HELP ME OUT GUYS! I NEED MORE PEOPLE TO HELP ME ALIGN AGI TO TROLLEY PROBLEM trolley.typememetics.institu…

switching to assembly doesn't give a 17x performance mate

With LLMs, we like to think of "smart" and "dumb" as one axis (because we think of humans this way). I'd like to argue against this framing. Instead, try to think of "smart" and "dumb" as two different axes. A model can be incredibly smart AND dumb at the same time. For a good example of this, look at a Gemini model. They are incredibly smart: you can bench them and see the capability. The amount of knowledge Google bakes into their models is incredible. Yet when you ask it to do work, the amount of stupid things the models will do is similarly incredible. They are "smart" and "dumb" at the same time. Similarly, look at a model like Fable 5.1. It is not quite as smart as Astra, but it is significantly less dumb. Astra is still the smartest model available today, but it is also one of the dumbest, regularly doing things that make literally no sense whatsoever. Finding the right balance of both "smart" and "dumb" is a challenge that everyone needs to figure out, from users to researchers. If we stop thinking of it as either-or and instead realize both can exist together, understanding model behavior gets much easier.

Rust is a good prompt compilation target for the moment, but so is C++. And soon assembler. Then microcode. Myopic to think we're going to stop the agentic drill bit until it reaches computing bedrock.

Asked Opus 5.5 to solve my latest challenge. It's currently running multiple commands concurrently searching my entire drive for a file named Continuation.c Incredible.

The reason why I'm working on that project on and off is that it's depressing as fuck. I hadn't done anything this depressing since before I wrote enterprise Java for a living. Code generation is alright, it works. But I also need the clanker to write prompts for subclankers and oh my god it's soul-sucking. The clanker doesn't understand the stochastic nature of AI. It feels like the subclanker will reasonably follow the prompt. It of course won't. So you explain it to the clanker and it just puts a wall of prompt legalese into markdown files, making the subclanker reward-hack and behave even worse. I've seen an entire ping-pong session where the clanker would encourage subclankers to do something, then forbid it completely, then encourage again etc, because each attempt would create a different set of issues. At every step your instructions get distorted to the point where they're unrecognizable and in the end the subclanker is doing exactly what it's not supposed to be doing. And on top of that there's the factor of LLMs being chaotic. There's never a causal relationship between the quality of the prompt and the quality of the final result. Maybe it just derailed this one time after the prompt was changed, despite the change being positive on average. So I never actually know if things are getting better or not. Plus, I have to rely on the clanker's judgement to determine how good the results are, because if I tried to analyze them manually, I'd just drown in all that data. So I can't even be sure that a good result is a good result. Literally nothing is certain or trustworthy. And the worst part is that all this nonsense and suffering is still worth it, so I can't just quit.

@dhh introducing himself as a 1000x programmer