In early 2026, we plan to release an @iroh.computer based discovery module for Exo with Tiles, enabling P2P LoRA training on consumer devices.

Exo Labs latest work: blog.exolabs.net/nvidia-dgx-s...

Inspired by SHAREit, we at Tiles are investing in the development of a fast, decentralized AI memory app for mobile devices, starting with an iOS TestFlight Q126 for iPhone 15 Pro+ devices.

www.ushareit.com/m/

Google Project Astra like technical execution beginning with high speed Macs and a USB drive, they even partnered with DARPA!

www.youtube.com/watch?v=PQqu...

LM Studio ships with better OCR support in its latest release: lmstudio.ai/blog/lmstudi...

"Sovereignty entails the ability to set one’s own rules in one’s own jurisdiction, and that is not something US cloud providers can offer Europe"

How clear does it need to be that Europe needs independent digital infrastructure before we actually take credible action? The need is as clear for cloud as it is for the rest of our critical digital systems.
www.project-syndicate.org/commentary/a...

At Tiles we are building our Rust based Modelfile SDK for universal version control and customisation for small local models.

It’s compatible with Ollama Modelfile syntax and in the beta builds we will use @hf.co Xet Core for model layers deduplication.

Work on model architectures, such as context optical compression with Deepseek AI’s OCR (3B model) and Google DeepMind’s Mobile Diffusion (500M model), presents a very attractive SLM deployment avenue.

- deepseek.ai/blog/deepsee...
- deepmind.google/models/gemin...

Identity and memory are two sides of the same coin. Tiles makes that coin yours: your user-agent.

This work by @jessylin.bsky.social et al. is phenomenal.

Memory layers are definitely worth studying. Given their parameter-efficient nature and applicability to smaller models, the authors reported promising results using a setup with a 1.3B model + a 1B memory pool.

jessylin.com/2025/10/20/c...

We’ve started sponsoring mzau, who’s building MLX Knife, an Ollama-like CLI tool for MLX models on Hugging Face (supports pull, rm, list, show, serve, etc).

github.com/mzau/mlx-knife

Thanks to Tom Dörr for highlighting this project.

The real bottleneck for useful AI isn’t just model quality.
It’s the walls around our data.
A thread on walled gardens, economic incentives, and why even the best models can’t help us if they can’t see us. 🧵

getting gpu rich journey