but in the cruder sense of "understanding which abstractions it's articulating."

A really basic example of where this is useful is Apple's advertised demo of "say the wrong thing to it and then go, oops, wait, I meant something else." For a computer to parse an input like that, it has to be able to comprehend that (a) the speaker made a mistake, (b) the speaker wants to modify what she just said, and (c) the speaker wants to modify one specific part of what she said. The semantic flexibility of an LLM lets it parse a statement like this, work out what various things were asserted within it—"oops" being an assertion here in and of itself—and figure out what's actually being requested in the end.

This is how LLMs do their funnest party tricks. If I say "rewrite Green Eggs and Ham in the manner of David Mamet," ChatGPT doesn't just recognize the proper nouns in that statement, and it doesn't just see that "rewrite" refers to one thing and "in the manner of" refers to another. It understands the relationship between the phrases "rewrite" and "in the manner of," without having been told what those specific terms mean. (And when it does the rewriting, it does so by parsing the semantic content of the Seuss story and some broad abstract sense of "what David Mamet's style is", both of which it achieves using a similarly flexible kind of semantic parsing.)

Ethical and environmental concerns aside—and both of those are pretty fucking alarming—this advancement is really neat! And its most useful applications aren't the bullshit "fire all your employees" use cases that the snake-oil salesmen developing this technology like to promote. (None of which it does remotely well.) What it's best at is literally handling imprecision. The more precise you need it to be, the worse it'll get, which is why it's terrible at researching facts and writing legal briefs. But there's a grey area where the computing side of things is very precise but the user-facing side of things isn't, and that's where this technology can be extraordinarily useful.

Apple described their method of letting Siri operate within their apps as something like: every app has been given a "cloud" of associated semantic terms: its mail app has messages, senders, folders, and what-have-you, it has associated verbs (like "send" and "search"), and so on and so forth. When Siri is given user inputs, it maps those requests to the semantic vocabulary associated with given apps: what requests likely go with which programs, and what actions are being requested within those apps. What's neat about that is that it winnows down the possible set of actions being looked at to "only" a few hundred or thousand, as opposed to millions or billions. Provided it can parse the user statements decently—and LLMs mostly can—it's operating within a narrowly-defined set of boundaries on the computing end, which drastically improves the odds of it knowing how to correctly do what I want it to do.

Because most things we do on computers are, conceptually, insanely simple. The hard part is working out a process that lets us do the specific things we want to do, because the things we want are usually two or three levels of complexity trickier than you can develop a simple user interface for, and because the things we assume computers understand are typically two or three levels dumber than we wish computers were.

If I want to find a old photo that my mother sent me of us outdoors somewhere, my options right now are to either scroll through my messages or my photo library, to guess at what exact year that photo was taken, or to try and search for something like "blue sky" and see if that yields any results. (The latter search already relies on machine learning, incidentally! It's a simpler version of the thing LLMs do that got rebranded and snake-oiled.) With a system-integrated LLM, I don't have to adroitly manipulate whichever app I'm trying to use: I can say "find me photos that my mom sent me with blue skies in them where one of us is wearing an orange shirt," let the LLM match those clauses to various app functions, and have it do the trickery for me. Or I can say "send Joe the notes I wrote last week about our algebra homework," and trust it to know how to parse "notes" for "algebra," find "Joe," and "send" him the thing it found. And while LLMs' capacities are vastly overstated, they honestly are smart enough to do tasks like that reliably, in part because these tasks are pretty simple by computing standards—the hard part is semantic, not complexity of function.

This use of LLMs also cuts back on hallucinatory risk, because these computers aren't generating data—they're just handling it. Those infamous Google queries that return results saying 17 presidents were born in Spain (or whatever) happen because answering a question like "which presidents were born in Spain?" requires a lot of layers of semantic processing: to comprehend the nature of the question, and to know where to find that data, and to evaluate that data, an LLM probably processes dozens if not hundreds of rounds of "figure Thing out;" a minor flaw in any of those steps will produce escalatingly dumber results, until you have a completely nonsense answer on the other side. It's much easier to, say, search a known quantity of things called Notes for a single assessable term like Algebra, and to find a person from a finite list of names, and to execute a basic operation like Sending, because the layers of operation are significantly simpler by comparison.

Honestly, this is why it's so fucking frustrating that Google's just spitting AI answers everywhere, or that shithead content managers are using it to blog or whatever. This technology has incredibly useful applications. What's more, those applications are flexible: Apple debuted literally a dozen new major UX schemas into its system, all at once, because the flexibility of LLMs mean you can implement them in a lot of different ways and have them be genuinely useful in most of them. This is a breakthrough—it's just not the breakthrough it's being sold as, because the shithead companies who raised enough money to develop this were looking to launch a trillion-dollar company instead of "only" a multi-million-and/or-lower-billion-dollar one. It's super neat, it has a lot of wonderful potential, and it's also capable of being fun in really silly joyful doofy ways: not "make an entire movie for you," that's bullshit, but the kinds of creative shenanigans that people use Photoshop for, only broader and better. But it has strengths and weaknesses, and virtually nobody building for it is playing to its strengths or even acknowledging its weaknesses. They want the Singularity, and because of that, they're completely obscuring the nature of the actual neat thing that they literally actually made. Because they're shitheads.

Apple's implementations of this are hands-down the most considerate, thoughtful implementations that any Big Tech company has gone for. And that's not getting into their privacy work, which predictably blows everyone else out of the water. Cull their implementation of this to the most thoughtful 20%, and you still have an extraordinary new feature; cull it to the most thoughtful 80%, and you still have a suite of new tools that are mostly useful and non-shitty—minus, again, the ethical and environmental concerns, which are still seriously concerning. I just hate that they also included an Everything Button, because ChatGPT's Everything Button sucks, in large part because OpenAI wildly misrepresents what it actually is and does. Apple seemingly did their best to handle this technology in a careful manner, which I appreciate. But where they whiffed, they whiffed, and it drives me up the fucking wall.
posted by Tom Hanks Cannot Be Trusted at 4:00 PM on June 11 [24 favorites]

Also, Ryvar, seconding everyone else who's telling you that your comments in all these AI threads are fantastic. Your informed commentaries and critiques are so goddamn appreciated.
posted by Tom Hanks Cannot Be Trusted at 4:01 PM on June 11 [8 favorites]

Apple Inc.
207.15 USD +14.03 (7.26%)today
Closed: Jun 11, 7:28 PM EDT • Disclaimer
After hours 207.19 +0.040 (0.019%)
posted by clavdivs at 4:29 PM on June 11 [1 favorite]

What’s funny is, Apple typically announces amazing new products or sales figures, and their stock goes down because PEOPLE ARE STUPID. Here their whole keynote relies on a shady buzzword (We’re adding AI!) and their stock goes up $10 per share because PEOPLE ARE STUPID.
posted by jabah at 5:23 PM on June 11 [4 favorites]

Like this makes me realize how much I would have LOVED to punch my 21 year old neophile self in the face.
(while still regretting that self didn't buy AAPL when it was like 15 cents a share so I can quit this job and parasitically fund a "Freedom Club")
posted by symbioid at 7:28 PM on June 11 [2 favorites]

So having this shit shoved in our face every day by every tech giant hopping the hype train isn't tiresome, but us calling it out is? Cool Cool.
posted by symbioid at 7:31 PM on June 11 [7 favorites]

it's worth noting that a major part of the benefit of running things locally (as the bulk of it is, according to the announcements) is much lower power consumption overall, especially given that Apple's been putting that ML-centric "neural engine" in their chips for years now

there are enough valid criticisms to be made that, I think, conflating unrelated things is ultimately counterproductive
posted by DoctorFedora at 7:40 PM on June 11 [4 favorites]

All my instances with AI have been asking Copilot how to disable it. Asking Facebook's (I'm not calling it meta cause fuck you Zuckerberg) bullshit how to disable it. Asking every time I've been introduced to it: HOW DO I TURN YOU OFF? And the answer is "You can't! But we understand it's upsetting. But you can't!". Fuck all technology. I want to burn every data server to the ground.
posted by downtohisturtles at 7:48 PM on June 11 [3 favorites]

Mod note: Couple of comments removed. Let's avoid rapid fire argumentative and doomer type comments. Expressing disagreement is totally fine, but there's no need to repeat that disagreement in such a quantity in a short amount of time.
posted by Brandon Blatcher (staff) at 7:54 PM on June 11 [1 favorite]

>So having this shit shoved in our face every day by every tech giant hopping the hype train

from what I can tell, we're still in the TRS-80 days of what the tech is going to grow into

I've said the following more than once here I guess, but having been raised not on TV but the 1965 World Book Encyclopedias and about 50 years of National Geographics, along with the Wallace/Wallechinsky Book of Lists and Peoples Almanacs, I hold great expectations for the further success in 'training' LLMs and the GPT approach of knowledge engineering.

The machines are learning.
posted by torokunai at 8:34 PM on June 11

Ryvar thank you for your comments and planned posts as well.

There are many things I can now do because of LLMs that I wouldn't otherwise. There are good opportunities for them to be used. There are many use cases people hype strongly that I don't get at all. There are some good or promising products. There are many bad and unpromising products. LLM-based products are especially capable of initially impressing and subsequently disappointing.

I tend to avoid AI posts, because many comments seem to assume there's no room for a reasonable person to have an opinion other than that LLMs are very useless and very harmful, from which follows (unstated and perhaps unintended) that someone who still finds them useful and uses them is therefore unreasonable and/or engaging in morally blameworthy behaviour or any number of other adjectives. I hope people will extend some good faith to those here who are interested in these features or find LLMs helpful. I have no interest in forcing anyone else to use them, nor evangelizing, but I appreciate space existing to share LLM-related news and projects such as is (sometimes) created by these posts.

To the point, the whole "what's on my calendar", "what's in my email" thing is hard to pull off well even with SOTA models, but I've heard they're fine-tuning their tiny models to be fit for each task which is interesting The features don't appeal to me that much, but if this kind of thing works it could make it easier to navigate computers. But we should probably design computers to be easier and that would help everyone.

Taking in uncontrolled input (emails, screen content) and permitting the model to take actions is a big security risk without a known 100% solution, so it will interesting to find out how that's being handled.
posted by lookoutbelow at 10:53 PM on June 11 [5 favorites]

The above discussion about the effectiveness, reliability and usefulness of this technology has been fascinating, but I'm also kind of curious about Apple's emphasis - in the WWDC presentation - on data privacy, such as their the notion of "on-device processing" versus cloud computing, and their animated data-with-a-lock logo.

As in Tom Hanks Cannot Be Trusted's example above, a In order to help someone find their family photo with blue skies and them wearing orange shirts, wouldn't the software have to index all their visual content? And if so would it not get an unhealthy access to other lifestyle and affiliation data in a way that could breach privacy?

A paranoid person would wonder if the software is also parsing variables like who else is in the photo, what race/religion/demographic markers do they have, what does it say on their orange shirt, what books and newspapers are lying around in the shot, etc? What stops a totalitarian country from forcing a customization where the Intelligence flags whoever they have as enemies of the state? Where you needed a VPN to access non-regime-approved websites, are we going to have to develop a "data-VPN" to safely parse one's family photos, let alone emails and chats?

Next thing you know, would the AI start placing ads for minivans because it saw a large family gathering in the photo?

And does Apple's approach mean that Siri would respect my privacy more than - say - Cortana or whatever Facebook's agent would?

Finally, at some point the convenience of the software starts being overshadowed by the risks associated with the misuse of the data. Are we reaching that point now? I'd be OK with having to continue laboriously scrolling through my own data to find what I need, if it meant not being constantly solicited with intrusive ads, let alone put on a list of enemies of the state who wear a colander as ritual headgear or [insert appropriate signifier here].

Can our more technically literate mefites chime in on the privacy aspects?
posted by Bigbootay. Tay! Tay! Blam! Aargh... at 10:58 PM on June 11 [3 favorites]

> solar... power to spare

Does absorbing solar energy and using it to run heaters (computers) and heat pumps (computers' coolers) have the net effect of increasing albedo of the area of Earth covered by panels?

I suppose our bigger concern (probably by several orders of magnitude) is increasing greenhouse gasses in the atmosphere, but if we can get that under control, isn't there some limit to how much solar and wind we can use without causing a more minor form of climate change? Won't we want to continue to put pressure on energy conservation (use the electricity for trains not AI and electric cars).
posted by ASCII Costanza head at 11:15 PM on June 11

torokunai: The future is cursorless ("if you see a cursor they're doing it wrong").

I want some UX indicators to help me spot where the tools and I are out of sync. Some hint where the information is going to, some way to change the intended context, some feedback from the ML tools that there's confusion about what my words might mean. (As a "proud parent cursor of child cursors" --an Oracle DB joke -- I hope cursors persist.)

About TFA: Apple aren't launching a General-Purpose Intelligence today, they must have had a product meeting where they had to accept that too many people would expect superhuman intelligence from the Apple implementation. It's not AI, it's some useful tools.
posted by k3ninho at 11:58 PM on June 11

Thanks for all the kind words and Tom Hanks Can Not Be Trusted that was an amazing comment. By itself worth overcoming the social anxiety that made this my first post to the Blue in (checks) ...18 years, 20 days. Christ.

lookoutbelow: for what it's worth my own bias is that I find neural networks absolutely fascinating as a technology, originally more on the RL / continuous training side, but now that deep learning/LLMs have scaled to where they're not just correctly encoding English grammar but (as Tom Hanks expressed so wonderfully) derivatives like conceptual relationships I think they've turned the corner into something genuinely interesting. That said: I do not trust technology corporations or late-stage shareholder capitalism to use this technology for purposes other than exploitation, harvesting wealth for billionaires and general abuse of the working class.

I take advocacy of open source / open weights models and legislation mandating open training set data seriously for that reason. This split stance seems to pass with most Mefites, even the most adamantly opposed to ML (many of them for with good reasons) rarely come after me specifically. And while there are occasionally slightly rough threads it's nothing a long-reformed old-school Internet troll is going to even blink at.

Sort of cross-fading into a reply to Bigbootay, now:
I particularly do not trust OpenAI to do anything but race ahead as fast as possible and use money / compute scale and a lack of ethics to continue outpacing everyone else. They've become the very thing they were founded to prevent, the true believers in their original mission have largely quit over the past year; senior safety/alignment team members in particular. Despite the "Open" in their name they're the only major LLM player with no open weight models published, let alone open training data. John Schulman comes off in the Dwarkesh Patel interview as the perfect hybrid of boy-genius programmer and Mr Rogers, but the truth is he's enabling Altman to achieve something hideous - in a capitalist class warfare way if not an existential threat way - and the workers rights / class consciousness proto-Marxist in the back of my skull won't stop shrieking. He says all the right things in the right way but at some feral level I am not willing to put down my rock and listen.

Because of their track record on rigorous best-practices implementation of things like the secure hardware enclave in every iPhone, because the FBI publicly complains about them and occasionally floats trial balloons about legislation to outright force them to create backdoors (without any publicly known success), like a lot of technically-inclined people it's fair to say that I actively distrust Apple far, far less than any of their competitors. It's not even close. This is not the same as trusting them, at all. Or maybe a better way to say it is: the amount of smoke-filled backroom arm twisting I believe it would take to get Apple to deliberately compromise my personal security is sufficiently high that I am comfortable predicting it will never be directed my way. I doubt anybody reading this is best pals with whoever is running ISIS these days, so that's probably true for all of us.

Finally, more specifically directed at Bigbootay's questions: the reason for Apple's emphasis of on-device processing whenever reasonable is that they want to keep all of the personally identifying data locked behind the secure hardware enclave. When small snippets of it are exported to the private cloud servers to solve a more difficult problem, they are not stored or paired with other identifiable information, and the hardware / software stack behind this has a verifiable cryptographic signature. More precisely: Apple wants to keep the cryptographic keys for accessing personal data within the enclave. Which is not perfect because no Turing architecture ever can be, but is close enough to prompt loud protests from the largest and best-funded intelligence agencies (which is, admittedly, an amazing marketing vehicle for Apple).

While Apple has cooperated with the CCP to greatly restrict App Store offerings in mainland China, to the best of anybody's knowledge the secure hardware enclave remains consistent across all mobile products for the past several years (decade?). It's not perfect: there are several well-sourced stories of intelligence agencies paying major security research or even criminal blackhat groups to crack iPhones of interest in the rare cases when new security holes are identified, but on the whole Apple is vastly better than any other implementations us normal consumers can access. To the question of: are we past the security/convenience inflection point - my first editorial comment kicking off this thread mentioned that after 33 years I am leaving the Microsoft desktop for Linux (I'd recommend a Mac for everyone who doesn't want to deal with Linux), while sticking with Apple for mobile after this announcement. Pretty much says it all.
posted by Ryvar at 12:58 AM on June 12 [5 favorites]

> So having this shit shoved in our face every day by every tech giant hopping the hype train isn't tiresome, but us calling it out is?

When we're all shuffling around in the same shoe-high shit soup, there's not many places you can conversationally go with "hey, this sucks right?"
posted by lucidium at 1:12 AM on June 12 [3 favorites]

I think that's correct that the big guys are steaming ahead whole hog without any concern for practicality or reasonable goals. My assumption is that they're all speeding toward the military, and all of this pop robot shit is manufacturing consent for when the military does start using it to kill people. "Well but the drone operators use game pads just like on Xbox!" "Hmm, I'd like to sign up for your newsletter."
posted by rhizome at 2:53 AM on June 12 [2 favorites]

As in Tom Hanks Cannot Be Trusted's example above, a In order to help someone find their family photo with blue skies and them wearing orange shirts, wouldn't the software have to index all their visual content? And if so would it not get an unhealthy access to other lifestyle and affiliation data in a way that could breach privacy?

The way this works with Apple OSes specifically is, your device already indexes this data (and you can search in the Photos app for things like "orange shirt" already). To some extent, Apple Intelligence's LLM is stored on-device, so it takes your input, searches its existing indexes, and returns the result (or does whatever it's been asked to do).

At some point, prompts switch over to being sent to the cloud. It seems like there's an icon that appears to indicate some kind of cloud operation. At that point, a query does get sent off-device, but that appears to be when generative requests are made. So the online server gets the prompt itself, but all the "data" being transferred goes from the server to the device.

What's more, the keynote demonstrated that, in instances where cloud queries are asked to run on personal data (such as photos), users are asked specifically if they want that specific data shared online. So while they've taken further steps to guarantee privacy on the server end—including a lot of new transparency disclosures from any third party whose servers they might use—they're also limiting how much data those cloud servers see to begin with.

If it's possible to run a query that invokes both local and cloud LLMs, it looks like all the personal-data-indexing stuff still gets run exclusively on-device. Anything that the cloud needs to see for a query is run explicitly by the user first.
posted by Tom Hanks Cannot Be Trusted at 3:23 AM on June 12 [1 favorite]

Rhizome: I’ve been cold-called by recruiters for a defense contractor seeking help with building sim environments for virtual training of autonomous military drone control systems in Unreal as a precursor to real world testing. Salary range double a very senior game developer’s, just in the initial scouting call (I very rudely declined). This was about two years ago, so that is definitely a thing that is happening - somebody took that salary. Like most robotics it’s a thing you would use reinforcement-based learning for, the only deep learning component would probably be a very fast, lightweight image classifier to identify hazards/threats/targets. No LLMs or generative anything.

Tom Hanks: I took away something slightly different from the demo and technical brief linked upthread. On-device / full personal context / “semantic index” for lightweight local inference, heavier inference gets seamlessly kicked over to Private Cloud with packaged personal data restricted to whatever’s strictly query-relevant and flushed afterwards as part of the stateless / no-data-retention policy. Didn’t see any indication that transition is brought to the users attention (I’m curious as to whether you can opt out of Private Cloud specifically).

On further review: this take matches Matthew Green’s review of the technical brief in the Mastodon thread graphweaver linked above, specifically part 17.

In the demo video it was only transmissions to third-party endpoints like ChatGPT prompting user permission to send out an identity-masked query. I’m assuming this is specifically for large generative or mixed text+image workloads where a multi-modal, plural-hundreds-of-billions parameter model like GPT-4o would offer significant additional value beyond Apple’s self-hosted PCC models. No reason to bother involving a third party, otherwise.

I could be wrong or missing something. Basic UX = I’d be surprised if there wasn’t a “No and Never Ask Again” categorical opt-out on third party endpoint queries at least. That may be wishful thinking, but I really hope not.
posted by Ryvar at 4:18 AM on June 12 [5 favorites]

“what all of us have to do is to make sure that we are using AI in a way that is for the benefit of humanity, not to the detriment of humanity.”

This is one of the most focus-grouped, nonsensical, meaningless statements I have heard in years. This statement would have the same weight as if he had used the word "phaser" instead of AI. It's a marketing spun version of "Don't be evil", yet somehow even more shallow even though it is longer.

I'll say it slow for you Tim Apple: What you have currently in the marketplace is not AI.

I can see the appeal from a marketing perspective, here you have something to advertise that already has "backstory" and people have some pretty strong feelings about it even though their only exposure has been fictional. It's the Crystal Pepsi/fat-free/low-carb/IF/hot trend for the 2020's. Calling what you are selling "AI" is shady misrepresentation at best and dangerously malicious fabrications at worst.
posted by Sphinx at 10:43 AM on June 12

Ryvar and Tom Hanks Cannot Be Trusted: Thank you so much for making these complicated issues so much clearer, and for the perspective on Apple vs OpenAI et al. !

To slightly paraphrase Ryvar,

Metafilter: By itself worth overcoming social anxiety to make a post!
posted by Bigbootay. Tay! Tay! Blam! Aargh... at 11:31 AM on June 12 [1 favorite]

I could be wrong or missing something. Basic UX = I’d be surprised if there wasn’t a “No and Never Ask Again” categorical opt-out on third party endpoint queries at least. That may be wishful thinking, but I really hope not.

And this is where I gotta give some of the detractors some credit - Apple hasn't exactly covered themselves in glory regarding stuff that you can opt out of or fail to opt into. Although I can't remember specifics at the moment, I think they're a little aggressive about getting people to set up Apple IDs, I think?
posted by Kyol at 11:57 AM on June 12 [1 favorite]

Today's Garbage Day has a take on this.

…In the 2010s, [Apple] shifted comfortably from a visionary tastemaker, at least aesthetically, into something closer to an airport lounge or a country club for consumer technology. They’ll eventually have a version of the new thing you’ve heard about, once they can rebrand it as something uniquely theirs. It’s not VR, it’s “spatial computing,” it’s not AI, it’s “Apple Intelligence”. But they’re not going to shake the boat. They make efficiently-bundled software that’s easy to use (excluding iPadOS) and works well across their nice-looking and easy-to-use devices (excluding the iPad). Which is why Apple Intelligence is not going to be the revolution the AI industry has been hoping for. The same way the Vision Pro wasn’t. The iPhone effect, if it was ever real in the first place, is certainly not real now.
posted by signal at 2:16 PM on June 12 [1 favorite]

Although I can't remember specifics at the moment, I think they're a little aggressive about getting people to set up Apple IDs, I think?

You might be thinking of the kerfuffle over Microsoft making it effectively impossible to set up a computer without logging into a Microsoft account? Apple certainly encourages people to have Apple ID accounts, but it's also easy enough to, say, not have one tied to every user account on every Macintosh — you just kind of click "not now" when it asks about that bit during setup. It'll continue to encourage having an Apple ID in the settings, but I think it stops a little short of "aggressive," if arguably not by much.
posted by DoctorFedora at 5:23 PM on June 12 [1 favorite]

Last time I looked, it was impossible to activate an iOS device without an Apple ID. Has that changed?
posted by flabdablet at 10:40 PM on June 12

Yeah, in fact I don't think that has ever actually been the case? You do need a valid SIM or eSIM at least, but I've done iOS device setup without an Apple ID countless times over the years. Maybe you're thinking of needing a valid SIM card to activate an iPhone?

It does at least have a setup screen that asks you to sign in with one by default, but there's always been a "skip for now" option on there too.
posted by DoctorFedora at 10:50 PM on June 12

Maybe you're thinking of needing a valid SIM card to activate an iPhone?

Quite possibly. It's been some years since my last bout with iOS setup.

I've done iOS device setup without an Apple ID countless times over the years.

How do app installations work in that case? I didn't know it was possible to do those without using an Apple ID to log into the app store first.

it's also easy enough to, say, not have one tied to every user account on every Macintosh — you just kind of click "not now" when it asks about that bit during setup.

Just for completeness: last time I set up a Windows 11 box, I was able to give it an initial local Windows account and no MS account by refusing to allow a network connection during setup. Dark pattern UX is a pain in the arse.
posted by flabdablet at 11:08 PM on June 12

yeah, TECHNICALLY you can set up a Windows machine without a Microsoft account, but you have to know the arcane cheat codes and manually prevent it from connecting to the internet, rather than just clicking the "not now" button

and, well, yeah, you can't download non-preinstalled apps without an Apple ID, because as you point out, those are how app store authentication works
posted by DoctorFedora at 11:15 PM on June 12

How do software updates work on iOS with no Apple ID? Are there different rules for the OS itself vs preinstalled apps?
posted by flabdablet at 11:40 PM on June 12

Preinstalled apps are actually part of the OS, like they're part of the cryptographically sealed boot disk image, which is why they only get updated with system updates (and when I say "preinstalled" I mean in the sense of, like, Safari or Notes, since carriers can't preload iPhones with third-party stuff).

Non-preinstalled stuff, even if it's from Apple like Pages or Keynote, gets installed through the app store and thus isn't part of the OS at an underlying level. This all also means that "uninstalling" system apps is just an illusion — it merely removes the icon and makes you go to the app store to "reinstall," say, Mail.
posted by DoctorFedora at 11:44 PM on June 12 [1 favorite]

Currently OOBE/BYPASSNRO still allows Windows 11 local account setup, but both the no network access and fake email address workarounds have been recently closed. This is all hearsay because I’ve never installed Windows 11 and I’m hoping I’ll never have to. The 4TB Samsung 990 Pro for my personal machine’s new primary drive arrived today - got a lot of backups and offsite archiving ahead of me before I install Nix, but it’s nice having a giant brand new SSD to look forward to.

Back to Apple UX snafus: not setting up Siri after a major version upgrade / new hardware leaves a little nag notification persistently active in Settings, and every 2 or 3 weeks (it’s pretty gracious honestly) you’ll get a brief pop-up reminder nag. Separately from that: CarPlay has a hard Siri-enabled requirement with no known workarounds. And it’s like: I just want the fucking rendered output from Google Maps on my center console display - not even interactive! - and also to pass the “Next track” button on my steering wheel back to Pandora. Literally nothing else. There is no fucking way Siri is essential for this (you wouldn’t necessarily think it from my comment history but I hate talking to computers).

Ultimately: Apple hasn’t entirely evaded the data harvest / metrics-driven-everything mindset of its competitors, but it is far, far less infected with engagement optimization behavior. They’re more about selling luxury hardware and lifestyle-brand image, the peace-of-mind aspects of which result in them sometimes building absolutely best-in-class security or privacy systems. It looks like ML on Apple hardware is in a good place to potentially become one of those. Fingers crossed.
posted by Ryvar at 12:08 AM on June 13

Curious: according to Bloomberg (original article link), OpenAI isn't getting paid for this partnership; they're doing it for the exposure. (I assumed Apple would be paying for this -- especially since OpenAI apparently won't get any user data out of it -- with a plan to uncouple when their own internally developed version eventually becomes robust enough.)
posted by nobody at 6:12 AM on June 13

I was a Mac user for a good decade or so but went back to Windows because the company just wasn't good anymore. I had a few iPhones but switched to Android because again the company wasn't letting me use my phone. If you use Apple products you use them according to and only according to the ways Apple wants you to use them. But Apple's way of doing things has just become soooooo bad. I don't miss Steve Jobs cause Steve Jobs was a piece of shit but he did run the company much better than Tim Cook. I have no idea what the fuck Tim Cook is doing.
posted by downtohisturtles at 6:54 PM on June 13