it’s a GeNeRaTiVe ArTiFiCiAl InTeLlIgEnCe feature
I need no further reasoning
A feature you don’t like might be useful for others. We do not want to maintain a set of apps with different subsets of features, so new features are implemented into the Sleep app, and it is up to the users, if they wish to use them or not.
If you won’t use this feature, nothing has changed for you.
3 Likes
Hello @evennight many thanks for your feedback. There are two reasons for this limit. I fact we have tough hard whether we shall allow any follow up questions at all when designing this but then decided to allow one question, to overcome this limitation you may again ask another predefined question which will give you another follow up… BTW what follow up questions would you typically ask. This is very interesting for us to know to make this feature better
-
We need to really limit the AI in answering ideally by predefined quations so mitigate the risk of any errors the AI tells you … e.g. the notorious “glue your pizza together example” with one follow up question which is define as only related to your sleep we hope to minimize such risks
-
Of course this also a question of the quota, we cannot give unrestricted access to a pro LLM model as every request means cost for us… so the idea is in the first few weeks we will see how is this feature used how well we are getting into the quota limits and based on that either extend or restrict the follow up quetions
Did that answer your question?
2 Likes
@b_j sorry for the issues. Could you please use menu > Support > Report a bug and I will investigate the problem…
Thanks for the answer. All I wanted to ask the AI were questions about definitions, like “What does ‘awake at night’ mean?” and further explanations. I think it wouldn’t hurt if you limited consequential questions to two or even three, if it’s a pro-exclusive feature. I like this feature, by the way. Great job:)
1 Like
From these comments, you can see how much people depend on your app.
I’ve chatted with you, and your staffers, about bugs and feature requests, for years now. In addition, I’ve purchased every sleep gadget that has been offered.
No other Sleep app holds a candle to Sleep as Android, and I have some more feedback. I’ll leave it here, rather than in-app submission, because it could warrant some discussion:
The Assistant was stymied, when I prompted about comparing the App’s sleep reporting to Wearables’ direct reporting.
For example, my Garmin Epix 2 Pro is linked and communicates with SaA (without needing the “alternative” addon). However, the Garmin’s sleep graphs report far different sleep cycles. My Deep Sleep, according to the Garmin’s reporting, only occurs within the first half of my sleep sessions. According to the device, the rest of my night is spent in REM, Light Sleep, or brief spikes of Awake. Last night’s Sleep Score is a 93, on the App, while the Garmin reports a 77.
Perhaps through distinguishing N1-REM instances, Sleep as Android reports regular cycles of Deep Sleep throughout the whole night. This highlights the differences in algorithms that determine the cycles as both the device and App are working through the same base data HR, HRV, Pulse ox, etc.
It’s difficult to discern whether users’ Wearables, which they depend on for all-day data (or just all-night) should be trusted, when comparing the App’s -vs- Wearables’ sleep reporting/graphs.
My feedback is to train the generative model further on the nuances of how the same input data can produce vastly different sleep graphs. Naturally, you’d need to balance this with protecting proprietary algorithms.
Like most users who pair Wearables with the App, the first thing I see in the morning is my sleep reporting on my phone, and then my Garmin watch.
Explaining the nuances of wildly different sleep reporting is an excellent opportunity for the new Assistant to shine. Disseminating these complicated calculations, especially in a concise (think ELI5) concept, showcases the potent lifestyle value of Sleep as Android and the power of its Assistant feature.
Thanks, Petr 
1 Like
@lenka-urbandroid My sleep duration goal in the settings is at 8 h 30 m.
@petr-urbandroid I can report it but I don’t really think this is a bug. This is just generative AI being generative AI. It has no place giving people health advice.
1 Like
@b_j I think this isn’t about generative AI, we use generative AI in a different way, first we feed the sleep score data, but we also explain what are expected ranges, what may be too low etc… and we basically use generative AI only to generate a human understandable summary of all the data we have provided. This is a very constraned approach and we hope to minimize any error caused by the generative AI this way… so what I wanted to check from your report is what prompt are we generating to see if we can improve on our end…
2 Likes
@evennight many thanks, I think we do not give it enough context to correctly answer the awake question … we will elaborate on this…
Hello @REM … many thanks for your suggestion. At the moment we do not use our own model, but we utilize Google’s Gemini LLM model.
But I think the differences you see in the data may step from the fact that there isn’t any standard way how differentiate between light sleep / deep sleep - where different vendors place the division line.
This is summarized in detail on this article on our website:
Not sure if and how an AI model could solve this.
Thanks, Petr.
I use their Gemini v1.5 Pro 0827 model for work, although I’m not sure which Gemini model Sleep as Android implemented.
The trainability of this model is strong. If not already in-progress, training the LLM on the wealth of knowledge from your blog articles empowers the Assistant to articulate in different ways. Users can then prompt the Assistant to explain in a way they know how to best understand and retain.
@petr-urbandroid that’s unfortunately not how generative AI works. you can’t constrain it in the ways you are seeking to.
Can this data be sent over MQTT or some other way I’d like daily briefing sent to Home Assistant could be awesome to TTS this in the morning
There seems to be an issue with the AI’s interpretation of the sleep disturbance value (according to the attached screenshot, this SHOULD be called AHI Apnoea Hypopnea Index… Is the AI giving wrong info here?).
The AI is telling me that as I have a Sleep Disturbance Value of 1, which is lower than the national average in my country, then it indicates possible issues.
I think that the logic here is reversed. The lower the Sleep Disturbance Value, the better. I use a CPAP machine, so I do expect the AHI to be lower than 5 (my target is lower than 2).
I mentioned my circadian rhythm disorder to the assistant and it gave incorrect treatment information.
Is there any way to limit its ability to provide poor medical advice?
Sure, it says to speak to your doctor, but that doesn’t justify poor information/advice. I was not given the option to correct it/inform it like you can with Chat GPT.
This feature won’t be useful for me, but I hope it helps others… But I mostly hope it doesn’t mislead anyone.
1 Like
I tried this and used the “How did I sleep last night” question. That made me want to k iw what it would say about Sunday’s sleep. There is not an option to check any past nights sleep. Would also be nice to maybe see a comparison of 2 different sleep cycles.
Also, a tad off topic, but it did mention my snoring numbers, how does the app seperate snoring from continuous allergic rhinitis that can sometimes sound like snoring? And yes, that is a thing.
Hi @ShoalBear,
EDIT: sorry, I was mistaken and deceived by the Assistant’s answer…
Regarding allergic rhinitis - this sound will probably confuse the algorithms into thinking it might be snoring. To design a new sound class for rhinitis, we would need hundreds of thousands of samples of confirmed rhinitis sounds - so the algorithms can learn to recognize them.
2 Likes
@b_j in fact you can, we use a grounding approach, we first restrict the answer to only relate to the data we have provided, then we explain what are correct ranges and what is good / bad or neutral so we specify very carefully the whole playground and we only ask the AI to summarize it in human language and translate it to user’s langauge…
This approach greatly reduces any risk of hallucination of the AI, but of course theme may be issues which can be addressed by further restricting the scope. We can work iteratively here and fine-tune, but for that I would need to see what promt did we send to the model so that I can first reproduce it and then modify the promt so that the problem is fixed…
1 Like
Thanks for the reply @lenka-urbandroid I appreciate it.
I am glad to hear that the app doesn’t differentiate snoring from rhinitis. Now I can relax a bit about that issue.
Maybe there could be an addition of that info on your page about snoring stats, etc?
When AI compares my sleep with others, it mentons stats about my snoring, heart rate variability and breathing - 3 things i am not tracking at all?
(I might have tried those options a few times in the past, but that was at least a year ago)
Also, if you ask to compare your sleep when others, which period does it look at? Last night, last week, last month?
