Christopher Moravec

This week we look at how far self driving cars have come by looking at my experience this week driving a semi-automated car around for our company retreat.

Sora2 means that deepfakes are easier than ever, making the possibility of misinformation more widespread too. Sure, making Bob Ross paint Excel is fun, but if someone can make Bob Ross talk about Excel, imagine what they could make YOU talk about, even if you don’t want to.

Two AIs are better than one - If we use multiple models, we can obtain different interpretations that we combine to yield a more reliable answer.

I built a system to help us collect information at conference booths in a fun and engaging way—building you a custom fantasy map.

The term of art for AI is called hallucinations, but I prefer to stick with Make Shit Up. The training incentives are flawed; we have been training models to hallucinate because they score higher on tests by guessing incorrectly than by admitting they don’t know.

We need a culture where it is okay to admit to using an AI and celebrate it, all while keeping the human component in mind. Even if AI models don’t improve beyond their current state, we already face numerous issues, anxieties, and superpowers!

LLMs do not learn on the fly, their training is compute-intensive and cannot currently be done dynamically. Instead, they can recall messages from previous conversations, making them appear to have learned things about you.

AlphaEarth is a set of pre-computed embeddings across the entire Earth (over time) that makes it possible to search for similar locations. But what does that mean for GIS?

AI is like a magic genie, incredibly powerful, but deeply literal.

I wasn't going to write about GPT-5 this week, but since it's all the internet can talk about, I guess I’ll join in. My hot take is this: It isn’t significantly better than the previous models.