Simone Salerno

I share tutorials on #tinyml and #esp32 for Arduino (from beginner level to pretty advanced). No TensorFlow unless strictly required.

[Simone Salerno @EloquentArduino]

Do you want to run @EdgeImpulse FOMO on your #esp32 camera using @arduino IDE?

Don't know where to start?

Look no further!

Here's the BEST tutorial you'll find on the entire internet.

Don't believe me? Keep reading 👇

https://t.co/P9cqUmpj6k

[Simone Salerno @EloquentArduino]

@MaximeRivest You're not sponsoring enough functai: it's awesome but it doesn't even show on Google when I search it (I have to resort to my history for the link). I would also appreciate some more content on the fine-tuning task as you were doing some months ago with Ettin charts 🙏

[Simone Salerno @EloquentArduino]

Everything is packaged into a single executable to invoke with 1 line. Read the full tutorial to replicate in 10 minutes: https://t.co/V8z3GJ6wyq

[Simone Salerno @EloquentArduino]

Turn your #esp32 p4 into a searchable knowledge base: convert your notes into embeddings and find top-K most relevant entirely on board.

[Simone Salerno @EloquentArduino]

@thdxr Type out code as in the old days

[Simone Salerno @EloquentArduino]

I (aka AI) ported @EspressifSystem esp-dl framework to @arduino for efficient NN deployment. Also wrote a companion Python library for easy model training and conversion to .espdl format. See KWS example (runs in 16ms on S3) 👉 https://t.co/Kmrg5Gahqo

[Simone Salerno @EloquentArduino]

@slimbuck7 @fulligin It's a shame this beautiful work gets so low exposure. Please don't stop creating

[Simone Salerno @EloquentArduino]

@MaximeRivest Looks awesome. Any chance you share this?

[Simone Salerno @EloquentArduino]

@PovilasKorop Pretty sure they don't review at all. Maybe they have tests that pass. Maybe the end product just "looks" good. I did a toy project, in one terminal, where I did no review: the tests (wrote manually) passed so it was fine. Could have opened 4+ tabs and I would've been fine too

[Simone Salerno @EloquentArduino]

@loretoparisi This only works on single-document scale. In their homepage "vector similarity != relevancy". True. Then in their hybrid search docs, to handle large volumes of documents "use vector search to filter relevant chunks". Either similar vectors are also relevant or they're not

[Simone Salerno @EloquentArduino]

@yongfook @emrahdma I use CC too, but am I the only one concerned about "let's fire a 1T LLM instead of 3 clicks?". I'm afraid we're trading energy efficiency for convenience/laziness too many times (since it's sooo easy to do)

[Simone Salerno @EloquentArduino]

@jeffrey_way As a Laracast user, I would still buy from you if the quality is on par with Laravel content. I guess there will be more competition in the AI space, but most will be low quality content from get-rich-quick people. Yours will still stand out

[Simone Salerno @EloquentArduino]

I published a Docker image to play with
@TencentHunyuan
#WorldMirror model easily. I'm still figuring out if it can replace COLMAP for #gaussiansplatting (OOM errors with tens of images)...

https://t.co/QRB91QfOUK

[Simone Salerno @EloquentArduino]

@JfkWhitlam Even cheaper: Freenove ESP32-S with camera. Add a 5$ microphone and a 5$ IMU if you want the full spectrum of sensors. XIAO is already assembled, though. But it gets hot in a few seconds.
https://t.co/7WhOR7EAki

[Simone Salerno @EloquentArduino]

@valeyakymenko 10ms is a really though costraint... https://t.co/Sz2rEukkeM

[Simone Salerno @EloquentArduino]

@DanyBittel What did you use for background masking?

[Simone Salerno @EloquentArduino]

@jonstephens85 @dreamingtulpa It's still a lot slower than other methods on A100. Much faster to do COLMAP + any 3dgs method in two steps

[Simone Salerno @EloquentArduino]

@rschu @theworldlabs After trying Hunyuan world I am skeptical on the usability of this kind of "magic", but I'd be happy to try this out and be proved wrong

[Simone Salerno @EloquentArduino]

I started fiddling with #esp32 WiFi CSI (Channel State Information) to detect motion/gestures totally wireless.

Contrasting results so far: clear signals sometimes, no noticeable changes other times.

Will continue investigating.. promising applications on the horizon...

[Simone Salerno @EloquentArduino]

Here's the generated panorama

[Simone Salerno @EloquentArduino]

Hey @TencentHunyuan, I've been playing with World-1.0 model but all my results exhibit this poor quality, no matter the input image (2048x1536, see first comment). Command is as per your Github README.

Can you suggest how to fix this?

[Simone Salerno @EloquentArduino]

I used labels_fg1=rocks, labels_fg2=crystals, clasess=outdoor