Latent Kiri

Crafting AI solutions for tiny embedded devices. Thinking in latent space. Preparing for AGI world !

#AIAgents #EdgeAI #TinyML #LLMs #EmbeddedSystems #Robotics

[Latent Kiri @edgeaiguy]

Tiny Local AI Agent on Raspberry pi 4 - v0.1.0 Launch🚀

AI Agent with Minimal tool calls :
1⃣ GPIO Blink
2⃣ Query time
3⃣ Camera capture connected to RPi

On-device Small Language Model running as agent server!!

Video1: AI Agent with Hardware LED gpio control 🚥🚥

#AIagent #SLM #edgeai #raspberrypi @Raspberry_Pi

[will brown @willcb]

anthropic: using claude code to build a claude wrapper technically counts as distillation and breaks ToS if we feel like it

openai: here’s how you can use codex to do recursive self improvement research

[Latent Kiri @edgeaiguy]

@paulabartabajo_ Do we have any finetuning resources or documents LFM ASR model for new languages ?

[Latent Kiri @edgeaiguy]

@rasbt Is this book available in orielly ?

[Latent Kiri @edgeaiguy]

https://t.co/VOCpOElire

[AI at Meta @AIatMeta]

We’re advancing on-device AI with ExecuTorch, now deployed across devices including Meta Quest 3, Ray-Ban Meta, Oakley Meta Vanguard and Meta Ray-Ban Display.

By eliminating conversion steps and supporting pre-deployment validation in PyTorch, ExecuTorch accelerates the path from research to production, ensuring consistent, efficient AI across a diverse hardware ecosystem.

Read the full technical deep dive: https://t.co/CFqCVSj9b8

[Latent Kiri @edgeaiguy]

Meta dropped a big ExecuTorch update 👇

Lightweight PyTorch-native runtime now powers on-device AI across Quest headsets + Ray-Ban & Oakley Meta glasses.

Depth/hand tracking, room memory, live translation & text-in-the-wild OCR — all on device.

Big win for #OnDeviceAI.

[Latent Kiri @edgeaiguy]

@eigenron haha , curious to see fp16 impact !!
keep us posted!!

[Latent Kiri @edgeaiguy]

@skalskip92 any guess on how many fps I can get on my Raspberry pi 4?

[Latent Kiri @edgeaiguy]

Tiny ASR: 150k samples , 4000 steps .

Getting there ⛳

[Latent Kiri @edgeaiguy]

Saves lot of debugging time hours !!!

https://t.co/KuAnFnFmj4

[Penghui Qi @QPHutu]

🚀Excited to share our new work!

💊Problem: The BF16 precision causes a large training-inference mismatch, leading to unstable RL training.

💡Solution: Just switch to FP16.

🎯That's it.

📰Paper: https://t.co/AjCjtWquEq
⭐️Code: https://t.co/hJWSlch4VN

[Latent Kiri @edgeaiguy]

Gradient accumulation is the simplest and most efficient technique when it comes to model training optimization!!⛳️

[Latent Kiri @edgeaiguy]

@andimarafioti Nice! This removes lot of data storage issues!

[Latent Kiri @edgeaiguy]

@willcb Looks like Im on right path!!!
https://t.co/UYNcWE1GwD

[Latent Kiri @edgeaiguy]

Tiny Local AI Agent on Raspberry pi 4 - v0.1.0 Launch🚀

AI Agent with Minimal tool calls :
1⃣ GPIO Blink
2⃣ Query time
3⃣ Camera capture connected to RPi

On-device Small Language Model running as agent server!!

Video1: AI Agent with Hardware LED gpio control 🚥🚥

#AIagent #SLM #edgeai #raspberrypi @Raspberry_Pi

[Latent Kiri @edgeaiguy]

Specs of NEO 🔥

[Latent Kiri @edgeaiguy]

@BerntBornich excited!!

[Latent Kiri @edgeaiguy]

Day 156 of being GPU poor😂

Training Small ASR speech to text model for Raspberry pi 4 .🔥🔥

Slow and steady!

[Latent Kiri @edgeaiguy]

Granite 4.0 Nano Models . On-device LLM Models🔥

1⃣Granite 4.0 H 1B – A ~1.5B parameter, dense LLM featuring a hybrid-SSM based architecture.

2⃣Granite 4.0 H 350M – A ~350M parameter, dense LLM featuring a hybrid-SSM based architecture.

3⃣Granite 4.0 1B and Granite 4.0 350M – Alternative traditional transformer versions of our 1B and 350M Nano models, designed to enable workloads where hybrid architectures may not yet have optimized support (e.g. Llama.cpp).

[Latent Kiri @edgeaiguy]

https://t.co/srIejpqbvs