No Sloptober
55 points by jeremiahlee
55 points by jeremiahlee
Onarheim's Law:
"Agents can only maintain or INCREASE entropy in a system. Humans are uniquely capable of decreasing it"
Is there a steelman of this? Because it seems false on its face unless you believe it’s impossible to control agents such that they, y’know, act as your agent by implementing your decisions, which LLM users do all the time.
I want this debated as well.
If human is not in loop I'd expect this to be definitely true. By the end of the day we are reading probabilities from a very large metric to emit next token and since prob adds up to 1, elements of this large matrix are very extremely small values, right? Large this matrix (context) less discriminatory power we have (e.g. slope)!
My experience is that this is clearly false assuming there are constraints
I find LLMs do really good at iteratively optimising problems, if you build a harness that keeps them from drifting.
Like: Optimise cost this PCB. The new version must meet this strict SPICE harness, revert any version that fails the harness, pick only components from the existing library.
Or: Reduce P99 wall time of this MILP problem. Solution must pass this property test suite that checks against slower oracle.
But without something like this, I think the premise is true and a lot of problems are hard or impossible to “tie down” like this
I think this is shown to be contradictory by just having an agent do exactly what a human tells it to at such a level of instruction granularity that it has no agency on its own.
Here’s a shot, though it’s a somewhat weaker claim:
Left to their own devices, agents will increase the entropy of any codebase. Only sustained and time consuming human attention will prevent that, not any simple recipes or prompt engineering or loops. Therefore, the huge speedups that agents can provide are only available if you’re willing to compromise on quality.
I mostly endorse this, but the kicker is that there are intermediate points between writing everything by hand, and just vibe coding. It does take work to steer agents and get any kind of quality, and the people saying they’ve increased velocity 100x are delivering slop, but that doesn’t mean you can’t use agents in a way that’s an overall win and keeps quality high.
Yeah, it's hard to be sure what some of these arguments mean since the author elides over some things they seem to accept as plainly true. This, "deskilling", etc.
Like, this statement is probably true (or close enough) in a "dark factory" situation but agents do not strictly increase entropy in a system in the way they are actually used by by people, in my experience.
I looked and couldn't find an "Onarheim's Law" either. The closest I could find is some ideas on creative expression and constraints by Balder Onarheim which I largely agree with but it's unclear exactly what this appeal to authority is appealing to.
I'm into this!
I'll join! Maybe difficult at work at this point, but for personal projects for sure.
One month seems like far too long, but perhaps that's all the more reason to do it.