Show HN: Spellwise – GPT-Powered iOS Keyboard

[teamsayana]

Hey everyone,

I'd like to invite you to test the beta of our text-editing iOS keyboard built on top of LLMs.

It can help you with grammar, tone, summarize or elaborate on your thoughts, and even add a sassy joke if you're in the mood.

As we're in the early stages, we'd greatly appreciate any feedback you can provide. Thanks in advance!

[aaronbrager]

In realms of app and UI, A tapestry of code so fine, There lies a typo small, unseen Within the onboarding scene.

Yet let us not forget, my friend, The privacy we must defend, To clearly state, with policy, The text discarded, logs set free.

And so, amidst these digital features, I find delight in poetry's creatures, Conversion, art, in cyberspace, A wondrous blend of code and grace.

[teamsayana]

Beautiful.

[smoldesu]

Does the model run locally?

[rvz]

I can guarrantee you that this sends everything you type to Spellwise for them to use OpenAI APIs.

There is no iOS/Android app that has a local LLM installed and the app size is less than 100MB. Even if it exists, it cannot be done commercially.

> When you use Spellwise, your crash logs, usage information and feedback will be sent to Apple and Spellwise, Inc.

'Usage Information' certainly means the text you input from the keyboard and submit to Spellwise.

I can only predict that LLM-based software keyboards will only take off if there is a LLM or equivalent small enough to be locally installed and Apple is waiting patiently for that. Google will follow afterwards.

[teamsayana]

That's correct, we process the input and send it through an API. However, as an editing keyboard, we don't track what you type. Moreover, in the UI, we've ensured that we only have access to the text you explicitly select. I understand that privacy is a major concern, so we're currently exploring ways to build a classifier that can locally remove/replace or truncate PII from any text you process, so that it doesn't reach our backend and the API provider.

[teamsayana]

Not yet, it's API-based. We're developing a smaller model for local grammar, rewriting, and summarization tasks. However, the smaller model needs to be tiny due to OS memory limitations on third-party keyboards. For tasks like co-writing, we'd still rely on the large models, therefore it'll be partially API-based.