Small sneak peek into what I have built
This is an open source hardware and a bunch of software to run an AI that always on and listens and transcribes everything around you. Then AI agents extract memories and create visualisations.
Right now most of the stuff is running out of my basement simply because otherwise it would cost ~$40 per user.
I am running them of one-two GPUs, one my favourite is RTX 4000 Ada. Which is costly but very low power while providing 20gb of VRam.
Recently it was finally approved by Google and now in both stores I have an app:
https://getbubble.org
This is an open source hardware and a bunch of software to run an AI that always on and listens and transcribes everything around you. Then AI agents extract memories and create visualisations.
Right now most of the stuff is running out of my basement simply because otherwise it would cost ~$40 per user.
I am running them of one-two GPUs, one my favourite is RTX 4000 Ada. Which is costly but very low power while providing 20gb of VRam.
Recently it was finally approved by Google and now in both stores I have an app:
https://getbubble.org
👍8❤3👏2🔥1🥰1💩1
Last weekend we won a big Lllama 3 Hackaton!
We built a smart glasses from scratch using off-the-shelf parts in $20 and it runs entirely on Open Source networks.
This kind of blown up everywhere - now everyone knows what we have done and we have 1500+ of preorders in the first day. Gizmodo took small interview, virtually anyone i am talking two knows about our glasses. I have built firmware and base app (everything up to LLM) and guys benchmarked different llms and picked the best approach.
What we found is that Moondream v2 + Llama 3 is capable to run similar to what OpenAI demonstrated. It can sense what's happening around, answering questions. You can get faster than realtime responses via Groq API for Llama 70B, TTS is also quite fast. STT is also 100x of realtime. Ofc openai is much better at this, but hacked in the weekend solution also works.
My goal getting street creds in bay area is slowly working!
We built a smart glasses from scratch using off-the-shelf parts in $20 and it runs entirely on Open Source networks.
This kind of blown up everywhere - now everyone knows what we have done and we have 1500+ of preorders in the first day. Gizmodo took small interview, virtually anyone i am talking two knows about our glasses. I have built firmware and base app (everything up to LLM) and guys benchmarked different llms and picked the best approach.
What we found is that Moondream v2 + Llama 3 is capable to run similar to what OpenAI demonstrated. It can sense what's happening around, answering questions. You can get faster than realtime responses via Groq API for Llama 70B, TTS is also quite fast. STT is also 100x of realtime. Ofc openai is much better at this, but hacked in the weekend solution also works.
My goal getting street creds in bay area is slowly working!
Gizmodo
Cerebral Valley Hackers Build $20 Open Source Smart Glasses
A five-person team at a San Francisco hackathon created an open source approach to Meta's AI-powered Ray-Bans.
🔥15🦄5❤4
My first audio pre-training finished!
Thankfully my small cluster of 4090 finished pre-training of my audio foundational model! It took about 30 days to finish. This model is akin to original GPT (not chat one) in essence that it is not really that useful without finetuning, but you can finetnue it to do denoising, speech enhancement, speaker separation, text-to-speech tasks. All this tasks are very very important for AI Wearables and every single company i have talked too are struggling with it.
Thankfully my small cluster of 4090 finished pre-training of my audio foundational model! It took about 30 days to finish. This model is akin to original GPT (not chat one) in essence that it is not really that useful without finetuning, but you can finetnue it to do denoising, speech enhancement, speaker separation, text-to-speech tasks. All this tasks are very very important for AI Wearables and every single company i have talked too are struggling with it.
👍6🤔3💅2👨💻1
This media is not supported in your browser
VIEW IN TELEGRAM
SuperVoice Enhance
My new model for speech enhance and denoising. This model processes noisy or damaged audio extracts high quality voice. Should work with most languages out of the box.
Model and sources: https://github.com/ex3ndr/supervoice-enhance
My new model for speech enhance and denoising. This model processes noisy or damaged audio extracts high quality voice. Should work with most languages out of the box.
Model and sources: https://github.com/ex3ndr/supervoice-enhance
🔥14❤2
Please open Telegram to view this post
VIEW IN TELEGRAM
🦄8😁4❤1🤮1🥴1💊1👾1