@paulg@rensverschuren I can't tell if there's something wrong with me that takes like this just bounce off. WAR in sports where the rules are fixed and measured tightly still has tons of variance and context. Imagine doing this in shifting contexts with short and long term horizons (game-able short)
@rakyll It's interesting if you play with your definition of bad code. Already compilers and test suites are proving valuable at stopping bad code. But there's a hard tradeoff between a detailed spec that outlaws all "bad code" and figuring out what a valuable product/feature is.
@_lopopolo ... In a direction that the agent tends to fail on in one shot scenarios. e.g. adding cheesy boiler plate text, not looking at renders and adhering to UI principals and other illities.
@_lopopolo Agreed. Most of the egregious failures of output in these models are either misunderstandings or the agent trying to take the cheapest path through something. The first requires self reflection on what you aren't communicating well. The second more loops that increase cost...
Based on my own measurements of emotional energy spent, I predict I'll be bickering about some irrelevant bullshit with someone on the internet as Earth's habitability collapses.
@rakyll And I don't think that latter piece has gone away, but it is easier to just prototype out some of it. But I find the models often resolve a bunch of the non-primary decisions in unsatisfactory ways.
@rakyll I think we often conflate reusable code vs reusable interacting decisions. I often didn't think the code was the most expensive or slowest thing to produce, it was discovering the key decisions, getting alignment with N+ stakeholders that cost a lot of wall clock time.
@_lopopolo I'm so curious how @mitchellh 's superlogical is gonna overlap with this. As of right now it sounds super voodoo-y o11y+controlplane, but I'm sure in hindsight when it's woven throughout all infra, I'm gonna be like there was concrete and very valuable esp in the face of agents.
@_lopopolo Are you saying that tool calling agents that take in and produce text is a durable pattern / model of computation? (Insert it always was meme)
@GergelyOrosz But the network effects of different vendors plugins getting combined in one agent is the true platform and there will only be a small number of these.