apple intelligence: my first impressions after using it for a week

ai · 2024-06-08 · 6 min

i used apple intelligence for a week on a real device, doing real tasks. not a staged demo, not a fifteen-minute hands-on at a launch event. a week of email, messages, notes, photos, and the small annoyances that make up a phone.

the short version: the good parts are quieter than the keynote, the annoying parts are real and predictable, and the honest verdict is more boring than either side of the argument.

what i actually used

notification summaries. a stack of notifications collapsed into one sentence. on a busy day it is genuinely useful. on a quiet day it summarizes a single message and adds nothing.

writing tools. rewrite, shorten, change the tone, proofread. this is the feature i used most, because it works inside the apps i already use and it does not require a new workflow.

photos cleanup. remove a distraction from a photo. it is fast, it is good enough for a casual shot, and it fails at the edges: hair, glass, a shadow that should stay.

genmoji. a custom emoji from a description. fun for ten minutes, useful for about two, and not something i would miss.

siri with chatgpt. for the questions siri cannot answer, it offers to hand them to chatgpt. it works. it also adds a confirmation step that makes a quick question feel like a form.

priority notifications. the phone decides which notifications matter. it is right often enough to be useful and wrong often enough to be annoying, which is the same trade-off every filter makes.

the good parts

the writing tools are the killer feature, and nobody talks about them. they are not flashy. they sit in the share sheet, they work in messages and notes and mail, and they do the boring part of editing: shorten this, make it clearer, fix the grammar. i used them every day, which is more than i can say for any other feature in the set.

notification summaries reduce the cost of a glance. instead of reading six notifications, i read one sentence and decide whether to open the app. that is a real improvement in attention, and it is the kind of improvement that does not demo well.

the integration is the point. none of these features require a new app, a new account, or a new habit. they appear where the work already is. that is apple's advantage, and it shows.

the annoying parts

the summaries are sometimes wrong in a way that reads as confident. a message that says "i can't make it" can be summarized as "i can make it" if the model loses the negation. the summary is shorter, the meaning is flipped, and the notification is the only thing you see. that is a bad failure mode for a feature whose job is to save attention.

the rollout is uneven. features arrive by language, by region, by device, and by account. a feature that exists in a keynote may not exist on your phone, and the settings do not always explain why. if you support users, this is a support burden.

the confirmation steps add friction. handing a question to chatgpt requires a prompt, a confirmation, and sometimes a wait. for a quick fact, it is slower than opening a browser. for a complex question, it is fine. the feature is not sure which one you are doing.

the privacy prompts are necessary and tiring. every new capability asks for a permission, and the permission is often "send this to a server". the prompts are the right design. they are also a reminder that the feature is not local, no matter how the marketing sounds.

the verdict

apple intelligence is useful glue. it does not change what the phone is, and it is not a reason to upgrade. if you already have a supported device, turn on the writing tools and the notification summaries and ignore the rest for a few months. if you do not, wait. the features will be better and more complete by the time you are due.

the bigger question is not whether the features are good. it is whether the assistant can act. summarizing a notification is a parlor trick compared with booking the thing the notification is about. apple has been careful about that step, and careful is the right word for it. the useful version of this feature is still ahead.

for the browser version of the same shift, see the year of the ai browser. for the model side, see chatgpt after one year.