[inventor7777]

I really, really, really love this concept. I think there is SOME feature creep, but it does seem more or less scoped well to IP-type protocols.

However, I don't think they need to be prioritizing the local AI features, which are cool...but models get far smarter when you run them on a proper Mac/external GPU vs a small battery powered Flipper device. I think it might be helpful on the go, in the field, etc, but the usability with no dedicated keyboard will be rather poor.

However, I think they should keep focusing on the Zero for a possible Zero 2 to match the capabilities of this One device. I love my Zero, but I think it is missing key features like full support for garage door and RFID rolling codes, and some other protocols. The WiFi dev board is very limited, and there is no simple way to capture/playback BLE remotes IIRC. Of course, it depends on whether you consider BLE to be layer 0 or layer 1.

[0xbadcafebee]

Re: the on-device AI, most people don't know what they don't know. And they don't know that there's dozens of on-device AI applications that already exist in the real world using tiny AI models.

ESP32-S3's have been doing on-device AI for years. That's a 240MHz processor with 512KB SRAM, 16MB PSRAM and no GPU, and AI works great on it.

[meatmanek]

> AI works great on it.

Define AI, and define "great"

[0xbadcafebee]

A $20 battery-powered camera that can detect animals, humans, vehicles, and trees, and store + send an alert in real time of which it detects (as well as a description). Or a mini robot that can navigate obstacles and perform tasks while offline. Or a tiny industrial sensor that can detect a motor degrading due to the sound of its hum deviating from trained examples. Or an RF detection algorithm that ignores common band patterns and isolates unusual spikes in usage among dense noise. Or a motion tracking sensor that uses radar and video to track an object in a room. I mean really there's thousands of use cases. And this little Flipper One has 10x-100x the power of the ESP32-S3

[krupan]

So, for better or worse, none of that is considered "AI" anymore. Actually, I don't think anyone ever had the gumption to call it AI. Machine Learning or Computer Vision are the terms I remember.

[aorloff]

I can't define AI, but I can define "great"

Great is the not-sucking.

If it is great, then it does not suck. If it sucks, then it is not great. To be great is not merely to be good, it is to actually not suck. Then you are great, in the most minimal, barely clawed yourself over the line way.

[aftbit]

One I've personally done is wake-word detection for a local voice assistant. Not sure what exactly "great" is as I don't benchmark unless someone's paying me, but for my personal use case, good enough precision/recall that it doesn't annoy me!

[Computer0]

Any ml model can be called AI now.

[aftbit]

A* used to be called AI, now it's just an algorithm. AI is anything that is hard to do and we barely understand. Once we grok it, it's just an algorithm.

[neumann]

It's typically the name for the front of computer programming that is 'doing human type stuff'. Once it's not on the front and the wave has passed, the AI name moves on.

[inventor7777]

True, but the use case that Flipper shows off is a chatbot style AI, which is different from the ML type stuff you're talking about. I find small AI chatbots to just not be very helpful for technical applications, thus my comment.

Of course, if you had a custom trained one that's different but then that brings me back to how are you gonna type with it easily?

[SV_BubbleTime]

>You can connect high-speed modules to Flipper One over PCI Express, USB 3.0, and SATA interfaces.

There is feature creep, and lost the plot. I feel like either this is the latter, or they vastly surpass my imagination.

Either way, I am not convinced enough people want in pocket PCIe that would not be contented with rasp pi or laptop form factors to make this worth it.