OpenAI’s ChatGPT for Mac is now available to all users
- Thread starter JournalBot
- Start date
Well, this is probably the place to mention https://manticore.nz/legacyai -- the unofficial ChatGPT client has been available for Macintosh System 7.0 through OS X 10.11 for quite a while now 
Last edited:
Upvote
22
(27
/
-5)
I think it’s weird that they rely on screenshotting the users screen to be able to give contextual help. If the file you are working in isn’t all visible you need to scroll to “show” the OpenAI app the contents.
There have been several third party apps already (like meetblue.ai, beam etc) that use the accessibility apis instead. That seems like a much better solution imho.
There have been several third party apps already (like meetblue.ai, beam etc) that use the accessibility apis instead. That seems like a much better solution imho.
Upvote
13
(14
/
-1)
PatientZero
Ars Centurion
I don't think that's weird at all -- it's clearly easier to just show an LLM something like a screenshot and have it "understand" what's there, than it is to try to tie into APIs for things like filesystems, etc.
Anyone's who's ever worked even once with an API would probably see where this is easier. (Not "cleaner," not "better" -- just easier.)
Upvote
4
(15
/
-11)
Giving an app far-reaching and hard to control access to whatever also doesn't have to be the better solution. Giving it just what you can see right now is easy to control in a quite intuitive way.
Upvote
30
(30
/
0)
Tell me a recipe like my grandmother used to about making API calls from the native MacOS client, reminding me how many steps there are and clear directions for each step of the recipe like which arguments to add on each step, so I can make API calls just like my grandmother taught me.
Upvote
-3
(14
/
-17)
Sonoma on Apple Silicon, Intel is not supported even if you're running the latest
I'm curious if there's a hard reason why since it seems much like the mobile and web apps, it's possible it's using the neural engine for some preprocessing to save what it has to send back and forth, but it's probably as likely they just didn't decide to support the waning platform.
Upvote
3
(7
/
-4)
Ha -- this is incredible!Well, this is probably the place to mention https://manticore.nz/legacyai -- the unofficial ChatGPT client has been available for Macintosh System 7.0 through OS X 10.11 for quite a while now
Upvote
6
(7
/
-1)
I'm very sorry for your loss. I can see that the API call recipe is very precious to you.
Upvote
21
(21
/
0)
Well, this is probably the place to mention https://manticore.nz/legacyai -- the unofficial ChatGPT client has been available for Macintosh System 7.0 through OS X 10.11 for quite a while now
So my Intel Mac on Sonoma is too old for the official client and too new for the unofficial one
Upvote
13
(14
/
-1)
Has Apple announced that everything you do will be OCR’d and saved in a non-encrypted SQLite database that can be trivially lifted from a compromised or accessible Mac? I think there is a reason for why Microsoft got roasted.
Upvote
79
(81
/
-2)
This isn't from Apple, and Apple have said when they integrate Siri with ChatGPT the user will need to given permission to use it on each occasion.
Upvote
45
(45
/
0)
Apparently you haven't used a decent LLM to explain why your code isn't working. It's reasonable to be skeptical about LLMs, but to completely dismiss them is just as silly as thinking they're perfect.
[EDIT: typo]
Upvote
28
(30
/
-2)
Or even to just reformat an email to make it sound more professional.
Just because it can’t do everything doesn’t mean it can do nothing. The black or white thinking on display is…. Odd.
Upvote
23
(24
/
-1)
A lie is an intentional falsehood. As there is no intention, there can be no lie.
Upvote
8
(11
/
-3)
Apple is also the ones who published a whitepaper at the same time detailing their seemingly practical and feasible approach to integrating AI in a privacy preserving manner, either running things locally or, in the event something does go remote, creating a highly secure container (from security of the hardware up) and running things transiently rather than retaining state.
I'm not an AI fan at all, but it's easy to see the differences in approach.
Upvote
31
(31
/
0)
This is why Apple is convinced that external models are a dead-end. This isn't just a hard problem to securely solve, it's an impossible problem.
Upvote
0
(0
/
0)
Microsoft got teased because they announced integrating it and then didn't show it doing anything useful.
We also tease Kelloggs when they announce their corn flakes are now fortified with blockchain.
I'm sure Microsoft will do something useful there, given their Copilot PC specs, but maybe they could have figured out what before rolling it out. Apple was also benefitted by not relying on OpenAI for their core technology and centering the whole announcement on user data relevancy and security.
Upvote
18
(18
/
0)
This is exactly the kind of "hallucination" you'd get from an LLM when you train it on sensationalist news headlines only so that it confidently draws a picture by connecting a few dots in a halfway plausible sounding, but actually utterly wrong way.
Upvote
19
(21
/
-2)
Upvote
-2
(7
/
-9)
Upvote
-7
(5
/
-12)
Depends on what you mean by securely solve. You would have to trust your compute provider, of course. That doesn’t mean external (off device) models are inherently insecure. They are as insecure as let’s say your gmail, and plenty of people store personal things there. Also, apple themselves are using external models right now because local compute is too weak.
You’re looking at several orders of magnitude improvement in efficiency before something like a GPT4 level model can be run locally. That’s not going to happen overnight. So within the foreseeable future, there still will be use cases where only external models work.
Upvote
0
(0
/
0)
Calling people who prefer accuracy "drones" is as stupid and lazy as calling anyone who praises something a company does "shills." People who come to Ars are not interested in mindless twitter-style noise. You're bringing absolutely nothing to the conversation.
Upvote
25
(25
/
0)
Sigh. At least you didn't use a "You are the weakest link. Goodbye!" mention.
Upvote
-1
(1
/
-2)
Ha, thank you, that's exactly what I was going for but my original joke doesn't seem to be landing very well. Maybe it was a little too nonsensical, or worse, people thought I was serious
Upvote
2
(2
/
0)
It's because AI is being sold as technomagic much in the same way that crypto is which causes a large amount of people to rightly become incredibly skeptical because of the obvious grift marketing towards know nothings who are in control of budgets. The well was poisoned from the beginning by OpenAI.
It's the same reason why I do not trust a single thing a religious leader says and go out of my way to verify anything they say.
So far, the only direct experience in my day to day life of these AI models is in monitoring software that now cuts additional tickets based on trends but gives details so vague that we cannot do anything with the tickets but close them out or spend ten hours figuring out the ticket was referencing a heart beat signal that had a single packet drop in a thirty second timespan so it looked like there was a flow with 100% loss to the model so everyone fucking panic. Who gives a shit about audio and video generation? We already are buried in a tremendous amount of an over production of legitimate and illegitimate media.
In my daily, lived experience, my only meaningful interaction with these models is that they're generating a lot of extra noise in a world that is drowning in noise, but now I have to deal with my boomer in laws thinking that everything is fake now and everyone is out to get them even more than they were before.
Upvote
3
(5
/
-2)
Mac Sonoma and M-series CPUs... there is a fine print saying so on their website! No Intel CPUs! I tried.
Upvote
1
(1
/
0)
I didn't expect that both Alfred and ChatGPT would respond to the option+space shortcut, just figured it would be one or the other. Gonna have to remap that before I end up asking ChatGPT to run Photoshop. 
…But I guess that's coming, eventually.
…But I guess that's coming, eventually.
Upvote
0
(0
/
0)
There does not, like the version of ChatGPT on iOS, appear to be any way to change the Model on the free version of ChatGPT for the Mac, in spite of what OpenAI said (or this article for that matter). In other words, no GPT4o.
Upvote
0
(0
/
0)
ChatGPT needs to be running for the Shortcut to workI didn't expect that both Alfred and ChatGPT would respond to the option+space shortcut, just figured it would be one or the other. Gonna have to remap that before I end up asking ChatGPT to run Photoshop.
…But I guess that's coming, eventually.
Upvote
0
(0
/
0)
Abhi Beckert
Ars Tribunus Angusticlavius
Well, this is probably the place to mention https://manticore.nz/legacyai -- the unofficial ChatGPT client has been available for Macintosh System 7.0 through OS X 10.11 for quite a while now
Note all your queries are sent unencrypted to a proxy server, before being encrypted and sent to OpenAI (because it won't accept an unencrypted API call). These days that likely means all of your conversations are being recorded and permanently stored, potentially by someone who might decide to use it against you one day.
Worse — fixing that is flagged as low priority in the issue tracker...
Upvote
-1
(1
/
-2)