Thoughts
Tried out devstral + ollama + aider. Accuracy isn't awful, it's just really really slow.
Right now it's easy to judge AI as bad because it's slow and big and inaccurate. Obviously a lot of people are focused and are betting on it getting more accurate, but I think it's almost equally important from a vibes perspective that it's small and fast. If a model of power similar to o3 was open source, small enough to run on a phone, responded in 0.1 seconds, then it's a very different ballgame. It just is.
AI is exciting and scary right now because it's around the size of a person. It takes the amount of time to answer that a human would to type out a similar length response. It takes up GB of space. Both AI-tech-bros and anti-AI people wouldn't be talking about AI if it was so small and fast that it was boring. You'd stop talking about AI replacing people and about AI having to compete with people if AI was so easy to use that it was ubiquitous. If the head of Abraham Lincoln on every penny talked to you with the brain with o3, then how do you argue that o3 is a disruptive and revolutionary product.
Anyways. The reason I'm writing this is so that in 50 years when some kid is running an LLM on their homemade breadboard computer they don't feel like they're betraying the wishes of the anti-AI gen Z. AI sucks right now because we can't run it on a breadboard computer and have to pay for a monthly subscription to ChatGTP plus or whatever.