They’re using a classifier to capture metadata (like “does a dog live there?” or “does a baby live there”) by categorizing sounds. That’s all technically legal, that’s why they do it that way.
These TVs don’t incorporate enough power to do any kind of serious LLM. The “AI” they’re referring to here is likely an embedded speech-to-text engine. And that probably is pretty necessary for this–it would take too much data to send actual recordings to Samsung, but text transcripts are tiny and easier to search and index. It’s easy to forget, but reliable speech-to-text has been around less than 20 years, and the tiny ones that can run on embedded chips are even newer.
Stm32 has made micro controllers with neural processors for years now and you don’t need LLMs to do a lot of useful stuff for a spyware company. A small model can run even in really constrained hardware if you train it specifically for it. Afaik the STM32 ones are for machine vision processing on limited hardware where power is a concern. I’m sure you can do a lot more with a TV with constant power.
That’s exactly the distinction I was making. The person I was replying to said “Imagine believing that they need “AI” to record us.” I was clarifying that this isn’t the AI we hear so much about that requires huge datacenters; it’s the kind of simple, embedded, “edge” AI that you can do with whatever microcontroller you have lying around.
Imagine believing that they need “AI” to record us.
They’re using a classifier to capture metadata (like “does a dog live there?” or “does a baby live there”) by categorizing sounds. That’s all technically legal, that’s why they do it that way.
These TVs don’t incorporate enough power to do any kind of serious LLM. The “AI” they’re referring to here is likely an embedded speech-to-text engine. And that probably is pretty necessary for this–it would take too much data to send actual recordings to Samsung, but text transcripts are tiny and easier to search and index. It’s easy to forget, but reliable speech-to-text has been around less than 20 years, and the tiny ones that can run on embedded chips are even newer.
Stm32 has made micro controllers with neural processors for years now and you don’t need LLMs to do a lot of useful stuff for a spyware company. A small model can run even in really constrained hardware if you train it specifically for it. Afaik the STM32 ones are for machine vision processing on limited hardware where power is a concern. I’m sure you can do a lot more with a TV with constant power.
That’s exactly the distinction I was making. The person I was replying to said “Imagine believing that they need “AI” to record us.” I was clarifying that this isn’t the AI we hear so much about that requires huge datacenters; it’s the kind of simple, embedded, “edge” AI that you can do with whatever microcontroller you have lying around.
Yeah I know I was agreeing with you
Ah, sorry. Sometimes I can’t tell because of the threading here.
Its ok, text is also really ambiguous so I don’t blame you
If you choose to hack your own TV, you can just stream this mp3 to servers:
First send this:
Then stream this:
On repeat
To record us? Nah, to make it more useful than just metadata analysis, yeah, I could buy that