that's the short version. a few years ago, running a language model on a phone was a parlor trick — a tiny model that could draft a barely-coherent reply if you were patient. today, apple ships a 3-billion-parameter foundation model on every iphone with apple intelligence. it's small by lab standards. it's enough to write a text message that sounds like you.

ignore uses that model. that's the entire architecture. there is no server. there is no api call. when you tap a suggested reply, the prompt and the response never leave the device.

what "on-device" actually means

most apps that say "private ai" still ship your message to a server. they encrypt it on the way there, they encrypt it on the way back, some even attest that the server is using the model they claim. but the data leaves your phone.

on-device means the model is sitting in your phone's neural engine — a chip that's been on every iphone since the a11 bionic, now powerful enough to do real work. when ignore reads your incoming text, it goes:

phone → phone

that's it. there is no second arrow.

why this is non-trivial

people sometimes ask why a small startup can pull this off when the big ai companies seem to need data centers. the answer is that we're not building the model. apple is. we're calling it.

apple intelligence exposes its language model to apps with the right entitlement. you ask the model to write a reply, it does. it's not as smart as gpt-5 or claude opus, and it doesn't need to be — it just needs to know enough about your tone to draft a text that sounds like you. that's a different problem, and a much easier one.

the hard parts, for us, are:

• learning what "your tone" means, on-device, without sending your chats anywhere
• deciding which messages to surface and which to let sit
• shipping all of that as a polished app you'd actually use

what we don't do

we don't have an account system. we don't have a cloud sync. we don't have a "premium tier" that processes things faster on our servers, because we don't have servers. if you delete the app, the model of your tone is deleted. there is no backup.

this is annoying in a few ways. if you get a new phone, you start over (we're working on icloud private sync to fix this, but it'll still be end-to-end encrypted with no server visibility). if a feature requires the model to be smarter than your phone allows, we can't ship it.

the trade-off is worth it. your texts to your mom are not training data for somebody's next-quarter model. they're not in a vector database. they're not in a logfile. they exist on your phone, get read by the model on your phone, and produce a reply on your phone.

who this is good for

if you're the kind of person who has 47 unread texts right now: this is for you. ignore reads what people sent, drafts a reply in your voice, shows it to you. you tap to send, or you edit it, or you ignore it (the whole brand, really).

if you're the kind of person who would die before letting a language model see your group chats: this is also for you. by design, no language model sees them — except the one running on your phone, which is the same one already auto-correcting your typing.

the worst part wasn't the number. the worst part was that they weren't from spam or people i didn't care about. they were from my mom asking how my week was. from a friend who had just gotten married. from my brother, twice, with no context, just "hey."

i wanted to reply. i would think about replying. i'd open the thread, type half a sentence, get pulled away, come back two days later and feel worse. by then it had been so long that "hey, sorry, i meant to write back earlier" felt insufficient, so i'd stall again. the longer i waited, the harder it got. you know how this goes.

so i built ignore.

the wrong solution

the obvious solution is "just reply faster." this is also the solution every productivity book has ever recommended. it does not work. or rather: it works for people who are good at replying, and those people don't have 92 unread texts.

the second obvious solution is "use ai to auto-reply." this is what i kept seeing apps do, and it always felt wrong. the replies sounded like a customer service bot wrote them. my mom would notice. she would ask if i was okay. she would be right to ask.

the actual problem is: i can write the right reply. i know what to say to my mom. the cost isn't the cost of composing — it's the cost of switching context, picking up the thread, finding the words, being on for that conversation. by the time i'm done with all that, i could've written the reply twice. but i didn't, because i never started.

the trick

what if a model read the incoming text, looked at how i normally write to that person, and drafted a reply that sounded exactly like me? then all i'd have to do is glance at it, tap, and move on. the cognitive cost drops to zero.

the model has to learn my tone per contact — i don't text my mom the way i text my boss. it has to pick up on my phrasing, my emoji habits, my contractions. and it has to live on the phone, so the chats never leave.

every part of that turned out to be possible. apple intelligence handles the language modeling. the per-contact tone learning is a fine-tune we do on-device. the priority detection — knowing which threads matter — is patterns we learn from how i actually behave with messages.

what surprised me

people don't want to reply more. they want to reply at all.

once you give someone a reasonable draft in their voice, they reply. not because they suddenly enjoy texting — they don't — but because the friction drops to one tap. the people in the beta started clearing weeks of backlog the first day. one user texted me afterward: "i talked to my brother for the first time in three months."

that one stuck with me.

what this is

ignore is the thing i wanted to exist. it's not trying to make you a better texter. it's trying to remove the part where you mean to reply and then don't.

i'm building it as a small subscription with no ads, no data collection, and no servers. if it works, more people get to stay in touch with the people they care about. if it doesn't, i'll have at least replied to my mom on time.

— dovas

every language has its own rhythm of texting. spoken language has rules. text doesn't. text has habits. these are the things ignore picks up on, per-language, per-person.

english

contractions everywhere. "yeah" instead of "yes." periods at the end of a text mean you're upset. ellipses mean you're not done thinking. "k" is a war crime. "lol" doesn't actually mean laughing — it's a softener. all of this varies wildly between contacts, and that's exactly what the on-device fine-tune is learning.

danish

low-key, very informal. "ja, helt sikkert" reads as warm; the same content in english would feel stiff. punctuation runs sparse — no apologies for it.

dutch

text-message dutch loves abbreviations: "ff" for "even" (a sec), "idd" for "indeed." word order can break for a short reply.

french

context-dependent formality. "tu" vs "vous" is non-trivial in text — ignore notices which one you use with which contact, and stays on the right side.

german

"hi" is universal, "moin" is regional, "hallo" is safe. compound words get written long even in shorts. surprisingly little abbreviation overall — germans like the full word.

italian

emoji-heavy but not random. there's a vocabulary of which emojis attach to which feelings. "ti amo" carries weight; "ti voglio bene" is everyday.

norwegian

shares patterns with danish but slightly more punctuation. "takk" appears more than english "thanks" — it's also used as a sign-off, not just a response.

portuguese

distinction between brazilian and european matters. brazilian text reads warmer, with more emojis and softeners; european portuguese is clipped and more formal.

spanish

regional difference is significant. mexican spanish in text is playful with "x" for "por" and "q" for "que." spain leans toward written-fully, with fewer abbreviations.

swedish

minimal capitalization. text starts in lowercase and stays that way (the swedish brand voice, and ignore's, by accident, for entirely different reasons).

turkish

richer suffixes per word, so short messages pack more meaning. the model handles the grammar fine; what it has to learn is whether you use the polite suffix forms or not.

chinese (simplified)

text often skips subjects when they're inferable. punctuation is lighter — you don't need a period to end a thought, a new line will do.

chinese (traditional)

similar to simplified at the casual register; the divergence is in formal writing, which doesn't show up in texts often anyway.

japanese

depending on contact, you'll swing between casual ですます and plain-form 食べる-style. ignore notices which contacts get which.

korean

honorifics are not optional. the difference between "고마워" and "감사합니다" is the difference between texting a friend and texting your boss's parents. the model handles it; your habit decides it.

vietnamese

tone marks matter for meaning. modern texting often drops them when context is clear, but not always — depends on who you're writing.

the punchline

the model is multilingual. your texting style isn't. you have habits in each language, and habits with each person, and those compound. ignore's job is to learn the second-order pattern — the rhythm of you, in that language, with that person.

a multilingual app shouldn't feel translated. it should feel like the version of you that speaks that language. that's the goal.