Show HN: OpenAI/reflect – Physical AI Assistant that illuminates your life
github.comI have been working on making WebRTC + Embedded Devices easier for a few years. This is a hackathon project that pulled some of that together. I hope others build on it/it inspires them to play with hardware. I worked on it with two other people and I had a lot of fun with some of the ideas that came out of it.
* Extendable/hackable - I tried to keep the code as simple as possible so others can fork/modify easily.
* Communicate with light. With function calling it changes the light bulb, so it can match your mood or feelings.
* Populate info from clients you control. I wanted to experiment with having it guide you through yesterday/today.
* Phone as control. Setting up new devices can be frustrating. I liked that this didn't require any WiFi setup, it just routed everything through your phone. Also cool then that they device doesn't actually have any sensitive data on it.
I also have been working with Daily on https://github.com/pipecat-ai/pipecat-esp32
I see so much potential if I can make hardware hacking + WebRTC easy. Not just for AI assistants but security cameras + robotics. If anyone has questions/ideas/feedback here to help :)
what is Daily?
https://www.daily.co/
You can use it to build lots of different real-time communication projects. Conferencing, Send your audio/video to GPU servers for AI, broadcasting and lots more.
It’s a super fun space to be in
this is very relevant to my interests. currently building a robot teleoperation project. i will have to investigate further.
Are there any cool demos that use daily I can explore?
https://www.linkedin.com/posts/thorwebdev_esp32-webrtc-activ...
https://m.youtube.com/watch?v=HbO18Elw9WY
Are two that I know of. Try it out, if you hit any roadblocks @ me on pipecat discord and would love to help
Why does this need hardware, other than the phone? Could just be an app on the phone couldn't it?
I was interested in the ‘hands-free’ idea.
If I put these devices through out my house it would allow me to switch AI personalities by proximity.
You can also use the device without your phone. These devices are also very cheap. I think you could do audio only for around ~5$
If you want to know what this is about, here’s the video they provided:
https://www.youtube.com/watch?v=G5OUnpPAyCg
It annoys me a lot that the current devices for controlling smart homes, such as Amazon Alexa or Google Home, lack the ability for lovely conversations the way OpenAI has.
I honestly can't tell if this comment is joking, serious, or AI lol
LLMs have a lot of advantages over humans for making conversation.
Even forgetting the main advantages (24x7 availability, and ability to talk about basically any topic for as much or little time as you want), they also get basically every obscure reference/analogy/metaphor and how it ties in to the topic at hand.
Usually when you're talking to another person, the intersection of obscure references you can confidently make (with the assumption your partner will understand them) is much more limited. I enjoy making those random connections so it's a real luxury to have a conversation partner that gets them all.
Another one is simply the time and attention they can bring to things a normal person would never have the time for. I'd not want to talk someone's ear off, unless I was paying them and even then, I don't want to subject someone to topics of only interest to myself.
(Edit: I suppose it's the final apotheosis of the atomised individual leaving all traces of community behind)
> final apotheosis of the atomised individual leaving all traces of community behind
It's not. In 10 years this is going to look as dumb as the biohacker wetware bros surgically embedding RFID chips in their hands. There's much more to communication (and life) than receiving pantomimed validation for your obscure references. You could be throwing away opportunities to connect with another person who would genuinely share your interests and help you grow as a person. Having a useless magnet in your fingertip is going to seem brilliant compared to ending up socially withdrawn and mentally unwell because you traded human companionship for a chat bot.
I think it's a much bigger social phenomenon already. Social talk will become even more a matter of performance, positioning and signalling, rather than something pursued for enjoyment of the thing itself.
Maybe I'm just weird but LLM conversations seem generally more interesting and enlightening, even in these early iterations of the technology.
> LLMs have a lot of advantages over humans for making conversation.
A lot of those advantages seem to be what enables an LLM to keep pushing people into delusion or triggering latent mental issues : https://www.psychologytoday.com/ie/blog/urban-survival/20250...
Of course, every technology comes with risks.
The way the Gemini Google Assistant rollout has been SO SLOW is utterly baffling to me.
I love seeing that hackathons are encouraged inside OpenAI and most importantly, that their outcome is also shared :)
It would be perfect if it could intelligently linkany device under authorization
Can you describe more? I would love to build/try it!
Somewhere in here there's a joke about how many tokens it takes to turn on a lightbulb.
It deserves a minor rewrite of the Black Mirror episode Fifteen Million Merits where people do menial labor like folding laundry and washing dishes to earn tokens so that their LLM will dispense their toothpaste and send Studio Ghibli stylized birthday cards to their friends.
inb4: When sama and co talk about UBI, they mean a variation of it based around a memecoin tethered/indexed on (tik)tokens.
https://en.wikipedia.org/wiki/World_(blockchain)
Probably 1,000 for the system prompt, 400 for the audio speech-to-text, 8 for the query, 180 for the thinking, 12 for the tool call, 33 for the response with a useless follow-up question
All to achieve something that could be done with a Raspberry Pi. You could do all this locally too.
https://gist.github.com/mgarratt/afb3b57a08e2eb2479eb6083a86...
https://www.xda-developers.com/ollama-ai-comparison-raspberr...
https://www.xda-developers.com/raspberry-pi-voice-assistant-...
https://www.youtube.com/watch?v=o1sN1lB76EA
This project isn’t tightly coupled with anything. Any service that supports WebRTC should work!
Also I was hoping to push people toward a RTOS. Better experience then a raspberry pi, I can cycle power and be back way quicker. Also cheaper/more power efficient.
I also think I unfairly like ESP because it’s an excuse to write C :)
Home Assistant integrates with WebRTC btw[0].
Also, why make the ESP32 the the hotspot? Why not just connect to the same network? Then you're not really range limited.
Is the comment about Home Assistant being python? Yeah, I can get that. Feels weird to be using slow scripting languages on lean hardware. Though of course you can write whatever routines in C and just patch it in to the interface.The ESPs are cheaper (here's the non-dev kit which has WiFi[1]), but way less powerful. I don't think you could get away with doing things on device. Though I wouldn't call that dev kit cheaper and that price point was context of my comment.
FWIW, I don't think there's really anything wrong with the project other than just that it comes off as doing things that have already been done before but presenting as if something novel was done. I'm all for reinventing the wheel. For fun, education, or even to improve. Just if I'm being honest, it came off with some weird vibes because of that. I imagine that's how some people are responding as well.
[0] https://www.home-assistant.io/integrations/homeassistant/
[1] https://shop.m5stack.com/products/m5stamp-esp32s3-module
> Also, why make the ESP32 the the hotspot? Why not just connect to the same network? Then you're not really range limited.
Because then they don't have to include the ability to configure wifi, which (while not that hard) is one more thing to do and for a hackathon that's not really contributing to the end goal.
I couldn't get on WiFi at the office at all. Corporate WiFi had a bunch of hoops to jump through that made ESP32 hard.
Once I got it working it felt really cool though. As a user I don't want to configure WiFi on the microcontroller at all. I would be really cool if I could walk up to a 'smart device' and set my phone next to it and do zero configuration.
I thought about your comment a lot. I worry that most people just say nice things (but think the opposite) so I appreciate you being direct.
-----
I don't expect you to know anything about me. It made me feel like you have written me off/dismissed me when you mention HomeAssistant + WebRTC. HomeAssistant uses Go2RTC and the WebRTC library it uses is Pion[0]. I created that and maintain it. Getting WebRTC on tiny devices is something I have been working on for years and always doing it Open Source/making it best for users.
-----
> comes off as doing things that have already been done before but presenting as if something novel was done.
I don't think 'Hardware AI Assistant' is a novel idea. What I hoped was a novel idea was putting it in an easy to use package for others to build upon and use. WebRTC + Hardware is something I have been trying to make easier for a long time https://github.com/awslabs/amazon-kinesis-video-streams-webr... [1] I wrote this code, but things still felt too difficult.
When ESP32s got powerful enough to do WebRTC I wrote [2]. Reflect inherits from that. So I am proud of that. I have been on this journey to make RTC + cheap hardware possible so that others can build upon that.
-----
Again I really appreciate your comment, and sorry to be so defensive. Someone I really respected (and I thought they respected me) said the same thing about my work not being novel. They said people have been building security cameras for years that use WebRTC, you are over inflating what you are doing. That has stuck with me. So part of me does fear that I am wasting my time trying to solve these problems.
I don't think what I am doing is novel. I do think that I am solving it differently because I make it accessible/Open Source. Most people solving these problems/building it just keep their code at work and don't try to help others use it.
If you are up for it shoot me an email sean@pion.ly and https://www.linkedin.com/in/sean-dubois/ I would love to have a friend that calls me out/is honest whats good work and what is just BS :)
-----
[0] https://github.com/pion/webrtc
[1] https://github.com/awslabs/amazon-kinesis-video-streams-webr...
[2] https://github.com/Sean-Der/embedded-sdk
Is it my browser, or does the video in the readme not have sound?
No sound! YouTube video in README does.
I was tempted to put Erik Satie in the README video. Didn’t want to risk copyright issues
Philips Hue is about to start a riot
I get that this is as-is, but I wonder if so many ultra-alpha products don't dilute the OpenAI brand and create redundancy in the product line. It feels like the opposite of Apple's well thought out planned product design and product line.
Let's see if it pays out.
This is just a hackathon project. Not a product in any way.
My primary work is on backend WebRTC servers. This was just an outlet/fun side thing to do client and embedded work. I love writing C and do microcontrollers. I just can’t seem to find a way to do it full time:(
We could help you find a pathway there :)
For a developer platform having examples is useful as a starting point for new projects.
Also, I’m not sure if it’s similar at OpenAI, but when I was at Google it was much easier to get approval to put an open source project under the Google GitHub org than my personal user.
They're selling shares at a $500B valuation. The market is telling them everything they are doing is amazing.
Is it possible to differentiate the feedback of the initial success of chatgpt from whatever came after it?
It's possible those investments are just the oai owners selling their 2023 chatgpt success and its profit share.
[dead]