ajuhasz

Some of the more magical moments we’ve had with Juno is automatic shopping list creation saying “oh no we are out of milk and eggs” out loud without having to remember to tell Siri becomes a shopping list and event tracking around kids “Don’t forget next Thursday is early pickup”. A nice freebie is moving the wake word to the end. “What’s weather Juno today?” becomes much more natural than a prefixed wake word.

Yes! We see a lot of the same things that really should have been solved by the first wave of assistants. Your _Around The House_ reads similar to a lot of our goals though we would love the system to be much more pro-active than current assistants.

Feel free to reach out. Would love to swap notes and send you a prototype.

> I hope the memory crisis isn't hurting you too badly.

Oh man, we've had to really track our bill of materials (BOM) and average selling price (ASP) estimates to make sure everything stays feasible. Thankfully these models quantize well and the size-to-intelligence frontier is moving out all the time.

Agreed, while we've tried to think through this and build in protections we can't pretend that there is a magical perfect solution. We do have strong conviction that doing this inside the walls of your home is much safer than doing it within any companies datacenter (I accept that some just don't want this to exist period and we won't be able to appease them).

Some of our decisions in this direction:

  - Minimize how long we have "raw data" in memory
  - Tune the memory extraction to be very discriminating and err on the side of forgetting (https://juno-labs.com/blogs/building-memory-for-an-always-on-ai-that-listens-to-your-kitchen)
  - Encrypt storage with hardware protected keys (we're building on top of the Nvidia Jetson SOM)

We're always open to criticism on how to improve our implementation around this.

> Are you okay with private and intimate conversations and moments (including of underage family members) being saved for replaying later?

One of our core architecture decisions was to use a streaming speech-to-text model. At any given time about 80ms of actual audio is in memory and about 5 minutes of transcribed audio (text) is in memory (this is help the STT model know the context of the audio for higher transcription accuracy).

Of these 5 minute transcripts, those that don't become memories are forgotten. So only selected extracted memories are durably stored. Currently we store the transcript with the memory (this was a request from our prototype users to help them build confidence in the transcription accuracy) but we'll continue to iterate based on feedback if this is the correct decision.

With cloud based inference we agree, this being just one more benefit of doing everything with "edge" inference (on device inside the home) as we do with Juno.

True, we focused on hardware embodied AI assistants (smart speakers, smart glasses, etc) as those are the ones we believe will soon start leaving wake words behind and moving towards an always-on interaction design. The privacy implications of an always-listening smart speaker are magnitudes higher than OpenClaw that you intentionally interact with.