AI, mobility, software, and the odd product lesson from real work. Sharp enough for builders, casual enough for a train ride. No fog machine.
2026-06-01
How LLMs Got Faster Without New Hardware
The training side of AI gets all the headlines. But in the last year the inference side has been quietly transformed. KV cache compression, rotation-based quantization, and multi-token prediction are why your local model runs twice as fast today as it did twelve months ago.
The NanoGPT Speedrun: From 45 Minutes to 81 Seconds
A GitHub repo nobody talks about might be the clearest window into how AI training actually got so much faster. Here is the story of modded-nanogpt and the people racing to train a language model as fast as physically possible.
It's 2019 and it is already midnight and I just wanted to get some sleep to be fit for work the next day but then I found this video on youtube which really got me thinking. So much, that...
The Self-Driving Revolution: The Journey to V12 and Beyond
Tesla is taking the world by storm with its groundbreaking advancements in self-driving technology. The release of V12 Full Self-Driving (FSD) marks a significant leap towards making auto...
Lights, Camera, Action! The Future of Entertainment: When Movies and Games Collide
The worlds of movies and games are colliding in a way that could reshape the entertainment industry. With the increasing use of 3D assets in filmmaking and the recent development of advan...
The Dawn of Artificial General Intelligence - How Language Models evolved into Intelligent Beings
A recent [study](https://arxiv.org/abs/2303.12712) by a Microsoft research team dived into the capabilities of GPT-4. The findings of this research are both exciting and concerning at the...
Are electric cars better than traditional cars? Not yet, but history proves they will be
Back in 2018 I wrote this article and never published it. I still think, that innovation S-Curves are one of the best examples of how a great visual representation can make a very complex...