My submission for the Liquid AI x DPhi Space 'AI in Space' hackathon: an on-board satellite AI pipeline that filters 98% of downlink bandwidth using LFM2.5-VL.

I added llama.cpp benchmarks across eight model groups to the local LLM benchmark suite. Speculative decoding with Gemma 4 hit 206 tokens per second on an M3 Max. Here's how it compares.

What started as a simple question (can you run agentic coding workflows against a local LLM on Apple Silicon?) turned into 57 configurations, 762 data rows, and a long weekend of benchmarking.

My 10-year-old and I spent 24 hours at Hackathon Heilbronn turning a 26-year-old TV game into a browser-based accessible gaming portal, without writing a single line of code ourselves.

A project assignment in San Jose, friends at Applied Intuition and Zoox, and a 13-year-old at the German International School. What three weeks in Mountain View taught me about where we actually stand with AI and autonomous driving.